Fabio Dias and Yvens Lima work out the constants OpenAI's ETH proof leaves out: its proof certifies no speed-up over brute force on any bounded-density formula that could be written down
What happened: Fabio Dias of the University of Surrey and Yvens Lima of the Federal University of Ceará take OpenAI's proofs as correct and ask how large a formula must be before its bound beats trying every assignment. The preprint states no constants: with a hidden constant of 2 the crossover is near 10³² variables, with 3 near 10¹²⁶. From the tensor-contraction proof's own constants, the time it certifies on bounded-density formulas stays above 2ⁿ until log₂(n+m) passes 10⁹.
Why it matters: If the proof holds, ETH is still refuted; what changes is what the result means. It guarantees no faster count for any bounded-density formula that could ever be stored. That separates the claim about theory from any claim about solving real formulas faster, and gives the review a number to check.
Evidence and limits: The authors say they do not assess whether OpenAI's proof is correct. Their limits on the clause-product route hold in a cost model they built to match what the preprint's evaluation pays, and the parity form and tensor contraction escape them. The work was done with Claude Opus 5.5 through Claude Code, refereed by a second instance, and is an unreviewed preprint.
BigHat finds that giving an antibody-design agent its tools makes it follow instructions better and optimise worse
What happened: Asli Alpman, Nick Bhattacharya, Aaron Solomon and Hunter Elliott gave six Claude and GPT models ten expert-checked antibody-editing briefs, each with must-have rules, and had code check whether every rule was obeyed. Without tools, 0.73 to 0.87 of runs obeyed all of them; with antibody tools for numbering, folds and exposed surface, 0.98 to 1.00. Those tools also cut the share of designs beating the starting antibody by 4.2 points, while past scored designs raised that share 10.5 points and lowered obedience 0.09.
Why it matters: Agents are sold to labs as designers that take an instruction and return molecules, and this measures what that costs. The two things a lab wants, doing what it was told and finding a better molecule, moved in opposite directions as tools and design history were switched on and off. Matching BigHat's own earlier Bayesian optimiser over ten rounds of 50 designs is the bar such a loop has to clear rather than a result, and the paper's ablation says its task agents could be dropped without losing performance.
Evidence and limits: Nothing was made or measured: every design was scored by a model BigHat trained on one of its own past campaigns, against one objective. Each must-have is a property the provided tools compute, so the test measures what those tools report. Claude models refused up to 60% of calls depending on the setting, and refused calls are left out of the rates. Not peer reviewed; the authors say lab validation is future work.
Princeton and Edison Scientific's ChefNMR2 agent works out the structures of four unreported natural products from simple NMR spectra
What happened: ChefNMR2 is a generative model that proposes a small molecule's structure from its formula and one-dimensional hydrogen and carbon NMR spectra, chemists' standard test of what a molecule is. Trained on 2.52 million molecules, it ranked the correct structure first for about 82% of 2,031 test molecules, against 60% for the retrieval method NMR-Solver and 52% for Claude Opus 5, the best language model tested. With an agent that adds reasoning and literature search, it solved four natural products from Mohammad R. Seyedsayamdost's lab whose structures had not been reported.
Why it matters: Working out a new molecule's structure by hand is slow expert work and holds up natural-product drug discovery. The team reports runs up to 2,170 times faster and 360 times cheaper than frontier language models, with a confidence score that tracks accuracy. Code and weights are public, and Edison has put the model into Kosmos for its partners.
Evidence and limits: A preprint, not peer reviewed. The four new structures were judged against the same lab's own chemists. For one of them the model alone ranked the right answer fourth or fifth, and the agent put it first in three of five runs. On well-known molecules such as taxol, the authors say language models may be recalling, not inferring. Planar structures only.
AuthorsZiyu Xiong, Mohammad R. Seyedsayamdost, Ellen D. Zhong
Recursion appoints Fred Hassan, former head of Schering-Plough and Pharmacia, as chair of its board
What happened: Recursion's board enlarged itself from nine members to ten and appointed Fred Hassan as chair, effective 31 October. Hassan ran Schering-Plough from 2003 to 2009 and Pharmacia from 2001 to 2003 as chairman and chief executive. Chris Gibson, chair until the June annual meeting, had told the company on 28 April that he would not stand for re-election, with no disagreement. The release does not say who chaired from June to October; it thanks Rob Hershberg, who stays on the board.
Why it matters: The board's new chair is someone whose career was running large drugmakers through late-stage development and launch. Chief executive Najat Khan said the appointment fits a move from clinical proof of concept toward approval and launch, which matches where Recursion is spending: a registrational trial for REC-4881 planned for 2027 and two programmes it wants partners to take on.
AlphaFold 3's open-source code gets a speed release, v3.0.5: Jackhmmer search about 2.5x faster, over 10x on a 24-core machine
What happened: Augustin Žídek at Google DeepMind released AlphaFold 3 v3.0.5, which the notes call a speed release. Jackhmmer, the search that builds the input sequence alignment, is about 2.5x faster and now scales with the CPU count, giving over 10x on a 24-core test machine; the alignment featurisation step is 10x faster, and the build and data scripts are faster too. It also carries bug fixes, more unit tests in continuous integration and better documentation, many from outside contributors.
Why it matters: Running AlphaFold 3 is dominated by the alignment search, so a faster Jackhmmer lowers the time and cost to fold a protein on the public code. The over-10x figure is tied to a 24-core machine because the step now uses all the cores it is given. This is maintenance on a tool the record already tracks, with no change to the model or its weights terms.
Evidence and limits: The over-10x speedup is one team's test on a 24-core machine; the general figure is about 2.5x. The numbers are the release's own, with no outside benchmark. The model and its weights terms are unchanged.
In a Zelinsky Institute test with 60 example molecules per task, GPT-6 Astra predicted five chemical properties better than eleven chemistry models
What happened: Stepan N. Ostarkov, Michael G. Medvedev and colleagues at the Zelinsky Institute in Moscow gave models 60 labelled molecules per task. Prompted with the examples, GPT-6 Astra beat the best chemistry model on solubility, hydration energy, lipophilicity and two quantum-chemistry targets: R² 0.85 against 0.53 on orbital energy, 0.97 against 0.63 on hydration. Chemistry models won only on BACE1 enzyme inhibition, the control. On orbital energy the best of them needed 1,000 molecules to pass Astra's score on 60.
Why it matters: Most chemistry projects have fewer than a hundred measurements, which is exactly the case chemistry foundation models were built for. Here a general model prompted with the data did better on the property-prediction tasks tested, outside the drug-like classification the chemistry models were trained around. If it holds, a lab with a small dataset should try a frontier model before fine-tuning a chemistry one.
Evidence and limits: All six benchmarks are public, so the general models may have seen the answers in training. With labels shuffled they still scored well, and the authors flag likely memorisation for Claude Opus 5.5. Each task used 100 molecules over 15 random splits; this is a preprint.
An NTU search that learns as it goes finds the lowest-energy atom arrangement in disordered crystals in 94% of runs, against 55% for standard methods
What happened: In a disordered crystal, several elements share one atomic site, and simulating it means first picking one arrangement out of a huge number. Martin Hoffmann Petersen, Zekun Ren and Kedar Hippalgaonkar at Nanyang Technological University fit a cheap model to the structures relaxed so far to choose the next. Across 14 test cases with known answers, the four learned searches reached the lowest-energy arrangement in 264 of 280 runs, and the best of them, Bayesian ridge regression, needed a median 4.3 times fewer relaxations than the best standard method.
Why it matters: Nearly half the experimentally verified crystals in the main inorganic database carry some site disorder, and they cannot be simulated until an arrangement is chosen. Cutting the search by about four times makes those materials cheaper to model. On a battery cathode material with two coupled kinds of disorder, the search found the arrangement that experiments report.
Evidence and limits: Energies come from a machine-learned potential, and the benchmark answers were found by enumerating with it, not by experiment. On the cathode material only one of five runs reached the reported arrangement.
Quantum 3xplorer, a plane-wave DFT code written by AI coding agents under scientists' direction, matches Quantum ESPRESSO's energies and runs 1.2 to 5.3 times faster on one core
What happened: Timur Bazhirov at Exabyte (Mat3ra), with co-authors at Berkeley Lab, Stanford, SLAC, NVIDIA, UC Berkeley and other institutions, released Q3, a density functional theory engine built from scratch. Claude Code, Gemini and Cursor agents wrote its 82,200 lines of C++ and over 1,800 tests in weeks; the scientists set the interfaces, equations and tolerances. On silicon and copper test cells its total energies agree with Quantum ESPRESSO's within 0.014 mRy. It is open source under Apache 2.0.
Why it matters: DFT calculations are the main source of training data for machine-learned materials models, and the codes that run them are decades-old Fortran that is slow to extend. If agents can write and test a new implementation in weeks while scientists check the physics, a new method can reach a fast, validated code much sooner. Q3 already runs on GPUs at NERSC and writes its results in BerkeleyGW's file formats for many-body calculations.
Evidence and limits: Accuracy and speed are checked against one reference code on small cells of up to 54 atoms that the authors chose, and outside review so far is black-box checks of its numbers. The paper gives no count of human hours, so 'weeks' cannot be checked.
Sogang and Ewha researchers use AlphaEvolve searches to find a family of entanglement-assisted quantum codes that, with a lifting step, closes 54 gaps in the published code tables, its construction checked in Lean
What happened: Jae-Hyun Baek, Byung-Sun Won and Jon-Lark Kim (Sogang and Ewha Womans universities) used Google DeepMind's AlphaEvolve to evolve programs that build entanglement-assisted quantum error-correcting codes. The programs produced small qubit codes whose pattern the researchers generalized into an infinite family with a proof, and Lean checks its algebraic core. With a lifting step the family closes 54 gaps in the archived code tables; ten nonexistence results come with checked solver proofs. The 115 verified codes and the search programs are released.
Why it matters: The paper records each step apart: what the evolved programs built, how the researchers generalized it, and what Lean and the solver certified. A reader can see which part came from the model and which from the mathematicians, and replay the search. On a secondary measure of a controlled test, programs refined over rounds worked at unseen code lengths more often than ones generated once (46.1% against 19.5% at length 13).
Evidence and limits: The controlled comparison covers one model, one researcher proposal and binary codes; its prespecified primary measure favoured refinement in five of eight paired blocks (p = 0.328). The distance guarantees assume ideal decoding without noise.
Flagship merges the platforms of Cellarity, Etiome and Empress Therapeutics into Etcelera, an AI drug discovery company led by Adrian Rawcliffe
What happened: Flagship Pioneering combined the platforms of three of its companies, which BioPharma Dive says had raised $374M between them, with its in-house AI unit, Pioneering Intelligence. Adrian Rawcliffe, formerly CEO of Adaptimmune, is CEO; Flagship senior partner Alec Reynolds chairs the board, and Noubar Afeyan is a co-founder. It starts with programmes in blood disorders such as myelofibrosis, in metabolism, including partnered MASH work, and in immunology. Cellarity's sickle cell drug CLY-124 goes to a separate company.
Why it matters: Each company held one kind of evidence: Cellarity read the state of whole cells, Etiome followed how a disease changes from stage to stage, and Empress showed what a molecule actually binds. Flagship is betting that the three together, plus data on how cells respond when a gene is silenced or a chemical is added, can tell which targets cause a disease, and so which drugs will work in patients before a trial. Reynolds framed the target as the roughly half of drugs that fail in Phase 2.
Evidence and limits: No new funding was announced. The 150-million-cell foundation model, which the release calls the industry's largest, and the more than 30 prediction models are the company's description; no result from them is shown.
ALLOX, a Barcelona spin-out co-founded by Ben Lehner, raises a $10.5M seed led by Asabys to find hidden control sites on hard drug targets
What happened: ALLOX closed a $10.5M (€9.4M) seed led by Asabys through its Sabadell Asabys II fund, with Spain's innovation agency CDTI, through Innvierte, and the Institut Català de Finances. It came out of Barcelona's Centre for Genomic Regulation; Júlia Domingo is CEO and co-founded it with André Faure, Pablo Baeza and Ben Lehner. Morten Graugaard, chief business officer of Orbis Medicines, becomes independent chair, and Asabys co-founder Clara Campàs sits on the board. The money goes to its mutation-mapping platform, its own AI models, a drug discovery team and small-molecule programmes.
Why it matters: Many disease proteins, such as transcription factors and proteins that work by gripping other proteins, offer no pocket for a drug to sit in. ALLOX looks for other spots on the protein that still change what it does, by measuring thousands of mutations at once. Faure, Domingo and Lehner published the method in Nature in 2022: neural networks fitted to those measurements mapped such sites across two common human protein domains. The round pays for AI models that predict these sites from the company's own data.
Evidence and limits: A financing, not a result: the release names no target or new data, and the company site names one preclinical inhibitor, ALX-001, with nothing shown. The 2022 paper mapped two small protein domains in the lab; nothing public yet shows a drug made against a site found this way.
EVENET, a model pretrained on 500 million simulated LHC collisions, finds the known Upsilon meson in real CMS data at 7.6 sigma, above the 6.4 sigma of the published method
What happened: Physicists at National Taiwan University, Shanghai Jiao Tong University, SLAC, the University of Washington and others pretrained a model of whole collision events on 500 million simulated events, then tuned it for four analyses. In simulated searches for new heavy particles and exotic Higgs decays it beat XGBoost, TabPFN and SPANet, most clearly with little data, and in simulation it measured top-quark entanglement 37% more precisely than a published benchmark. In real 2016 CMS data it found the known Upsilon meson at 7.6 sigma, against 6.4 for the published method.
Why it matters: Collider physicists today train a separate model for each analysis. EVENET tests whether one pretrained model can serve many: one large pretraining run, on 512 GPUs at NERSC, then cheap fine-tuning for each new question. The gain is largest in searches for rare signals, where each new-particle hypothesis comes with only thousands of simulated signal events. The pretraining data and three of the four downstream datasets are public on Hugging Face under the MIT licence.
Evidence and limits: Only one test uses real collisions, and it finds a particle already known; the other three run on simulated events. The authors say the model does not yet carry detector and theory uncertainties through, and that use in a live LHC analysis is the decisive test.
With analysis steps written by the researchers, 11 AI models processed a catalysis dataset that none finished without them, and the experiments they proposed were run
What happened: Fanxu Meng, Kedar Hippalgaonkar and colleagues at UC Berkeley's Singapore centre, Nanyang Technological University, the National University of Singapore, Monash University and Lawrence Berkeley National Laboratory built SKIER: AI models call analysis procedures the researchers wrote, and plain code does the calculations. On 49 records from the group's own published study of methanol oxidation on platinum, all 11 AI configurations finished the data processing with the shared procedures and none without. The models then proposed three mechanisms and tests to tell them apart; the team ran the tests.
Why it matters: AI models are increasingly asked to interpret experiments and propose mechanisms, and their claims are often hard to trace to the measurements behind them. SKIER splits the work: researchers fix how each measurement is analysed, the model chooses and reasons, and each conclusion points back to its data. In the case study the loop closed once: isotope-labelling, residence-time and oxygen-exchange tests argued against two of the three mechanisms and placed C-O bond cleavage before formic acid forms.
Evidence and limits: One reaction system, from the authors' own earlier study. The 100 single-choice questions on which GPT-5.6-sol scored 100% were expert-designed, and the paper does not say by whom. The authors note that intermediates below the detection limit and pathways they did not consider are not ruled out. A preprint.
Edison Scientific agrees with Pfizer to deploy its AI scientist Kosmos across Pfizer's research and development pipeline
What happened: Edison Scientific, the FutureHouse spin-out led by CEO Sam Rodriques, announced an agreement to deploy the Edison AI Platform and Kosmos across Pfizer's research and development. Kosmos will draw on Pfizer's data in oncology, internal medicine, immunology and other areas, and Edison's forward-deployed scientists and engineers will build it into specific workflows with Pfizer's research teams.
Why it matters: This is the third Kosmos deployment Edison has announced this year, after Incyte in May and Population Health Partners in June. Here the AI scientist reads a large drugmaker's internal records, including programmes that did not work, which published papers rarely report. Rodriques describes the aim as letting what each programme learned inform the others.
Evidence and limits: Only Edison announced it: no Pfizer release was found, and no Pfizer person is quoted. No money, duration or scope is stated. It is a deployment, not evidence that Kosmos improves Pfizer's decisions.
Mistral releases the open weights of Large 4, a trillion-parameter model, and reports a potential for uplift to help attacks on live targets
What happened: Mistral released the weights of Mistral Large 4 ("Le Chonk"), a 1-trillion-parameter model with 52 billion active, under a modified-MIT licence, free to commercialise except for model-as-a-service providers. It reports a tie with the top open-weight model GLM-5.3 on a coding-agent index (48%) at 47% fewer tokens, and a 96.6% refusal rate on malicious cyber requests, the highest it measures among open models. It also says that in live traffic it saw a potential for the model to help conduct real attacks, and is changing its guardrails.
Why it matters: Mistral's pitch is the open frontier: a model researchers and companies can download and run on their own machines, trained in Europe for what it calls sovereign AI. The open weights are the part that bears on open science tooling. The self-disclosed finding is the one to note: a lab releasing a capable model's weights while saying the same model could help attack live systems, and shipping anyway with guardrails it is still revising.
Evidence and limits: Every benchmark figure is Mistral's own, and it says a full technical report is still to come. The uplift finding is Mistral's own observation from its live traffic, with no number attached.
Opentrons, with more than 11,000 pipetting robots deployed, raises a $40M Series E led by Avego to have AI run experiments on them
What happened: Opentrons Labworks, a New York maker of open benchtop lab robots, closed a $40M Series E led by the healthcare investor Avego, with BroadOak Capital Partners; SoftBank Vision Fund and Khosla Ventures are existing backers. It says it has placed over 11,000 robots in more than 80 countries. CEO James Atwood says the money goes to robotics, software and data tools, AI-run experiments and new markets; James Flexner, Avego's managing partner, says every AI-driven discovery still has to be tested in a real lab.
Why it matters: The company's case is that models now propose experiments faster than labs can run them, so the robot that does the pipetting is where the queue backs up. Opentrons has a large installed base and an open programming interface, which is why outside teams have already had Claude drive its Flex (Tetsuwan Scientific) and GPT-5.6 Terra drive its OT-2 (Duke's PEARL). The round is money betting that accessible, open robots like these become the usual way AI models reach a wet lab.
Evidence and limits: A financing, not a result. The robot count is the company's own figure, and the release gives no valuation, no total raised and no measure of how much of its robots' work is run by AI today.
Nomic Bio and Parse Biosciences start building Cascade, 100 million immune cells read for RNA with matching secreted-protein profiles, and GSK, Pfizer and Sanofi signed on
What happened: Cascade will treat blood immune cells from 12 healthy donors with cytokines, chemokines and drugs, alone and in combination, and read them at several time points: 100 million single-cell RNA profiles from Parse's GigaLab and more than 11,000 profiles of over 1,000 secreted proteins from Nomic's Omni 1000, across more than 150,000 perturbation measurements. GSK, Pfizer, Sanofi, Graph Therapeutics and Turbine are founding partners. Data generation has started, and the set will be licensed.
Why it matters: Models of how cells respond to drugs, the virtual cell, are short of data that follow a response over time and measure more than RNA. Cascade measures the RNA inside each cell and the proteins the cells release, from the same samples and time points, so a model can be tested on whether cells' responses predict the proteins they secrete, or held out on unseen donors and perturbations. GSK, Pfizer and Sanofi signing on before any data are released is a sign of how much drugmakers want this kind of data.
Evidence and limits: An announcement: no data have been released yet, and the companies give no date, price or partner terms. The set is licensed, not open.
A Wisconsin team's AI agents test findings from 121 single-cell cancer studies in other cancers, and a lung cancer risk signature they flag also tracks survival in three more
What happened: Lingxin Cheng, Jack Shireman, Christina Kendziorski and colleagues at the University of Wisconsin-Madison built SMART, about 20 Claude agents that read a published single-cell RNA study, list its findings, then write and run analyses testing each one in another dataset. Across 121 studies of six cancers it wrote 726 reports: 46% of the cross-cancer reports found a transfer, against 75-88% within one cancer. In 10 cases checked by hand, 92-97% of its outputs were correct, and fed summaries with every finding reversed, it rated all 15 low or middle-low.
Why it matters: A published finding is rarely tested outside the setting it came from, because checking it means rereading the paper and redoing the analysis. SMART makes that check routine and points to what is worth following up. ORACLE, a 23-gene lung cancer risk signature, is cited by 161 studies and none tested it in another cancer; after SMART flagged it, the authors found high scores linked to worse survival in public liver (hazard ratio 1.92), pancreatic (1.73) and mesothelioma (3.04) cohorts.
Evidence and limits: Every check reuses existing public data; no new experiment was run. ORACLE showed no significant link to survival in ovarian, brain or colorectal cancer. The team audited its own system, and the preprint is not peer reviewed.
PerturbMind, founded out of a Tongji University lab to model how cells respond to drugs, raises angel and angel-plus rounds of tens of millions of yuan led by Qingsong Capital
What happened: Qingsong Capital led both rounds, with Tongyang Fund and several drug companies following. It builds AlphaCell, a model meant to predict how a cell's gene expression shifts when a gene or a drug perturbs it, and released a 1-billion-parameter version that it says can be cut to 300 million or scaled to 50 billion. Founder and CEO Guohui Chuai did his PhD in the Tongji lab that chief scientist Qi Liu heads. It plans to sell on-device service subscriptions, private deployments and co-run drug programmes.
Why it matters: A model that predicts how a cell responds to a gene knocked out or a drug added would let a lab test far more interventions on a computer than in a dish; that is the aim of virtual-cell models such as Westlake University's ProteinTalks. This round puts a team out of Qi Liu's Tongji lab into that direction with its own model, drug companies among the backers, at the earliest stage of funding.
Evidence and limits: The amount is given only as tens of millions of yuan, with no valuation and no named drug company. The AlphaCell preprint of March 2026 claims the model predicts responses in cell types it never saw, and the report shows no drug target or molecule found with the model.
Edison Scientific's EDB-Bio: asked how likely a biology hypothesis is given published results, frontier models score no better than a flat 50% guess; its trained Ibis0 model does better
What happened: Conor Igoe, Siddharth Narayanan and colleagues at Edison Scientific and MIT CSAIL built 96,000 episodes from drug-response RNA data, CRISPR gene-dependency screens and GWAS results. A model reads a run of related results and states the probability that a held-out result supports a hypothesis. GPT-5.6-sol and Claude Opus 5 rank true above false (AUROC 0.737 and 0.752) but are underconfident, with Brier scores no better than a flat 50%. Qwen3-30B post-trained on scoring rules is called Ibis0; its classifier version reaches a Brier of 0.177 and AUROC 0.820.
Why it matters: An agent that chooses experiments needs to know how far a result should move its belief, not only which answer is likely. The benchmark measures that, and the training improves it. In a simulated task of picking the next experiment, Ibis0 paired with the paper's information-gain estimator got about 2.5 times the information from its first pick that the best prompted frontier model did; prompted alone it did no better. Code, data and models are released.
Evidence and limits: The episodes are structured summaries of published data, with yes-or-no hypotheses balanced at 50%, and experiment choice was tested against a simulator, not in a lab. The experiment-choice gain shows in two of three domains. One model family at one size was trained, on the company's own benchmark.
A machine-learning screen of two materials databases finds 33 compounds made years ago that fit the MXene pattern, and the team then makes 11 new rare-earth MXenes
What happened: Ali Saffar Shamshirgar, Johanna Rosen and colleagues at Linköping University, Tallinn University of Technology and Aalto University trained a classifier to recognise the layered structure of MXenes, thin metal carbide and nitride sheets, and ran it over the Materials Project and OQMD. With database tags and literature checks, it turned up 33 compounds reported as made but never recognised as MXene candidates. The team then made seven MXenes with a self-sustaining combustion reaction that needs no furnace and finishes in minutes, and 11 new ones built on seven rare-earth metals.
Why it matters: Most MXenes are still made by etching with hazardous acids such as hydrofluoric acid, in slow multi-step processes. Here the model's job is not to invent materials but to find ones already in the literature under other names, and those finds pointed the team to rare-earth carbides with halogen surfaces, a region where almost no MXene had been made. The combustion route is a candidate for making them at scale.
Evidence and limits: The classifier only sorts structures already in the databases; whether each of the 33 really behaves as an MXene is not tested. The new phases are multilayer powders, not delaminated sheets, band gaps were measured, but the magnetic states come from calculations. One group's work.
TU Dresden's Morpheus.AI agents build virtual tissue models for 15 published studies on their own, meeting 83-86% of the scoring criteria with GPT-6.1 Sol
What happened: Prerana Chandratre, Lutz Brusch, Simon Syga and colleagues at TUD Dresden University of Technology paired OpenAI language-model agents with Morpheus, the open-source multicellular simulator their group develops, and had them build, run and revise models of 15 published studies with no human in the loop. With GPT-6.1 Sol, the setups that could run the simulator averaged 83-86% of hidden, study-specific criteria, in about 105 to 110 minutes and US$3.74 a model. Five studies had no published model, and the agents' own mechanisms scored about as high.
Why it matters: Building a virtual tissue model, turning observations of cells into a simulation that reproduces them, takes biology, mathematics and programming at once, so it is slow, specialist work. The deciding factor here was letting the agent run its own simulations: without that, scores fell to 0.55 with the larger model and 0.11 with the smaller. The larger model cost 10.8 times as much per run and scored 0.85 against 0.66, the price-for-quality choice a group adopting such agents would face. Because the agent writes only model files that the simulator runs, a modeller can check what it built.
Evidence and limits: The scorer is itself a language model, one of the two the agents used; it agreed with hand scores by three of the authors at r = 0.93 on an earlier version of the benchmark. Costs are computed at OpenAI's list prices. The authors set the criteria, the models reproduce published behaviour rather than predict new experiments, and the full run transcripts are not public.
AuthorsPrerana Chandratre, Lutz Brusch, Simon Syga
BioGeometry closes an A+ round led by China Life Capital, taking its A series past $100M for molecule and virtual-cell models
What happened: Existing investors Shanghai Biomedical Innovation Fund and Fortune Venture Capital put in more; Guokai Kechuang, Ruili Investment and Guofang Capital joined. The Beijing company, founded by Jian Tang of Mila and HEC Montréal, gives no size for this round, only that its A series, which includes a June round of several hundred million yuan, now tops $100M. The money is for GeoFlow, its molecule model, GeoCell, a model of how cells respond to perturbations, a larger automated lab and its own pipeline.
Why it matters: The round funds a move from designing a molecule to choosing what it is for: GeoCell is meant to say which target, which disease and which patients a design suits, and the automated lab is to feed results back into both models. Qiushan Guo, a partner at the Shanghai fund, which co-led the June round, said the company has carried its models from molecules to cells. The only revenue it puts a date on comes from synthetic biology, not drugs: it says biosynthesised borneol, 80% cheaper than plant extraction, will bring in revenue at scale in 2027.
Evidence and limits: The $100M is the A series as a whole, not this round. The IND filings for several bispecific antibodies in 2027, the borneol cost cut and the June claim of a nearly 20% hit rate across more than 20 antibody targets are the company's statements, with no data shown.
Formation Bio, an AI drug developer, sets up a $600M fund of its own to buy and develop 10 to 15 clinical-stage drugs
What happened: The New York company buys or licenses Phase 1 and Phase 2 drugs from other companies and runs their development itself. Its new Drug Development Accelerator Fund has $600M committed by research foundations, endowments, hospitals and philanthropies; Formation Bio is the general partner through a subsidiary and puts in money of its own. Each drug sits in its own company. CEO and co-founder Ben Liu, chief business officer David Steinberg and chief scientific officer Michael Ehlers spoke in the release.
Why it matters: It is another way for an AI company to pay for drug development: Formation Bio raises money against individual drugs, and investors buy into the drugs, not the platform. Liu said it opens paths to liquidity without selling the parent company and should lower its cost of capital over time. The claim to test is that one shared, AI-supported development team takes each drug to a clinical result more cheaply, within 2 to 4 years of buying it.
Evidence and limits: This is a fund for drug programmes, not an equity round in Formation Bio. No limited partner and no drug the fund will buy is named, and the release shows no case where the platform shortened a programme.
Anthropic extends its AI for Science credits to funders, institutes and projects, among them the NSF, the Sanger Institute, EMBL and gnomAD
What happened: Anthropic highlighted 20 organisations in an expansion of its AI for Science programme, 18 named across three groups that get free Claude credits. Six funders and societies, among them the Sloan and Simons foundations, Schmidt Sciences and the US National Science Foundation, pass credits to researchers. Six institutions, including the Wellcome Sanger Institute, the Francis Crick Institute, EMBL and the Space Telescope Science Institute, use them in-house. Six projects get them directly: FlyWire, Columbia's LEAP climate centre, the Broad Institute's gnomAD and the Tau Ceti Lean library.
Why it matters: The programme began by giving credits to individual scientists. Routing them through funders and institutes hands the choice of who uses Claude to organisations that already decide which research gets money, or that run shared resources such as gnomAD and EMBL's databases. Anthropic says many of the projects should produce open datasets, software and standards.
Evidence and limits: The post gives no amount of credits and no eligibility terms, and says most of the collaborations are still at an early stage.
Google DeepMind releases Gemini 4 Carbon, a frontier model built for cybersecurity, first to cyber defenders only and into the US government's pre-release testing
What happened: Google DeepMind announced Gemini 4 Carbon, which it calls its strongest model, announced by Koray Kavukcuoglu (SVP, Google DeepMind). It reports Carbon first on version 5.0 of the Artificial Analysis Intelligence Index at 59, and leads on three cybersecurity benchmarks it cites: finding software flaws (SEC-bench Pro, 88.2%), exploiting them (ExploitGym, 54.8%) and patching them (CWE-bench, 71.0%). It resists every attempt on an indirect prompt-injection test (0.0%), against 0.7% for the earlier Argon. Pricing matches Argon at $2 and $10 per million input and output tokens.
Why it matters: Carbon continues the Gemini 4 line the record already tracks, but where Argon shipped with science case studies and a free-access plan for scientists, Carbon's stated focus is cybersecurity and long agentic tasks. Google is gating it: cyber defenders first through its Fairwind programme, developers and consumers later with no date, and it says it is in the US government's voluntary pre-release testing. The rollout and the government testing are the governance step to watch as these models gain offensive and defensive security skills.
Evidence and limits: Every benchmark figure is Google's own, with no independent check and no technical report cited. The internal coding and secure-code gains are on its own teams and tests. Access is restricted, so nobody outside the vetted partners can test the claims.
Anthropic widens its free-credit AI for Science program from single researchers to funders, institutes and projects, the US National Science Foundation among them
What happened: Anthropic named the partners of its AI for Science program, which gives free Claude credits for research. Six funders and societies will pass credits on to researchers: the Sloan and Simons foundations, Schmidt Sciences, Renaissance Philanthropy's Open Source for Science Fund, the US National Science Foundation and CAS, the American Chemical Society's chemistry index. Six institutes, including the Space Telescope Science Institute, the Wellcome Sanger Institute, the Francis Crick Institute and EMBL, use them in-house; six projects, including FlyWire, gnomAD and the Lean library Tau Ceti, get them directly.
Why it matters: The program began by backing individual scientists. Now funders such as the NSF and private foundations will choose which researchers get frontier-model time, as they already choose who gets grants. Several recipients run resources many labs depend on, such as gnomAD's reference of genetic variation from over 800,000 people and EMBL-EBI's databases, so work done with the credits can end up in shared tools.
Evidence and limits: The post gives no credit amounts, selection rules or dates, and Anthropic says most of the collaborations are at an early stage; no recipient has yet reported what it did with them.
An AI agent at MIT runs a standard low-temperature measurement system on two quantum materials, choosing each next measurement and recovering from an instrument overload itself
What happened: Eunbi Rha, Mingda Li and colleagues at MIT built QUANTA: an agent that draws on literature and past runs, plans transport measurements on a Physical Property Measurement System and revises the plan as data arrive, while a rule-based layer checks every command against hard limits. On the Weyl semimetal TaP it recovered from an amplifier overload, chose a current that passed its noise checks and rejected the faster averaging setting, though none fully passed. On Co2MnGa it tested a nonlinear Hall signal's origin in four steps.
Why it matters: Measuring a quantum material means many small calls, such as which current, how long to average, and whether an odd signal is the sample or the setup, and that know-how mostly lives in people. Here an agent makes those calls on a real instrument and real samples, with the reasoning for each step logged. It also stops where the data stop: it left the Co2MnGa signal's origin open, electronic or thermal, because the fit could not tell the two apart.
Evidence and limits: One instrument, two samples, run by the team that built the agent, with no comparison against a researcher doing the same campaign. People still mount samples and change wiring, and high-risk actions need human sign-off. The agents ran on OpenAI's Astra Ultra model. It cleared the overload by widening the amplifier's input range, without finding the cause. Neither campaign produced a new physical finding. A preprint.
Toronto's El Agente Quntur agents pick the orbitals for 32 excited-state calculations and halve the error of the best fixed protocol against published references
What happened: Juan B. Pérez-Sánchez, Varinia Bernales, Alán Aspuru-Guzik and colleagues at the University of Toronto, with co-authors in Chile and Spain, had El Agente Quntur, a team of agents built on OpenAI's GPT-5.6 models, choose the active space, the small set of orbitals a multireference calculation treats in full, for 32 excited states of 31 molecules. Against published CASPT2 values the agents' energies were off by 0.239 eV on average, against 0.459 eV for the best of six fixed Active Space Finder protocols.
Why it matters: Choosing the active space is the step in these calculations that still takes an expert, and it is why some chemists say the methods are not a proper model chemistry. Here the agents did more than pick orbitals: they checked which excited state a calculation had actually found, noticed orbitals lost during optimisation and changed course. The prompts, agent traces and calculation files are in a public replication dataset, so the decisions can be audited.
Evidence and limits: The reference is another calculation, CASPT2 from earlier papers, not a measured spectrum. For four cases the agents computed a different excited state from the reference, which the average does not penalise, and case by case the agents and the best fixed protocol each won 16. The 32 cases used 764 million tokens, searches included, and repeat runs of one prompt took different routes.
Crystoria adds a vibration stability check to 2.6 million crystals from five public databases and finds about two thirds of them stable
What happened: Ke-Xin Pang and Yan Gao at Yanshan University and Ze-Feng Gao and Zhong-Yi Lu at Renmin University of China took low-energy crystals from Alexandria, Materials Project, JARVIS, OQMD and NIMS HeuslerDB, removed duplicates, added mirror-image twins of some structures, relaxed each with a machine-learned interatomic potential, EquFlashV2, and computed its phonons. Of 2,649,640 structures across 89 elements, 1,793,353 (67.68%) came out dynamically stable: no vibration would make the lattice distort on its own. The data are on Zenodo and the code is under the MIT licence.
Why it matters: Most crystal databases say whether a structure is low in energy, not whether it holds together when its atoms vibrate, because the phonon calculation costs far more than the energy. Generative models trained on those databases inherit that gap. Crystoria puts both labels on the same millions of structures, so a model can be trained on, or scored by, candidates that pass both tests.
Evidence and limits: Every label is computed. Checked against DFT phonons for 9,958 nonmagnetic semiconductors, on the same fixed structures, the potential's stable or unstable call matched 95.11% of the time (91.22% balanced accuracy); that test leaves out errors from its own relaxation step, and it covers only one class of material. The test is harmonic, and no structure was made in a lab.
AuthorsKe-Xin Pang, Ze-Feng Gao, Yan Gao, Zhong-Yi Lu
Edison Scientific's writing model, trained only on real NIH grant abstracts, fools biomedical readers in 55% of side-by-side comparisons
What happened: Edison Scientific's Reinforced Adversarial Learning trains two models together: a generator that writes, and a judge that sees only the final text and estimates whether it came from a set of real examples. The judge's estimate is the generator's only reward. Given a grant's title, the trained model wrote NIH abstracts that evaluators with or pursuing a biomedical PhD mistook for the real one in 55% of pairs, against 44% before training and 21 to 37% for frontier models.
Why it matters: Most ways of training a model for a task need a scoring rule, and good scientific writing has none an expert can write down. This method learns from good examples alone, and Edison sells Kosmos products that draft regulatory documents. The result also bears on reviewers: more than half the time, trained readers could not tell the machine's abstract from the real one.
Evidence and limits: A preprint from the company. Abstracts only, not full proposals, and nothing tests whether the content is sound or fundable. Five language-model judges were fooled about 35% of the time, less often than the people.
Basel's Atlas Bio raises a $10M seed led by Creandum to run drug trials on virtual patients before Phase II
What happened: Creandum led the seed, and its partner Katharina Wilhelm joins the board. The company also disclosed an earlier $2.5M pre-seed led by First Momentum Ventures and LEA Partners, for $12.5M in all. Simon Romanski (CEO) and Sebastian Bayer (CTO, who led language-model research at Google) founded it in 2024. Its model, trained on about 2 million RNA readouts from human tissue, predicts how each patient in a virtual cohort responds to a treatment. The money goes to hiring, validating the platform and new pharma clients.
Why it matters: Phase II is where most drugs fail, and where a developer learns which disease and which patients a molecule suits. Atlas wants to answer that before the trial, by running the same virtual patients through several treatments and diseases, a comparison no real trial can make. It is the second in silico trial company to raise money in two days, after adsilico's virtual patients for heart devices; this one is for drugs, starting in immunology. Wilhelm calls clinical development the second wave of AI in drugs, after drug design.
Evidence and limits: The case studies on the company's site, including a predicted Phase II miss that the trial then confirmed, are its own, with methods and cohorts shown to partners only under NDA. That a Basel client skipped animal tests on the strength of its simulation is the company's statement, and the client is unnamed.
BigHat and Anthropic improve two antibody leads over seven rounds, after finding one thermal measure did not measure what its name said
What happened: BigHat Biosciences and Anthropic ran two lead-optimisation campaigns on BigHat's Milliner platform, about 1,010 designs over seven rounds in all, four on one molecule and three on the other, with Claude in the loop. Reading the raw instrument curves, they found a thermal measure on each scaffold did not measure what its name implied: on the nanobody, the clumping-onset temperature. They fitted a physical model of the curves, chose new measures and changed the protocol. Both leads beat their parents on binding, heat stability and clumping.
Why it matters: A closed-loop lab optimises the number it is given, so a measure that stands in badly for what a drug needs cannot be rescued by a better model or more data. The gain here came from people and a model reading the instrument signal and changing the experiment, not from more rounds of search. Published lab-in-the-loop work mostly takes the reported assay value as a correct label; this campaign did not.
Evidence and limits: Both companies' own report, not peer reviewed, every measurement in BigHat's lab: two molecules on one platform. The report says each lead's values rest on one or two replicates and that confirmation runs come first, and the new measures were chosen inside the campaign that produced them. The leads are antibody fragments, not drugs, and nothing is tested in animals or people.
Generate:Biomedicines cuts a CAR's signalling tail to 45 amino acids, under a third of the clinical design, and reports stronger function
What happened: Over three rounds of design and measurement with AI agents, the company built a 45-amino-acid signalling tail for a CAR, less than a third the length of the 4-1BB-CD3zeta domain in approved therapies. Against that domain it raised receptor levels on the cell surface and antigen-driven IFN-gamma release, with none detected when antigen was absent; it kept killing going longer after mRNA delivery, carried four receptors on one mRNA dose, and depleted more B cells in donors' blood cells after lentiviral and lipid-nanoparticle delivery.
Why it matters: Shortening the tail was expected to cost function, because more activation motifs were read as stronger signalling. The short tail made more receptor per unit of mRNA, which is why the team could put four receptors on one dose, and the report suggests compact tails may lower the dose needed to make CAR T cells inside the body. It is also a worked case of agent-guided design: the agents proposed designs and revised them as measurements came in.
Evidence and limits: The company's own technical report, not peer reviewed, and all of it in cells: human blood cells from a few donors, no animal test. The report says how compact tails perform in a living animal, and how many cells carry all four receptors, remain unknown; exhaustion was not measured, and whether agent-guided design beats expert design, or needs fewer rounds, is untested against a matched baseline.
Across its runs, Enveda's document reviewer raised all 11 issues human reviewers found in 1,061 pages of batch records, and 7 they had not
What happened: Enveda tested CDP, its Claude-based checker for drug manufacturing batch records, on four executed records from one clinical campaign: 1,061 largely handwritten pages. Human reviewers had raised 11 actionable issues; CDP raised nine as direct matches and two judged intended matches, across its runs rather than in one pass. Single runs on Claude Opus 4.7 added 7 actionable comments no reviewer had raised. A review holding CDP's comments took 9.4 times less time.
Why it matters: Batch-record review is required before a batch of a drug is released, and Anthropic's post says Enveda's manufacturing chemists had been spending up to 75% of their time on paperwork. This is the regulated-paperwork end of drug development, which rarely gets measured at all, and the yardstick here is what human reviewers caught on real records rather than a score on a benchmark.
Evidence and limits: The company's own report, not peer reviewed: four records from one campaign, checked against reviewers' comments, two of the four reviewer arms being outside consultants. No CDP comment cited content the records do not contain, and each run also posted one misread at Major, and the human reviewers could consult earlier-step records CDP was not given. Whether the 7 extra comments matter is Enveda's judgment. The time figure comes from a review holding CDP's comments in advance.
Anthropic buys a one-person drug company and starts running drug programmes of its own
What happened: On its Science page Anthropic said it has acquired Mithril Medicine, which was developing treatments for two rare childhood muscular dystrophies, and that its founder, Cody Tranbarger, has joined the company to continue that work and start further programmes. The same post reports research partnerships with six organisations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate:Biomedicines and The Column Group, which is setting up a company, Pleiades, to run drug programmes with Claude.
Why it matters: A model supplier has become a drug developer. The partnerships put Anthropic's own scientists inside other companies' programmes, and the acquisition gives it programmes it owns, so it now carries the risk and would hold the result. The post also says Anthropic plans to scale this work up substantially.
Evidence and limits: No terms were disclosed for the acquisition. Anthropic says that if the candidates continue to hold up it expects a drug to be ready to start human trials as soon as late 2028, which is its own forecast, and the mouse results behind it are in a technical report it published itself, not peer reviewed; that report studies two tool molecules in a pilot study, not a candidate. Pleiades has no disclosed programme yet.
A Claude-steered campaign designs 20 antibody arms that pull IgE off its receptor; of the eight titrated, three remove more than the approved drug does in the same assay and none clearly beats the parent
What happened: Omalizumab, the only approved biologic for food allergy, neutralises circulating IgE but does not effectively remove the kind bound to FcεRI on mast cells. Guided by its research team, Claude ran 22 phases toward a two-armed antibody that would prise it off. Of 91 disruptor designs screened at one concentration, 20 beat every negative control; eight were purified and titrated, three of them removing more than omalizumab from 25 to 400 nM. Three of 57 anchor designs bound IgE, at apparent affinities of 3.4 to 235 nM.
Why it matters: Pulling bound IgE off the receptor needs a precise geometry, which display and immunisation libraries cannot specify, so designing from a structure is the point here rather than a convenience. Activity survived as many as eight substitutions from the parent, and hits came from every design scope and method the campaign tried. Divergence says the two arms came in weeks rather than the months screening takes.
Evidence and limits: Divergence's own report, not peer reviewed, all biophysics on a sensor chip: no cells, no animals, the two-armed molecule not yet made. In the titration the report says the variants did not appear to improve on the parent sdAb026, the best of them matching it at most concentrations, and Anthropic's post summarising the work does not mention the parent. The anchor affinities may be confounded by avidity, and the method needed an existing disruptor's sequence and solved structure.
DeepMind releases Genie 4 to trusted testers, a world model it pitches as a training environment for robots and agents
What happened: The model generates video together with music, ambient sound and interactive dialogue, which DeepMind calls its first real-time world model to generate audio natively, and it takes natural-language edits while a simulation runs, to change the weather, move the camera or add a person. It builds on the Gemini family, with Gemini Omni anchoring real-world appearance, and every output will carry a SynthID watermark.
Why it matters: DeepMind's stated purposes include developing training environments for robots and agents, and testing AI systems in different environments before taking them to the real world. One clip is generated from the prompt "Universal Robots UR5 holds a pipette and dispenses liquid into a lab plate, triggering tiny electric sparks", the only lab scene in a post whose testers are creative industries, journalism, education, gaming and engineering.
Evidence and limits: No benchmark, score or comparison with Genie 3 appears: it is clips, captions and an internal safety test with no result given. The two-minute consistency figure in press coverage is in neither post; neither gives a session length, and the post says object permanence can drift the longer a simulation runs. Trusted testers only, no public access.
Omabit raises a $2M seed led by Bright Ventures for an AI that analyses drug developers' own trial and genomic data
What happened: The San Francisco company, in Stanford's StartX summer 2026 cohort, puts a drug developer's trial, RNA-seq, variant and other data beside published studies, and runs and writes up the analysis a scientist asks for in plain language, starting with teams planning the next trial from completed ones. CEO and co-founder Elena Cassidy, Bright Ventures founder Lenore Champagne Beirne and scientific adviser Michael Snyder spoke in the release. The money goes to the platform and to hiring scientists and engineers.
Why it matters: It is a bet that the slow step in clinical development is analysis, not data: companies already hold the trial and genomic results that could explain who responded, but they sit in separate studies and systems. Omabit sells the joining-up and the analysis, including predictive models built on the company's own data, rather than new data.
Evidence and limits: The site's claim that an analysis takes a day instead of three months is the company's, with no case shown. No customer is named.
Biographica's CROP-1 predicts a crop gene's activity in each tissue and condition from its DNA plus which regulators are on, and designs edits aimed at one context
What happened: In a bioRxiv preprint, Paride Antinucci, Dominic Hall and colleagues at Biographica, a London crop-AI startup, feed a model a gene's DNA plus the measured levels of transcription factors, the proteins that switch genes on, in each tissue or condition. On held-out data from six plant species its median per-gene correlation across contexts is 0.831, against 0.486 for the same model without that input, and 57-246% above three fine-tuned genomic foundation models. It then designed edits of up to 20 DNA letters in soybean and tomato.
Why it matters: Many crop genes affect more than one trait, so an edit that helps in one tissue or condition can cause unwanted changes in others. A model that predicts expression per context lets a breeder rank candidate edits by the change wanted in one place and the change everywhere else, before any plant is edited. Biographica sells gene-editing targets to seed companies.
Evidence and limits: The edits exist only as predictions: none has been made in a plant and measured. The test data come from the public PlantExp collection and were scored by the company's own team, and the preprint is not peer reviewed.
Singapore puts S$150 million into MARVEL-AI, an A*STAR platform that pairs materials AI models with self-driving labs, with CuspAI among its partners
What happened: Deputy Prime Minister Gan Kim Yong announced MARVEL-AI at SWITCH 2026: a platform led by A*STAR, starting with S$150 million of government money, in which AI models pick candidate materials and design experiments and self-driving labs make and test them. It is anchored at A*STAR's Institute of Materials Research and Engineering, whose executive director, Xian Jun Loh, also leads MARVEL-AI. Named partners include ST Engineering, Heraeus Electronics and Foxconn Industrial Internet on materials, and CuspAI and Google Cloud on the AI side.
Why it matters: The speech names the gap it is meant to close: models such as Google's GNoME and Microsoft's MatterGen propose materials, but a material still has to be made, tested and manufactured before industry can use it. MARVEL-AI puts the models and the labs in one national programme and opens its equipment, testing and researchers to companies, startups and SMEs included. CuspAI, which designs materials with AI, will connect its system to the platform's automated labs.
Evidence and limits: This is an announcement: the speech and reports give no split of the S$150 million, no timeline and no material yet. The figure of up to 470 alloy samples a day, more than 400 times conventional throughput, describes an existing lab process Enterprise Singapore cited, not a MARVEL-AI result.
Crystal-structure prediction finds all seven solved forms of acridine among 15 picks, and of the machine-learned potentials tried, only one tuned on acridine gets their stability order right
What happened: Adam Lahouari, Jutta Rogal and colleagues at New York University and the Flatiron Institute studied acridine, a small molecule with at least eight crystal forms. Melt experiments solved the structure of form VIII for the first time. Their CrystalMath method kept 15 of 341,683 generated structures without computing any energy, and the 15 include all seven forms with known structures, form VIII among them. A MACE potential fine-tuned on acridine, with full free-energy calculations, reproduced the switch from form II to form III as temperature rises that experiments show.
Why it matters: One molecule can crystallize in several forms that behave differently, which matters for drugs, organic electronics and agrochemicals, and crystal-structure prediction with machine-learned potentials is how the field hopes to rank the forms before making them. Here the ranking step needed tuning: four off-the-shelf foundation potentials put the forms in the wrong order, and simpler free-energy approximations missed the observed crossovers. The study also grows acridine from the melt with 18 additives and finds the two strongest effects match computed defect energies.
Evidence and limits: One molecule. The predicted crossover between the two most stable forms sits near 290 K, against about 318 K measured, some observed transformations disagree with the free-energy curves, and form IV was left out of them. The additive link rests on two of the 18 additives. A preprint.
A normalizing flow computes the lattice free energies used to rank drug crystal forms, matching the standard method with about 2.5 times less computing time on average
What happened: Edgar Olehnovics, Matteo Salvalaglio and colleagues at University College London, AbbVie and the University of Colorado Boulder train a normalizing flow that maps molecular-dynamics snapshots of a crystal onto a simple reference state, which gives its free energy directly, without the chain of intermediate simulations the Einstein crystal method needs. On succinic acid and two drug molecules, Veliparib and Mivebresib, the results agree with that method, in crystals of up to about 3,000 degrees of freedom; across all systems it needed about 2.5 times less computing time on average.
Why it matters: Crystal-structure prediction can propose hundreds of possible packings for one drug molecule, and ranking them by stability at room temperature needs free energies. The standard method is accurate but too slow to run across a whole landscape. A method that needs one ordinary simulation per crystal form makes that ranking cheap enough to automate, and because it needs fewer energy evaluations it suits slower, more accurate force fields.
Evidence and limits: Checked against another computed method, not against experiment. The ranking depends on the force field: for Veliparib a generic force field and a tailor-made one disagree. The speed-up is uneven: for succinic acid it was faster only in the smallest crystals, and exact runtimes were not recorded.
GlycoWeave, a diffusion model built for sugars, places glycans in protein pockets far more often than AlphaFold 3 on a 698-complex benchmark
What happened: Huifeng Zhao, Tingjun Hou, Yu Kang and colleagues at Zhejiang University and Peking University built GlycoWeave, which predicts the shape of a sugar chain bound to a protein by adjusting its bond torsions rather than moving atoms freely, which keeps the sugar's chemical structure intact. On 698 recent protein-sugar complexes, 71.9% had a high-quality pose among its top 25, against 19.2% for AlphaFold 3 and 16.9% for the docking tool Vina-Carb. Each pose took 0.34 seconds.
Why it matters: Sugars on proteins shape immune recognition, infection and how some drugs act, but they are flexible and thinly covered in experimental structures, so general structure predictors handle them poorly. Here only 41.0% of AlphaFold 3's top predictions passed the structural-validity check, against all 698 of GlycoWeave's (the docking tool Vina-Carb reached 99.7%). A model built for one awkward class of ligand measures how much accuracy a general co-folding model gives up on molecules unlike its training data.
Evidence and limits: Checked against known experimental structures, not new wet-lab work. The headline compares the top 25 poses; the paper's main text does not state AlphaFold 3's own sample count for that comparison. The protein pocket is given as input, so this is not blind docking. A preprint.
In an Anthropic trial, analysts working with Claude Code went with a lean secretly assigned to the agent on 86% of their analysis choices
What happened: Amaç Herdağdelen, Hailey Joren and Alex Tamkin had 50 analysts trained in causal inference judge a fictional farm programme with Claude Code. On each of four choices with two reasonable answers, such as how to measure productivity, a coin flip set which way the agent leaned, unknown to the analysts. They matched the lean 86% of the time, against 50% if it made no difference. Afterwards 77% gave the agent's arguments as their reasons in at least one answer; 15% credited it.
Why it matters: As agents take over the hands-on work of an analysis, the choices left to the researcher, what to measure and how to compare, are where judgment enters. This trial measures how much of that judgment moved to the agent: the coin flips alone predicted 82% of final recommendations, against 61% by always guessing the commonest one. It bears on any lab using agents to analyse experiments, and on review, since a reviewer working with an agent is an independent check only if the calls are their own.
Evidence and limits: The task was fictional and the agent was told to argue its lean, so real agents may sway people less. The 11 who worked without the agent made more serious errors, but they were fewer and less experienced, so the comparison is weak. The study is Anthropic's own.
Lam Research widens its research partnership with imec to 3-D chip stacking and to its dry resist for the newest EUV lithography
What happened: Lam Research, which makes deposition and etch tools, has added two programmes to its long collaboration with imec, the Belgian chip research centre. One readies 3-D integration for factories: thick insulating films, stress compensation, plasma dicing, and charge-trap materials and low-resistance wiring for stacked memory. The other tests Lam's Aether dry resist in imec's high-NA EUV research environment, aiming at sharper printing with fewer random defects. Earlier joint work printed 20 nm-pitch ruthenium lines in one exposure, all electrically sound on bridge tests.
Why it matters: Stacking chips and the next lithography step each bring materials problems that must be solved before volume production: films whose stress stays under control, materials that trap charge in stacked memory, and resists that print lines 20 nm apart without random defects. An equipment maker and a research centre are naming those problems and putting joint research on them, which shows where materials and process research will be asked to help next.
Evidence and limits: This is Lam's own blog post, with no budget, timeline or new result; the yield figures come from earlier joint work. The post says nothing about using AI.
Leeds spinout adsilico raises a £10M Series A to test heart valves and stents on AI-built virtual patients
What happened: PXN Ventures led the round, with existing investors Northern Gritstone, which put in £5M, and Parkwalk. adsilico was founded in 2024 by Sheena Macpherson, its CEO, Nishant Ravikumar and Zeike Taylor. Its Virtual Chimera platform builds virtual hearts and blood vessels from anonymised scans with generative AI and runs physics simulations of a device inside each. The money goes to the heart models, neurovascular devices and a team growing from 14 to 34.
Why it matters: A device trial enrols a few hundred patients, so many body types are never tested. adsilico's answer pairs a generative model of anatomy with physics simulation, an in silico trial: the device is tested on thousands of virtual bodies before real ones, in a field where regulators decide what evidence counts.
Evidence and limits: The company names completed projects with heart-valve and neurovascular device makers but no customer, and no validation of the virtual patients against real outcomes is cited. That PXN led comes from Tech.eu; Northern Gritstone's own post lists the investors without saying who led.
Anthropic files to list as ANT and names scientific research and lab robotics among the new markets frontier AI may open
What happened: The S-1, filed on 26 October, shows $16.3bn of revenue for the first half of 2026 against $4.6bn for all of 2025, and an operating loss narrowed to $693m from $2.1bn. The $80.7bn net loss is almost all a non-cash $72.4bn remeasurement of convertible notes, derivatives and SAFEs. Compute commitments include at least $114.5bn with Google, $110bn with Amazon and $44.9bn with Microsoft, owed regardless of usage.
Why it matters: For science, the filing says frontier AI 'may extend' to robotics and to scientific research, including drug discovery and materials science, and that Claude with the Model Hardware Standard can 'now' operate microscopes, liquid handlers and robotic arms. Its vision of 'Powerful AI' adds tasks running for weeks and 'compressing decades of scientific progress into years'. In the accounts, though, there is one segment, and the life sciences lab appears among the risk factors.
Evidence and limits: The June figures are unaudited interim ones. The three cloud commitments are three of seven named, with $125bn to $150bn more TPU compute estimated outside the table, and the aggregate line is blank. The top five customers were about a quarter of revenue. A Founder LLC's new share class carries 50.1% of the vote, and the safety trust's own class elects a majority of the board.
A designed protein scaffold binds a metal photocatalyst, and five evolved mutations turn it into an enantioselective enzyme reaching 50,000 turnovers
What happened: Jared Lewis's group at Indiana University, with David Baker, tested 96 designs; 63 expressed, and one scaffold called DE3 bound one mirror image of an iridium photocatalyst at 94% enantiomeric excess. As designed it gave low selectivity in catalysis. Screening mutations reached a 3:97 enantiomeric ratio, with turnovers above 50,000 in air and the protein reusable over three cycles.
Why it matters: Generative design supplied a pocket that holds a cofactor nature has no protein for, which is what the field could not get from existing folds. On this reaction the enantioselectivity came from screening mutations afterwards; a second design campaign later reached 90:10 on a different reaction with no evolution at all. Metal polypyridine photocatalysts had resisted both routes until now.
Evidence and limits: One scaffold carried through, out of five folds tested for binding. The pocket was designed around a ruthenium cofactor and bound only iridium analogues, and no cofactor-bound structure was obtained. Other reactions in the scope are only moderately selective. Designed and evaluated by the same groups. An accelerated preview.
Texas A&M gets $2M to build a robot-run foundry that designs, melts and tests new metal alloys
What happened: A grant from the WoodNext Foundation, the philanthropy of Roku founder Anthony Wood and Susan Wood, funds the Aggie AI Foundry in the university's materials science department. AI trained on physical models and past alloy work proposes compositions, and robot arms mix, melt and test them at up to 3,000 °C. Department head Ibrahim Karaman and Raymundo Arróyave lead it; the university says it will cut alloy development from months to a week or less.
Why it matters: The university says existing self-driving labs handle materials that are easy to make and process, while new alloys have to be melted and tested at very high temperatures; Arróyave adds that no computer can yet predict how one will behave. A lab that runs that whole loop for metals would test whether AI-led, robot-run discovery can reach the structural materials used in energy, aerospace and medicine.
Evidence and limits: This is a grant to build the lab, not a working lab: no opening date, throughput or result is given. That it is the first lab of its kind in academia, and the months-to-a-week speed-up, are the university's own claims.
OpenAI and Anthropic each claim a subexponential 3-SAT algorithm refuting the Exponential Time Hypothesis, each checked in Lean, neither checked by anyone outside
What happened: OpenAI's manuscript, dated 17 October and revised 25 October, went into its public repository on the 26th with a Lean formalization and gives a deterministic algorithm that counts the satisfying assignments of fixed-width CNF formulas in subexponential time. Anthropic's 49-page paper, dated 26 October and pushed on the 27th, claims the same thing and prints its author as Claude, Anthropic. It dates its model's discovery to 16 October.
Why it matters: The hypothesis has organised decades of lower-bound work: Anthropic's paper says a literature search by Claude found more than 800 papers proving a result of their own only under it or something stronger. If either proof holds, those conditional results have to be read again, and Ryan Williams, who is helping organise the review, says that if true it should move every computer scientist's weight a little toward P=NP.
Evidence and limits: Nobody outside either company has verified either proof. Lean checks that a proof follows from a formal statement, not that the statement means what the paper claims; Anthropic writes that down and adds that at the time of writing no person is known to have checked its Lean. OpenAI's scope note formalizes a weaker bound than its manuscript, on canonically encoded inputs, still subexponential. Anthropic calls it purely asymptotic.
OpenAI arranged its examiners before announcing, and the group it arranged wrote to its field to say the involvement is not certification
What happened: OpenAI told three complexity theorists it had the result and would release it, and asked them to organise its examination. Russell Impagliazzo and Ramamohan Paturi at UC San Diego and Ryan Williams at MIT wrote to their field the same day: "Our involvement should not be taken as certification of the claim." Williams said on his own account that the group signed no NDAs and takes no pay, and that OpenAI agreed to release its raw data and sponsor a conference.
Why it matters: Who certifies a machine's result is what the last six weeks have been about, and here a lab named its examiners before releasing the claim. The letter is them refusing the role that implies while keeping their leverage: Williams says they asked for the raw data and a sponsored conference and OpenAI agreed, and the letter says the effort should confer ownership on no company and no chosen few, themselves included.
Evidence and limits: OpenAI's own post confirms the terms Williams described: no NDAs, no pay, the raw data public and a conference it funds. It also names a second group, led by David Gross, that is holding back the materials for a physics result; nothing from that result is public. The letter itself is undated.
A claimed proof of the local smoothing conjecture in every dimension carries the Kakeya conjecture with it, where the published record had reached three dimensions
What happened: OpenAI's 59-page paper, its author line reading OpenAI, dated 25 October and added to its public repository the next day, proves Sogge's conjecture for every spatial dimension at least two, with arbitrarily small Sobolev loss and none at p=2. From it the paper deduces the Bochner-Riesz conjecture, sharp bounds for Stein's square function, and the Kakeya and Nikodym maximal conjectures. OpenAI says it is releasing the model's output directly.
Why it matters: The published record had the plane from Larry Guth, Hong Wang and Ruixiang Zhang, and three-dimensional Kakeya from Hong Wang and Joshua Zahl last year; this claims every dimension at once, and cites two OpenAI preprints from September that it supersedes. Unlike the complexity result released the same day, the Lean here covers the headline claims: the local smoothing bound in every dimension, and that a set holding a unit segment in every direction has full Hausdorff dimension.
Evidence and limits: Nobody outside OpenAI has checked it, and OpenAI's own repository calls these manuscripts raw research data rather than polished papers. The restriction estimates for the sphere, paraboloid and cone hold only inside the conjectured range, so that conjecture is not settled here. The same repository withdrew three manuscripts on 7 October over a sign error.
POISK, 89,074 antibody-antigen records mined out of patents, asks whether structure models can pick the real binder: the best manages 72%
What happened: A Roche team at Prescient Design and Large Molecule Research used a language-model pipeline to pull 89,074 antibody-antigen records out of 10,110 US patents, then built 11,247 pairs: an antibody a patent says binds, with no solved structure, against a sequence-similar decoy. Six structure models were scored. Protenix-v2 led at 72% of pairs and 87% on its most confident half, and picked the true binder out of 50 candidates 29% of the time. The best model touched a listed epitope residue in only 62.8% of cases.
Why it matters: Teams screening designed antibodies read a model's folding confidence as a stand-in for binding. This puts a number on what that is worth: much better than picking at random, nowhere near enough to name the one candidate to make. And telling a binder from a decoy does not mean the pose is right, so reading an epitope off these models is a separate claim these numbers do not support.
Evidence and limits: Nothing here was assayed: the binders are what patents claim, the non-binders are repertoire antibodies assumed not to bind, so the accuracy is against a constructed negative set rather than measured affinities. The structures were not in the models' training data; the patent text may have been. The preprint is CC BY and says the dataset is released, without stating its licence. Not peer reviewed, and the full text was not read.
AuthorsDaria Tupikina, Andrea Roncoli, Alexander Bujotzek, Brennan Abanades
ChronoSyn learns the order in which materials were first reported, and picks out real materials more often than the same model with the years shuffled
What happened: Junkil Park and Aron Walsh at Imperial College London gave a composition-only synthesizability model a second task: rank known materials by the year each was first reported in the ICSD crystal database. On about 10,000 held-out reported materials it recovered 52.4%, against 47.0% for the model it extends and 48.4% with the years shuffled, all at one calibrated threshold where chance is 8.1%. Trained only on materials reported up to 2018, it put 7 later-reported compositions in its top 100, against 3 for the baseline.
Why it matters: Generative models now propose candidate materials far faster than labs can try them, so the question is which to make first. Synthesizability models usually treat every reported material as an equal example. This one uses the reporting order as free extra supervision, on the authors' reading that the order carries which precursors, methods and applications made each material reachable. The shuffled-year control is what ties most of the gain to the order itself rather than to the added task.
Evidence and limits: All checks run on existing databases. The authors say learning from history may favour chemistry already studied and score unexplored families lower. It reads composition only. The authors say the largest gaps come in the top 100 at the later cutoffs, where the statistical uncertainty is also larger. A preprint.
Axiom Math's agents formalize the classification of finite simple groups in Lean: about 43 million lines in 24 days
What happened: The startup released what it calls the first complete computer-checked proof of the classification, whose published proof runs to some 15,000 pages in over 500 articles. Dozens of agents, directed by its mathematicians and engineers, wrote about 43.1 million lines of Lean, following the first-generation proof and reusing 1.2 million lines from the Qiuzhen-CFSG project. Group theorists Richard Lyons and Ronald Solomon advised. The proof is public under Apache 2.0.
Why it matters: Results across algebra, number theory and combinatorics cite this theorem in a line; now every step under it can be checked by machine. The work also found 34 places where the published proofs needed repair. The largest, in John Walter's 1986 argument for the Unbalanced Group Theorem, was closed by another route: drawing on Solomon's 1976 work, the company's prover proved a new theorem that the team wrote up in a paper. Claude's Fermat's Last Theorem formalization in September ran to 13 million lines.
Evidence and limits: The figures and the list of repairs are Axiom's own, and no group theorist outside the project has reported checking the result; people still have to read the formal statement to confirm it says what the classification says. Building the proof takes close to a day and relies on changes to Lean's build tool that are not yet upstream.
OpenAI says its models proved that four-dimensional Yang-Mills theory and quantum chromodynamics exist as rigorous mathematical theories, and leaves the release to a group led by David Gross
What happened: OpenAI says its models produced a natural-language proof that pure Yang-Mills theory exists on 9 September and one for quantum chromodynamics on 18 September, building on Balaban's approach, and that it finished a Lean formalization of the Yang-Mills part on 16 October. The proof also covers the simplest grand unified theory containing the Standard Model's gauge fields and fermions. A group led by David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, will examine it.
Why it matters: Pure Yang-Mills existence is part of a Clay Millennium Prize problem, and the result, if it holds, would put the theory of the strong force on a rigorous footing. It is also a test of how a lab releases an AI result: here the outside group decides, and it chose to write the work up itself before any AI-produced material is released.
Evidence and limits: Nothing is public: the proof and the Lean files are with the Gross group, which is writing the work up itself, so no one outside it has checked the claim. OpenAI reports a Lean formalization for the Yang-Mills part only, none for quantum chromodynamics. All of this is OpenAI's own account.
A small add-on layer corrects a foundation interatomic model's melting points without retraining it: gold goes from 844 K to 1346 K, against 1337 K assessed
What happened: Sarath Menon, Braulio Sánchez Ortega and Ralf Drautz at ICAMS, Ruhr University Bochum, add a linear correction on top of GRACE, a foundation interatomic potential, fitted to free energies computed in simulation so that phase transitions land on the assessed SGTE values. Across 11 metals and 14 transitions, every corrected temperature falls within the set tolerance: 15 K for melting, 50 K for solid to solid. Gold melted at 844 K in GRACE and at 1346 K after correction. For titanium it took about 55 H100 GPU hours.
Why it matters: Foundation interatomic potentials learn from DFT data, and an error of a few meV per atom in which phase is more stable can move a melting point by hundreds of kelvin. This layer fixes the balance between phases without touching the base model: it needs only the energies and structures the model produces, not its parameters or training data, so in principle it fits any potential, open or closed. Copper's melting curve, fitted only at zero pressure, then follows measurements within 15 K up to 4 GPa.
Evidence and limits: The corrected temperatures are fitted to the SGTE values, so hitting them is calibration, not prediction, as the authors say. The layer can only rebalance phases the base model already describes and cannot add missing physics such as magnetism. Where the correction is large it moves other properties too, mostly towards experiment: liquid gold and silver become 3 to 6% denser, but vacancy energies move away from experiment for vanadium and aluminium.
Insilico's AI-designed TEAD inhibitor produced partial responses in 4 of 22 evaluable mesothelioma patients in its first human trial
What happened: At ESMO 2026, Melina Marmarelis of Penn's Abramson Cancer Center presented first human data for ISM6331, a pill Insilico designed with its Chemistry42 generative platform to block TEAD, proteins at the end of the Hippo growth pathway. By the 14 August cutoff 36 people had taken 15 to 130 mg a day. Of 22 evaluable mesothelioma patients on 30 mg or more, 4 had a partial response and 68% disease control; at 130 mg, 2 of 5. No dose-limiting toxicity; 8.3% had a drug-related grade 3 event.
Why it matters: TEAD has long been hard to block with a small molecule, and mesothelioma has few treatments once the first ones fail. A drug Insilico says its generative models designed is now shrinking tumours in people who had up to six earlier treatments, and falling blood levels of CCN2, a protein TEAD switches on, suggest it reaches its target in patients. Subject to further data, Insilico plans a randomised Phase 2a to choose the dose.
Evidence and limits: A single-arm Phase 1 still raising its dose, with no control group, and only the company's summary of the talk is public. The 40% headline is 2 of 5 people. The release does not say what the models did beyond naming Chemistry42, and a Phase 1 response rate is not evidence of longer life.
ByteDance Seed and Peking University build exact total spin into neural network wavefunctions, and resolve the spin ladders of iron-sulfur clusters
What happened: Ruichen Li, Ji Chen, Weiluo Ren and colleagues at ByteDance Seed and Peking University split a real-space wavefunction into a spatial part, computed by a neural network (LapNet), and a spin part built to carry one exact total spin. Earlier neural network methods added a spin penalty to training, which can leave the wavefunction stuck mixing spins. The method matched experimental singlet-triplet gaps of biradical molecules about as well as the penalty method, gave an antiferromagnetic spin ladder for [Fe2S2] and a spin gap for an [Fe4S4] cluster.
Why it matters: Iron-sulfur clusters are central to biological nitrogen fixation, and their many spin states of nearly equal energy are where standard methods such as DFT, and even CCSD(T), can go wrong. With spin built in by construction, a calculation cannot drift into the wrong spin state, and chemists' own ideas (core and active orbitals, a first guess at how the spins couple) can be written into the network. With 10 active orbitals it reproduced the local spin trend on [Fe2S2(SCH3)4] that DMRG needed 32 to show.
Evidence and limits: Everything is computed. The iron-sulfur results have no experimental reference: they are compared with other calculations (FCIQMC, DMRG), and for [Fe4S4] the authors say the order of the three methods' gaps is not tied to any one cause. The November 2025 preprint called that gap more accurate; the published version drops the claim. Only the biradical gaps are checked against experiment.
The UK funds an AI Biosecurity Lab with £100m, alongside up to £200m in credits from OpenAI and Anthropic
What happened: The Cabinet Office announced the AI Biosecurity Lab, an independent non-profit it calls the world's first dedicated AI biosecurity organisation, meant to detect biological threats earlier and speed up vaccines and other countermeasures. OpenAI is putting in £100m of compute and model credits through its Rosalind Biodefense programme, Anthropic up to £100m in credits, access and support. Google DeepMind and Basecamp Research join as research partners, not funders.
Why it matters: The £300m headline is £100m of public money; the rest is credits and access rather than cash, and the research partners give access, not money. It also puts frontier labs inside a national function: a National AI Biosecurity Initiative puts the lab with UKHSA, APHA and Dstl, and informs the MHRA's regulation of AI-designed medicines.
Evidence and limits: No budget breakdown, no timeline and no named staff; the scientific advisory committee is still to be announced. Anthropic's £100m is an 'up to' figure, and the page gives no period for either.
Calibrated uncertainty carries only unevenly out of distribution, and in active learning it mostly beats neither random sampling nor no calibration at all
What happened: A Toronto, NIST and KAUST team compared deep-ensemble uncertainty against loss-landscape estimates: ALIGNN on band gap and formation energy, a multilayer perceptron on solubility. Calibration methods generalised only variably from in-domain to out-of-domain data, and calibrated uncertainties generally failed to cut the data an active-learning campaign needed on out-of-distribution data. Random forests and XGBoost on the same tasks showed the same poor uncertainty.
Why it matters: Active learning picks the next sample by uncertainty, and a campaign aimed at new chemistry is out of distribution by construction. The same failure in tree models as in the network points at the data: at least partly intrinsic to it, the authors say, rather than model capacity alone.
Evidence and limits: All computed, on three public property tasks. The claim is about these architectures and calibration methods on these datasets, not a proof that no calibration can work. Peer reviewed.
Pathos AI raises a $237M Series E led by Bristol Myers Squibb at about $3B, and takes its trial-steering AI from cancer into Alzheimer's with BMS and Tempus
What happened: Bristol Myers Squibb led; the company puts the post-money valuation at about $3B, up from $1.6B at its $365M Series D in May 2025. Pathos, a New York company co-founded in 2022 by Eric Lefkofsky and Ryan Fukushima and run by CEO Iker Huerga, licenses drugs other companies made and uses its agent platform, Foundry, to pick patients and adjust trials. BMS gets non-exclusive access to Foundry and three-year exclusive rights in neuroscience; Tempus will build a neurology patient dataset that Pathos licenses for $55M over five years.
Why it matters: Pathos puts its AI on the decisions that decide whether a trial succeeds: choosing which existing drug to run, in which patients, and when to change the trial. It does not design molecules. A drugmaker leading the round and taking access to the platform shows that one large pharma company will put money behind this use of AI, and the deal carries the method from cancer to Alzheimer's, a field with long trials and many failures. Simon Davies, who heads data and quantitative sciences at BMS, spoke for it in the announcement.
Evidence and limits: The round, the valuation and the deal terms are all from the company's two releases of 23 October; neither names the other investors or says whether BMS pays for its access beyond its stake. That Foundry identified JSKN016, an antibody-drug conjugate licensed from Alphamab, and has improved any trial's odds is the company's own account: no trial result from its approach has been published.
Xaira built a camera-guided arm for instruments made for human hands, and over seven days it loaded 18 plates into an HPLC, 80% more than its model of weekday hours
What happened: Bo Wang, Ci Chu and colleagues describe X-Automate in a companion post to X-Scientist: an off-the-shelf six-axis arm with a camera on it, machined grippers and control software written in-house, directed by the same agent system. Its first target is an Agilent HPLC, which has no motorised door and a drawer that shifts when touched. Over seven days in the company's own lab, a mobile robot and the arm loaded and analysed 18 plates, holding the HPLC busy 97.2% of the time against 53.6% in its model of weekday hours.
Why it matters: Protein-design groups can now make thousands of candidates a week, and the measurement that tells them which ones worked is the step that waits for a person to open a door. Xaira's answer is vision on the arm rather than taught positions, which is what lets one arm move between instruments that were never meant to be automated, and it sidesteps the instrument maker's paid interface. The company says adapting the arm to another instrument now takes weeks rather than months.
Evidence and limits: A company blog post, no paper, with the company reporting its own demonstration and calling the week a technical proof of concept. One instrument, one week, and the 80% is against a schedule the company modelled rather than ran. Four interruptions in plate storage upstream needed a person to recover them, with no errors it attributes to the arm or the agents. The plates were analysed but no scientific finding is reported.
Fudan's EdenGNN predicts charge density and the occupancies that go with it, so band structures come out at PAW accuracy
What happened: Hongjun Xiang and Xingao Gong's group at Fudan built an equivariant graph network that predicts both the charge density and the projector augmented-wave augmentation occupancies. Earlier universal models gave only a pseudo charge density, which cannot rebuild the electronic structure consistently. A trainable radial basis and a delta-learning step for charge transfer make it 50 to 150 times faster than earlier charge-density models; over 1,000 unseen materials its median band-energy error is 0.045 eV.
Why it matters: Charge density is what density functional theory solves for, so a model that reproduces it at the accuracy real codes use can stand in for the expensive self-consistent step rather than just predicting one property. The augmentation occupancies are what makes that possible here.
Evidence and limits: All computed, by the authors. The universal model covers non-magnetic structures. The speed is measured against two other machine-learning models and plane-wave DFT on one GPU. Peer reviewed.
Merging 55 crystal databases shows a MatterGen run's stiff hits all contain osmium, and 82 of 100 repeat known compositions
What happened: A Hong Kong University of Science and Technology team matched the crystals of 55 databases by structure into one set of 14.6 million crystals and 131.8 million property values, only 4.5% of whose crystal-property pairs carry a value. Given 100 verifications each, MatterGen conditioned on 400 GPa returned 100 stiff hits against screening's 55; but every generated hit contained osmium, 82 repeated a recorded composition, and screening's best hit was stiffer, 433.8 against 375.4 GPa.
Why it matters: A novelty share measures the reference set as much as the generator, and the usual reference is Materials Project alone. Against all 55 sources, GNoME's matches go from one to 18,234, its new share from 100% to 96.7%. The forward screen turned up 53,813 ultra-incompressible candidates with no recorded bulk modulus; 62 have been made, among them ReN2, measured at 428 GPa.
Evidence and limits: All computed, plus a literature audit: ReN2's 428 GPa is a published measurement, not a new experiment here, and the screen had predicted 349 GPa, 18% below it. The MatterGen runs are the authors' own, from released weights. A preprint.
Simulated coatings for a sulfide solid electrolyte are made and measured: the lithiated ones keep 40 to 55% of the initial ionic conductivity after processing
What happened: An EPFL and Fribourg team fine-tuned the PET-MAD potential on r2SCAN data and ran large-scale molecular dynamics on Li6PS5Cl, screening carboxylic, sulfonic and phosphonic acid coatings by chain length and coverage. Simulation said lithiated head groups adsorb more strongly and that long-chain lithiated carboxylates form dense brush-like hydrophobic layers, with protection setting in near 8 Å. Infrared, impedance and H2S measurements followed.
Why it matters: The loop half closes: simulation set the chain length, and the H2S measurements run the same way, longer carboxylate chains suppressing it better. It also names the conflict rather than hiding it, since a thicker coating blocks lithium transport, so the best coating is a middle length.
Evidence and limits: One electrolyte, Li6PS5Cl, and one potential fine-tuned by the authors. The 40 to 55% figure is what is left after processing, not a gain. The experiments cannot rank the head groups, and the paper says so. A preprint.
AlphaGenome is turned on 17 million human-lineage substitutions and predicts 71,345 places where human gene regulation changed, nearly every fixed one older than the split from Neandertals
What happened: Tetsuya Yamada and Henrik Kaessmann at Heidelberg, with Svante Pääbo at the Max Planck Institute for Evolutionary Anthropology, predicted the regulatory effect of every single-nucleotide substitution on the human lineage across 305 tissues and cell types. Most of the 71,345 gains sit outside the regions evolution researchers had flagged as fast-changing, and more often than not they come from a regulatory element being reused in a new tissue rather than a new one appearing.
Why it matters: The question of which non-coding changes made humans human has been stuck on data nobody can get: great-ape tissue panels are thin, and no functional measurement is possible on an extinct hominin at all. A sequence model sidesteps that, because it scores a reconstructed ancestral allele as readily as the living one. The payoff here is the archaic comparison: nearly every gain now fixed in humans was already there before the split, so the modern-human-specific ones are few, and the paper names which ones move gene expression most.
Evidence and limits: The map is prediction, not measurement. Where it could be checked, against accessibility measured in four great apes' cell lines, AlphaGenome beat Enformer and ChromBPNet and 82.7% of its strongest calls held up, against 31.1% of controls, but it recovers only a quarter to a half of the measured human-specific changes depending on the threshold, and misses clustered gains more often than isolated ones. A preprint.
Refining the reconstruction with a neural network recovers atom-by-atom electron tomography at a sixth of the electron dose, on one re-read tilt series
What happened: Jihan Zhou's group at Peking University put a two-stage network after the reconstruction step: a ResUNet-3D variant corrects distortions across the volume, then a set transformer reads the shape of the electron density around each atom to name its element. Trained on 17,340 simulated samples and benchmarked on 42,588, at typical noise it cut the median coordinate error from 0.244 to 0.100 angstrom and raised atom recovery from 93% to 99%, elements called right 99.5% of the time.
Why it matters: Electron dose is what has kept atom-by-atom tomography to materials that survive the beam, which leaves out most of what people want mapped. The gain here is in post-processing rather than in the microscope, so a tilt series already sitting on disk can be re-read: thinned to a sixth of its dose, a published Pd@Pt series matched the full-dose conventional reconstruction on peak sharpness and repeatability.
Evidence and limits: No new experimental data were taken for this paper: the one demonstration on real data thins a published series, halving its projections too. Every other number is simulated, and the claim that this reaches beam-sensitive materials such as perovskites rests on simulation alone. Peer reviewed.
Eleven ways of drawing cell boundaries in spatial transcriptomics are benchmarked side by side: none wins across tissues, and the choice changes which cell types and which cell-to-cell signals come out
What happened: A KAUST and Karolinska team ran image-based, transcript-based and multimodal segmenters, from Watershed to Cellpose, Baysor and three versions of Segment Anything, over 15 datasets from 5 platforms and 12 human and mouse tissues. On their overall score Segment Anything led at 0.58 and Watershed came last at 0.27, but count how many cells each found and SCS leads instead; the ranking also reorders by dataset: SAM did well where cells look distinct, SCS where they look alike.
Why it matters: Segmentation is the step every spatial result stands on, because it decides which transcript belongs to which cell. This benchmark shows the choice carries all the way through to cell-type proportions, how much signal bleeds between neighbours, and inferred ligand and receptor pairs. So a published spatial finding depends on a step with no standard, and the pipeline released here is an attempt to make the comparison repeatable.
Evidence and limits: There is no true cell boundary to score against, so the metrics are biological proxies: expression purity, marker genes, cell-type proportions against a reference. The authors say platform design, image quality and gene panels cannot be fully separated from the algorithms, and that their gene-set sensitivity test ran on one dataset only. Two of the 11, Swin-S and Cascade Mask R-CNN, were left out of the scored comparison. Peer reviewed.
Iambic reports partial responses in 36% of 53 patients on its AI-designed HER2 drug, and in 9 of 22 whose tumours had spread to the brain
What happened: At ESMO 2026 Iambic gave updated Phase 1/1b data for IAM1363, the HER2 inhibitor its own models designed. By the 31 August cutoff 122 people had received it. Among 53 with targetable HER2 changes at 960 mg or above, 36% had a partial response and 38% stable disease; of 22 whose tumours had reached the brain, 9 responded and 12 were stable. Of the 68 at the chosen 960 mg daily dose, 5.9% had a drug-related grade 3 event and none a grade 4 or 5.
Why it matters: A drug a company says its models designed, and took from project start to the clinic in about two years, is now in its second year of human testing. These patients had been treated a lot already, a median of three regimens, 61% with trastuzumab deruxtecan and 36% with tucatinib, which is the context the response rates have to be read in, and the brain figures matter because those patients have few options. Iambic's prospectus puts about $75M of its $135M net IPO proceeds into this drug.
Evidence and limits: A single-arm Phase 1/1b with no control group, and only the company's summary of a conference poster is public: no median duration of response, and the durability it highlights is named patients. The rates come from subsets it chose, 53 of the 122 and 22 for brain tumours, and the intracranial responses after both prior drugs are three people. The release says nothing about what the models contributed, and a Phase 1 response rate is not evidence of longer life.
Benchling has eight AI models plan 65 molecular cloning experiments, and the top scorer, GPT-6 Astra, gets 30.6% right
What happened: Benchling and DNAForge built BenchBench-Cloning: in 50 tasks, taken from real lab notebooks kept on Benchling, a model writes the protocol for a requested plasmid from up to eight starting plasmids, and a simulator runs it; in 15, from published studies, it decides which constructs an objective needs. Across eight model setups GPT-6 Astra scored 30.6%, GPT-6.1 Sol 22.2% and Claude Opus 5.5 20.5%. Given ten tries, Astra produced a working protocol for 36 of the 50.
Why it matters: Scientists already hand models single cloning steps; this asks for the whole plan, and the failures shown are the kind a person only finds at the bench, days or weeks in: a missing control, two primers that cannot work in the same reaction. A failing plan, the authors write, reads as complete and confident, misleading the users most likely to trust it, so expert review of model-written protocols is still needed. Benchling sells the notebook software these tasks came from, and on 15 October launched Science Foundries, where agents run a lab process's steps and scientists approve key decisions.
Evidence and limits: Benchling and DNAForge built, scored and published it. Protocols were run in a simulator, not at the bench, which the authors say misses variation in reagents, handling and growth. Each task is one turn, with no clarifying questions or feedback. Ten of the 15 construct-design tasks are partly scored by a model judge, which was GPT-6 Astra, the top scorer; the paper releases those 15 and does not say whether the other 50 will follow.
Kevin Weil, OpenAI's former head of science, launches Warp Sciences with $215M led by Lightspeed, to build a US chemistry lab it plans to automate
What happened: Lightspeed led; Premji Invest, Mayfield, IVP, Lux, Conviction, Menlo Ventures and 12 other firms joined; angels include OpenAI's chief research officer Mark Chen and its CFO Sarah Friar. The South San Francisco company is a contract research organisation: it will make variations of a lead molecule for drug hunters, with chemists guiding the work. Its co-founders are CEO Kevin Weil, chief scientific officer Sol Reisberg (medicinal chemistry, Terremoto Biosciences), CTO Manohar Paluri (Meta Superintelligence Labs, a founding member of FAIR) and COO DJ Patil, the first US chief data scientist.
Why it matters: Weil's case is that AI has sped up software and mathematics but not chemistry, because chemistry moves only when someone makes a compound and tests it. A lab that sells that work to drugmakers also produces the experimental data better models need, and Warp says it will have no drugs of its own. The other half of the pitch is location: the release cites the BIOSECURE Act, signed at the end of 2025, which frames limits on federal procurement and funding tied to designated biotech companies, and gives customers more say over where their research and IP sit.
Evidence and limits: The round is the company's own announcement, which names no stage. The $865M post-money valuation comes from press reports citing the company and Weil, not from the release. Nothing is automated yet on the public record: the release describes labs being built and a first service guided by chemists, and the site names no customer, only a small number of design partners.
DeepMind maps the world's forests by type at 10 metres, as a baseline for the EU's deforestation rules
What happened: Google DeepMind and Google Research, with the World Resources Institute and IIASA, released Forest Typology 2020, which labels every 10-metre patch of land as one of six types, from primary forest to plantation, tree crops and other land. Built on AlphaEarth Foundations embeddings, it is 90.2% right overall against 8,190 expert-labelled sites, and is open on Earth Engine.
Why it matters: The EU rule turns on what a plot was in 2020, not just whether it had trees, and a tree-cover map cannot tell a natural forest from one logged and replanted as plantation. This map makes that distinction.
Evidence and limits: The 90.2% is carried by the large other-land class; per class, F1 runs from 82.9% for primary forest down to 58.2% for planted forest. Accuracy was checked on the Geo-Wiki platform of IIASA, a co-developer.
Microsoft's MAI-DxO, a panel of AI agents that orders tests and revises its diagnosis, is published in Science: 83% on hard NEJM cases against 36% for generalist doctors
What happened: Harsha Nori, Eric Horvitz and colleagues at Microsoft AI built a benchmark where a case starts from a short presentation and the diagnostician must ask questions and order tests, each with a cost, before naming a diagnosis. Their orchestrator, MAI-DxO, has language-model agents play a panel of doctors. It was right on 83.2% of New England Journal of Medicine clinicopathological cases and 80.3% of MIMIC-IV emergency admissions; generalist physicians scored 36.1% and 46.8%. Mustafa Suleyman, Microsoft AI's CEO and a co-author, called it a step toward medical superintelligence.
Why it matters: The benchmark, first posted as a preprint in June 2025 and now peer reviewed, tests diagnosis the way a doctor works: not picking an answer from a full case write-up, but deciding what to ask and which tests are worth paying for. Scoring cost beside accuracy lets a system be judged on whether it orders fewer tests, not only on whether it gets the answer. The Science version adds emergency-department cases and newer models.
Evidence and limits: The cases are published records run in simulation; no patient was seen. The doctors worked alone on cases usually handled by specialists, who were not compared. TechCrunch reports much of the gain comes from the underlying model: GPT-5 alone scored 78.9% on the hard cases, and on emergency cases the orchestrator added little.
Hugging Face releases Carbon-X, an open 14B model that answers questions about DNA in plain English and scores close to GPT-6.1 Sol on five DNA benchmarks
What happened: Carbon-X, built by Hugging Face with Zhongguancun Academy and TIGEM in Naples, starts from Qwen3-14B, is trained on 302 billion more tokens (55% of them raw DNA), then tuned to answer questions about sequences pasted into a chat. On five DNA benchmarks it averages 66.7%, against 67.5% for GPT-6.1 Sol and 60.9% for Opus 5.5, and leads every model tested on coding-variant effects (74.9%) and next-base prediction. The weights are out under Apache 2.0 and run on one 80 GB GPU.
Why it matters: DNA models usually output scores a scientist has to interpret, while a chat model may explain what a gene does yet struggle to judge a change in its sequence. Carbon-X puts both in one open model a lab can run on its own hardware, so unpublished or patient sequences need not go to an outside API. The team estimates that scoring ClinVar's 2.4 million variants of uncertain significance would cost about $300 with Carbon-X, against $32,000 to $4 million with the frontier models it tested, a range that wide because it spans models and reasoning settings.
Evidence and limits: Every score is the team's own, and it built two of the five benchmarks. Nothing was measured in a lab or clinic, and the team says the model is not clinically validated. Frontier models still lead on insertions and deletions, on single-base variants in long sequences and on sequence-feature questions, where Carbon-X can also miscount bases. The costs are not like for like: Carbon-X at measured throughput on a $5-an-hour H200, the frontier models at list API prices.
Reflector Bio comes out of stealth with a model that predicts what a compound does to living cells, built by Yuge Ji in Fabian Theis's lab
What happened: The New York company was founded by Yuge Ji, who built drug discovery models at Flagship Pioneering's Cellarity and developed the model as a PhD student in Fabian Theis's lab. Rather than predicting how a molecule binds to a protein, as structure tools such as AlphaFold and Boltz do, it predicts what happens in cells: whether a compound kills cancer cells or is toxic. The release names one partner, KiraGen Bio, which uses the model to rank gene-edit combinations for a glioblastoma cell therapy, and announces no funding.
Why it matters: Most AI drug design scores how a molecule fits one protein, while the failures that sink candidates, such as off-target effects and toxicity, show up later in cells and animals. A model trained on millions of past cell experiments aims to flag those before synthesis, and to make phenotypic screening, finding drugs by what they do to cells, cheap enough to run widely; the company says money, time and lab capacity are what kept it from scaling.
Evidence and limits: No round, investors or valuation were disclosed. The release's 80% hit rate is four of five compounds tested, from the founder's own preprint, and its line that top-20 drugmakers have used the platform names none of them.
Scinetics reports ScienceRSI, where small expert models' verified gains are written into a target model, and opens an 8B science-reasoning checkpoint
What happened: The company's research note and preprint set out a four-step cycle: train compact experts on separate skills, break each answer into checkable requirements and accept an expert's contribution only where it beats the current one, write the accepted parts into the target model, then specialise again from it. In the 4B and 8B settings it ranks first among the methods compared on all five scientific benchmarks, scientific reasoning at 55.0% against GRPO's 53.0%. With 8B experts and a Qwen3.5-27B target, all seven benchmarks improve, the mean from 87.3% to 94.0%.
Why it matters: Most science models are trained once and then stand still, and a lab that wants a new skill in one trains a new model. The claim here is that a smaller model's strengths can be moved into a larger one, which the 8B-to-27B run is the test of, and that the filtering is what makes it work: without the verification step the same supervision drops scientific reasoning from 55.0% to 50.6%, and scattering it at random costs 6.3 and 12.1 points. The open 8B checkpoint lets someone outside the company check what the recipe produced.
Evidence and limits: Benchmark scores only, all of them the company's own, scored in-house and not peer-reviewed. The note says its own report does not show that repeated cycles keep accumulating gains, and the 27B run also changes model generation, which it reports separately; the model card says the open checkpoint renames and republishes an existing one rather than a fresh training run. The rankings are first among the methods compared, not against the field.
A model that learns equivariance from data instead of building it in matches the equivariant reference models on molecular dipoles and NMR shielding tensors
What happened: Michele Ceriotti's group at EPFL applied the Point-Edge Transformer, which drops explicit O(3) algebra for edge-to-edge attention, to molecular dipoles and polarizabilities, NMR chemical shielding tensors and auxiliary-basis expansions of the electron density. A bigger backbone improved accuracy and equivariance error together. It matches or improves on the released models at lower inference cost, the density only when trained on a better-posed target.
Why it matters: Strictly equivariant networks pay for their guarantee in tensor products that grow steeply with angular order, and can fail to represent a target if truncated. Targets like shielding tensors and basis expansions are where that bites. Learning the symmetry instead removes the trade-off, if approximate equivariance is good enough.
Evidence and limits: All computed, by the authors. Equivariance here is approximate rather than guaranteed, and the authors say only that it affects most observables little, not all. Electron density is the weakest case: its equivariance error grows with angular order, and above order two it falls with backbone size no faster than the symmetrized error does. A preprint.
Independent DFT databases disagree by more than the gap between the top universal potentials, and three reported model failures shrink under the right control
What happened: Gus Hart at Brigham Young University and Jacob Pursley at the DOE's Kansas City National Security Campus tested 14 universal machine-learned potentials on 1.45 million ordered intermetallics against AFLOW, a database none of them trained on. For half the structures AFLOW, Alexandria and OQMD differ by under 6 to 8 meV per atom, but a minority differ far more, raising the average to 16 to 30. Re-referenced, the best potential agrees with AFLOW as closely as Alexandria does.
Why it matters: Leaderboards rank these potentials by formation-energy error against one database, as if it were ground truth. The disagreement between independent databases sets a floor under any such number, and the top models now sit at it. Screening out broken elemental references aligns the energy and stability rankings, and the leader's 7 meV per atom margin shrinks to 0.4 when both are scored at a common geometry.
Evidence and limits: All computed: this compares databases and models, with no new measurement. Intermetallics only, chosen as a region where DFT is relatively accurate. A preprint.
Whether AI weather models beat the physics model on extremes depends on which metric you pick: the AI models miss extremes, the physics model raises false alarms, and each metric punishes one of those
What happened: Statisticians at Karlsruhe and Marburg put Pangu-Weather, GraphCast and Aurora against ECMWF's HRES under one experimental design. HRES came out ahead when scoring was restricted to events that actually happened; the AI models came out ahead under weighted scoring functions. A contingency table shows why: the AI models forecast extremes less often than they occur, so misses dominate their scores, while HRES forecasts them more often than the AI models do, so false alarms dominate its.
Why it matters: Published studies disagree about AI models on extremes, and this says the disagreement is largely in the protocols rather than in the models. The decomposition puts the score gaps down to calibration, not discrimination: every model tells an extreme from a normal day about as well, Pangu-Weather least of them, and the table shows the AI models are simply too conservative about saying so. That is a fixable problem, and a different one from not seeing the event.
Evidence and limits: All computed, against the analysis that initialised all four models, which the authors say slightly favours HRES; station recordings would test it. One physics model and three AI models, on the variables and lead times chosen here. A workshop paper at NeurIPS 2026, so shorter and less reviewed than a journal article.
A deep latent-variable model with do-calculus inside it picks out which regulators may drive a cell-state change and which merely travel with it
What happened: Chengming Zhang and Luonan Chen, with Kazuyuki Aihara at the University of Tokyo, split the latent space of a hierarchical model into causal and spurious factors and apply do-calculus to the causal half, then recurse to rebuild the regulatory layers. On synthetic data, stem-cell differentiation, a CRISPR activation screen of 31 single and 26 paired conditions with some held out, a type 2 diabetes multiome and gut microbiome data, it beat the methods compared on causal triggers in most settings.
Why it matters: Most regulatory networks read off single-cell data are built from correlation, so a gene that moves with a state change is indistinguishable from one that drives it, and that is the difference between a target worth perturbing and one not worth the bench time. Encoding the intervention itself, rather than the association, is a different route to the problem, and the output is a ranked list of candidates to test.
Evidence and limits: All computed on public datasets, with no new experiment. The authors say the inferred triggers may still contain spurious biological signal, that the in silico perturbations are a qualitative read of direction rather than a predicted magnitude, and that simple additive baselines still do well on part of their own perturbation benchmark. Its results also depend on the regulatory hierarchy assumed. Peer reviewed.
Designing an RNA to bind a protein, with no binding site given, recovers the native complex on 16% of its authors' benchmark
What happened: A Nanjing University, CUHK and NUS team split the problem in two: generate an RNA shape in a pretrained structural latent space, then predict the contacts that place it against the protein, rather than generating shape and pose together. On 124 protein-RNA complexes held out by date, it leads ODesign, RFDpoly, RNAFlow and NACraft on complex recovery, with DockQ 0.131 against 0.116 for the next best. Twenty of the 124 clear DockQ 0.23.
Why it matters: RNA that binds a chosen protein is a route to reagents and to drugs, and unlike protein binder design it has stayed out of reach, partly because so few protein-RNA structures have been solved. What this paper pins down is where the state of the art stands when the binding site is withheld: five methods, and the best of them recovers the native pose on one target in six.
Evidence and limits: All computed, no synthesis and no binding measurement, and the authors say structural recovery establishes neither affinity nor activity. The headline ranking uses the native interface after generation to pick which of RiboBridge's own pockets is scored, so it is not a blind result. RiboBridge also clashes more than two of the baselines and does not lead on sequence recovery. A preprint.
Full Phase 3 data for Moderna and Merck's individualized melanoma vaccine: risk of recurrence or death cut 42% (HR 0.58), with no difference in the first six months
What happened: In INTerpath-001, 1,137 patients with resected stage IIB-IV melanoma were randomised 2:1 to intismeran, an mRNA encoding up to 34 neoantigens chosen from each patient's tumour sequencing, plus KEYTRUDA, or to placebo plus KEYTRUDA. After a median 24.6 months, two-year recurrence-free survival was 82.2% against 70.0% (HR 0.58), and the risk of distant metastasis or death fell 38% (HR 0.62). The data go to ESMO 2026's Presidential Symposium and are accepted by NEJM.
Why it matters: Until now only the topline was public. The two arms part only after six months, once the course of up to nine vaccine doses is given: the hazard ratio is 0.99 in the first six months and 0.42 after, in a prespecified sensitivity analysis. That is the timing expected if the benefit comes from an immune response the vaccine builds against each patient's chosen neoantigens. Merck says it is now talking to regulators.
Evidence and limits: Merck and Moderna released these numbers ahead of the talk and the paper. Overall survival is immature: the release reports a 'favorable trend' and no number. The trial tests the whole product, so it cannot say how much the neoantigen selection adds over another way of choosing antigens. Grade 3-5 treatment-related events were 14.2% against 13.4%.
A single protein-design model from Baker's lab and NVIDIA beats the best specialist tools on four design tasks, and its designs were made and measured in the lab
What happened: RFD4-Proteina, from the University of Washington Institute for Protein Design and NVIDIA (David Baker corresponding), designs proteins, DNA, RNA and enzymes in one shared latent space, using reasoning and guided search rather than a separate model per task. On AlphaFold3-refolded designs it reports a 10.6% pass rate for protein binders against BindCraft's 5.9% at 42 times the throughput, and higher rates than every method tested on small-molecule, enzyme-scaffold and DNA-binding tasks. In the lab, 158 of 188 designed TrkA binders bound below a micromole.
Why it matters: Protein design has used a different specialist tool for each job: backbones, binders, enzymes, DNA. RFD4-Proteina puts them in one model that also covers modified amino acids, sugars and nucleic acids, and still comes out ahead on all four. The part that matters is the wet lab: designs across many target classes were made and measured, including a designed enzyme that converted 79% of an intermediate in making the HIV drug lenacapavir, a de novo catalyst doing a real step of drug chemistry, further than designed enzymes had reached.
Evidence and limits: A preprint, not peer reviewed, and all the wet-lab work is the authors' own, so nothing is independent yet. In silico a 'pass' means an AlphaFold3 refold matches the target (a model checking a model), and the pass rates are low in absolute terms (3.5% for DNA binders). The binder and enzyme successes are counts from the authors' own assays, not an outside lab's, and the enzyme result is a single proof-of-concept.
Google's personal health agent hands health questions to three AI specialists; in Nature, experts rate it above one agent and above the three answering separately
What happened: A. Ali Heydari, Daniel McDuff, Xuhai Orson Xu and colleagues at Google Research built an orchestrator over three Gemini 2 agents: one writes and runs analysis code on Fitbit and blood-test data, one checks health knowledge against sources, one coaches. The data came from the WEAR-ME study of 1,165 people. Across 11 tasks and 1,350 hours of rating, 45 users put it first 46.7% of the time; three health experts scored its overall quality 72.8%, against 64.3% for the three agents answering separately and 54.4% for one agent.
Why it matters: Questions about your own wearable data need code and statistics, while coaching needs a different way of talking, and the authors found one model struggles to balance both. Splitting the work let each part be tested on its own: the data agent's analysis plans scored 75.6% against 53.7% for plain Gemini, and its code passed 75.5% of expert unit tests on the first try against 58.4%. The framework and its benchmark are open under Apache 2.0, so other groups can swap in their own models.
Evidence and limits: Everything was rated in conversation, not measured in health: the live users played study personas instead of asking about their own data, and no one tested whether behaviour changed. The data are Fitbit only, some scoring used model raters, and a response took 205 seconds and 6.5 model calls, against 36 seconds for one agent. The authors call it a research prototype, not a diagnostic tool.
Mirendil says one researcher working with AI agents pretrained Mirendil-Base, a coding model of 100B to 500B parameters, in 3.5 months for under $10M
What happened: Mirendil reports that one researcher built Mirendil-Base from an empty repository: agents wrote over 99% of the code, ran the experiments and diagnosed failures, while the researcher chose what to try and what to trust. It learned only from public code, web pages and datasets, with no distillation, and the cost covers all compute, the agents' tokens included. On seven code repositories published after training it scored 0.222 bits per byte (lower is better), behind Kimi-K3's 0.193 and ahead of GLM-5.3, DeepSeek-V4-Pro-Base and Nemotron-550B-Base.
Why it matters: Pretraining a competitive base model has taken a team of specialists in architecture, data and chips. If one person directing agents can do it, the scarce inputs become compute and judgment, and a group whose expertise is in its own science, such as a robotics lab or a hospital, could train a model on its own data. It is the first result from a company that raised on the pitch that labs using AI in biology or chemistry should not have to become AI labs.
Evidence and limits: Every number is Mirendil's own, in a blog post with no paper, no weights and no named researcher. The only score is how well the model predicts unseen code, not whether it solves coding tasks, and Mirendil notes that Kimi-K3 and GLM-5.3 were scored after post-training, which usually raises this loss slightly. The two larger models are still in training.
Duke researchers gave an off-the-shelf language model a pipetting robot and a webcam, and it designed its own experiments when it found it was missing information
What happened: Jeonghyeon Kim and Philip A. Romero connected GPT-5.6 Terra to a standard Opentrons OT-2 pipetting robot and a webcam through PEARL, five model roles sharing one record of the experiment. Given only a goal, it refused an impossible instruction in all ten runs, read from the colours that a human had turned the plate 180 degrees mid-run, and in seven of ten diluted dyes to see their colour first. Asked to characterise an enzyme assay, it set concentrations and timings, discarded a spiked well, and repeated runs missing early readings.
Why it matters: Lab automation runs the procedure it was given, and a moved plate or a saturated reading breaks it unless someone wrote a check for that case first. The paper's own comparison locates the behaviour: in a plain tool loop none of the 80 execution runs passed, one agent keeping a persistent record of the experiment passed 60%, and the five-role version passed all 80. So the judgment is mostly the general model's and the scaffolding buys reliability, which is what would let a lab with an ordinary liquid handler try this rather than build a platform.
Evidence and limits: Eight scenarios on one bench, run by the authors: colour mixing and one horseradish peroxidase assay, whose kinetic constants matched their own human-run reference and previously reported values, not an outside lab's. The scoring of planning, adaptation and execution is theirs, on their own rubric, and every perturbation was one they introduced, so nothing tests what a working lab meets unplanned. Colour was read off a webcam calibrated against a plate reader. It is a preprint.
A KAUST and Bestzyme team redesigned a laundry protease with reinforcement learning, and three of their variants cleaned soiled cloth better than the commercial enzyme in a standard wash test
What happened: AlphaVariant pairs a sequence generator trained on related natural proteins with activity predictors retrained after each round of results. On savinase, a detergent protease, two rounds produced libraries in which 75.4% and then 83.8% of the variants tested beat wild type, as the positions it could change went from 25 to 67; its best variant was 7.1 times more active than wild type on a synthetic substrate. A third round also scored heat tolerance and patent distance, and of 32 variants built, three beat the commercial protease Uno 100 L.
Why it matters: Directed evolution gets its gains mostly one position at a time, because the combinations explode and the measured data is far too sparse to fit a model over them. Here the useful variants carry changes at eight to sixteen positions at once, and the final round needed only 32 measurements, which is the scale at which a design method is worth a company's screening budget rather than a benchmark table. Putting patent distance in the reward is the unusual part: freedom to sell was optimised alongside performance, and it is a constraint any enzyme has to satisfy.
Evidence and limits: Every assay and the wash test were run by the authors with the industrial partner Bestzyme, which has filed a patent application on the variants; no outside lab has tested them. The wash test is a Chinese standard method on one soiled-cloth type, not a product trial. The 7.1-fold figure is kinetic fitting on a peptide substrate, and round three used casein. On the six public landscapes the clear wins were the three scored by a surrogate. Preprint, CC BY-NC-ND.
Xaira's X-Scientist agents picked genes to perturb in liver cells, and 7.1% of its picks reversed the TGF-beta signal against 6.9% for its own scientists and 0.9% at random
What happened: Bo Wang, Ci Chu and colleagues describe X-Scientist on the company's blog: a lead agent plans and checks the work of worker agents that call bioinformatics tools and Xaira's own models. It nominated 381 genetic perturbations to undo what TGF-beta does to HepG2 liver cells; the company also tested 321 its own scientists picked and 433 drawn at random, measured all of them by Perturb-seq, and counted a hit at three standard deviations back toward untreated cells. Nominating took 20 minutes against more than three hours by hand.
Why it matters: TGF-beta drives scarring in chronic liver disease, and which genes to perturb against it is the kind of choice a biologist spends a morning on. The measurement here is of the model's picks, not of the model's own score: every nomination was made, perturbed and read out in cells. Three of its hits outside the known TGF-beta pathway beat the median of the ones inside it, and none were in the scientists' list, which is what the loop is for, since a list of known pathway genes is one a scientist already has.
Evidence and limits: A company blog post, no paper and no peer review, and Xaira both chose and scored the comparison. One cell line, one exposure out of the twelve it ran, and the 7.1% against 6.9% gap is well inside what 381 and 321 perturbations can resolve. The benchmark scores beside it compare harnesses on the same models, not biology. Nothing has gone back into the models yet.
One LoRA adapter per space group, not per compound, cuts the cost of high-throughput thermal conductivity by a projected five to seven times
What happened: Alexander Shapeev's group at Skoltech fine-tuned the MatterSim 1M potential with low-rank adaptation on a small set of compounds sharing one crystal symmetry, then solved the Boltzmann transport equation for lattice thermal conductivity. On half-Heuslers, space group 225 and spinels, the adapted potential came two to three times closer to the reference than the unmodified one, matched per-compound fine-tuning, and cut fine-tuning time 1.5 to 2 times on the 28 compounds actually run.
Why it matters: Universal potentials soften the energy surface, which spoils the derivatives thermal conductivity depends on, so people fine-tune per compound and pay for it. Tying the adapter to a space group instead spreads that cost over every compound sharing the symmetry.
Evidence and limits: All computed. The reference is other potentials, moment tensor and polyMLP, not measurement. The five-to-seven speedup is extrapolated from 28 compounds; the paper's own conclusion says five to six. Three space groups, and the paper does not show transfer across symmetry classes. A preprint.
AuthorsRustam Arabov, Konstantin Sozykin, Nikita Rybin, Alexander Shapeev
Astellas signs AI Proteins to design miniprotein drugs against its targets, in a deal worth more than $400M with milestones
What happened: AI Proteins will design and optimise miniproteins from scratch against several targets Astellas picks. Astellas, through its affiliate Astellas Innovation Management, takes an exclusive worldwide licence to any candidate and runs development, regulatory work and sales. Upfront and milestone payments come to more than $400M, plus royalties. Chris Bahl, the Boston company's founder and CEO, and Morten Sogaard of the Astellas Innovation Lab spoke for the deal.
Why it matters: This is AI Proteins' second large-pharma deal, after Bristol Myers Squibb's in late 2024, also valued at up to $400M. Miniproteins are small proteins built from scratch rather than taken from nature. Bristol Myers Squibb took an option on a licence; Astellas takes the exclusive licence outright, so this time the designs are bought as future drugs from the start.
Evidence and limits: The release does not break out the upfront payment or say how many targets are involved, and no designed molecule from the deal is shown. One trade report described the $400M as upfront; the release counts it as upfront and milestones together.
Pinal, a model that writes a protein from a description in words, is published in Nature with four kinds of protein that worked in the lab
What happened: Fengyuan Dai, Fajie Yuan and colleagues at Westlake University trained a 16-billion-parameter model on 1.7 billion protein-text pairs: a description of the function wanted goes in, a sequence comes out, with less hand-written structural specification. Four tasks were run, a fluorescent protein, a PET hydrolase, an alcohol dehydrogenase and a metabolic H-protein, and each gave the intended function, with catalysis measured for both enzymes. In a reconstituted CO2-fixing reaction a designed H-protein raised product yield 1.7-fold over the natural E. coli one.
Why it matters: Protein design normally starts from a natural protein to optimise, or asks the designer to restate the goal as geometry and sequence rules a model can take. Here the goal is a sentence, and what came out of it catalysed reactions rather than only folding as predicted. The H-protein is the sharper part: a designed carrier protein beat the natural one in a reconstituted CO2-fixation reaction, with no rounds of experimental optimisation in between.
Evidence and limits: The per-task numbers are in the extended data and they are small: 94 PET hydrolase designs assayed out of 2,416 filtered gave one hit, 8 alcohol dehydrogenases tested gave 5, and 20 H-proteins picked from about 1,000 gave 5 active. All the wet-lab work is the authors' own. The inference code is Apache 2.0 and the weights repository states no licence; the full training corpus was withheld for biosecurity and given only to the reviewers.
Westlake University's Pinal turns a sentence describing the job into candidate proteins, and candidates for four different jobs worked in the lab
What happened: Fajie Yuan's group at Westlake University, with HKUST (Guangzhou), the Shenzhen Institute of Synthetic Biology and Zelixir, trained Pinal, a 16-billion-parameter model, on 1.7 billion machine-written pairs of a protein and a description. It drafts a structure from the words, then a sequence to fit it. Designs for a fluorescent protein, a PET-degrading enzyme, an alcohol dehydrogenase and an H-protein each showed the function asked for; in a CO2-fixing pathway the designed H-protein gave 1.7 times the product of E. coli's own, with no lab optimisation.
Why it matters: Design models typically start from a natural protein to improve, or need a person to turn the goal into shape and sequence constraints first. Pinal takes the goal as a sentence, which is a different interface: the biologist says what the protein should do and gets candidates to test. The weights are on Hugging Face and the code is Apache-2.0, so the claim can be tried by anyone.
Evidence and limits: Each hit came out of a funnel the headline hides: about 151 designs assayed across the four functions, out of thousands generated, with the PET-degrading enzyme leaving one standout among 94 tested. The measurements are the authors' own, and the training corpus is withheld over dual-use risk. Fajie Yuan told Science and Technology Daily that language-driven protein design is still at the proof-of-concept stage, with finer instructions, such as setting a fluorescent protein's emission band, open for the field.
Recursive releases Flint, an open 40B coding model for laptops, which it says its automated research system helped build
What happened: Flint is a mixture-of-experts model with 40 billion parameters, 2.5 billion active per token, released under Apache 2.0 on Hugging Face. Recursive says it leads models under 50 billion parameters of its kind on agentic coding benchmarks, including SWE-Bench Pro v2 and Terminal-Bench 2.1, and runs at over 125 tokens a second on an M4 Pro MacBook Pro. It says a small team with its Eureka Machine built it from scratch in under four months.
Why it matters: Recursive calls Flint a practical test of its automated research system: the Eureka Machine searched the architecture, found the quantization recipe and wrote the hardware kernels. That moves the system's work from benchmarks onto a model people can download and use. What it shows for science is indirect; the system's targets so far are AI's own stack.
Evidence and limits: The benchmark and speed figures are Recursive's own. How much of the work the Eureka Machine did, against the team, is the company's account, with no breakdown.
Eka Robotics, founded by MIT's Pulkit Agrawal and ex-DeepMind Tuomas Haarnoja, raises a $150M Series A for robots that grip by feel
What happened: a16z and F-Prime led the round; seed investors Obvious Ventures and E14 Fund returned, with MFV Partners, Bison Ventures, Glasswing Ventures, Samsung Next and unnamed Fortune 500 strategics. TechCrunch reports a $1.45B post-money valuation. Pulkit Agrawal runs MIT CSAIL's Improbable AI Lab; Tuomas Haarnoja co-created the Soft Actor-Critic algorithm and trained humanoid behaviours at Google DeepMind. Eka's vision-force-action model is trained by reinforcement learning in simulation.
Why it matters: Its model uses the forces a gripper feels on contact, not only what a camera sees, to adjust a slipping grasp or a part that resists insertion. That is the skill a lab robot needs to handle vials, pipettes and samples it has not seen before. Eka itself aims at general manipulation, from packing produce to assembling server racks, not at labs.
Evidence and limits: The 1,500 picks an hour (against about 600 for a person) and the transfer from simulation to new objects are the company's and a16z's own claims, with no outside test; the valuation is TechCrunch's report, not the company's.
Bob McGrew's Arda raises a $160M Series A at a $1.5B valuation for AI that turns product designs into assembly lines
What happened: Arda, started in 2025 by Bob McGrew, OpenAI's former chief research officer, launched on 20 October. McGrew announced a $160M Series A at a $1.5B valuation; CNBC, citing the company, gives $230M raised in all, led by Altimeter Capital, Accel and Khosla Ventures. Augustus Odena, a co-founder of Adept, is co-founder and chief AI officer. Anduril has been its design partner for a year.
Why it matters: Its software reads CAD files, PDFs and product requirements and produces the build plan and work instructions for a manual assembly line, updating them when the design changes; next it plans lines that robots can be reconfigured on as quickly as people. That is engineering, not discovery, but it automates the same step a self-driving lab needs: turning a design into physical work.
Evidence and limits: The $160M and $1.5B are McGrew's own figures; how CNBC's $230M splits across earlier rounds, and the other investors, were not read at source. No customer result is public beyond the company's word that it runs on lines at a large US manufacturer.
Thore Graepel's Metis Reasoning launches with four founders, a 12-person team and a white paper
What happened: Metis Reasoning came out of stealth on 20 October with a team of 12. The founders are Thore Graepel, Piers Ingram, Michael Warren and David Berman; the white paper, bylined Graepel, Berman, Warren and Ingram, gives the company as London, Singapore and Los Angeles. The paper proposes an AlphaGo-inspired system that keeps an explicit record of what it knows and accepts a claim only when evidence warrants it. Ingram told Tech.eu the company has been funded more than adequately to date, without commenting on the reported raise.
Why it matters: In September the record had only Bloomberg's report of fundraising talks. Now the company names its founders and sets out its approach in its own words: reasoning modelled on the scientific method, with auditable traces, aimed at science, engineering and robotics, where an answer has to be checkable.
Evidence and limits: A position paper and a launch, not a result: no model, benchmark or customer is public, and the 'encouraging' reasoning results Ingram mentions are not shown. The raise is still unconfirmed and no investor is named.
OpenAI acquires Orchestra Bio, whose agents track what a drug programme owes and when, and folds its team into Rosalind with no price disclosed
What happened: OpenAI says it has acquired Orchestra Bio, a platform for coordinating drug development, and is bringing its team and technology into Rosalind. Its own example: a failed manufacturing batch can delay the start of a clinical trial, and Orchestra's agents work out which activities and timelines that hits, gather the relevant information and compare ways to respond, after which scientists and programme leaders decide. The same page names Novo Nordisk and Moderna as Rosalind partners.
Why it matters: Rosalind has been a model and a workbench; this adds the layer that tracks what a programme owes and when. OpenAI puts the case in its own numbers: roughly 10 to 15 years on average from target discovery to approval in the United States, with teams coordinating across internal departments and outside partners on disconnected systems. So what is being sold here is coordination rather than prediction, and the division of labour is stated outright: agents monitor and track, people decide.
Evidence and limits: No price, and OpenAI's page names nobody from Orchestra. The reported funding, about $20 million in total, reaches here through Decrypt citing Caplight rather than from the company, and the seed round's size is not given. Neither company says how many customers Orchestra has, what happens to existing contracts, or when the agents reach Rosalind. Orchestra Bio is not Orchestra BioMed, the unrelated Nasdaq device maker.
Radical Numerics reports that phage variants its Omnii-1 model designed were given to one patient with a drug-resistant infection, under compassionate use
What happened: Aniketh Iyengar, Benjamin Siranosian and colleagues report it on the company's blog. Omnii-1 rewrote the DNA for gp17, the tail-fibre protein the T7 phage uses to recognise E. coli. At 11 to 15 amino-acid changes from the natural protein, 69.9% of its designs were recovered after assembly against 4.8% of random controls, and 2,161 of 6,694 variants with measurable activity were enriched above the natural phage. A cocktail containing its designs cleared one patient's resistant bacteria in a dish, then was given to that patient.
Why it matters: A model's design has gone into a person's treatment, not only into a dish: the company says this is the first clinical use of AI-generated phages, under an authorised, physician-led compassionate-use pathway. What the model wrote is one region of a known T7 genome, not a genome. It also ranks which designs to test against a given bacterial strain (on the hardest of three library strains, 24% of its top 100 picks beat the activity threshold against 4.5% of the library), and that ranking is the step a therapy matched to one patient has to get right.
Evidence and limits: A company blog post, no paper and no peer review, and every measurement is the company's own. One patient: no clinical course and no outcome, citing privacy, and the company says a single patient cannot establish safety or efficacy, so nothing here tests whether the treatment worked. The post does not name Tabula, which The New York Times later reported had done much of the lab work and supplied some of the cocktail's phages.
Anthropic has Claude run a protein-binder design campaign against scorpion toxins; outside labs find 17 of 211 designs bind
What happened: The team chose five toxins; from one prompt and four follow-ups, an internal research model then worked mostly on its own for 64 hours, picking the sites to aim at and generating 318,484 candidates, 211 of which were made. Adaptyv Bio tested all of them for binding, Twist Bioscience 210 against four toxins. Seventeen met Anthropic's criteria, 15 of them at both labs. The tightest, against the Mexican toxin Cn2, was single-digit nanomolar. All 211 and their data are released.
Why it matters: This is the company's earlier binder work pushed at a target nobody has money to work on (antivenom for a neglected disease), and the loop is what to look at: the model set its own quality thresholds, decided how much to lean on each design tool the prompt offered, and sent sequences to a vendor, with people steering four times. Among designs aimed at a single toxin, 11 of 60 bound, which is what an ordinary protein-design campaign gets, so the claim is about who ran the campaign, not about a better method.
Evidence and limits: Binding is all that was measured, against toxin proteins grown in the lab, not taken from venom; nothing was tested in cells or in an animal, and Anthropic says it has not made an antivenom. None of the 151 designs aimed at more than one toxin met the bar against more than one. The criteria were set after the data came in, and no design baseline was run, so the model's own contribution cannot be separated from its tools'.
An agent framework simulates 14 nickel grain boundaries, then shows that leaving the boundary's own defects out of the cell moves the predicted sliding strength either way
What happened: Yazhuo Liu, Ting Zhu and colleagues at Georgia Tech, the University of Osaka and USC build GBClaw: an orchestrator and five agents, routed to two GPT-5.6 models, that set up atomistic shear simulations in LAMMPS and analyse them. Across 14 asymmetric tilt boundaries in nickel the strength at which sliding starts ran from 14 megapascals to 9.3 gigapascals. Four were rebuilt with arrays of secondary misfit dislocations, the defects a real boundary has: one fell from 1,074 to 530 MPa, two by four orders of magnitude, one rose.
Why it matters: Most interfaces inside a real metal are these mismatched boundaries, and they stay understudied because their structure is hard to pin down in both experiment and simulation. What the sweep shows is about the models, not the agent: if the predicted strength depends this much on whether the boundary's own defect structure sits in the cell, then screening boundaries on simplified cells (with a first-principles or machine-learned potential, at real cost) is reading the cell as much as the material.
Evidence and limits: All of it is simulation with one empirical nickel potential: nothing was made or measured, and the authors call their numbers athermal instability thresholds for the cells they modelled, not universal upper bounds. The cells are small and periodic, which suppresses the long-range defect interactions a wide boundary has, and the energy paths carry no temperature. The {111}/{110} anomaly was traced only after a researcher directed the follow-up runs.
CrystalAR writes a crystal as one sequence, from space group down to atom positions, and nearly matches the best structure-prediction baselines on MP-20
What happened: A team from the National University of Singapore and Southern University of Science and Technology puts a crystal's whole description into one sequence for a single autoregressive model: space group, then which element sits on which symmetry site, then the cell, then positions. Each position is chosen coarse to fine, by repeatedly halving the region that site may occupy. Given composition and space group, one guess recovers 62.92% of MP-20 structures (CrystalFormer 29.62%, OMatG 63.75%) and 25.25% of the larger MPTS-52 structures, where its position error is also the lowest.
Why it matters: Generators that write symmetry sites out explicitly can explore new structural templates, but they have lagged at producing stable, accurate crystals. Here one model does both, and its own ablation shows why its structures come out valid: choosing positions in symmetry-aware coarse-to-fine steps gave 90.12% valid structures, against 68.16% for plain position bins and 41.64% for continuous coordinates. That representation is something other crystal generators can reuse.
Evidence and limits: All computed, by the authors. In the prediction test CrystalAR is given the space group with the composition, while generic baselines such as DiffCSP and FlowMM get the composition alone, as the paper's appendix notes. With 20 guesses it trails CrystalFlow, 74.55% against 78.34%. Stable new crystals stay rare: 0.9% of generated structures after pre-relaxation.
A screen of 2.8 million computed compounds for electrical conductivity finds none that beats copper, but several that rival aluminium on paper
What happened: Miguel Marques and Silvana Botti's group at Ruhr University Bochum ranked about 2.8 million near-stable compounds in the Alexandria database with two machine-learning models and a fitted formula, kept about 1,200, added about 12,800 compounds already computed, and narrowed the combined set to 52 for full electron-phonon transport calculations. Tested on elemental metals, those calculations land mostly within 30% of measured values. The best compounds, Be2IrPt, H3NaNi, LuIr and LiBePt2, come out at about 34 to 38 million S/m, near aluminium's measured 36.59.
Why it matters: Chip wiring and power lines need metals that conduct well, and conductivity is a property materials databases rarely compute because the full calculation is expensive. This shows a workable route: rank millions cheaply, then spend the full calculation on a few dozen. It also explains why silver, copper and gold stay ahead, and points to a mechanism: electrons from light elements can push a compound's high-scattering d-states below the Fermi level, as in LiBePt2.
Evidence and limits: The match with experiment is for elemental metals, not the new compounds, none of which was made or measured here. Even among elements, calcium comes out at 7.98 against a measured 28.99 million S/m. Spin-orbit coupling was left out; adding it cut Be2IrPt and LiBePt2 by 8% and 9%. Several leading candidates contain platinum, iridium or rhodium.
Scored by what a lab would actually discover, protein models' zero-shot accuracy explains under a fifth of the difference in what they find, and 50 to 100 measurements beat them
What happened: A Sydney, Columbia, BAAI and Tsinghua team re-scored 95 predictors over 217 ProteinGym assays and 2.47 million variants by what each would find under a matched budget, on two objectives: finding top variants, and ranking the landscape. Across 20 scores fitted to 20 measurements of the target assay, prior ranking ability explained under a fifth of the differences in what they recovered (r = 0.44). A plain additive model from measured single mutants then overtook them: at 50 to 100 measured singles for top variants, around 500 for ranking.
Why it matters: This changes what a protein-model leaderboard is for. A campaign's question is which variants to measure next, and the paper shows the quantity that answers it is not the one benchmarks report: a score that ranks a landscape well need not direct a budget well. The practical reading is that a model prior is worth most in the first few dozen measurements, after which the lab's own data wins, and that closed-loop design should be scored by what gets discovered at a given budget, not by correlation on a fixed dataset alone.
Evidence and limits: All retrospective, re-scoring measurements that already exist, with no new experiment and no closed loop run. Substitution landscapes only; the authors say insertions, deletions and de novo design would need their own controls and stated measurement costs. The crossings are additivity predicting double mutants from measured singles; on variants with three or more substitutions the models led instead. They come from deep mutational scans, where every variant is already measured, which is not a lab's candidate pool. A preprint.
Electric Plant comes out of stealth with an $11.2M seed round led by Onto Ventures and Climactic, to sell sensors and a model that read plants' electrical signals
What happened: Electric Plant was founded by Justin Mast, who built the online plant shop Bloomscape, and the neuroscientist Lee von Kraus. TechCrunch reported the seed round: Onto Ventures and Climactic led, with Ponderosa, Endeavor Catalyst, Ludlow, Michigan Rise, The Fund, Schox VC and Dug Song. Its sensors screw into a stem or trunk and send the plant's electrical signal to Bombadil, the company's model, which reads early signs of stress and what the plant needs, such as water. Mast told TechCrunch a sensor costs tens of dollars.
Why it matters: Growers judge what a plant needs from soil probes, handheld leaf meters or satellite images. Electric Plant wants to read the plant itself, all the time, from a cheap sensor, and has trained a model on that signal across more than a hundred species. The round backs a model of a living body signal taken from plants rather than people, aimed at farming, forestry and plant care.
Evidence and limits: The round is from TechCrunch's report, with Mast on record; the company's Business Wire release could not be opened, and no valuation was given. Mast's claim that the system can cut water use by 20% or more is his own, with no field data published, and he did not name customers.
Radical Numerics reports its Omnii-1 model scoring above AlphaGenome and Borzoi at telling which changes in non-coding DNA actually do something
What happened: Matthew Carter and colleagues post the numbers on the company's blog. On AlphaGenome's own panel of 21,948 test cases, each pairing a DNA change with a gene and a tissue, half of them changes statistics pinned as the cause and half it ruled out, Omnii-1 scores 0.810 against Borzoi's 0.705 and the 0.713 AlphaGenome's paper reports. On a splicing panel of 14,458 cases it scores 0.814 against 0.721 for Pangolin. It also predicts methylation, a chemical tag on DNA, one site at a time, agreeing with measurements at 0.879.
Why it matters: Most DNA changes tied to disease sit outside the stretches that code for proteins, and working out which of them does anything is the bottleneck. The earlier models read DNA alone. This one also reads lab measurements of how the DNA is packaged and used in a given tissue, and can take those measurements as an input rather than only predict them. Its lead over Borzoi grows the further the change sits from the gene, from 0.05 within 500 bases to 0.19 beyond 50,000.
Evidence and limits: Every score is the company's own, in a blog post with no paper and no peer review. It says it cannot run AlphaGenome itself, so those numbers are read out of the published paper rather than measured side by side, and AlphaGenome's higher supervised score on the same panel, 0.799, is not the one compared against. The labels are statistical inferences from population data, not variants tested in an experiment, and the model is behind a request form.
Radical Numerics and Aptamino run three rounds of lab-measured antibody design with Omnii-1, and two designs bind tighter than the test can read
What happened: Max Highsmith and colleagues report it on the company's blog. The campaign began with 1,620 designs from earlier pipelines, 150 of them with measured binding. Since mid-September, Omnii-1 rewrote positions in the best of those, ranked its own proposals, and Aptamino made and measured a selected set, three times over. The designs push out the measured trade-off between how tightly a single-domain antibody binds and how much of it a cell makes; two from the third round bind below the test's floor.
Why it matters: This is a loop where the model's own ranking decides what gets made, and each round's measurements go back into it. Antibody sequences were not a domain of this genome model's original pretraining; they came in afterwards, as public repertoires and the partner's own measured designs. What is being tested is whether a general biological model can be turned to one partner's narrow design problem. Ranked on the same 327 measured designs, what the model has learned predicts expression and potency a little better than protein-specific models, with a plain one-hot encoding of the sequence close behind on potency.
Evidence and limits: A company blog post, no paper and no peer review, and both sides report their own work. One target, single-domain antibodies, three rounds, with the target's identity and some measured properties held back by the partner. Two designs beating the test's floor is a limit of the test, not a measured strength; the antibody post gives no number for that floor, and 0.2 nM comes from the launch post. The gains over the other models sit inside overlapping confidence intervals.
With Lawrence Livermore, Radical Numerics reports its biology model predicting which rare earth metal a protein prefers better than the other methods tested
What happened: Armin Thomas, Gautam Machiraju, Benjamin Siranosian and Christine Parthemore report it on the company's blog. Livermore measured how 609 natural proteins split their binding across the 15 rare earth metals, on a plate assay it built this year. Given two of the metals, Omnii-1 ranked proteins it had not seen by which one they favour, at a median rank agreement of 0.340 over all 105 pairs, against 0.318 for the strongest public model, ESM-C 6B, and 0.254 for the best of three methods that go by sequence similarity.
Why it matters: Telling these metals apart is a refining problem: China refined 91% of the world's supply in 2024 on the International Energy Agency's count, and the elements differ so little in size, 0.013 angstroms between neighbours, that the standard route needs harsh acids and a lot of energy. Proteins can distinguish them, and the question is whether a model can say which protein will, before anyone makes it. This model had no rare earth binding data in its training, and the lab is now testing the candidates the model generated and ranked.
Evidence and limits: Prediction only: the proteins the model designed are at Livermore awaiting testing, and no result from them is reported. The company's post, its own evaluation, no paper and no peer review. On the 17 pairs closest in size, the hardest cases, the company puts the gap to ESM-3 at 0.006 against a paired error of 0.005 and declines to rank the two. AlphaFold2 scored 0.191 here, below the simplest sequence-similarity baseline, 0.198.
Sanofi signs ScionxBio to redesign some of its mRNA drug candidates with a deep-learning platform so they keep making protein for longer
What happened: ScionxBio, the name Eleven Therapeutics of Cambridge, England now operates under after a split, signed a worldwide collaboration and licensing deal with Sanofi. Its TERA platform screens large combinatorial libraries of chemically made mRNA regulatory elements and uses deep learning on the results to design molecules that resist breakdown and keep making protein longer. ScionxBio optimises the chosen Sanofi candidates up to animal tests; Sanofi then runs all later development and sales. CEO Adi Gilboa-Geffen and chairman and co-founder Yaniv Erlich announced it.
Why it matters: The release aims the platform at protein deficiencies and chronic disease, where a drug has to keep making protein, which standard mRNA, less stable, does not do for long. A large drugmaker is handing its own candidates to a model trained on screening data to redesign them, rather than starting from the company's molecules. Undecimal Bio, which Eleven spun out, launched with an $18M seed on 13 October.
Evidence and limits: No upfront payment, milestone or disease area is disclosed, and the release shows no data on how long its designs last.
Insilico reaches the first milestone in its oncology deal with Servier, ten months in, and Servier pays an undisclosed sum
What happened: Insilico Medicine says the first of two cancer programmes it runs with Servier met criteria set in advance for an early discovery stage, triggering a milestone payment of undisclosed size. Both programmes, on unnamed targets, have completed target nomination since the deal was announced in early 2026 (up to $888M, $32M of it upfront and near term). Insilico designs the molecules with its AI platform; Servier shares the costs and takes over clinical work. Founder and CEO Alex Zhavoronkov spoke for the company.
Why it matters: A milestone payment is the partner paying because a programme met terms both sides set in advance, which is a firmer sign of progress than a deal's headline value. For a company that earns most of its revenue from such partnerships, this is the first money from Servier tied to reaching a stage rather than to signing.
Evidence and limits: The release names no target, molecule or criterion, and gives no payment amount. The claim that AI shortened the timeline is Insilico's own.
Ligo Biosciences' Artios designed esterases from scratch, and 17 of 88 were active in the company's own screen
What happened: Ligo Biosciences tested four sets of Artios designs in its own lab, at most one 96-well plate each: of 88 designs for a short-chain ester, 17 were active, the best at 291 per molar per second for the acylation step; a second 88 aimed at an eight-carbon ester gave one hit, the purified ArCut1, at about 1,800. Luciferases came in at 7 of 85, then 9 of 45 on a later version. An all-atom diffusion model builds the backbone around the catalytic atoms, and a Potts model writes the sequence.
Why it matters: Directed evolution can only improve an enzyme that already exists, so a reaction with no natural starting point has none. Past de novo luciferases came from huge screens: the paper notes LuxSit was one of three hits among 7,648 designs. Artios reports hits from a single plate, after about a day of computation that generates up to a million backbones, a week for the DNA and two days of assays. Mutating either catalytic residue of its best luciferase to alanine dropped the light to the no-enzyme level, so the activity comes from the designed site.
Evidence and limits: Ligo designed, ran and scored every experiment, and no outside lab has repeated it. The paper calls selection its principal limit: AlphaFold2 and Protenix choose which designs get made, and cannot judge catalysis. The screens read one well per design against a threshold three standard deviations above the no-enzyme controls. The activities stay well below natural enzymes and lab-optimised designs, and no design has an experimental structure yet.
MATARIA, a Shanghai company led by Zheng Liu of Singapore's Nanyang Technological University, raises a seed round of tens of millions of yuan led by Legend Capital
What happened: MATARIA, founded in Shanghai in September, announced the seed round on 19 October: Legend Capital led and Xiaomiao Langcheng joined. Zheng Liu, a chair professor of materials science and engineering at Nanyang Technological University in Singapore, is the founder; Bijun Tang, his former doctoral student, is CEO; Bo An, a presidential chair professor at the same university, leads training of the materials model; Yushan Xiao is CTO. The company says its model predicts the strength of commercial composites to within ±1.2%.
Why it matters: A lab in Singapore putting its company in Shanghai, and aimed at engineering materials rather than new molecules: composites for aerospace first, then semiconductor and infrared materials. Its pitch is data from instrumented experiments (furnaces with sensors, growth watched under a microscope) fed back into one model. Its plan is a deployable model in 18 months and pilots in three to five industries.
Evidence and limits: The ±1.2% figure, the infrared-material licence and the customer count are the company's own; no paper or outside test is cited, and the outlets disagree on customers (nearly 20 vs nearly 10). No amount beyond tens of millions of yuan and no valuation.
On 131 protein-structure tasks, three of four frontier agents look up recorded metadata better than they compute geometry from coordinates
What happened: Matt Turk, James Golden and colleagues at Protege DataLab and Heimdall Bio ran 4,249 rollouts of four frontier models, plus 4,486 trials across 26 models. Claude Opus 5 (0.816), GPT-5.6-sol (0.812) and Grok 4.5 (0.785) are not separated by the intervals; Gemini 3.1 Pro Preview scores 0.707 and every gap to it excludes zero. Opus 5 averages 0.907 on metadata lookup and 0.716 on numeric calculation. Numeric tasks are 28 of the 131 but six of the twelve tasks all four models fail.
Why it matters: A researcher picking an agent to measure distances and angles in a structure needs the numeric score, not the overall one, and for Opus 5 the two differ by 0.19. The paper also shows how much of a leaderboard is the setup around the model: on the 99 tasks with eight full runs, one run per task instead of eight puts a different model first in 23.9% of resampled task sets, and swapping each provider's own command-line agent for the reference harness moves scores by -0.071 to +0.086.
Evidence and limits: The four experiments use different task subsets, so model, harness, access and task mix cannot be ranked against one another, and the harness comparison can only detect differences of about 0.12 to 0.20. A task all four fail is not shown to be unsolvable. The 131 tasks lean to biophysics, and the scores describe those model endpoints in the collection windows the appendix gives. A preprint, not peer reviewed.
AuthorsMatt Turk, Brandon L. Neel, Diana M. Lopez, James B. Golden
Replayed over 203 archived spectra from a self-driving lab, an NTT agent decides when the fitting script can be nudged and when no number should be reported
What happened: Yuki K. Wakabayashi and Yui Ogawa put GPT-5.6 Sol around the unchanged Python script that extracts Urbach energy, an optical-quality measure, from transmittance spectra, and ran it over 203 archived spectra from NTT's own sputter-grown gallium-oxide film campaign. The script had returned a value for 183 and failed on 20. The agent kept all 183, recovered 16 of the failures by shifting the fitting window or relaxing the fit-quality threshold within set bounds, and returned NaN for 4 whose spectra do not support a reliable Urbach tail.
Why it matters: When a fitting script fails inside a closed loop, someone has to say whether the measurement can still yield a number, and that call is what stops the loop and waits for a person. Here the agent never touches the data or the physics (only parameters the script already exposes), and the number is still computed by the script. Abstaining is treated as the right answer rather than a failure, which is the part a loop that optimises on whatever number it gets cannot do.
Evidence and limits: Nothing new was grown or measured, and there is no independent value to check the agent's decisions against: whether a repair or an abstention was right is the authors' own reading of the spectra. Every spectrum went to the agent, but it was only ever asked to intervene where the script had already failed, so it is untested on an analysis that finishes and returns a physically wrong number. One descriptor, one campaign, one lab. A preprint, not peer reviewed.
A team audits the single-cell perturbation harness it built and finds an identity predictor, returning the input unchanged, scores a perfect perturbation-effect correlation on all six splits
What happened: Manuel Gallone, Shyam Ravidath and Srihith Madireddy built an out-of-distribution harness (two datasets, six dataset-split combinations, scVI, CPA and a Ridge baseline, three uncertainty methods), then retrofitted null baselines it had never included. The harness fed each test cell to the model and scored it on reproducing that same cell, so an identity predictor reaches a perturbation-effect correlation of 1.0 everywhere. Against a constant train-mean predictor, scVI wins on 99.7% of Norman cells and 0.2% of Replogle cells.
Why it matters: Perturbation prediction is one of the directions this record follows most closely, and the audit says metric names there can outrun what is measured. The six failure modes are mechanical rather than statistical (scoring reconstruction as prediction, holding out sequencing batches instead of perturbations, silently swapping an unseen perturbation for a control), and each comes with a diagnostic that needs no retraining, so another harness can be checked largely from what it already stored. The one model never handed the target cell, the linear baseline, is the only one that beats its null on a genuine counterfactual.
Evidence and limits: It is the authors' own harness, not a published benchmark; PerturBench is cited but never audited. The audit needed no retraining, though models were re-run counterfactually, and everything is computed on two datasets with no new experiment. One result survives: split-conformal coverage holds at 0.900 in distribution and falls to 0.847 across cell types, while on Norman that nominal 0.900 covers 0.271 to 0.342 of differentially expressed genes.
Six foundation interatomic potentials are pushed onto disordered carbon, where low energy and force errors do not reveal the failures that binding curves and dynamics expose
What happened: Joseph Hart, Samuel G. H. Brookes, Fabian L. Thiemann and Christoph Schran screen six models on density-functional energies, forces and binding curves, then put four through melt-quench-anneal dynamics at seven densities, five runs each. Every pure-carbon run finishes. Adding oxygen and hydrogen separates them: MACE-MH-1 completes all 35 runs, ORB-v3 33, MACE-MP-0b3 25 and MACE-OMAT-0 18. UMA never reached this stage (its carbon-dimer curve has two spurious minima) despite low errors on flat-carbon environments.
Why it matters: A foundation potential is sold on transferring to systems it never trained on, and the usual evidence is a table of energy and force errors. Here those errors do not say which models break: UMA scores well on flat carbon, then is thrown out for a dimer curve no potential should have, while models sharing a training set share ring statistics across architectures. All four that ran reproduce the main experimental peaks for nanoporous carbon, so the question is where they stop, not whether they work.
Evidence and limits: One element family and one workflow. Stability is scored in 216-atom cells; the structures compared with experiment are larger, at one density, and matched to published measurements of different samples. The first author co-authored both the model that comes out best and the scoring framework the error table uses, and the paper declares no competing interests. Two of the six never reached dynamics, one on cost alone. It is a workshop paper.
METiS TechBio will design lipids for Eli Lilly with its own AI platform, under a collaboration worth up to about $1.1B in milestones
What happened: METiS announced a strategic collaboration with Lilly on 19 October to research lipid nanoparticle delivery that reaches one tissue rather than everywhere. METiS runs its own AiLNP platform, which designs the ionizable lipid inside the particle, and Lilly provides its drug discovery, development and commercialisation expertise. METiS takes an upfront payment of over $10M, is eligible for up to about $1.1B in development, regulatory and commercial milestones, and single-digit tiered royalties on net sales.
Why it matters: The record holds METiS's Hong Kong listing, its mouse data for a liver-cancer mRNA drug and a Phase 3 formulation result. This deal is about the platform rather than one drug asset: Lilly is paying for the design step, which is where METiS says its AI sits. Where a nanoparticle ends up in the body decides what an mRNA or gene therapy can treat, and it is still mostly found by trial and error.
Evidence and limits: A deal, not a result: no target, no programme count, no timetable, and nothing public about how well AiLNP picks lipids. All the money beyond the upfront is contingent on programmes advancing, and the upfront is stated only as a range.
MOFFieldBench computes electron density and electrostatic potential for 46,000 metal-organic frameworks, so models can learn from the fields as well as the atoms
What happened: A Zhejiang University team ran DFT calculations, about a million CPU core-hours, on 45,909 frameworks from six databases, compressed the resulting fields about 26-fold and turned them into point clouds. Models reading the fields beat a structure-only model on 12 of 13 properties, all the adsorption and separation ones among them; the structure model kept band gap. Combining the two did better still on every property. Electron density predicted from structure is off by 0.51% on average, and starting DFT from it cut run time by 13.59%.
Why it matters: MOF models mostly learn from atom positions, while gas adsorption depends on how a guest molecule feels the framework's charge. Aligned electronic fields at this scale were scarce because they are costly to compute and store; the authors say the dataset will be released when the paper is published. The results say the fields carry information structure-only models were not using, most of all for adsorption and separation.
Evidence and limits: All computed, by the authors. The structure baseline is one model, CGCNN from 2018, so the comparison with structure rests on it. The electrostatic potential is reconstructed far worse than the density, 8.42% off and worse in large cells. The 13.59% saving leaves out the time spent predicting the density, and is an average: some frameworks ran slower.
Electric Plant's Bombadil learns plants' electrical signals from 4 million sensor-hours across 100 species, and follows hand-measured leaf stress with correlations of 0.27 to 0.71
What happened: James Kirk, Lee von Kraus and colleagues trained a Transformer, without labels, on over 4 million hours of recordings from electrodes in more than 1,400 plants of 100 species, in labs, greenhouses, orchards and forests. A second part, trained on over 41,000 handheld leaf readings, estimates seven stress measures such as how efficiently leaves photosynthesise and how warm they run. On held-out plants of six species under drought, its estimates moved with the readings at correlations of 0.27 to 0.71.
Why it matters: Earlier models of plant electrical signals were each trained for one task, usually on one species. One model that carries across species would let a grower read a crop nobody has labelled. On species left out of training the estimates still moved with the readings, less closely, and on plants stressed by salt or a herbicide, which the model never saw labelled, they shifted the way the literature expects.
Evidence and limits: The data, sensors and evaluation are all the company's own, in a preprint. The correlations follow one plant over time; about half the error is a fixed offset per sensor, and on the error measure, three of the seven estimates did worse than guessing each species' average. One of the six densely labelled species read at or below zero, and nothing yet shows a grower acting on it.
DeepSeek opens the Search and Fetch APIs behind its own search engine to every API user, at $0.75 per thousand calls
What happened: The two endpoints return ranked web results with a relevance score and an authority tier, and a page's content as Markdown. DeepSeek says its in-house engine has long served the search in its web app and API, handles hundreds of millions of requests a day and indexes hundreds of billions of pages, with filters for SEO spam, GEO manipulation and low-quality AI-generated text. Calls made through the server-side tools in DeepSeek Harness, Claude Code and Codex stay free until 31 October.
Why it matters: A literature agent needs a search backend it can call by program, and almost all of them sit on Bing's or Google's index. This is one of the few that does not, from DeepSeek, whose open models many AI4S groups already run, priced per call rather than by tier. Whether its index covers journals, preprint servers and Chinese-language sources well enough for science is for anyone building on it to test.
Evidence and limits: The request and index figures are DeepSeek's own and unaudited; the English release note gives the daily requests and the Chinese one the index size. No coverage or freshness comparison against other search engines was published, and the endpoints were not called here.
DeepSeek updates its Flash model to V4.1-Flash-1017 under the same name and price, with MIT-licensed weights and higher scores on reasoning and agent benchmarks
What happened: DeepSeek's model card calls it a minor update to September's V4.1-Flash: Humanity's Last Exam without tools 42.6 (from 36.8), Terminal-Bench 4.0 42.3 (31.2), DeepSWE v1.1 77.0 (74.2), Apex Bench v1.1 72.1 (64.6), with shorter runs and outputs. It serves as deepseek-flash on the web, in the app and on the API. The base is September's 552B-parameter mixture-of-experts model, a causal encoder-decoder that activates 8B parameters per token in prefill and 16B in decode; the checkpoint on Hugging Face holds 763B.
Why it matters: Humanity's Last Exam is written by subject experts across the sciences and mathematics, so a near-six-point gain there at an unchanged price changes what an open-weight model at this cost can answer. The weights are MIT-licensed: a group can run it on its own machines, with data it cannot send to a hosted model.
Evidence and limits: Every score is DeepSeek's own run, with harnesses it chose; no independent evaluation had been published by 17 October. Nothing here tests a science task beyond exam questions, and the September technical report was not read.
Google Research runs AI agents in a closed loop with a wet lab for three rounds, and designed aptamer hit rates rise from 0% to 42% on thrombin
What happened: Zachary Cutts, Pooja Rao, Annalisa Pawlosky and colleagues gave Gemini-based agents literature search, folding tools and a Python environment, and had them design 1,000 single-stranded DNA aptamers per round against two proteins. Each pool was synthesised, pulled down on beads and sequenced, and the agents read the counts before designing the next. Corrected for bead background, hit rates went 0% to 26% to 42% on alpha-thrombin, or 37% counting only designs free of published motifs, and 10% to 4% to 34% on myoglobin, at about $3 a candidate.
Why it matters: Earlier studies of reasoning agents in molecular design tested their designs once, at the end. Here the measurements came back three times, and what the agents did with them is on the record: after a counter-screen showed that 97% of the cross-reactive sequences were G-quadruplexes, which stick to charged surfaces whatever the target, they moved the myoglobin designs to pseudoknots and hairpins, and specificity there went from 19% back to 72%. The authors picked aptamers because the designed sequence is the molecule, so one cycle is 10 days instead of the 3-6 weeks a protein needs.
Evidence and limits: Binding is measured in the authors' own pooled assay. Single-clone titrations covered eight thrombin designs only, did not reach saturation, and the authors call those affinities preliminary: nanomolar for reworked literature motifs, above a micromolar for de novo ones. Round 1's 0% on thrombin was a construct artifact the agents then fixed, the harness changed after it, and the authors say control arms with no feedback would be needed to credit the loop itself. A preprint, not peer reviewed.
A plug-in that adds differentiable restraints to the denoising step wipes out ligand chirality and cis/trans errors in six structure predictors, AlphaFold3 among them
What happened: Tatsuki Hori, Yoshitaka Moriwaki and Ryuichiro Ishitani at Institute of Science Tokyo and the University of Tokyo release RGI-Toolkit, which corrects the denoiser output at each reverse-diffusion step with a differentiable restraint loss, as a backend-agnostic library wrapping six predictors: Boltz-2, Protenix-v2, AlphaFold3, Chai-1, OpenFold3 and ESMFold2. On 209 protein-ligand complexes, conformer restraints lifted chirality agreement from about 50% to 100% and cis/trans from 70-80% to 100% for all six, with lower bond-length and bond-angle errors. No retraining. MIT-licensed.
Why it matters: Chemically invalid ligand geometry is a known weak spot of AlphaFold3 and its kin, and this fixes it from outside the model, at inference, without retraining it or changing a weight. It does not touch where the ligand sits: the authors report pose accuracy is neither helped nor hurt, so the gain is a clean stereochemistry patch, not a better binder prediction. Geometric and custom restraints can also push sampling toward a chosen conformational state.
Evidence and limits: The chirality and cis/trans figures are computed agreement rates against each ligand's ideal conformer, on 209 complexes with one seed each, not a physical measurement. The restraints only work as well as the reference conformer: a bad RDKit embedding misguides them. The two earlier RGI papers this builds on were in Boltz; the new work is extending that to six predictors and packaging it. A preprint, not peer reviewed.
One agent per perturbation: 1,946 copies of GPT-Rosalind re-derive most of a whole-brain Perturb-seq study's findings in 32 minutes, without the paper
What happened: Scripps Research, OpenAI and PerturbAI gave one agent to each of 1,946 perturbations in an in vivo whole-brain CRISPR Perturb-seq atlas, with the source manuscript withheld and published after the model's stated knowledge cutoff. Of 55 findings derivable from the expression data alone, the agents matched or refined 48. They adjudicated 1,519 high-confidence findings against the literature, reading 39,185 papers and keeping 16,445 citations. The run took 32 minutes.
Why it matters: The bet is coverage through many parallel agents rather than one stronger agent, and the paper measures it directly: at 893 perturbations a single agent cited 11 references while the distributed run kept 711 times more. Blinded experts preferred the distributed readings. This is a throughput claim about interpretation, the step that now limits what a large screen yields, not a claim that the agents found new biology.
Evidence and limits: All computed, on one atlas the same group produced, and every corresponding author has a stake: one at OpenAI, whose model this evaluates, two who co-founded PerturbAI, which hosts the data. Recovery is scored by the source study's own authors, 14 of 71 findings uncovered. Held-out predictions reach 0.705 balanced accuracy: 82% of the silent perturbations called, only 59% with a real response. The expert preference ranks readings, not their correctness. A preprint, not peer reviewed.
Claude, checked by an automatic evaluator, beats the best published quantum circuits on six of eight standard tasks and widens the trade-off on the eighth
What happened: Andi Gu, an Anthropic STEM Fellow, with Alek Kemeny of Anthropic and Harvard's Susanne Yelin and Mikhail Lukin, had Claude Mythos 5 write circuits in Python, each checked and scored by an evaluator. In six tasks the circuits cut both Toffoli and qubit counts against the best published design; a seventh had no published reversible design to beat. In the eighth, a new reversible modular inversion, its key step formalised in Lean, gives one design 17 times fewer gates than the smallest previous one, for 12% more qubits.
Why it matters: Most of the fall in what useful quantum chemistry is expected to cost has come from redesigning circuits, not hardware: the FeMoco estimate dropped about 150,000-fold over eight years of human work. A loop where the model writes circuits as code and a program checks every one speeds that redesign up: Anthropic sets the elliptic-curve design, from a single 53-hour run, against eight years of field-wide work. The same loop cut the costliest step of the DQI optimisation algorithm, its decoder, about 12-fold.
Evidence and limits: All of it is computed: the circuits have not run on any quantum computer, the counts come from the team's own evaluator, and the arXiv preprint is not peer reviewed. People wrote the briefs and at times told Claude to search more broadly.
Anthropic is circulating a confidential draft AI safety bill on Capitol Hill under which a model could be stopped only on a billion-dollar or fifty-casualty risk from three named threats
What happened: A model could be blocked only on a foreseeable, imminent risk of over $1 billion in damages or of killing or seriously injuring more than 50 people, from a biological weapon, a model evading its developer's controls, or a cyberattack. Commerce would ask a federal judge in Washington to order it; no other agency could. Evaluators would register with Commerce, may be lab-paid if the fee does not turn on their findings, and must report an imminent risk within 72 hours. Some state AI laws would be preempted.
Why it matters: A frontier lab has written the text that would govern it. Politico notes its threshold language mirrors California's SB 53, and Anthropic calls the bill its Advanced AI Framework, pushed since June. For science the category that matters is biological weapons, one of three grounds on which a model could be stopped: an evaluator's report to Commerce would be the first step toward a court order there.
Evidence and limits: The draft is confidential and Politico did not publish it, so every provision here is Politico's account of a text it obtained; three people with knowledge are its sources for the pitch itself. The article does not name which lawmakers received it. An unnamed Anthropic spokesperson endorsed the bill as the company's framework without publishing the text.
Google DeepMind's Gemini 4 Argon safety report finds no significant speed-up in its own AI progress, though most of its production code is now over 75% AI-written
What happened: Google DeepMind fitted its own capability index, GCI v1, to 123 models on 92 benchmarks. Public Gemini models gain about 0.1 points a month: 0.091 before 1 April 2026, 0.118 after, a difference that is not statistically significant. Most production code is now over 75% AI-written, yet fully delegated pull requests rose only from 7.5% in April to 11.9% in September; about 88% still need multi-turn human steering or human-written code.
Why it matters: Whether AI is speeding up its own development now has a frontier lab's internal measurements behind one answer, set against the September working paper co-signed by Geoffrey Hinton, Yoshua Bengio and OpenAI's chief scientist, which estimated that fully automated research could make progress tenfold faster. A survey of 12 senior researchers puts automation highest on well-specified, verifiable tasks and lowest on open-ended scientific ideas and research taste. Argon sits above the trend, which the report suspects is due to more compute and organisational investment.
Evidence and limits: The index is DeepMind's own and built on benchmarks: a steady slope means measured capability grew steadily, not that no research step got faster, as the report itself notes. The post-April slope rests on few models (0.053 to 0.171 at 95%), so it cannot rule out a speed-up either. The survey and the agent-use sample, graded by Gemini Flash 3.8, are internal, and the finding supports the report's own conclusion that the risk is low.
EZPro-Fit, a model that designs an enzyme's protein and its DNA together, proposed cellulase variants; of 13 tested, two were both more heat-stable and more active
What happened: Ran Xu, Liangzhen Zheng of Zelixir, Dongyang Liu, Jingjing Guo and colleagues at Macao Polytechnic and Nanjing Agricultural universities trained a model on the protein structures and gene sequences of heat-loving microbes, then tuned it to E. coli. From structures, it recovered 55.54% of residues, 17.4% more than ProteinMPNN in relative terms. For the cellulase cel5E, its candidates passed stability and conservation filters and 13 were tested: N291D raised the melting point from 63.16 to 69.58 °C with about 35% more activity, D259E reached 72.35 °C with 17% more.
Why it matters: Most protein design stops at the amino-acid sequence; this model also writes the DNA the host cell will read, and the lab test asked for two properties at once, heat stability and activity, which often pull against each other. The double mutant D218E/S262I shows that pull: the most stable variant at 78.12 °C, it kept about a third of the activity.
Evidence and limits: One enzyme and 13 variants, with no comparison against variants picked by another method or at random, so the hit rate cannot be judged; the candidates also passed energy and conservation filters, so the model's own share is unclear. The DNA design was scored only in software. A preprint, not yet peer-reviewed.
Blitz cuts AlphaFold 3's structure-sampling stage from 200 steps to eight with little loss of accuracy, and shows how averaged scores hide the defects that come with fewer steps
What happened: Vinay Saji Mathew and colleagues at Penn State and the University at Buffalo distilled the diffusion stage of AlphaFold 3 and Boltz-2, the part that turns a predicted layout into atoms, into an eight-step student. On about 2,300 Boltz-2 benchmark targets, the AlphaFold 3 student scores 0.8635 complex lDDT against its teacher's 0.8615, and predicted ligands passing PoseBusters' validity checks fall from 45.0% to 43.0%. For Boltz-2 the cost is larger: ligand validity falls from 60.9% to 52.0%, with structures carrying any geometry violation rising from 30.3% to 41.9%.
Why it matters: Protein design and screening pipelines call a structure predictor tens of thousands of times, so the sampling stage's cost adds up. The paper also argues that speed-up claims have been scored too loosely: earlier few-step samplers were judged on similarity scores with the chemistry checks switched off, which stay high while bonds, angles and clashes go wrong. Blitz reports both, under one protocol for every method, and shows a one-step sampler losing about 32 points of lDDT once the checks are on.
Evidence and limits: Computed, by the authors, on public benchmarks. Only the sampling stage gets faster: the paper says the trunk, the network that runs before sampling, is unchanged and still costly, so eight steps is not a 25-fold faster prediction. No end-to-end timing is reported.
AuthorsVinay Saji Mathew, Soundar Kumara, Gretta Kellogg, William K. M. Lai
HighFold-Lasso gets AlphaFold 3 to thread all 52 lasso peptide structures in its test set correctly by starting it from a lasso shape
What happened: Lasso peptides are short molecules whose tail is threaded through their own ring, which makes them stable and hard to predict. A Macao Polytechnic University team led by Hongliang Duan teaches AlphaFold 3 the ring's unusual bond and starts its sampling from a lightly perturbed lasso shape. On 52 experimental structures it gets the threading right every time, with 2.375 Å mean backbone error over the whole peptide. Plain AlphaFold 3 threaded 67.3% at 2.72 Å; LassoPred, a lasso-specific tool, threaded all of them at 3.50 Å.
Why it matters: Lasso peptides are scaffolds for peptide drugs, and one, lariocidin, is a broad-spectrum antibacterial that has worked in mice; engineering them starts from their 3D shape. General predictors often return the right bonds with the tail unthreaded, which is in practice a different molecule. Building the known topology into the starting point fixes that while keeping AlphaFold 3's accuracy.
Evidence and limits: Computed, by the authors. The threaded starting shape comes from LassoPred, so the 100% shows the method keeps a threading it is handed, not that it finds one, and it leans on the tool it is compared with. Of the 52 test structures, 50 come from LassoPred's benchmark. The gain over LassoPred is 3.50 to 2.375 Å; over AlphaFold 3, about a third of an ångström.
A Toronto system given a goal, photos of the room and the instruments' interfaces built a digital twin of its lab, taught its robot arm the plate transfers and ran a 96-well quantum-dot screen
What happened: Jiaru Bai, Alán Aspuru-Guzik and colleagues at the University of Toronto describe agents that start from the equipment in a room, rebuild it as a simulated twin, learn missing robot skills by trial, first in simulation, and write the workflow. In the one physical test, after a human approved the plan, the system wrote a 236-command liquid-handler protocol, had the arm carry the plate to a sealer and then a plate reader, and an 8-hour cadmium sulfide quantum-dot screen found the ligand and pH range where emitting dots formed.
Why it matters: A self-driving lab today runs only the procedures it was built for, and adding a new one takes specialist engineering. If agents can build the missing pieces, what a lab can try is limited by its equipment rather than its code, and researchers set goals instead of programming robots. The authors also turn each solved skill into a script that runs without the language model, because the agent loop was too slow for the chemistry.
Evidence and limits: One physical workflow on known chemistry, run by the authors; coordination across several robots was shown in simulation. The paper gives no count of human interventions or of engineering time saved, and says people still had to settle disagreements between agents and supply physical facts the instruments could not sense. It is a preprint.
Accel is set to lead a $60M round for Nando de Freitas's Revolution Labs, below the $100M+ reported in September
What happened: Sifted reported on 15 October that Accel is set to lead a $60M round for Revolution Labs, founded by Nando de Freitas after about a decade at DeepMind and a stint at Microsoft. Dealroom's summary of the report adds that GV and NEA are said to have joined. The company builds diffusion models, extending them from images and audio to text. In September Bloomberg had reported talks to raise at least $100M initially.
Why it matters: The September figure was a target in talks; this is the first report of a round sized and led. At $60M, the first close is smaller than the $100M+ floor reported then, though Bloomberg also said the raise would likely come in several tranches. It is one more DeepMind alumnus raising for a bet on a different kind of model than today's language models.
Evidence and limits: Media-reported and not closed: Sifted's headline says Accel is set to lead, and Sifted's text could not be read here. GV and NEA come only from Dealroom's summary. No stage, valuation or company statement; no model has been published.
AstraZeneca pays insitro $40M upfront for AI-led target and drug discovery in MASH and cardiomyopathy, with up to about $1.46B in milestones
What happened: insitro will use its Virtual Human platform to find disease-causing biology in both diseases and take targets AstraZeneca nominates from discovery to a development candidate. The collaboration is worldwide and exclusive, and AstraZeneca runs clinical development, regulatory work and sales. Beyond the $40M upfront, insitro can earn up to about $1.46B in discovery, development and commercial milestones, plus single-digit royalties. Daphne Koller, insitro's founder and CEO, announced the deal.
Why it matters: A large drugmaker is paying for an AI company's reading of disease biology at the target stage, the earliest and riskiest decision in drug R&D, and not only for molecule design: insitro's platform looks for what causes the disease, then takes targets AstraZeneca picks through to a drug candidate. insitro already works with Bristol Myers Squibb, Lilly and Gilead; AstraZeneca is the fourth.
Evidence and limits: The $1.46B is a ceiling that depends on milestones being met; only the $40M is certain. The release names no target, no number of programmes and no timeline, and gives no data from the platform.
CERN and Mistral sign a statement of intent to build one foundation model for collider physics, trained on billions of simulated LHC collisions
What happened: The two plan a long-term partnership on foundation models for science, starting with high-energy physics. The goal is a single pretrained Large Physics Model that could reconstruct collisions, identify particles, flag unexpected signals and help evaluate future detector designs: jobs now spread over many separate tools; reconstruction alone runs through a chain of specialised algorithms. Enrica Porcari, CERN's chief information officer, and Johannes Brandstetter, Mistral's vice-president of AI for engineering, spoke for the two.
Why it matters: Particle physics has used machine learning for years, but one model per task. The plan is for a frontier model company to learn across the whole reconstruction chain, at the lab holding about an exabyte of LHC data, before the High-Luminosity LHC raises the collision count tenfold around 2030. Brandstetter co-founded Emmi AI, the physics-simulation start-up Mistral bought in May, and worked at CERN before that.
Evidence and limits: A statement of intent, not a contract: CERN's release gives no budget, timeline, compute figure or terms on who owns or can use the model. Nothing has been trained yet, and 'the first foundation model for collider physics' is the release's own description.
The European Commission asks companies and foundations to pledge models, computing and software licences to Europe's AI-for-science researchers; seven sign up first
What happened: Launched at the European AI in Science Summit in Dublin under RAISE, the Commission's AI-for-science programme, the initiative collects in-kind help (access to frontier models, computing, software licences, data, training) beyond what public funding provides. First pledges came from the Novo Nordisk Foundation, NVIDIA, Elsevier, Sanofi, Frontiers, Mistral and Google DeepMind. Elsevier's own release lists its share: 4,000 free licences, 1,000 of them for its LeapSpace research assistant, and about EUR 1M a year of in-kind support and funding.
Why it matters: The initiative is part of the EU's AI in Science Strategy of October 2025, and asks companies and foundations for support beyond what public funding alone can provide. A pledge has to show support beyond the pledger's direct commercial interest, and the Commission will present pledges and progress at each year's summit, so what was promised can be checked against what was delivered. It counts CERN and Mistral's new partnership on a physics foundation model as part of the initiative.
Evidence and limits: The Commission's release gives no total value and no terms for six of the seven pledges; Elsevier says it will refine its own details with the Commission in the coming weeks. Pledges are voluntary and in kind, and some, like Elsevier's licences, are free use of the pledger's own products.
OpenAI opens GPT-Rosalind to US university labs through a Rosalind Academic Plan: up to 30 free ChatGPT seats for up to a year, no API
What happened: OpenAI's terms: up to 30 ChatGPT Business Standard seats at no charge for up to twelve months, for eligible labs at US universities, with research data on those accounts kept out of model training by default. A seat does not unlock GPT-Rosalind. That needs a separate trusted-access review of the lab head's identity, employment and authority, the research purpose, and a named institutional safety contact. Subscription seats only, so no API, and a university holding an Enterprise agreement applies through the school.
Why it matters: Until now the model went to organisations through trusted access, with Amgen, Moderna, the Allen Institute and Thermo Fisher as its named partners. This is the first way in for a university lab on its own, and it says where OpenAI expects life-science demand to come from next. Without the API it stays something a researcher types into rather than something a lab wires into its own pipeline, which is the difference between a tool and infrastructure.
Evidence and limits: What counts as an eligible lab, how long the trusted-access review takes, how many labs will be admitted and what happens to access after twelve months are all unpublished. Benchmark scores for GPT-Rosalind are circulating from a single secondary source and are not on OpenAI's page, so they are left out here.
XtalPi names its first siRNA drug candidate, for IgA nephropathy, nine months after the programme began on its Kodexia platform
What happened: The company says generative design found a stretch of the target's mRNA nobody had used, and set the siRNA sequence and its chemical modifications together. In its own tests the candidate silenced the target at sub-nanomolar concentrations, with an IC50 tens of times below two clinical-stage molecules against the same target, and one dose kept the protein suppressed in monkeys for months. XtalPi says toxicology showed a broad safety window and will take the candidate toward the clinic.
Why it matters: On 6 October XtalPi said this programme had reached monkey data in seven months; this is the next step it promised then, a candidate chosen for development. It is XtalPi's first named candidate in siRNA, a modality new to it, and the release says the model's ranking of the designs held up in animals. For a platform company that sells design work, a candidate of its own is an asset it owns.
Evidence and limits: Every figure is XtalPi's own, in a press release with no data, figure or paper; the target and the two reference molecules are not named. A preclinical candidate comes before IND-enabling studies and any human dose.
OmicsBench breaks 97 bioinformatics studies into 662 dependent subtasks, and no model and harness combination comes close to saturating it
What happened: Gideon Shaked and Siddharth Nagisetty, both at the University of Michigan and AfterQuery, build each task as a directed acyclic graph of subtasks over a real biological dataset, posed one per conversational turn and scored against an answer key by 58 graders that match entities through reference databases, not by string. All twelve models ran on one harness, two on two more. Means ran 0.43 to 0.57 with no feedback and 0.54 to 0.67 when a failing answer could be retried; the slowest agent averaged 3.6 hours a task.
Why it matters: Most bioinformatics benchmarks ask one question in one prompt, which cannot show whether an agent carries intermediate results through a real analysis. Here every combination scores lower after its first turn, and lower again on subtasks that build on its own earlier answers, with feedback winning back only part of it. That shape of failure is exactly what a single-prompt score cannot see.
Evidence and limits: Scores measure agreement with an answer key, not whether the analysis is biologically sound. The authors call the retry mode open to reward hacking by design and do not quantify how much of its gain that explains. 94 of the 97 tasks are private, so the numbers cannot be reproduced from outside, and harness effects cannot be separated from model effects. The failure labels come from Opus 5, one of the models judged. A preprint, not peer reviewed.
Holyvolt raises $45M to wire battery AI models into the high-throughput lab it bought, and take a nickel- and cobalt-free cathode to market in 2027
What happened: Zone II and CC.VC backed it, with angels the release places in finance and Fortune in defence and finance. Fortune reports the round is all equity and values Holyvolt at $245M, and later struck Fifth Wall from its own list of this round's investors after what it calls a disclosure error by the company; the release names it as an existing investor. The lab is Wildcat Discovery Technologies, bought in 2026: 18 years of runs, more than 500,000 cells cycled. Mathias Ingvarsson is CEO, Steven Prast, formerly of Tesla, COO.
Why it matters: Most AI-for-science companies build a model and then look for data to feed it. Holyvolt bought the lab and the 18 years of cycling data first, then trained on them, and its models now pick what the lab makes and cycles next. The timing comes from procurement: from 1 October 2027 the Pentagon cannot buy batteries from CATL, BYD and four other named Chinese makers, and Holyvolt plans to sell its DRX cathode into drones first.
Evidence and limits: Nothing here is a result. The ten-times figure comes from one customer case study, and the hundred-times figure for the closed loop is the company's own. DRX data has been shown at scientific conferences, the International Battery Conference in March 2026 among them, and the suppliers are undisclosed. The $245M valuation rests on one person Fortune does not name.
Google DeepMind shows its Antigravity agent hunting antimicrobial peptides, and reaches one candidate now in wet-lab testing with no results yet
What happened: A post by Pengfei Tian, Maxwell L. Bileschi, Lucy J. Colwell and colleagues walks one search through Antigravity, which they call an agentic science workbench. The agent reads UniProt entries through ProtNLM, pulls structures from the AlphaFold database, runs BLAST at EMBL-EBI and Foldseek, and shows the result in a 3D viewer newly embedded in the product. It lands on a 97-residue bacteriocin found in Streptococcus pseudopneumoniae, UniProt A0A0U0D591, whose predicted activity is against the pathogens behind respiratory-tract infections.
Why it matters: It is a capability update and a workflow demonstration, not a finding: what it shows is several databases and services folded into one session a biologist drives. The post's own framing is that the wet lab is still the test. It marks where DeepMind is pointing a general coding agent, and the one candidate it names has reached an assay.
Evidence and limits: The post reports no experimental result: no count of candidates generated, none tested, no hits, no potency. The named candidate and several others from the same pipeline are described as in wet-lab validation, with results framed as still to come. No paper or preprint is linked. Every step shown is the team's own walkthrough of its own product.
Made to invent each atom's electronic fingerprint alongside the structure, a crystal generator triples its hit rate on the hardest two-property target
What happened: Ibuki Okuda, Izumi Takahara and Teruyasu Mizoguchi extend Microsoft's MatterGen so one network denoises the crystal and, for every atom, either its Bader charge (how much charge it carries) or its atomic density of states (which energies its electrons occupy). Of 1,024 structures per setting, counting only stable, unique, novel ones, 1.1% of the baseline's landed inside both a 2 eV band gap and a -4 eV per atom formation energy, against 3.5% of the density-of-states model's. Single targets gain more in points and less in proportion.
Why it matters: Asking a generator for a target property costs structural quality, and the electronic variable makes it cost less: at that two-target setting the share of stable, unique, novel structures falls from 33.3% to 21.8% with the descriptor and from 33.6% to 15.4% without it. The control experiments are the part to carry away. Corrupting the charges, or attaching a random number that means nothing, also raised the unique-and-novel share, so a novelty gain on its own shows only that something random was added, and it takes stability, validity and the target hit rate to tell physics from noise.
Evidence and limits: All of it is computed: structures are generated and scored by DFT, and nothing was synthesised or measured. A hit means landing within 0.5 eV, or 0.5 eV per atom, of the target. The validity filter is the authors' own caveat (about 30% of their training structures fail its charge-neutrality check), so both the stricter and the looser counts are reported. Several single-property gains are a few points, and some do not reach significance at 1,024 samples.
A fine-tuned Aurora reproduces a chemistry-climate model's ozone, aerosol and radiation about 100 times faster, and satellite ozone in its training corrects biases it would have inherited
What happened: James Weber (University of Reading) and Makoto Kelp (University of Utah), equal authors, fine-tune Microsoft's Aurora on the UKCA chemistry-climate model, in three variants. The first reproduces 23 of 27 evaluated fields at R-squared of at least 0.85 in troposphere and stratosphere, with annual mean biases under 4% for ozone, CO and temperature; a simulated year takes 5.5 minutes on one GPU against 9.5 hours on 576 CPU cores. A second, trained on paired emissions scenarios, reproduces ozone's response to NO and CO. A third is trained on satellite ozone.
Why it matters: Chemistry is the most expensive part of a climate model, so it is usually simplified or fixed in advance, and trying many emissions scenarios needs it fast. The third variant is the interesting one: trained on Aura MLS satellite ozone and tested on years held out, it cut the error in ozone averaged around each latitude band by 23%, from 0.37 to 0.28 ppm, roughly 10 to 25 km up. That correction sits in the weights, so it carries to runs where no observation exists: an emulator can be better than the model it learned from.
Evidence and limits: Most of what is measured is agreement with UKCA, so a well-reproduced field inherits UKCA's own errors, and the species predicted are a subset the authors chose of UKCA's 100-plus. The satellite gain holds only in the latitude-band average: the same model is 9.6% worse at individual retrievals. The emission-response variant has its winds and temperature prescribed from the control run, and the 50-year run recycles one year of forcing.
Terray and NVIDIA fine-tune a Nemotron model to read chemical reaction diagrams, setting a new best on the RxnCaption benchmark
What happened: Terray Therapeutics, with NVIDIA, fine-tunes NVIDIA's Nemotron 3 Nano Omni and Nemotron 3.5 Super-VL to turn chemical reaction diagrams (the route drawings in papers) into structured reactions. On RxnCaption's own 15k test set it reports a HybridMatch F1 of 63.4 against the previous best of 59.8, clearing every published number across the sets tested; a single Super-VL run beats the matched Omni model. Zero-shot, frontier models score far lower on that same set (Claude Opus 5.5 at 28.9, Gemini 3.1 Pro at 10.5).
Why it matters: Synthesis planners like Terray's own TerraSynth need training data drawn from how chemists actually describe routes, and most of that sits locked in diagram form in papers. A model that parses those diagrams turns the literature into structured reaction data a planner can learn from, and lets an agent read a chemical document on the fly. The gain over the prior best is modest (about four F1 points), but it is on a public benchmark scored with the original authors' code.
Evidence and limits: A benchmark score from a company post, not a paper, and self-reported: Terray ran the public benchmark with the RxnCaption authors' scoring code, but no outside party checked it. It measures parsing accuracy on a 400-diagram test set, not any downstream improvement to synthesis planning or a drug program. No model weights are released.
Fields Medallist Jacob Tsimerman becomes full-time scientific director of the Mathematical AI Safety Institute instead of joining OpenAI as an employee
What happened: Jacob Tsimerman, a University of Toronto mathematician and 2026 Fields Medallist, said in July he would take leave to work on AI safety at OpenAI. On 15 October he wrote that he will instead work full-time as founding scientific director of the Mathematical AI Safety Institute (MAISI), an independent non-profit, not as an OpenAI employee, though he will still collaborate with it. OpenAI will give MAISI $10 million in API credits and access to its latest models, and join a multi-lab AI safety workshop MAISI is organising next year.
Why it matters: A Fields Medallist is moving his main work from number theory to the mathematics of AI safety, and doing it at an independent institute rather than inside a lab. In July he explained the plan to join one by the gap in his machine-learning background. MAISI's stated aim is mathematical foundations for the safety of powerful AI systems, and its advisers include Timothy Gowers. Its main funding is a $10 million grant from Coefficient Giving; OpenAI's part is credits and model access.
Evidence and limits: The change of plan rests on Tsimerman's own post; OpenAI has made no statement that was found. MAISI has announced its aims and funding, not research results.
The Department of Energy releases Genesis Agent Skills, a public set of instructions that teach AI coding agents how national labs do their work
What happened: DOE released what it calls the first set of technical resources from ModCon and the American Science Cloud under the Genesis Mission. Each skill is a markdown file a coding agent such as Claude Code, Codex or Gemini CLI loads when a request fits: jobs on the Aurora, Frontier and Perlmutter supercomputers, dataset documentation, model evaluation, safety scans, plasma simulation codes, literature search. Court Corley of PNNL, the mission's interim technical director, called it "an early example" of the mission turning research into capability people can use.
Why it matters: It is a piece of the Genesis Mission a researcher can install today: lab practice for running jobs on DOE machines, describing data and evaluating models, written so the commercial coding agents researchers already use can follow it. The code is BSD-licensed and open to outside contributions.
Evidence and limits: A release, not a result: DOE gives no measure of whether agents with the skills do the work better or faster. The public copy read on 8 November held 37 skills, and its own README warns that some depend on a particular agent's features or on outside tools.
Sanofi options the first development candidate from its Recursion collaboration, an oral immunology molecule, and pays a $10M milestone
What happened: Sanofi optioned the first development candidate from its multi-programme research collaboration with Recursion: an oral small molecule aimed at a pair of targets in inflammatory disease, which neither company named. Sanofi has added it to its pipeline, will pay for and run development, and expects Phase 1 in 2027. The option triggered a $10M payment, taking Sanofi's upfront and milestone payments to $144M. Recursion can earn more than $300M in further milestones on this programme, plus low double-digit royalties.
Why it matters: A drugmaker choosing to pay for clinical development is a buyer's decision, made on data Sanofi saw and the public has not. It is the first molecule from this collaboration that Sanofi has taken into its own pipeline, after five earlier progress milestones across five programmes. For AI drug-design companies selling discovery work to pharma, an exercised option is the step where the partner starts spending its own development budget.
Evidence and limits: The targets, the candidate's data and Sanofi's reasons are not public, and the account is Recursion's release; a Sanofi statement was not found. Recursion says it used several computational design methods over repeated rounds of prediction, synthesis and testing, with Sanofi's chemists and biologists; the release does not separate what the models contributed. Being optioned is not evidence the molecule works in people.
Recursion sells $200M of new shares at $3.25 and says its cash now lasts into 2029
What happened: Recursion agreed to sell 61.54 million Class A shares at $3.25 each in an underwritten offering, with Leerink Partners, Goldman Sachs, Morgan Stanley and Evercore acting for the underwriters, for about $200M gross; it closed on 16 October with net proceeds of about $187.2M. The company named Adage, ARK, Driehaus, Invus, Janus Henderson and Longaeva among the buyers. Cash stood at $561.5M on 30 September, about $748.7M with the offering, and Recursion says that funds it into 2029.
Why it matters: Recursion is a listed company built on the bet that AI and automated biology make drug discovery repeatable, and how long it can fund its trials decides whether that bet gets tested. It raised the money on the day it reported first responses for REC-1245, a Sanofi option and new REC-4881 data. It also used $77.4M of operating cash in the third quarter, against $117.4M a year earlier, and will look for partners for two programmes, REC-617 and REC-3565, instead of funding them itself.
Evidence and limits: The runway is the company's own forecast and assumes no further financing. The cash balance includes about $86M raised in the quarter by selling 23.7 million shares through its at-the-market programme.
Recursion reports three partial responses among 12 patients on REC-1245, a cancer drug whose target and molecule it says both came from its platform
What happened: Recursion gave the first efficacy data from DAHLIA, its Phase 1 dose-escalation trial of REC-1245, a pill that degrades the splicing protein RBM39. Among 12 heavily pretreated patients at 45 mg a day or more, three had a partial response: two confirmed, in endometrial cancer, and one unconfirmed, in liver cancer; 11 of 12 had disease control. It picked 45 mg daily for expansion cohorts. At that dose, in 10 patients, it saw no dose-limiting toxicity. The data cutoff was 21 August.
Why it matters: Recursion says its platform contributed all three core elements of this programme: the target (degrading RBM39 mimics losing CDK12), the molecule and the first guess at which patients would benefit. Chief executive Najat Khan called it an early example of the whole engine working inside one programme. For partners and investors, that is the claim the platform is sold on, and it now has a small number of patients behind it. Expansion data in endometrial cancer are due in the second half of 2027.
Evidence and limits: Twelve patients, open-label, no control arm, and only the company's release: no duration of response, and one of the three responses is unconfirmed. Disease control in a Phase 1 trial is not evidence of longer life. The 15 October release says the molecule was found after synthesizing only 204 compounds; the 2 November one says after screening 204. The release does not show what the models did at each step.
Designed from scratch, 13 of 69 cysteine proteases cut the peptide backbone where told, and designs cut their targets in cells
What happened: Hojae Choi, Samuel Pellock and David Baker at the University of Washington's Institute for Protein Design, with Donald Hilvert at ETH Zurich, report enzymes that use an activated cysteine to hydrolyse the polypeptide backbone in a sequence-dependent way. 13 of 69 designs worked, 19%, with rate enhancements over the uncatalysed reaction of up to 3x10^7. The folds are unlike natural proteases (TM score below 0.50), six crystal structures sit within 1.2 A of the design models, and the designs cleave their targets in mammalian cells.
Why it matters: A peptide bond in neutral water lasts hundreds of years, and de novo enzyme design had mostly managed reactions with low barriers. Cutting one where the designer chooses turns protein design into a way to make programmable scissors: pick the sequence, get an enzyme that cuts it. That is a different kind of tool from a binder, which is what designed proteins have mostly been.
Evidence and limits: Activity is still short of natural cysteine proteases: the best design, after optimisation, reaches 413 M-1s-1 against 350,000 for the Ulp1 the motif came from, which the paper puts down to imperfect active-site preorganisation and substrate positioning. In cells the proteases cut a substrate the authors designed and transfected in with them, not an endogenous protein, and the work is the designing lab's own.
AuthorsHojae Choi, Samuel J. Pellock, Donald Hilvert, David Baker
Gaowei Yiyao closes a Pre-A round for AI target discovery, with its first candidate for Notch-driven cancers in Phase I
What happened: Huatai GenScript Fund and Orient Jiafu co-led; Guofang Capital, Ruili Investment and Chuchang Fund joined, and existing investors Xinghua Dingli, West Lake Innovation and GF Xinde followed on. Qifeng Capital was exclusive financial adviser. The announcement gives no round size, saying only that the Hangzhou company has raised several hundred million yuan so far this year, and no valuation. The money goes to the AI target platform and partnerships, the multi-centre Phase I study of GW-01, and preclinical work on new programmes. Professor Xu Li, the founder, spoke for it.
Why it matters: This is a round about targets, not molecules: the company's pitch is that its GROOVY model mines protein-interaction data to find targets others call undruggable, and GW-01, a small molecule for cancers driven by an activated Notch pathway, is the first one it has carried into a clinical trial, now enrolling at several hospitals at once, under three years from target to clinic, by its own account. Two more platforms, ACUMEN for pairs of neighbouring surface proteins and NEXUS for tumour neoantigens, were launched recently.
Evidence and limits: Everything about the platform is the company's own account: no paper, no target list, no data on GW-01 beyond that Phase I has started enrolling; 'inhibitor' is PEdaily's own summary box, not the announcement, which says only that the drug targets Notch-pathway-activated cancers. The round size is not stated, and the 'several hundred million yuan' figure covers this year's rounds together, not this one.
Zinc proteases designed with RFdiffusion cut chosen peptide bonds, and some cut TDP-43 and amyloid-beta
What happened: Anqi Chen, Kejia Wu and David Baker at the same institute, again with Donald Hilvert at ETH Zurich, published the same day, used RFdiffusion to put a water-activating zinc site next to a peptide-binding pocket. Of 135 designs tested, 36% cut their target bond; the best beat the uncatalysed reaction by more than 10^8-fold, and four substitutions took it past 10^10. Designed proteases selectively cut the disease proteins TDP-43, amyloid-beta and serum amyloid A, up to 10^9-fold, and switch on caged cytokines.
Why it matters: The disease targets are the point: TDP-43 and amyloid-beta are proteins drug makers have tried to clear, and cutting one is a different kind of intervention from binding it. With the cysteine-protease paper beside it, the lab now reports two independent chemistries for the same hard reaction, one metal-based and one not.
Evidence and limits: Rate enhancements are measured against the uncatalysed reaction, not against a natural protease: out of the computer the designs run 10^3 to 10^4 below the best natural ones, two orders after four substitutions. The disease proteins were cut as purified protein in vitro, and nothing reaches a cell interior, an animal or a patient. The designers have filed two provisional patents on the methods and designs.
AuthorsAnqi Chen, Kejia Wu, Donald Hilvert, David Baker
Ai2 releases HiRO-ACEv2, which generates global 3 km rain, wind and pressure for about 1/300 the cost of the simulation it learned from
What happened: The system pairs ACE, which runs the atmosphere at 100 km, with a new downscaler, HiROv2, that sharpens it to 3 km and now predicts several fields at once rather than rainfall alone. Given coarsened X-SHiELD data it matches that simulation's high-resolution statistics to within 3%, even for events that occur once in a million data points, and estimates tropical-cyclone energy for storms it never trained on. Over a 10-year run the hurricane energy is comparable to X-SHiELD's. The models, notebooks and a comparison demo are public.
Why it matters: Work on extreme-event risk needs many years of storm-resolving data, and Ai2 puts generating it with X-SHiELD at roughly 300 times the cost. At about a three-hundredth of the cost, a decade of global 3 km rain and wind becomes something a research group can generate itself, several times over, which is what estimating rare events takes.
Evidence and limits: Every number is measured against X-SHiELD, the simulation HiRO learned from, not against observations. Ai2 says HiRO only works with the ACE version tuned to X-SHiELD, so it cannot yet produce a warmer climate or reanalysis like ERA5, and winds and pressure around hurricane eyes stay more lopsided than in X-SHiELD.
A year of blind methane releases tests 12 satellites and 15 analysis teams: instruments built for methane catch 92-100% of releases, and AI methods produce 79% of the false alarms
What happened: A team led by Frances E. Reuland and Adam R. Brandt at Stanford released metered methane in Arizona as satellites passed overhead: 481 releases of 9.5 to 1,184 kg an hour and 265 blank controls, with the teams not told which was which. Fifteen teams from nine countries made more than 2,250 estimates from 12 government, commercial and non-profit satellites. Instruments built for methane detected 92-100% of releases with 0-11% false positives; repurposed ones 62-88%. Teams applying AI methods to Sentinel-2 and Landsat images made 79% of all false positives.
Why it matters: Satellite plume reports already feed leak alerts and national methane rules, and the EU's proposed import standard for gas leans on them, yet Carbon Mapper's Tanager-1, EMIT, GHOSt and the AI detection methods had had little or no blind testing. This one gives each a measured detection rate and error against known releases, and leaves an open dataset for training the next detectors.
Evidence and limits: Teams using AI scored 0.62-0.77 accuracy, better than chance, but made most of the false alarms, and the authors say AI methods need more work before routine use. Releases were at a known site: the authors' conclusion holds when the source location is roughly known, and cloud cover over 20% degraded every system.
Iambic prices its Nasdaq IPO at $16 a share, the middle of its range, raising $150M before costs
What happened: Iambic sold 9,375,000 new shares at $16 for $150M gross, and the underwriters can buy 1,406,250 more within 30 days. The shares started trading on 15 October as IAM, opened at $16.15 and closed at $18.91, 18% above the offer price; on 16 October they closed at $23.81. J.P. Morgan, Jefferies, BofA Securities and Citigroup ran the books. Tom Miller, co-founder and CEO, and Fred Manby, co-founder and CTO, lead the company.
Why it matters: Public investors have now put a price on an AI drug-discovery company whose most advanced drug, the HER2 inhibitor IAM1363, is still in Phase 1. The prospectus puts the money, with cash on hand, toward finishing that trial and starting Phase 2 and 3 trials, taking IAM217 and IAM-C1 through Phase 1 dose escalation, and the AI platform, and says it lasts into the fourth quarter of 2028. By the prospectus's share count, $16 a share values Iambic at about $760M.
Evidence and limits: $150M is before fees; the offering was due to close on 16 October and the extra-share option is unexercised. The price says what investors will pay, not whether Iambic's AI-designed drugs work.
Gilead pays Vir $90M upfront to design three masked T-cell engagers for solid tumours, using an AI-assisted protein-engineering platform
What happened: Vir Biotechnology will design and optimise three dual-masked T-cell engagers against undisclosed solid-tumour antigens, using its dAIsY platform and its PRO-XTEN mask, which keeps the drug inactive until tumour proteases cut it free. Gilead takes an exclusive licence and runs development. The $90M upfront is $65M in cash and a $25M share purchase at a 30% premium; up to $45M more comes in near-term milestones and up to $1.74B later. Marianne De Backer, Vir's CEO, announced it.
Why it matters: Gilead is paying for design work, not for a molecule that exists: Vir is to make three candidates to order, with an AI-assisted engineering platform and a masking technology already in the clinic. A day earlier Zealand Pharma likewise paid Generate Biomedicines upfront, $30M, for AI-designed proteins not yet made.
Evidence and limits: How much of the design the AI does is not stated: the release calls dAIsY a platform for multiparameter optimisation and gives no data from it. The $1.74B depends on milestones, and the targets are undisclosed.
Google.org gives $36M to 17 organisations using AI in science, from fusion plasmas and sodium-ion batteries to African drug-metabolism genes
What happened: The Impact Challenge: AI for Science drew applications from 30 countries. Recipients include MIT (agents that read microscope images for materials discovery), the Max Planck Institute for Plasma Physics with TU Eindhoven (physics-informed AI for fusion plasmas), Cambridge (forecasting sodium-ion cathode wear), the Allen Institute (DNA sequences aimed at vulnerable brain cells, shared openly) and the University of Cape Town (variants in drug-metabolising genes). Each can join a six-month Google.org accelerator with engineering help and Google infrastructure. Maggie Johnson, head of Google.org, announced the awards.
Why it matters: Almost all the money goes to universities and research institutes: eight projects in climate, energy and materials, nine in health and life sciences, and four at institutions in Chile, Singapore, Australia and South Africa, plus AfriClimate AI. Alongside the grants, the accelerator puts Google engineers and Google infrastructure behind the teams, so part of the support is access to Google's own infrastructure.
Evidence and limits: Google's post gives no amount per project and no project leads, and each aim is one line; none has a result yet. AfriClimate AI's flash-flood warning time, from under three hours to over six, is a target, not a measured gain.
Leash Bio releases Minos: its pre-measured protein-binding screens as worlds where an AI agent picks molecules to test and gets the real binding result back at once
What happened: Minos turns Leash's DNA-encoded-library screens, millions of molecules measured against many proteins, into environments where an agent chooses molecules within a budget and is told what was measured. The flagship task gives no starting molecule and a budget of 100 molecules over 20 rounds, across 33 targets with about 6 million molecules each. Leash reports about 100,000 agent trajectories in six weeks, releases a few environments and thousands of reasoning traces, and finds that stronger models do better and weaker ones stay too long in poor regions of chemical space.
Why it matters: Because every molecule was measured in advance, an agent's choices cost compute rather than lab weeks, and the results of the molecules it did not pick are known too, so its decisions can be judged, not only its final answer. Leash CEO Ian Quigley's argument is that data collected this way is reusable experience, in the way game moves are for a game-playing model.
Evidence and limits: All results are Leash's own; agent performance is shown in figures this page did not read values from, and the preprint was not opened. The environments cover binding to one target among molecules already screened, with no pharmacokinetics or toxicity, and the published ones are a small sample of what Leash holds.
A machine-learning correction on top of ECMWF's forecasts tops ECMWF's 2025 real-time contest for weather three and four weeks ahead, ahead of six operational forecast centres and 34 teams
What happened: Probabilistic bias correction (PBC) learns from a forecast system's past predictions and the weather that followed how to shift its probabilities. Over 2016-2024 it improved ECMWF's already debiased dynamical model on 91% of pressure, 92% of temperature and 98% of precipitation targets; on the 2025 test year it doubled the skill of ECMWF's AI system. Entered as MicroDuet in ECMWF's AI Weather Quest, it scored best for every variable and both lead times in the September-November 2025 season, and best again in the winter season that followed.
Why it matters: Two to six weeks ahead is where farmers, grid operators and flood planners need a forecast and where models do worst, often no better than the long-term average for the date. PBC is a layer that works on top of any model, and its contest scores come from forecasts issued before the weather happened, set against operational centres' own systems. Hannah Guan and Soukayna Mouatadid led it, with Judah Cohen of MIT and Lester Mackey of Microsoft Research; Cohen says Kenya's meteorological office has asked about using it.
Evidence and limits: The contest covered two seasons, and ECMWF, which runs it, has two authors on the paper. For precipitation, all six centres' debiased forecasts scored below the climatological average, so the gain starts from a low base. Better weekly-average scores do not yet show a decision made better, and the paper reports no operational use yet.
Danaher and NVIDIA say they are working to put accelerated computing into Danaher's instruments, starting with a select group of them and with no terms disclosed
What happened: Danaher's release says the two are working to combine its instruments, automation and lab workflow software with NVIDIA's computing: edge hardware such as DGX Spark, RAPIDS for analytics, and real-time inference and foundation models including the BioNeMo Open Models. Kimberly Powell of NVIDIA states the loop the release describes, experiments improving models and models designing experiments; Julie Sawyer Montgomery of Danaher says AI is reshaping how science is done. Initial efforts cover a select group of priority instruments.
Why it matters: Danaher sells instruments that sit in a great many labs, so where its boxes compute matters: an autonomous lab needs the measurement and the inference close together, and edge hardware at the instrument is one way to get there. The release names the parts rather than a product. It stays an intention until something ships or a customer says what changed.
Evidence and limits: No money, no equity, no timeline and no product. Which instruments are in the select group is not named, the wider portfolio is only a possibility, and the Cosmos world models and Omniverse digital twins are described as being explored. NVIDIA has issued no release of its own. Danaher's separately announced autonomous antibody lab is still at the plans stage, expected to run at scale in early 2027.
Scimplify, which develops specialty-chemical processes in its own labs and coordinates 500 partner plants with agents, raises $25M led by Hitachi Ventures
What happened: The Bengaluru company, founded in 2023, takes a customer's material requirement, develops the manufacturing process with its own scientists in Bengaluru and Hyderabad, and moves it to one of more than 500 partner factories. Its platform, Scinode, runs separate agents for research, manufacturing and quality that help coordinate the work, picking routes and plants and tracking regulation. The all-equity round was led by Hitachi Ventures, with customers Henkel and Teknor Apex as new investors and Accel India, Bertelsmann India Investments, Omnivore and 3one4 Capital following on.
Why it matters: The money is going to the step after discovery: a material counts only once someone can make it to spec, and that is where months turn into years. Co-founder Sachin Santhosh says that in some cases a job that would normally take 12 to 18 months has gone from a customer's requirement to delivered material in two. About $7.5M of the round goes to expanding in the US, including a manufacturing network there, and $2M to Scinode. Two of the new investors buy from the company, which is one way to read what they are paying for.
Evidence and limits: The $250M post-money valuation comes from a person familiar with the deal, not the company, which declined to give one; TechCrunch alone reported the round, with Santhosh on the record, and no release was found. The revenue run rates, the two-month timeline, the customer names and the $80M total raised are all the company's own figures. This is manufacturing coordination with agents, not a claim that AI found any of the materials.
Google Research folds only the patch a designed protein binds, cutting 88% of the GPU time for a 719-residue enzyme and getting single-digit nanomolar binders
What happened: Across 15 targets and 3,666 designs, folding only the contact shell plus a margin reproduced the full target's interface to a median 0.37 angstrom, and cut GPU hours by 88% on the 719-residue outer part of fibroblast activation protein alpha. With a researcher setting the strategy and a Gemini 3.1 Pro agent choosing the crops and parameters, the in-software hit rate rose from 5.06% to 9.33% while saving a further 42%. Yeast display then gave four clones at apparent 3.0 to 9.1 nM.
Why it matters: AlphaFold3 charges by token bucket, so most of the compute in screening designs goes on target residues far from the binding site, and a big outer domain is what makes a target impractical to design against. The paper treats cropping as something to test rather than a free saving: whether the interface geometry survives, whether the ranking that separates binders from non-binders survives, and whether the saving crosses a bucket boundary. Left to pick its own strategy from a map of the whole target, a fully autonomous arm converged on the same catalytic-rim approach the researchers had specified.
Evidence and limits: Cropping cost ranking: mean interface ROC-AUC fell from 0.777 to 0.729, and on one target the ranking collapsed, recovered only by dropping the target's alignment, which helped deep alignments and hurt a shallow one. Seven of the 15 targets were small enough to gain nothing. Affinities are apparent values on the yeast surface, not purified protein; one target carried both the wet lab and the agent arms, nobody outside repeated it, and this is a technical report.
Zealand Pharma pays Generate Biomedicines $30M upfront to design multispecific proteins for obesity
What happened: Generate takes $30M upfront, up to $35M in early preclinical milestones and milestones of nearly $1.2B in all, plus tiered royalties, for two initial programmes with an option on a third. Generate runs research through lead-candidate selection; Zealand takes clinical development, regulatory filings and worldwide commercialisation. Utpal Singh, Zealand's chief scientific officer, says proteins complement its small molecules and peptides; Mike Nally, Generate's chief executive, says the teams will define the target profile first.
Why it matters: Obesity is where the metabolic money is, and Zealand is buying protein design rather than building it: its own 25 years are in peptides. For Generate it is the third pharma collaboration after Amgen in 2022 and Novartis in 2024, and the first since it listed on Nasdaq in February, which is the kind of revenue a newly public platform company needs.
Evidence and limits: The $1.2B is what could be paid if programmes reach development, regulatory and commercial milestones; $30M is what is paid now. No target, candidate or data is named, and the work starts at design.
Albert raises a $55M Series B led by TCV for AI that searches and reasons over chemical companies' own R&D data
What happened: New investors Delta-v Capital, Socium Ventures, HighSage Ventures, Coral Capital and Legion VC joined, and Coatue, J.P. Morgan Growth Equity Partners, Index Ventures and Applied Ventures followed on. CEO and co-founder Nick Talken says the money expands Albert's chemistry AI; Morgan Gerlak of TCV spoke for the lead. The partners the release names include AkzoNobel, BASF, Henkel, Kenvue and Clorox.
Why it matters: Albert sells to the makers of coatings, adhesives and consumer goods: AI that searches their own formulation records, suggests candidate formulations and designs experiments, rather than discovering new molecules. The evidence that its model helps is Albert's own: a white paper says that on the hardest set of a benchmark Albert built, its retrieval model Valence put the right passage from about a million drawn from US patents in its top 100 results 81.8% of the time, against 2-4.5% for OpenAI and Voyage models. No valuation was given.
Gamow Labs' AI pipeline re-reads 26 hard newborn lung-disease genomes blind, matches all 19 solved cases and finds a rearrangement in an unsolved one that Baylor confirms
What happened: Pawel Stankiewicz's lab at Baylor College of Medicine gave Gamow Labs de-identified whole genomes from 26 infants whose lung disease first-line labs had not diagnosed; the lab had solved 19. Gamow's George-0.1 pipeline, built on frontier models including Gemini 4 Argon, matched all 19 and flagged candidates in 2 of the other 7. In one, read evidence pointed to a balanced translocation separating FOXF1 from its enhancer, consistent with alveolar capillary dysplasia; Sanger sequencing in Stankiewicz's lab confirmed it.
Why it matters: The step forward is that a candidate the AI proposed went on to be confirmed in a lab: Baylor's own sequencing verified a chromosomal rearrangement that had gone undiagnosed for three years. That is the step from flagging a variant to a diagnosis a family can act on.
Evidence and limits: It is a small study, 26 cases, disclosed by the company in a post rather than a paper, and the outcome of the second candidate is not given. Argon is one of several models in the pipeline, so the result belongs to the pipeline, not to Argon alone.
Google opens Gemini 4 Argon to all, publishing science case studies and a free-access plan for scientists
What happened: On 13 October Google opened Gemini 4 Argon to developers, enterprises and eligible subscribers, and invited 10,000 scientists to apply for six months of free access to Antigravity's top usage tier. Its science case studies cover enzyme design, quantum computing, fusion research and rare-disease genome analysis.
Why it matters: What stands out is less the model's scores than frontier AI beginning to leave traceable results in different research workflows. The fusion work has a public preprint, and the speed-up of the quantum chemistry package xTB has code submitted to its repository; Gamow Labs reports that its AI pipeline found a previously undiagnosed chromosomal rearrangement, which a partner lab confirmed by sequencing.
Evidence and limits: The cases rest on different kinds of evidence. The fusion paper is still a preprint; the three xTB pull requests are open, not merged, and a submission is not an independent check of the speed-up; Gamow's study is the company's own account, and Argon is one model in a multi-model pipeline. The enzyme activity and some quantum results lack public material an outside group could check.
Undecimal Bio, spun out of Eleven Therapeutics, launches with an $18M seed to find molecules that carry siRNA drugs to tissues beyond the liver
What happened: 4BIO Capital and IQ Capital co-led the oversubscribed round, with Verve Ventures, Harel Tech, Entree Capital, Kerna Ventures and Saras Capital. Its DELiveri platform screens large combinatorial libraries of candidate ligands in cells that report when siRNA actually gets inside, and AI agents read the resulting structure-activity data to pick conjugates. The money is for the platform, its first kidney and blood-vessel programmes, and hiring at its labs in the Cancer Research UK Cambridge Institute. Yaniv Erlich, chief executive, and Shaul Ilan co-founded it.
Why it matters: siRNA drugs work mainly in the liver, because a ligand that liver cells take up is known; finding one for other tissues has been slow, the company says. Undecimal treats that as a data problem: measure delivery in cells at scale, then train models on the screens. The release calls the platform validated in animals and mentions undisclosed collaborations with large drug companies, but shows no data, names no partner and gives no valuation.
Same force error, different rates: a preprint bounds how far a machine-learned potential can get a reaction rate wrong, using single-point energies on reaction paths
What happened: Xiyao Yu of Nanyang Technological University shows that the usual test, force error on equilibrium snapshots, barely sees the barrier region that sets a rate, so worst-case rate error grows exponentially with barrier height. Weighting energy errors by a 'leverage density' concentrated on reaction paths instead gives an estimated interval for the rate. In alanine dipeptide, five designed perturbations of matched force error changed the rate 0.49 to 3.5 times; in Li3OCl the intervals ordered six models' ion-hop rates, but five measured rates sat about 7% above the upper end.
Why it matters: Machine-learned potentials are judged by force error but used to compute diffusion and reaction rates. This offers a check costing a few hundred reference energy calculations, reported as two numbers beside force error, and picking training data by it cut rate errors even as force errors rose. The Li3OCl agreement holds only once hops are sorted by mechanism; on the pooled set the intervals do not track. Every rate tested against is the author's own simulation, not experiment, in an unreviewed single-author preprint.
Tianxu Intelligent Technology raises tens of millions of yuan for a wind-and-solar weather model and satellites of its own, with Houxue Capital the only investor
What happened: Houxue Capital invested alone in the Pre-A+ round; Yirong Capital is the long-term exclusive financial adviser. The money goes to the Ningkong satellite constellation, to the CMA-WSP-AI energy weather model and a sector world model, to hiring, to key customers and to groundwork abroad. Founder Li Liu has a meteorology doctorate from the University of Chinese Academy of Sciences and started the company in Jiaxing in October 2024. It sells weather data, a power-trading platform and deployments inside the customer's own network to generators and to both state grids.
Why it matters: The weather models on this record come from large labs. This is a small company selling one slice of one: wind speed and irradiance, the two variables that set how much a wind or solar plant generates, and so what it earns in China's power market. Its answer to the data bottleneck is to make the data itself, by putting the model on satellites: one has been in orbit since July, with 18 planned.
Evidence and limits: The performance figures are the company's own and unpublished: it says its 1km wind forecasts beat EC-HRES and AIFS on error and correlation, and that it came first in a four-event extreme-weather AI contest, with scores 39.6% to 351% above the baseline method. The amount is given as a range rather than a figure, so none enters the totals, and no valuation was given.
Drylab AI raises a $3M seed led by Vertex Ventures Southeast Asia & India to run computational biology analyses as agents inside a customer's own machines
What happened: The Singapore company, founded in 2025 by John Phan, from AI research at ByteDance's Seed team, Huy Phan, formerly of Landing AI, and Nhi Nguyen, a biomedical PhD candidate at Nanyang Technological University, raised $3M led by Vertex Ventures Southeast Asia & India with PatSnap. Its agents plan an analysis, pick from over 800 tools and databases, get the compute and work through failures, while a model of its own checks each analytical decision and flags method errors. It runs on the customer's own machines.
Why it matters: The bet is on where agent work in biology actually gets blocked in a drug company: not the model, but whether the analysis can run on data that cannot leave the building, and whether a method error gets caught before someone acts on it. The company has been selling since February 2026, so the claim is about deployment rather than a demo. Genping Liu, the Vertex general partner who spoke for the firm, put the deal as Vietnamese technical talent going global.
Evidence and limits: No customer is named, no revenue given, and nothing has been published on how well the checking model catches method errors or how often the agents' analyses are right.
A Stanford verifier that hunts for errors in AI maths proofs roughly halves the invalid solutions a human must read, and reports real mistakes in 11 OpenAI preprints
What happened: Fan Nie, Haotian Ye and colleagues split verification into proposing candidate errors and then judging whether each is real, with the threshold set so that almost no valid proof is wrongly rejected. Across three maths benchmarks that cuts the invalid solutions a reviewer must read by 46% on average against a voting baseline at the same false-rejection rate, and by 57% once the validator is trained. Pointed at Xinwen Zhu's 2017 Annals of Mathematics paper, it reported 45 issues where the same model reviewing directly found 11.
Why it matters: Expert review becomes the scarce resource once a solver produces more candidate proofs than anyone can read, and this treats that capacity as a budget: reject as much as possible while letting through almost every valid proof. The authors of the papers it checked confirmed the errors and revised. Paired with a solving agent it also helped settle three open conjectures in relative Langlands duality, which experts then confirmed.
Evidence and limits: The 11 preprints' main conclusions hold after the corrections. The confirmations come from the papers' own authors, and for the preprint credited with 88 issues one of them is Shurui Liu, an author of this paper; he also wrote up the three Langlands proofs, co-wrote the Manin constant paper and co-wrote the solver OSVerify is paired with. No formal proof checker is used anywhere.
Nine agents write 135 follow-up papers to order for national-lab scientists; a judge finds 114 of them fail an integrity check
What happened: Carnegie Mellon, Minnesota and Stanford researchers asked ten scientists now or formerly at five US Department of Energy labs for an unpublished follow-up direction from their own work, plus the conditions a paper on it would have to meet. Nine agents each wrote 15 follow-ups across quantum computing, plasma physics, computational chemistry, materials and scientific machine learning. Graded against those conditions by an automated judge, none of the 135 met all of them, and 114 failed the integrity rubric. Six of the nine agents had no submission pass it.
Why it matters: The failures are not mostly about capability. The commonest flags are a paper that does not match its own code (90 of the 114), numbers contradicting the recorded evidence (85), and work claimed but never done (68). So an evaluation that reads only the write-up cannot tell a finished result from a described one, and an internal benchmark of research throughput needs the logs checked against the conclusions. The scientists' conditions are written before the work, which is what makes grading possible without a reference answer.
Evidence and limits: The grading is itself an agent's, GPT-6-Astra run four times per submission, and GPT-6-Astra is one of the nine agents graded, and the best of them on every column. The authors check it against experts: the rubrics cover 96.97% of the weaknesses reviewers raise and agree with them on 80.34%, but on scientific quality and impact agreement falls to 68.68%, and the authors say that part still needs expert verification. Fifteen research problems, all with computational experiments. A preprint.
DeepSeek is in talks to lead a seed round for Winverse AI, a robot-AI startup founded by former Tencent scientist Han Hu, Bloomberg reports
What happened: Bloomberg, citing people familiar with the matter, says the round is worth tens of millions of dollars and would value Winverse at about $400M after the money, with Meituan's Long-Z Investments also committed. The deal is not final and its terms could change. Han Hu founded Winverse this year after leading Tencent's vision-language and embodied foundation model teams; he was earlier a principal researcher at Microsoft Research Asia. The company is still in stealth.
Why it matters: It shows a model lab putting money into the AI that lets robots act, not only into robot hardware: DeepSeek already took a stake in Unitree when the humanoid maker listed in Shanghai. Bloomberg's point is that robot bodies are already built well, while the models that turn instructions into reliable actions lag. Nothing has closed, none of the parties commented, and Winverse has shown no product or result.
Matter Machine, a Liverpool spin-out searching for inorganic crystals, raises a GBP 4M pre-Series A led by IP Group, and Oxford Nanopore's Gordon Sanghera joins to lead its commercialisation
What happened: Matt Rosseinsky of the University of Liverpool's chemistry department founded the company. It sits in the university's Materials Innovation Factory and pairs AI prediction with experimental and characterisation expertise, aimed at inorganic crystals for batteries and catalysts. IP Group led the oversubscribed round, with Northern Gritstone, the University of Liverpool Enterprise Fund, Norrsken Evolve, LYVA Labs and the Parkwalk-managed Northern Universities Venture Fund alongside. The money is for the team, the platform and commercial partners.
Why it matters: Matter Machine represents a different commercial route for materials AI: pairing model predictions with a university's existing synthesis and characterisation capability. Its founder's Liverpool lab has repeatedly made new materials that computation pointed to, and the round adds an executive the release says helped build Oxford Nanopore.
Evidence and limits: The round funds the team and the platform. The company has not yet published material data of its own from which its technology could be judged.
Britain extends its use of frontier AI to find holes in government systems, and says it will explore reporting duties for serious AI incidents while work begins on responding to ones involving AI agents
What happened: Four departments and agencies say pilots have helped identify and fix over 1,000 potential vulnerabilities in government systems, and the programme will grow. The government will explore using the Cyber Security and Resilience Bill to have serious AI incidents and near misses identified and reported, and to cover loss-of-control risk, where autonomous systems evade human oversight. The AI Security Institute and the National Cyber Security Centre are already looking at how to respond to incidents involving AI agents. AI Minister Kanishka Narayan announced it.
Why it matters: AI safety governance is extending from testing models before release to reporting incidents and responding after deployment. That matters for the science agents the record follows, which are starting to run code and lab equipment on their own.
Evidence and limits: The vulnerability pilots are already running, and the 1,000 figure is the government's own, with no method or model named. Mandatory reporting of serious incidents is still being explored, with measures to be worked through with parliamentarians, technical experts and operators.
AriadneX splits a powder diffraction pattern into the phases already known and what is left over, to point chemists at the samples holding something new
What happened: Zhuoran Yang, Vladimir V. Gusev, Matthew J. Rosseinsky and colleagues at Liverpool write it as a mixed-integer optimisation under crystallographic constraints: fit each pattern from a library of known phases, then rank samples by the signal left unexplained. Hide one of four known phases and the samples holding it rank near the top: all ten for SnO2, 11 of 14, 12 of 16, 16 of 18 elsewhere. On 31 patterns from a past Ba-Y-Si-Ru-O campaign, the leftover signal reproduced the main reflections of the two phases that campaign found.
Why it matters: High-throughput and autonomous synthesis produce more diffraction patterns than experts can read. AriadneX does not identify new crystals; it picks out, from many patterns, the samples most worth an expert's look, and shows what their unexplained peaks look like.
Evidence and limits: The results come from retrospective tests on data whose answers were known; the open-world test is one campaign of 31 patterns, assigned afterwards by the same group's experts. The next step is a new synthesis campaign showing it actually cuts expert analysis time. The code is MIT but needs Gurobi and licensed structure databases.
Where in a transcript an RNA sensor should bind: a model trained on 7,383 measured sensors picks working target sites in a virus gene it had never seen
What happened: Jihoon Han and Seth L. Shipman at the Gladstone Institutes and UCSF tiled six RNA targets at single-nucleotide steps, measured all 7,383 sensors, and trained a model that compares two candidate sites rather than scoring one. Held-out transcripts: its top sites were enriched for high-performing sensors. Frozen and run on the SARS-CoV-2 N gene in E. coli, with that transcript supplied alongside, 4 of its ten top sites gave working sensors against 0 of 20 random ones by qPCR, 5 against 2 of 20 pooled.
Why it matters: Toehold switches can be programmed for any RNA, and most work has gone into the switch itself; where in the transcript to aim has had far less, though the authors find neighbouring sites differ nearly as much as random ones. The gain comes from a base-pairing rule in the sensor stem that transfers between transcripts. Numbers this small (4 of 10, one transcript, one lab) measure enrichment, not a design rule. Toehold-VISTA, which works in its own setting, did not rank these sensors; the authors leave open whether the architecture, the readout or both explain that.
Friedrich Schiller University Jena finds language models read which atoms a material contains but not how they are arranged
What happened: Nawaf Alampara, Santiago Miret and Kevin Maik Jablonka built synthetic datasets in which a target property depends by a tunable amount on positions in space or on category labels, the stand-ins for atomic coordinates and element types. Language models fine-tuned on text versions learned the category part and largely ignored the coordinates, behaving much like simple n-gram models. The gap held for models up to 70 billion parameters, datasets up to 2 million structures and nine text representations of crystals in their open MatText framework.
Why it matters: Many groups now feed crystal structures to language models as text to predict properties. The paper says that for properties set mainly by structure, graph neural networks built for geometry consistently do better, a gap it calls the GNN-LM wall. So a team choosing a model for materials screening should first ask whether composition or structure drives the property.
Evidence and limits: The synthetic targets come from a made-up potential function, so the clearest result is on constructed data, and the materials benchmarks run in the authors' own framework. Peer reviewed and open access. Miret's affiliation on the paper is Intel Labs, with Lila Sciences given as his present address.
AuthorsNawaf Alampara, Santiago Miret, Kevin Maik Jablonka
A Brucella vaccine candidate designed from predicted epitopes raised cytokine release and shifted T-cell subset proportions in donor blood cells, with no animal protection test yet
What happened: Xuanji Zhang and colleagues at Xinjiang Medical University's First Affiliated Hospital picked 16 cytotoxic T-cell, 10 helper T-cell and 3 B-cell epitopes from the Brucella proteins Omp28 and GroEL, using Xinjiang HLA frequencies, and linked them to a dendritic-cell-targeting peptide. AlphaFold3 predicted its structure, docked to MHC molecules and TLR4. Expressed in E. coli, it bound antibodies in patients' serum. Donor blood cells given it averaged 145.5 pg/mL of interferon-gamma and 87.53 of interleukin-4, against 10.91 and 12.03 untreated and 66.84 and 54.44 without the peptide.
Why it matters: Epitope-prediction papers usually stop at the computer, and this one goes a step further: the designed protein was made and put in front of human cells. What the cells show is a reaction, not protection. No animal was infected, so efficacy is untested, and the authors say so. Five donors a group for the cytokines and three for the T-cell subsets, one region's HLA profile, and surface staining alone, which the paper notes cannot show the T cells responded to this antigen rather than to anything else.
The ART preprint lands, and reports that ten reruns of the same agent campaign all reached the locus and none saw the repeat array
What happened: Peter H. Yoon, Nicholas T. Perry, Matthew G. Durrant and colleagues at Anthropic put the September announcement on bioRxiv. Claude Code workers and supervisors surveyed reverse transcriptases across 1.9 billion protein clusters in 949 sessions and 119 tasks, 98 of them opened by the agents, over 21.5 hours with nobody stepping in. In phage SA1 infection, RNA from the array reaches 8% of phage transcripts 15 minutes in. The authors then reran the campaign ten times: nearly all reached ART loci, none read the DNA upstream, none found the array.
Why it matters: It shows a weakness of science agents that is easy to miss: being able to analyse evidence is not the same as going and finding it. With the DNA in the prompt, the four strongest Claude models described the array in at least 90% of attempts; when they had to open files themselves, as few as 32% did. How tools and data are put in front of a model decides whether a discovery happens.
Evidence and limits: In ten reruns of the full campaign the agents reached the locus but none read the DNA upstream or found the array. Whether the original find was Claude's alone is disputed, and the preprint does not address that.
Twenty-three ways of deciding when to trust a chemistry model are scored against held-out error: the model's own uncertainty cuts error by 65%, the best similarity measure by 32%
What happened: Ivan Zlobin and colleagues at the Russian Academy of Sciences' Kurnakov Institute predicted which of two metals a ligand binds more strongly, from 88,525 selectivity values the authors derived from the Critical Materials Institute's compilation, holding out whole ligands rather than rows. Keeping the safest half of predictions, the model's own predictive uncertainty cut mean-of-pairs error by 65%, the best descriptor-space distance by 32% and the best ligand-only distance by 18%, while fingerprint nearest-neighbour and learned-embedding distances ordered error no better than chance.
Why it matters: An applicability domain is usually asserted: a prediction is called reliable because the molecule resembles the training set. Here that assumption is tested against held-out error, and similarity mostly fails it while the model's own uncertainty holds up. The same split reframes the published model for this compilation: its reported 0.599 mean absolute error on binding strength reproduces, but that test set shares 94% of its ligands with training, and a ligand-disjoint retrain gives 1.484. One task, one compilation, and the selection exercise replays fixed data rather than running an experiment.
Scinetics releases Chain of Structures, where a language model plans an enzyme in four steps and a protein generator builds each step before the next is written
What happened: Qingran Cai and five colleagues at the company, with founder Zaixi Zhang at HKUST corresponding, pair a reasoning language model with an all-atom protein generator. From the reaction the enzyme should run, the reasoner writes four rounds (the amino acids that do the chemistry, the pocket around them, the chain's shape, then sequence and structure), and each round steers the generator, whose output becomes the next round's context. The preprint reports it ahead of the baselines compared on five of six measures.
Why it matters: This is the company's first public method since its first round in September, and it ships with the weights and the scripts that produced them, so the procedure can be inspected rather than taken on trust. Its design choice is to make the model think in structures: each step has to be built as a structure before the next one is written, instead of a chain of reasoning that stays in text.
Evidence and limits: Everything here is computed, and the headline numbers are narrower than they read. The 57.0% is a predictor's enzyme-function code for the design overlapping the target's at three levels, on the 2,435 test proteins shared with the baselines; the authors' own script says all 2,801 score about 8 points lower. The 0.729 TM-score refolds the design's sequence against the model's own structure, not a real enzyme. Nothing was made or assayed. The weights are Apache-2.0, the code carries no licence.
Litai Zhiyao raises a seed round of nearly 50M yuan to design peptide drugs with generative models
What happened: iFlytek Ventures and the L2F fund co-invested with an industry investor the announcement does not name; no lead is named and no valuation given. The money goes to renting compute, core pipeline R&D, patents, hiring and operations. The Beijing company runs a platform-plus-pipeline model: partner work now, its own oncology, autoimmune and metabolic programmes later. L2F's Xuequn Huang is the one who describes the design space reaching mirror-image D-peptides. Founder Yinjun Jia, Jingming Xu of iFlytek Ventures and Xuanle Zheng of L2F also spoke for it.
Why it matters: Peptides sit between small molecules and antibodies, which is why a generative model has room there, and D-peptides are a space no natural library covers. What the announcement shows is money and intent: it says some partner projects have reached preclinical work, names none of them, and shows no designed molecule or assay. Compute rental heading the use of proceeds is the honest shape of an AI-native drug startup's first round: the model is the asset being bought.
Random forests trained on 105 lab specimens predict how swapping crushed stone for natural gravel weakens asphalt, and place the steepest loss between 12% and 16% gravel
What happened: Mohamed Ragab, Mostafa M. Arnaout and Mokhtar F. Ibrahim of Suez University replaced 0% to 24% of the crushed dolomite in an asphalt mix with rounded natural gravel and measured Marshall stability, flow, indirect tensile strength and wheel-tracking rut depth. From 0% to 24%, stability fell from 2,890 to 1,940 lb and rut depth after an hour rose from 4.2 to 9.7 mm. Random forests fitted to the 105 measurements scored test R² above 0.95 with errors under 2.5%, and SHAP ranked gravel share far ahead of every other variable.
Why it matters: It shows what explainable machine learning can add to a small materials experiment: the model's predicted step-by-step changes put the sharpest loss in flow, rut depth and tensile strength between 12% and 16% gravel, a threshold the authors suggest checking in the lab, since no new specimens tested it. The accuracy says little beyond this setup: seven gravel levels from one source, one binder, one gradation, and gravel share alone correlates between 0.95 and 0.97 with each of the four results.
AuthorsMohamed Ragab, Mostafa M. Arnaout, Mokhtar F. Ibrahim
TypeSafe AI raises an $870M Series A at a $7.5B valuation for Jev, a model that hands software typed decisions instead of text
What happened: Andreessen Horowitz led; Sequoia Capital and existing investor DCVC joined, and Martin Casado of a16z joins the board. CEO Diogo Almeida worked on RLHF and InstructGPT at OpenAI and was at Google Brain; his co-founders are Sasha Sheng, a former research engineer at Meta FAIR, and Erik Gafni, who founded Ravel, an AI company for DNA sequencing. TypeSafe came out of stealth on 15 September with a $40M seed led by DCVC.
Why it matters: Jev writes no text: it answers yes or no, picks an option or gives a score, each with a probability, and by TypeSafe's and a16z's own figures is one to two orders of magnitude faster and cheaper than frontier chat models at this. a16z calls it the first of a new class of models built to be called by code, not talked to, a bet that decisions in software reach further than chat. Neither names a science use; ranking candidates or checking an agent's work is where it could reach a lab, and nobody has shown it.
GSK takes Chai Discovery's design models across its pipeline after its own lab found binders to every target it tested
What happened: Chai Discovery announced a collaboration giving GSK access to its protein folding and design models for discovery across GSK's pipeline. It follows an evaluation in which Chai's models designed molecules for GSK targets without target-specific training, and GSK's own wet lab found binders to all the targets it tested. Christopher Austin, GSK's head of R&D technologies, said the models complement GSK's internal ones, which it retrains on its own lab data. No financial terms were given.
Why it matters: It is Chai's sixth drugmaker deal on this record this year, after Lilly, Pfizer, Novartis, argenx and BMS, and the release says GSK ran the wet-lab test itself before widening the deal. That makes the check external, but it gives neither the number of targets nor any binding figure: the designs bound, and GSK calls the affinities strong without a number. In January GSK also licensed Noetik's tumour models, so it is buying outside models while training its own.
A general chatbot, given formation energies as feedback, found 96% of a materials benchmark's low-energy compositions in 5,000 tries, more than twice what purpose-built generators found with the same budget
What happened: Hedda Oschinski, Christian Carbogno, Karsten Reuter and colleagues at the Fritz Haber Institute had GPT-5.4, with no fine-tuning, propose elpasolite compositions one at a time and fed back each one's formation energy from a precomputed set of machine-learned predictions. Targeting below -2.26 eV per atom, a class holding under 0.2% of the space, it found 3,577 of 3,740 targets in 5,000 proposals across three runs. With the same 5,000 attempts, the GAN, VAE and reinforcement-learning models it is compared against found 1,726, 1,511 and 1,569.
Why it matters: The gain comes from the loop, not the chemistry knowledge: asked for all 5,000 at once with no feedback, the same model did far worse. It reads its own growing history of hits and misses, which the one-shot generators it beats cannot do, and the authors say that asymmetry should be weighed when comparing. Everything happened in software against a lookup table. The paper is explicit that the figure is recovery of a region an oracle defined, not a count of stable or makeable materials, and that 262 of the target compositions contain noble gases.
AuthorsHedda Oschinski, Maximilian L. Ach, Konstantin S. Jakob, Christian Carbogno
A neural-network potential proposed where the hydrogen atoms X-rays cannot see sit in a copper cluster, and the cluster itself drove a range of photocatalytic reactions
What happened: Xiaona Yang, Hui Shen and colleagues at Inner Mongolia University, the Qingdao Institute of Bioenergy and Bioprocess Technology and the Nagoya Institute of Technology made the copper-hydride cluster Cu13H6 in one pot at 21.37% yield on copper, and resolved the hydride positions with stochastic surface walking global optimisation driven by a neural-network potential, in the LASP code. The cluster oxidised alkyl aromatics and alcohols to ketones under mild light, and also drove defluorination, deuteriation, cyclopropane ring-opening and dehydrogenative coupling.
Why it matters: Hydrogen is nearly invisible to the X-ray diffraction these clusters are normally solved by, and pinning it down usually needs neutron diffraction and a large enough crystal. The paper's abstract twice says the computation only tentatively identifies the positions, so the structure stands as a joint computational and experimental inference, not a direct determination; the count of six hydrides, though, is settled by mass spectrometry. The chemistry either side of it is physical: the cluster was synthesised and the reactions run and measured.
Ten GPT-5 setups were put on 161 clinical cases: thinking harder helped, running an agent made it worse, and search sank the weakest
What happened: Shaohui Zhang, Jing Huang and colleagues at Pennsylvania, Lehigh, Stanford and Children's Hospital of Philadelphia scored ten configurations on 161 cases drawn from the MedXpertQA benchmark, each with a narrative, findings, at least one image and a reference diagnosis. Accuracy rose with reasoning strength, 32.9% to 46.0%. Web search cut the routing mode from 38.5% to 26.7% while the reasoning modes each gained a little. Agent and deep-research modes scored 36.6% and 29.8%, below the plain model, and deep research cited nothing in 29 cases.
Why it matters: It is the rare benchmark where added machinery is measured and found to cost accuracy, which is the opposite of how agent systems are usually reported. The failure labels come from GPT-5 reading each trace, with physicians confirming 827 of 928 labels on 51 cases, a check the paper says measured precision, not recall: images misread, evidence not integrated, no reasoning to audit. The cases come from an existing benchmark rather than a clinic, nobody was treated, and the top score is under half.
AuthorsShaohui Zhang, Jiarong Qian, Kai Zhang, Jing Huang
Oganov's group runs evolutionary crystal-structure search on a landscape that includes quantum nuclear motion, using machine-learned potentials to make it affordable, and recovers LaH10's observed cubic phase
What happened: Daniil Poletaev and Artem Oganov at Skoltech searched for LaH10's structure at 150 GPa and 300 K with USPEX, ranking candidates by a free energy that includes quantum anharmonic effects (SSCHA), which DFT alone makes far too slow. Potentials trained on the fly, the universal potential MatterSim-5m and DPA3 effective potentials trained on SSCHA data each found the cubic phase seen in experiments; MTP-based effective potentials did not. Of seven universal potentials, only MatterSim-5m relaxed a test structure correctly, and it needed fine-tuning to approach DFT accuracy.
Why it matters: Standard structure search ranks candidates at 0 K, and for LaH10 that picks the wrong answer: the same search with 0 K DFT chose lower-symmetry structures over the cubic phase at 11, 44 and 88 atoms per cell. The effective potentials made the 88-atom search affordable. The test is one compound whose answer experiment already gave, and direct SSCHA search still costs about a hundred times a 0 K DFT search, even with a universal potential.
The National Hurricane Center leaned on Google DeepMind's AI model for Isaias from the first advisory, which called a Category 2; the storm made landfall near Destin, Florida, as a Category 2
What happened: From its first advisory on 6 October, when Isaias was still an unnamed depression, NHC forecast a Category 2 hurricane, shaping its forecast to match DeepMind's model, whose newest version is WeatherNext 3, CNN reported; many physics-based models then showed a weaker storm on a different track. That first discussion put the NHC track in best agreement with the DeepMind and ECMWF AI aids. Isaias peaked near 100 knots offshore, then landed near Destin, Florida, late on 9 October at 90 knots, east of where that first track pointed.
Why it matters: This is the model used in practice, not scored afterwards: NHC director Michael Brennan says 2026 is the first season WeatherNext is fully part of the centre's forecasting aids, after it guided last year's Hurricane Melissa call, and that NHC helped train it on its own best-track data. One storm is not a verification. The formal storm report comes months later, and the verdict that DeepMind's track was best so far comes from reporters and a TV hurricane specialist, not from NHC.
ARPA-H puts up to $28.6M into NATIVE-ID, a project to measure how 50,000 proteins clump in cells and train AI to predict it
What happened: NATIVE-ID, led by Brad Ringeisen of the Innovative Genomics Institute at UC Berkeley, is part of ARPA-H's BIOGAMI programme on harmful protein aggregation, run by programme manager Shannon Greene. Randal Halfmann's lab at the Stowers Institute gets about $4.1M over two years to take more than 10 billion single-cell aggregation measurements across 50,000 proteins in yeast. The team, with Brown, Emory, Johns Hopkins, Texas A&M and others, starts with frontotemporal lobar degeneration.
Why it matters: It funds the data before the model: roughly a third of proteins have no stable structure, so structure predictors say little about them, and the bet is that measured aggregation at this scale lets a model learn the sequence rules. The 10 billion measurements are still to come. This is Phase 1, datasets and models; testing therapies and earlier detection would come in a second 24-month phase.
An audit of Google's GNoME crystal data finds most predicted structures sit within 0.05 Å of a known crystal's geometry
What happened: Daniel Widdowson, Olga Anosova and Vitaliy Kurlin (University of Liverpool) and Tatiana Kurlina (UCL) compared every crystal in two public GNoME versions (384,938 entries in 2023, 554,054 in 2024) by continuous geometric invariants. Matching atom positions only, ignoring elements, 63-78% of entries lie within 0.05 Å of a crystal already in the ICSD or Materials Project, depending on database and metric. About a fifth contain radioactive elements. Of 1,133 groups of exact duplicates in 2023, 1,010 were deleted whole in 2024.
Why it matters: The audit reopens a basic question: in what sense is an AI-predicted material new? GNoME's 2023 Nature paper is the record's reference for how many new stable materials AI has found, and the audit argues a crystal within 0.05 Å of a known geometry should not count. Similar geometry does not mean the same composition or properties, so structural novelty, compositional novelty and whether a material can be made need to be judged separately.
Evidence and limits: The analysis is computational and not yet peer reviewed. Most near-matches carry different elements, so whether a new composition on a known framework counts as a new material is the point in dispute, not a settled finding.
AuthorsDaniel Widdowson, Tatiana Kurlina, Olga Anosova, Vitaliy Kurlin
Prompteina designs all-atom protein structures from written prompts, and nine of its GM2A designs bind in the lab, the strongest at 107 nM
What happened: A University of Washington, Ai2 and NUS team, with David Baker among the authors, text-conditioned the Complexa structure generator on 1.8 million structures paired with 52.4 million captions, written by a language model from database annotations. It leads the baselines on 15 of 18 structural tasks and on at least 13 of 20 binder targets, raising a computed interface-confidence score 14% over the strongest. Against GM2A, which has no reported de novo binders, Adaptyv Bio assayed 89 designs picked from a thousand: nine bound, six below a micromolar.
Why it matters: A designer can now ask for a fold, a disulfide count and an interface size in one written sentence, instead of a separate input format per property. That matters most for the agent loop the paper tests: a language model researches a target, writes the specification in words, reads the scores and rewrites it, with no translation step into a model's own labels. The GM2A binders are what separate this from a benchmark result.
Evidence and limits: Binding is not function: Adaptyv Bio, commissioned by the authors, measured apparent dissociation constants by surface plasmon resonance, fitting eight of the nine, and nothing shows they do anything to GM2A. The paper's 6.7% hit rate counts the six sub-micromolar binders against 89 designs the authors picked from a thousand. The 14% and the task counts are the authors' own runs of the baselines, and the preprint is not peer reviewed.
ACED-Bench scores agents on choosing the right experiment and knowing when to stop, with a hidden causal model supplying every answer
What happened: Yiwei Chen and colleagues at UC San Diego and Amazon put each task in a natural-language scientific scenario backed by a hidden probabilistic graphical model: a language model writes the readable surface, while code builds the causal graph, fixes the intervention semantics and computes the oracle answer, so runs are graded without human labelling. Agents get four turns to request observational or interventional samples. The main board is 180 tasks, with a separate 150-question yes/no check; seven current models, two older ones and three active agent designs were run.
Why it matters: Most science benchmarks hand a model the evidence and ask what it concludes. This one makes it decide what to collect. The paper's evidence ledger then says where a run broke: it tested the wrong intervention, lost evidence it had already won, or kept querying after the answer was supported. Those are the parts of an investigation that a final-accuracy score hides.
Evidence and limits: Every world is synthetic, with fully observed categorical variables and known intervention semantics, so no latent confounding, continuous variables, measurement error or open-ended experiment design. The authors say success here should not be read as readiness for real scientific or policy settings. Claude Opus 4.6 is both one of the evaluated models and the fixed parser between every agent and the simulator.
Roche and ETH Zurich let one model choose which borylation reactions to run, and a second, trained on the result, called where the reaction lands on ten new ring compounds
What happened: Mason Minot, Gisbert Schneider and colleagues at ETH Zurich and Roche had a tree-based model pick which of 22,253 Roche compounds to try in C-H borylation, a reaction chemists use to modify an already finished drug molecule. Three rounds tested 50 new substrates, taking the dataset to 6,865 reactions on 568 substrates, with site labels for 812. Graph neural networks trained on it then predicted the borylation site for ten unseen substrates, and the top-ranked site matched the experiment in all ten.
Why it matters: Reaction models for drug chemistry lack data, above all failed reactions and labels for where a reaction lands; this loop produced both, 106 negatives among them, and the data is public. The ten were run after the predictions, by the team itself, which is what makes it a test. The site model reads no reaction conditions, though ligand and solvent can move the site, and on unseen scaffolds the graph models scored only 0.43 to 0.50 MCC, where 1 is perfect, at telling whether a reaction works. The authors ask for more drug-like molecules: accuracy fell on the hardest.
AstraZeneca trained a toxicity model on liver-cell proteomes, traced why some prostate-cancer degraders kill liver cells in a dish, and picked a near twin that did not
What happened: Shaon Basu, Kevin Moreau and colleagues at AstraZeneca, Charité Berlin and the Francis Crick Institute measured the proteomes of liver cells exposed to a library of 204 androgen-receptor degraders; the cells lack the receptor, so every response was off-target. A gradient-boosted model trained on them predicted toxicity (holdout AUPRC 0.78) and pointed to respiratory complex I. It scored AstraZeneca's compound 1 at 91 of 100 and two analogs differing only in the linker region at 47 and 35; those barely touched complex I and spared primary human liver cells.
Why it matters: Degraders join two drug parts, and the joined molecule can be toxic where neither part is; this catches that in cells before animals. The call was then tested with new experiments: cell respiration, an enzyme assay on heart mitochondria, and primary human liver cells: all cells and isolated enzymes, the team's own work, from one cell line and one target class. A 28-day mouse study, seven treated animals, showed the safer analog still shrank tumours, but had no liver endpoint, so it tests the drug and not the toxicity call. The analogs were already in the library.
DOE picks four national-lab projects to build shared robotics testbeds for autonomous science: $30M in all, $2M of it in fiscal 2026
What happened: Argonne's MAESTRO is to build lab robots that improve with use; Brookhaven's DART tests when robot autonomy is reliable and whether its scientific results hold; Oak Ridge's TRACE links robots, instruments and AI agents through standard interfaces so a run can be repeated in another lab; SLAC's SPIRE, led by Angelo Dragone, brings autonomy to X-ray and electron facilities. DOE expects open software, reusable robot skills, benchmark tasks and datasets.
Why it matters: Most autonomous-lab money on this record builds one organisation's own lab. This is a federal bet on shared parts and on checking: one of the four exists to find where robot autonomy fails, another to make runs repeatable across labs. It is small beside the $159M of Genesis awards the same day, only $2M is set for fiscal 2026, and a selection is not yet an award.
Isomorphic Labs is in early talks to raise money at a valuation of at least $40B, five months after its $2.1B Series B, Bloomberg reports
What happened: People familiar with the effort told Bloomberg the round could value Alphabet's drug-design company as high as $50B. The amount, the lead and the terms are not reported, the financing has not closed, and Isomorphic did not comment. Demis Hassabis, who in August moved from running Google DeepMind to chairing it and became Alphabet's chief scientist, remains Isomorphic's CEO. The May Series B gave no valuation.
Why it matters: It shows what investors may pay for a drug-design engine before its own drugs have results: the record holds Isomorphic's pharma deals and models, and no clinical readout. It is talks, reported by unnamed people, so the figure can move or fall away; nothing here says the engine works.
Iambic sets its IPO at $15-17 a share, about $150M at the midpoint, and two investors signal interest in up to $60M of it
What happened: An amended S-1 offers 9,375,000 shares on Nasdaq as IAM, with J.P. Morgan, Jefferies, BofA and Citigroup running the books. At $16 Iambic expects $135M net, and about $75M of it goes to IAM1363, its HER2 drug: finishing Phase 1, then Phase 2 trials aimed at accelerated approval and a Phase 3. ARK, an existing investor, and Duquesne Family Office have indicated interest in buying up to $60M. Tom Miller, co-founder and CEO, and Fred Manby, co-founder and CTO, run the company.
Why it matters: The public market is asked to price an AI drug-discovery company on one drug's early data. By the filing's share count, the midpoint values it at roughly $760M, after $461.8M raised privately. Indications of interest are not orders, and the price is not set until the offering prices.
Machine-learned potentials were run over 317,457 experimental crystal structures to find bad entries, and caught two nobody had reported
What happened: Justinas Slepavicius and colleagues at FTMC and Vilnius University, who run the Crystallography Open Database, relaxed its entries with universal potentials to test physical plausibility, a check that had relied on expert judgment. Of 522,086 entries, 317,457 passed a quality filter and were relaxed; 99.0% of those converged without a large volume change. Hand-checking 327 structures showed no single threshold on energy, force or displacement separates valid from invalid. On 524 hard cases, M3GNet distorts cyclopentadienyl rings and PET models do not.
Why it matters: The models disagree in ways that depend on chemistry, not on quality: each has its own blind spots, so a large change after relaxation can be the potential's failure rather than the structure's. The paper says its criteria are not calibrated against ground truth, so how many bad entries they miss cannot be estimated. Their own conclusion is to combine several models rather than trust one. Two coordinate errors they found, in CIFs rather than the original papers, are now corrected, and the relaxed structures are deposited so a reader can see whether relaxation confirms a published one.
A world model learns to skip ahead in atomic simulations by predicting which atoms move and how much time passed, with the physics bolted on as hard constraints
What happened: Ziming Pan, Kun Li and colleagues at Tsinghua's Institute for AI Industry Research train on simulator trajectory segments: kinetic Monte Carlo for three reactor pressure vessel steels, molecular dynamics for a Cu-Zr glass and a Li3N electrolyte. Per macro step, the model predicts a sparse set of structural edits plus the physical time elapsed. Edits are restricted to sites the teacher's rules can reach and species counts projected to stay conserved. Macro-step inference runs 1,000 to 10,000 times faster than event-by-event simulation.
Why it matters: The constraints are the result: on held-out steel the model violates reachability and conservation in no case and scores 0.91 on predicting which sites changed, while the same architecture without them keeps only 45.2% of edits reachable and 59.2% of counts conserved, and its clock drifts to 0.35 of the true elapsed time. But the ablation drops all three constraints at once, so which one earns what is unknown, and the teacher is a simulator rather than a measurement. A separate billion-atom comparison runs the baseline on a Sunway supercomputer and this model on A100s.
Anthropic commits $150M over three years to the Genesis Mission, one of eleven company pledges the White House totals at $2.4B
What happened: Anthropic says the money will make Claude available to more than 15 agencies in the federal programme, naming NASA, the National Institutes of Health and the National Science Foundation. Over three years it will give Claude, Claude Code and API credits to several hundred Genesis Mission projects, partner with agencies and national labs on fusion energy and quantum computing, and train scientists on those projects. A White House fact sheet the same day lists it beside NVIDIA at $1B, AMD at $500M, OpenAI at $200M and Google at $150M.
Why it matters: The named deliverables are product access, credits and training, and Anthropic never says the $150M is cash; the White House totals the eleven pledges as science tools and compute credits, so the figure is a valuation of what is given, by whoever gives it. No agency has confirmed a share, Anthropic does not say which 15 it will serve, and the Department of Energy's own Genesis pages did not mention it the next day. Fusion and quantum are named as priorities to partner on, with no project, lab or date attached.
Nine general-purpose atomic potentials were scored on how carbides stiffen and soften between 300 and 1200K, and most came out too soft
What happened: Miroslav Lebeda and colleagues at the Czech Technical University in Prague ran five binary and two high-entropy refractory carbides through stress-strain molecular dynamics with nine pretrained potentials, none fine-tuned, and compared them against published ab initio molecular dynamics from another group. MACE-MH-1 had the lowest mean absolute percentage error, 5.7%. Most models underestimate stiffness, with equilibrium volumes larger than the reference's. Errors track the metal, not the chemical complexity: 6.6% for ZrC against 20.1% for TaC, and the high-entropy carbides match their constituent metals.
Why it matters: Absolute stiffness and the rate of softening with heat are separate skills, and the paper reports them separately on purpose: a model can get the softening trend right while sitting at the wrong stiffness, so one averaged score hides which. The reference is another group's ab initio calculation, not a high-temperature measurement, so this says how well the potentials reproduce that calculation. The data is available from the authors on request rather than deposited.
AuthorsMiroslav Lebeda, Jan Drahokoupil, Simon Svoboda, Petr Vlcak
Anthropic's new usage policy, effective 12 November, requires a person who can stop any machine its models drive, and exempts lab automation inside enclosed instruments
What happened: Where model output moves hardware without human approval, the policy now requires a qualified individual able to watch the equipment and stop it at any time, equipment that stops or holds a safe state if that person intervenes or the connection drops, and operating limits (speed, force, temperature, pressure, dose) enforced by the equipment or a controller independent of the model. It lists hazardous energy and materials and acting on the human body among the triggers, and excludes laboratory automation handling non-hazardous materials inside enclosed instruments.
Why it matters: The company's Model Hardware Standard, on this record since August, set out how Claude may drive lab equipment; this turns that into a rule with a date. The exclusion marks the boundary: an enclosed instrument with harmless reagents is out, a bench that heats, pressurises or dispenses hazardous chemicals at levels that could injure is in, and its limits must be enforced somewhere the model cannot reach. The policy sets no rule about how long an agent runs, despite how the update was framed.
DOE funds 12 more Genesis Mission Phase II projects with $159M, from a fusion digital twin to enzyme design and autonomous accelerators
What happened: The department announced 12 Phase II awards totalling $159 million, taking Phase II to 14 projects; Darío Gil, its Under Secretary for Science, is the official quoted. Commonwealth Fusion Systems leads a digital twin of its SPARC device; UC San Diego will expand an RNA structure database five-fold for model training; the University of Washington works on sequence-structure ensemble modelling for enzyme design; Berkeley Lab's MOAT-Core is to run and design particle accelerators autonomously. MIT, Fermilab, Oak Ridge, Harvard and Argonne lead five of the rest.
Why it matters: Phase I was selection; this is money, and the list says which bets the programme is paying to build. Three of the twelve are the site's own beats (enzyme design, autonomous accelerator operation, inverse design of magnets), and the fusion digital twin puts a company's own machine at the centre of a federal project. None of it is a result: the release states no target any of these projects has to hit, which the earlier PNNL awards did.
NVIDIA puts $1B over five years behind US science, the largest of the eleven Genesis Mission industry pledges
What happened: Jensen Huang announced it at the Science: A New Golden Age event in Washington on 8 October. NVIDIA describes the $1 billion as commitments over five years for super intelligence for science and for quantum computing, through higher-education research institutions, quantum investments and cloud providers serving government missions, and says it collaborates on several Phase II awards announced the same day. The release restates its existing DOE work: the department's largest scientific supercomputer at Argonne, and seven new systems across Argonne and Los Alamos.
Why it matters: The $2.4B of industry pledges to the Genesis Mission consortium is already on this record, and this is the biggest single piece of it. What the release does not say is what the $1 billion is: it is "commitments valued at" that sum, spread over five years and across three channels, with no figure for any of them and no cash changing hands on the day. Read it as a statement of where the compute will be pointed, not as a grant.
OpenProblemBench gives AI agents 82 open problems from mathematics and theoretical physics, and model judges rate GPT-6 Astra's answers solved 14% of the time
What happened: Kun Chen of the Institute of Theoretical Physics, Youjin Deng of USTC and colleagues collected 46 mathematics and 36 physics problems still open in the literature, with context and prior progress but no reference answers. Seven models worked offline, up to 12 hours each; four of them also judged every submission. GPT-6 Astra's mean judged solve rate was 14.0%, every other configuration under 8%. All four judges agreed on eight Astra solutions, seven of them physics problems; many physics items ask to extend a given construction.
Why it matters: A score on problems nobody has solved measures something benchmarks with known answers cannot, but every verdict here is a model's: the authors say expert review is still needed to show any 'solved' answer is correct or new, and that the judges agreeing does not estimate how often they are wrong. So 14% is a model-judged figure, not a count of new results, and the eight unanimous answers are where to check. HorizonMath and FrontierMath: Open Problems check answers with computational or problem-specific verifiers instead.
Academia Sinica researchers had Claude Code supervise model building and refinement for 24 cryo-EM and X-ray structures; re-refined cryo-EM models gained better geometry, and the researcher kept the final call
What happened: Kuen-Phon Wu and Chuan-Kai Chueh of Academia Sinica and National Taiwan University had Claude Code with Claude Opus 5.5 run Phenix, Coot, ModelAngelo and other established programs, read the validation and pick the next step, in about 36 hours on one workstation. On 15 deposited cryo-EM models the median MolProbity score went from 2.62 to 1.76 and clashscore from about 21 to 6. Three crystal structures built from diffraction data reached Rfree 0.182, 0.207 and 0.238, against 0.173, 0.197 and 0.229 published.
Why it matters: This tests an agent as a supervisor of existing programs, judged against the data each structure was measured with, not as a new refinement method. It did not win everywhere: CaBLAM outliers rose in seven of the 15 cases, one ligand fit its map worse, one in-house model was kept as deposited, and the crystal structures came close to the published ones without beating them. The investigator steered throughout, ligands and ions had to be supplied, and nothing was compared with human experts or other agents.
Claude derived, coded and fitted a new equation of state for polar fluids from a one-page brief; it fits ethers within about 1% and needs its polar term all but switched off for ketones
What happened: Werner Müller, Benoit Creton and Carlos Nieto-Draghi of IFP Energies nouvelles and Erich Müller of Imperial College gave Claude Opus 5 three papers and three prompts. Over three sessions it derived SAFT-VR IMP, wrote about 600 lines in the Julia package Clapeyron.jl with 97 verification tests, and fitted it to DIPPR data for six solvents. For the ethers and the sulfide, errors on the fitted data stayed within about 1%. The authors say none of its errors was caught by the researchers; its own tests caught them.
Why it matters: It is a derivation-to-fit chain a reader can check end to end, and the checks are real: exact reduction to the parent model, analytic limits, derivatives and the Clapeyron equation. But the 97 tests were written by the same model as the code, so they cannot catch a misreading of the physics that both share; the authors say auditing this much output is the real limit. And for the three ketones the data rejected the measured dipole: the fits only worked once the dipole was refitted to near zero, which removes the physics the model was built to add.
AutoDrude fits a polarizable force field straight from DFT, and catches the lithium transport in a polymer electrolyte that charge scaling misses
What happened: Artur Lyssenko, KyuJung Jun and Rafael Gómez-Bombarelli at MIT make the Drude relaxation differentiable, so bonded, electrostatic, polarization and screening parameters are fitted together against DFT energies, forces and dipoles, with no empirical tuning, at 1.20 times the cost of a fixed-charge simulation. On PEO-LiTFSI the dipole-constrained model matches the measured ionic conductivity about as well as the best charge-scaled model, 17% against 16% mean absolute percentage error.
Why it matters: Charge scaling is the standard way to fake the missing electrons, and on conductivity alone it still wins narrowly. The case for real polarization is what it gets that scaling cannot: the cation transference number in the dilute limit, its rise with salt concentration, and the absence of lithium-TFSI contact pairs at low concentration. A second result cuts against anyone fitting potentials: the variant that matched the DFT data more closely underestimated conductivity by up to an order of magnitude.
Evidence and limits: One polymer and one salt. Conductivity, the cation transference number and ion-pairing trends are all checked against published experiment, but the model reproduces only the dilute-limit transference and its rising trend, not its size: it flattens near 0.17 above 1.7 molal where experiment reaches about 0.5. Training is 5,942 DFT configurations of seven species, 10% held out, no separate validation set. A preprint, not peer reviewed.
Computing the Coulomb repulsion from first principles takes about 40 K off Mg2IrH6's predicted superconducting temperature
What happened: Kieran Bozier and Chris J. Pickard at Cambridge compute the Coulomb parameter in the random-phase approximation for six hydrides and for the ambient-pressure metallic elements. For Mg2IrH6 they get 0.62, and a pseudopotential near 0.39, far above the 0.1 to 0.15 that Tc predictions conventionally assume; using it costs about 40 K against the standard 0.16. The parameter closely tracks the density of states at the Fermi energy, and its energy dependence accounts for a factor-of-two gap between two earlier Tc estimates for that material.
Why it matters: Hydride searches rank candidates by predicted Tc with this repulsion held at a number borrowed from elemental superconductors. If it instead rises with the density of states at the Fermi level, the ranking is skewed exactly where the field looks hardest, because a high density of states is also what makes a candidate look promising. The paper's answer is to compute the parameter, or to correct for the density of states, rather than keep one value for everything.
Evidence and limits: Six hydrides and one set of elements, all computed. Mg2IrH6, YSbH6 and LaSc2H24 are predicted materials with no measurement to check against. For H3S, where Tc has been measured, including the computed repulsion moves the prediction toward experiment, which the paper reports as a trend rather than a number. A preprint, not peer reviewed.
A doubly metastable synthesis method makes four phases, three of which match no structure in the databases its authors searched
What happened: Kyuho Lee, Joseph G. Checkelsky, Heather J. Kulik and colleagues at MIT stack two metastable methods, diffusion and epitaxy, into diffusion-activated epitaxy: a substrate sits in a lower-melting flux and reacts at its surface while its bulk stays inert. In a narrow synthesis window they stabilised four distinct phases on oxide substrates, confirmed single-crystalline by X-ray diffraction and dark-field STEM. Three could not be matched to any structure in the experimental and computational inorganic-structure databases they searched.
Why it matters: Closed-loop materials discovery, from GNoME to DeepMind's robotic synthesis, mostly predicts stable structures first and then tries to make them. These three phases were made but sit outside the computed databases those predictions are drawn from, which is evidence that synthesizable space runs ahead of what the databases cover, so a predict-then-make loop misses metastable phases by construction. The paper's suggestion is a branch that makes first and solves the structure after.
Evidence and limits: The structures of the three unmatched phases are not solved: the claim is that no database structure fits their diffraction and microscopy, not that they are new, which would need a solved structure. One narrow window of one flux system, four phases. No AI method is used in the work; the bearing on predict-then-make pipelines is the reason it is recorded here. A preprint, not peer reviewed.
Brice Ménard uses Claude Science to build a full-sky ultraviolet map, a third of it predicted from other wavelengths
What happened: Brice Ménard, an astrophysicist at Johns Hopkins and Anthropic, had agents in Claude Science gather and cross-calibrate the public UV surveys, among them GALEX's 38,000 observations, and fill the third of the sky never observed in UV by learning how UV brightness tracks visible, infrared and radio maps, then add over 100 million Gaia stars. On measured patches hidden from it, the model came within about 10% of the real UV. Each pixel is marked measured or predicted, with an uncertainty.
Why it matters: A complete UV sky was missing because filling it takes weeks of pixel-level calibration that nobody puts first; here it took a few days of a scientist steering agents between his own projects. The map is a teaching tool, and the post's point is that most fields have a backlog of such resources that is now much easier to clear.
Evidence and limits: The predicted third is a statistical estimate, checked only against held-out patches that were measured; no new UV observation tests it. Ménard caught a calibration artefact that two rounds of agent review had missed, and the post cites no paper or outside review of the map.
OpenAI withdraws three of its model's maths manuscripts after a sign error, and revises 14 others
What happened: OpenAI's changelog for its public maths repository says a sign error in one manuscript, on Weil classes of split abelian eightfolds, breaks a cancellation argument and a construction that two K3-surface papers depend on. All three are withdrawn, one of them a claimed proof of the rational Hodge conjecture for products of K3 surfaces. Fourteen others got proof repairs or corrected statements. OpenAI counts 300 of 719 top-line results formalised in Lean.
Why it matters: It is the first withdrawal from the 722 manuscripts OpenAI posted on 6 October, a day after the release. One wrong sign took out two more papers built on it, which shows how an error travels through a family of machine-written results that cite each other. The repository keeps the withdrawn texts with notices, so readers can see what was claimed and where it failed.
Evidence and limits: The changelog does not say who found the error or how. The 14 repairs are OpenAI's own account, and its count that day left 419 of the 719 results without a Lean proof.
Biohub, the US DOE and NIH, and Google DeepMind, Isomorphic Labs and Meta commit $1.8B to open data for virtual-cell models
What happened: Biohub's Virtual Biology Initiative, launched in April with $500M of its own, adds three blocks: DOE invests more than $500M over five years through the Genesis Mission, NIH brings datasets and repositories built with more than $500M of past federal money, and Google DeepMind, Isomorphic Labs and Meta put in $300M together. The goal is open, standardised data on how cells respond to interventions, across far more cell types. Alex Rives of Biohub, Max Jaderberg of Isomorphic and Pushmeet Kohli of Google DeepMind speak for it.
Why it matters: It bets that data is what holds virtual-cell models back, the same limit Xaira's X-Cell paper reports. The $1.8B is not all new money: NIH's share is past spending on data that already exists, and DOE's is spread over five years. No first dataset, date or access terms for the company-funded data are given. It shows where public and big-tech money is going, not that a virtual cell will work.
Matchpoint Therapeutics raises a $150M Series B to take its oral covalent T-cell drug MPT-062 into the clinic in 2027
What happened: Nextech Invest and Norwest co-led; Invus, BB Biotech, T1D Fund and BOLD Longevity Growth are new, and Access Biotechnology, Atlas Venture, Sanofi Ventures and Digitalis Ventures followed on. Kanishka Pothula of Nextech and Brian Matesic of Norwest join the board; Andre Turenne is CEO. Nathanael Gray and Tinghu Zhang of Stanford, Jianwei Che of Dana-Farber and Edward Chouchani of Harvard Medical School and Dana-Farber founded the company. No valuation was given.
Why it matters: Matchpoint finds covalent binders with chemoproteomics and uses machine learning to choose targets and guide compound design. The money takes its first drug, an oral T-cell modulator for autoimmune disease, to clinical proof of concept, with trials to start in 2027. No clinical data exist yet, and the release does not say how much of MPT-062's design came from the models. Novartis already funds one of its programs under a July 2025 option deal worth up to $1B.
Rivercell launches in Paris with a $25M seed to generate data on how cells respond to treatment over time, and train a virtual-cell model on it
What happened: HV Capital led; HCVC, Alven and Bpifrance Digital Venture joined. CEO Yann Fleureau, who co-founded the cardiac-AI company Cardiologs that Philips bought in 2021, started it in summer 2025; Eric Durand, formerly Chief Data Science Officer at Owkin and director of oncology data science at Novartis, joined this year as co-founder and CSO. Charlotte Bunne of EPFL and Fabian Theis of Helmholtz Munich join its scientific advisory board, and Maxi Pethö-Schramm spoke for HV.
Why it matters: Rivercell bets on the same gap as the Biohub initiative announced the same day: virtual-cell models lack data on how a cell changes over time after a drug or a genetic change, measured several ways at once. Its answer is a purpose-built data platform and an automated wet lab in Paris of its own, not public data. The money pays for that and a first model; no dataset, model or result has been shown yet.
Meta fits a scaling law to how well a robot world model imagines what happens next, and it holds for the two largest models, which were left out of the fit
What happened: Artem Zholus, Mahmoud Assran and colleagues at FAIR train predictors from 22M to 8B parameters on a frozen V-JEPA 2.1 encoder, from 23 public manipulation datasets across 12 robot platforms. Imagination error is the per-token distance between predicted and encoded features over a nine-step rollout. It follows a second-order power law in compute. Fitted on the 22M to 2B frontier, it matched the held-out 4B and 8B checkpoints more closely than three rival forms. Planning toward one goal image, grasp success rose from 30% at 22M to 67% at 8B.
Why it matters: The law was fitted without the two largest models and still describes them, which is more than a curve drawn through every point can claim. What it does not do is reach past 8B: the authors report an irreducible error the fit is already near, and say more compute alone will not move it without more and more varied interaction data. The robot runs are the authors' own, 30 episodes per model and task, scored by hand, on a platform whose data was in training. The promised code and checkpoints were not public two days on.
Ten extra structures per material sharpen a pretrained potential's phonons, but most ways of fine-tuning it then get phase transitions more wrong than not fine-tuning at all
What happened: Jonas Grandel, Philipp Benner and Janine George at BAM Berlin and Jena fine-tuned MACE-MP-0b3 separately for each of 53 materials, against their own new VASP reference calculations. Regularising towards the pretrained weights while training only the node embedding and first interaction layer cut the median phonon error to 0.05 THz against the foundation model's 0.27. On recovering the phase a soft mode leads to, across the nine materials DFT finds unstable, multi-head scored 0.43 and ordinary transfer learning 0.56, both below the untouched foundation model's 0.63.
Why it matters: The finding that matters is the gap between the two scores: getting harmonic frequencies right, and even classifying which structures are unstable, does not mean the model has the energy surface right along the unstable direction. The authors say as much in their introduction and conclusion. So a recipe picked on phonon error can quietly be worse at the thing phonons are usually computed to predict. Everything is model against DFT, in software; the best recipe's margin over doing nothing is 0.66 to 0.63.
Danaher plans its first AI-run autonomous lab, at Abcam, to design, make and test antibodies and other reagents, at scale from early 2027
What happened: The lab uses AI to propose affinity reagents such as antibodies, builds and tests each one, and feeds every result back into the model. It strings together tools from Danaher's own companies (Beckman Coulter Life Sciences, Cytiva, Genedata, Integrated DNA Technologies, Molecular Devices), with robotics built with Automata. CEO Julie Sawyer Montgomery spoke for the company; chief science officer Jose-Carlos Gutierrez-Ramos says it is designed for up to 8x faster discovery and up to a tenfold rise in output, from tens to hundreds of reagents a year.
Why it matters: Danaher is building a closed design-make-test loop out of instruments its own companies already sell, as the first step of a stated push toward smart instruments and autonomous labs. It is a plan: the release says 'plans to launch', though a Simply Wall St piece, also run on Yahoo Finance, said 'launched', and the 8x and tenfold figures are design targets with no data behind them yet. What it would prove is whether such a loop speeds up reagent discovery inside a working business.
Simon Batzner, a co-author of GNoME, leaves Google DeepMind for Anthropic
What happened: Simon Batzner, who built the NequIP interatomic potentials and was a co-author of GNoME and GenMS at Google DeepMind, announced on X that he has joined Anthropic as a Member of Technical Staff, thanking colleagues at DeepMind and Google Brain.
Why it matters: With him, four of GNoME's six authors now work elsewhere; three of them co-founded Periodic Labs. The team that produced DeepMind's best-known materials result has largely dispersed, while DeepMind builds a materials lab in London with different people. It follows John Jumper's move to Anthropic in June.
Evidence and limits: His post does not say what he will work on at Anthropic. His own website still described him as at DeepMind on 2026-10-22.
OpenAI releases 722 maths manuscripts from the unreleased model behind its Navier–Stokes claim
What happened: OpenAI posts 722 manuscripts in 372 families from an internal model it gave about 4,000 open problems, at about three hours of ChatGPT Pro thinking per result on average. Sam Altman singles out four: a zeta zero-free region above real part 7/8 (the quasi-Riemann hypothesis), the Unique Games Conjecture and the free group factors problem, which he says have Lean proofs, and the Hodge conjecture for CM abelian varieties, without one. OpenAI says many but not all results are formalised, and unformalised ones could have issues.
Why it matters: Navier–Stokes was one disputed claim; this is hundreds from the same model at once, none confirmed outside OpenAI. A Lean proof checks each step, not that the formal statement is the open problem, and only some results have one. Mathematicians split: Alex Kontorovich (Rutgers) said the quasi-Riemann result would be "an instant Fields Medal" for a human; Álvaro Lozano-Robledo (UConn) said he was not sure what it is trying to achieve "other than, again, a massive PR stunt". The IAS advisory group, whose request to stop testing open problems on private models OpenAI declined, said its advice was no endorsement.
Resolution's agents formalize Hironaka's resolution of singularities in Lean: about 600,000 lines in a week, for an estimated $135,000 in tokens
What happened: Chris Elliott of Resolution, a lab that studies how AI models generalize, worked with agents on a 112-page step-by-step plan, then had an orchestrator agent, mostly Claude Fable 5.1, carry it out. The proof of Hironaka's 1964 theorem and its modern forms took about a week and some 30 interventions from him: about 100 billion input and 300 million output tokens and 600,000 lines of Lean. Billy Snikkers and Adam Newgas then spent two weeks refactoring it.
Why it matters: It gives formal mathematics a price tag a reader can cite, reached outside a frontier lab with little scaffolding beyond the plan and some agent skills from earlier efforts. Resolution needs the theorem for singular learning theory, which ties the shape of a model's loss landscape to how it generalizes, and says the library already lets it automate more of that theory.
Evidence and limits: The figures, including the token cost, which is an estimate, are Resolution's own. The proof is public with a checker, Comparator, that confirms it proves exactly the stated theorems, but the post reports no outside check. It formalizes known mathematics; the learning-theory results it is meant to enable are promised for the coming months.
WhiteLab Genomics raises $26M to test its AI-designed gene therapy delivery vehicles in animals
What happened: AVP led the Series B, with Yaday Health and Blast Club new and Omnes Capital and Debiopharm following on; François Robinet of AVP and Daniel Teper of Yaday Health join the board. The Paris company, run by co-founder David Del Bourgo, designs AAV carriers with its ALFRED platform and is extending it to lipid nanoparticles and synthetic promoters. It says AI-designed AAVs crossed the blood-brain barrier in in vivo work with the Paris Brain Institute, with no liver signal detected.
Why it matters: The money pays for the in vivo experiments that show whether computed designs deliver genes in living animals, and for building designs it can own and license, not only selling services. The brain result is the company's own claim, with no paper or numbers in the release; more data is promised at the ESGCT congress.
Multiply Labs raises a reported $75M for robots that automate manual steps in making cell and gene therapies and other biologics
What happened: Per The AI Insider, Patrick Soon-Shiong and NantWorks led the Series B, with AstraZeneca, Lingotto, Teradyne and Strange Ventures new and Founders Fund and Lux Capital among returning backers; total raised passes $100M. The company says the money expands manufacturing capacity, speeds up product development and grows its engineering, regulatory and commercial teams; Fred Parietti is co-founder and chief executive.
Why it matters: Parietti's case is that AI now designs more therapies than anyone can make, so the bottleneck moves to the factory; AstraZeneca, already testing the robots, now invests too. The 74% lower cost per dose and up to 100 times the throughput are the company's own figures, and the round rests on one outlet's report.
Vinci raises $250M at a $1.5B valuation for a physics model meant to replace simulation checkpoints in chip design
What happened: Advent, Temasek and Xora Innovation co-led the Series B; AMD Ventures and Madrona joined, and Eclipse and Khosla Ventures followed on. The $1.5B valuation is the company's own. Vinci says the money takes it from thermal, thermo-mechanical and convective-fluid work in semiconductors to memory, advanced computing, vehicles, aircraft and satellites, on a physics model it says needs no per-design training and reads designs of hundreds of millions to more than 15 billion degrees of freedom in minutes. Hardik Kabaria is co-founder and CEO.
Why it matters: This is the simulation bet that materials and chemistry have been making with learned potentials, moved to engineering: if a model can stand in for a finite-element run and stay accurate, physics stops being something you check at a few points and becomes something you ask while the design moves. AMD is now in through AMD Ventures. The published comparison with conventional solvers is Vinci's own, at EPTC 2025 and on its site; the customers it says benchmarked it are unnamed, and none of their results is public.
A Peking University team and coding agents formalize the Poincaré conjecture in Lean: about 2.8 million lines, a little over two weeks for the main phase, about $25,000
What happened: Following Morgan and Tian's account of Perelman's proof, mathematicians wrote a blueprint and designed about 90 milestones; GPT-6-Astra put the milestone statements into Lean and, through coding agents working in parallel, built most of the proofs, with Fable-5.1 helping review statements. The result proves both the topological and smooth theorems, passes a full Lean build and a check against target statements written only with Mathlib, and is rechecked by an independent kernel. Bin Dong and Chunlei Liu are corresponding authors.
Why it matters: After Fermat's Last Theorem, another famous proof is machine-checked, and the paper says where the work went: preparation began in May, and people chose definitions, reviewed the milestone statements and unblocked the agents. The kernel settles that each proof step is valid; whether the statements say what Perelman proved still rests on human review. An independent formalization by Ziyang Qin, Bennett Chow and others was announced at the same time.
AuthorsZhiyuan Zhang, Axel Delaval, Chunlei Liu, Bin Dong
A model that reads the retina's vessel network predicts preeclampsia months early, and held up at a second hospital with 13 cases
What happened: Srilaxmi Bearelly, Cassandra Areff and Liat Shenhav, at Columbia and NYU, turned retinal photographs into measurements of the vessel network's shape and branching, not image embeddings. Trained at Columbia on 1,267 women, 55 of whom developed preeclampsia and 54 of whose images were usable, it reached AUC 0.91 and average precision 0.81. With no fitting on NYU outcomes it scored 0.81 and 0.68 on the 66 analysable NYU Langone women, 13 of them cases, against 0.73 and 0.50 for the Fetal Medicine Foundation's calculator. RETFound was near chance there.
Why it matters: Preeclampsia is diagnosed late, and trials find aspirin started before 16 weeks cuts preterm delivery with preeclampsia by up to 62%, so an early signal is worth something. What this tests is transfer: different cameras, workflows and populations, with NYU's features rescaled to Columbia's but no refitting. It is a second site, not a second party: NYU's lead is the model's senior author, so this record counts it internal. The authors' caveats: 54 and 13 cases, NYU's 17.7% preeclampsia rate against the hospital's usual 8%, calibration and thresholds still open.
XtalPi debuts a closed-loop siRNA design platform and says six of its own programmes now run on it, the first reaching monkey data in 7 months
What happened: Kodexia pairs a model XtalPi calls siFormer (built on RNA thermodynamics and strand-loading structure, not sequence correlation alone) with automated labs it says run over 500 in vitro and 30 in vivo experiments a week. XtalPi reports a nearly threefold gain in design efficiency over conventional workflows, and says more than half of first-round designs across several programmes beat the positive control in vivo. Its lead programme, in IgA nephropathy, reached monkey efficacy data in seven months, exceeding a clinical-stage molecule against the same target.
Why it matters: The claim being tested is that a design loop can carry a new modality end to end: XtalPi's evidence has been in small molecules and antibodies, and siRNA's hard parts (extrahepatic delivery, cell-to-animal translation, patent space) are the ones it says Kodexia designs against. Monkey efficacy in seven months is the figure to hold it to. But every number is measured in-house with nothing shown (no figure, no dataset, no paper), and the clinical-stage molecule it says it beat is unnamed, so none of it can be checked from outside.
insitro puts models trained on Lilly's decades of preclinical data into Lilly TuneLab, for member biotechs to predict how a molecule behaves in the body before making it
What happened: The model reads only a molecule's structure and predicts clearance, volume of distribution and mean residence time after intravenous dosing in mouse, rat, dog and monkey. On a held-out test set split by chemical cluster, predicted and measured values correlated at roughly 0.6 to 0.8 and 69-91% of predictions fell within three-fold. For mouse and rat clearance it reached a correlation of 0.54-0.62, against about 0.5 for the standard extrapolation from lab assays. TuneLab runs by federated learning, so members' structures are not shared.
Why it matters: Animal pharmacokinetics is one of the measurements that decides which molecules go forward, and drugmakers' decades of it are rarely seen outside the company. Here one drugmaker's data reaches smaller companies as a model they can query rather than as data, which is how TuneLab is meant to share what large companies hold.
Evidence and limits: All figures are insitro's own internal test, and the post warns they may not hold for new chemistry. The baselines' correlations were computed on the whole dataset, not the same test split. The data's size is given only as hundreds of thousands of molecules overall and a few thousand in vivo studies.
Gemini 4 agents propose, and the authors re-derive, exact 3D plasma equilibria that break stellarator symmetry, widening the counterexamples to Grad's conjecture
What happened: Opal Issan, Jonathan Citrin and Brendan Tracey of Google DeepMind, with Matt Landreman of the University of Maryland, extend two families of exact 3D magnetohydrodynamic equilibria that Landreman found weeks earlier as counterexamples to Grad's 1967 conjecture. A new parameter breaks stellarator symmetry and gives back the earlier solutions when set to zero. The paper says Gemini 4 agents in Antigravity proposed the generalised solutions within hours; the authors chose the problem, steered the agents and re-derived every result.
Why it matters: It is a model contributing a new closed-form result in mathematical physics, with its role written into the paper and every step re-derived by people. It stays on paper: the solutions are offered as benchmarks for stellarator codes such as VMEC and DESC, and none has been run on them yet. The first counterexamples were Landreman's own; this paper widens them.
Arc Institute releases PIE, an open model that predicts how cells respond to a perturbation by reading what is already known about the gene, drug and cell line
What happened: PIE turns curated sources such as Gene Ontology, DepMap, PubChem and Cellosaurus into inputs and predicts which genes change, in which direction and by how much, in one pass with no fine-tuning. On the Replogle-Nadig Perturb-seq data across four human cell lines, it reached 1.2 to 3.2 times the AUPRC of the strongest competing method in each setting, with the largest gain on perturbations it had never seen. Code, weights and data are public.
Why it matters: A virtual cell is only useful if it predicts experiments nobody has run yet, and PIE tries to get there by borrowing what biologists already wrote down. Arc says accuracy drops when both the cell line and the perturbation are new, and every score is on existing public data, computed by Arc itself.
Italfarmaco signs an AI drug-discovery deal with VeriSIM Life on an HDAC target, potentially worth over $100M in fees, milestones and royalties
What happened: The two will design and optimise small molecules against an HDAC target for neuromuscular, fibrotic and rare diseases. VeriSIM brings AtlasGEN, which proposes synthesisable molecules, and BIOiSIM, which scores them for pharmacology, exposure and safety; Italfarmaco brings HDAC biology, chemistry and clinical development. Christian Steinkühler, Italfarmaco's chief scientific officer, and Jo Varshney, VeriSIM's founder and chief executive, spoke for the companies. No upfront payment was disclosed.
Why it matters: A private drugmaker that already has an HDAC drug on the market is paying an AI company to find the next one, which shows where such companies earn money: service fees now, milestones later. Over $100M is the most it could be worth, not money paid.
Google DeepMind and Caltech's AlphaProtein Novo designs an enzyme that makes piperidine at 99:1, where none of 188 natural and engineered variants beat 30:70
What happened: Zachary Wu, Josh Abramson and colleagues run a diffusion model that writes an enzyme's backbone and sequence around a catalytic motif, then filter candidates on AlphaFold 3 predictions. Frances Arnold, who won the 2018 Nobel for directed evolution, is on the byline and her Caltech lab did part of the wet work. Of 188 natural and engineered heme proteins screened, nine were active and none passed a 30:70 piperidine-to-pyrrolidine ratio; design GDM_NT_270 reached 99:1 at 94% enantiomeric excess over 22 turnovers.
Why it matters: Designs were expressed, purified and measured, and knocking out active-site residues abolished activity in 14 of 19 serine esterase mutants, so the catalysis is the designed one. They win on selectivity and durability, not speed: a designed DEHP esterase is 14-fold more active at 90 °C than at room temperature, where two natural ones are dead, but at room temperature it reaches only 6% of a natural enzyme, and the paper says de novo rates stay orders of magnitude below natural or evolved enzymes. The 360,000 M-1s-1 esterase has a natural serine-esterase fold; the best with none is 4,900.
Iambic files to start a Phase 1/2 trial of IAM217, a KIF18A inhibitor it says it designed with AI and without a drug-bound crystal structure
What happened: The IND asks the FDA to let Iambic test IAM217 in triple-negative breast cancer, ovarian cancer and other tumours with unstable chromosomes. Iambic says its NeuralPLexer model supplied the structure where no inhibitor-bound crystal structure existed, and Enchant and earlier Iambic models took hits from 24 screening plates from about 100 nM to about 10 nM potency while tuning oral uptake, drug interactions and brain entry. In a preclinical brain-metastasis model, it reports established tumours shrank by about 90%.
Why it matters: If the FDA clears it, IAM217 becomes Iambic's second drug in human trials after IAM1363, and the IPO prospectus already names it as a use of the money. It is also a concrete case of a claim AI drug companies often make: that a predicted structure can stand in for a crystal structure when designing a drug.
Evidence and limits: An IND is a request, not a clearance. The potency, brain-entry and tumour figures are Iambic's own, given in a press release without the data, and no person has yet received the drug.
A Vals researcher and Claude agents compute two magnet candidates that sort electrons by spin with no net magnetism, one of them first made in 1999
What happened: Geby Jaff of Vals AI and a team of Claude Opus 5.5 agents used density functional theory to look for Luttinger-compensated magnets: zero net magnetism, yet electrons sorted by spin, as spintronic memory needs. They designed a new oxide, YBaMnFeO5, predicted to be a semiconductor with wide spin-sorted windows, and pointed out that KV[Cr(CN)6], made in 1999 and measured as magnetically ordered up to 376 K, is this kind of magnet, which earlier papers had not said. Inputs, outputs and code are public.
Why it matters: A 2025 study named a room-temperature magnet of this kind as the open goal. The second candidate already exists as a sample, so the next test is short: make it again and measure whether its electrons really sort by spin.
Evidence and limits: Everything is calculated. The designed oxide needs its manganese and iron in a checkerboard that the agents' own simulation shows scrambling near 950 K, while such oxides are made at 900 to 1300 C, so it may be hard to make in useful form. The only 1999 sample is a powder with water in its pores; the two methods disagree on whether that cuts its spin window by more than half, and nothing has been measured.
The US and 16 other countries endorse the Kyoto Vision, a non-binding statement that they will try to put AI to work in science
What happened: Science ministers meeting in Kyoto alongside the STS Forum on 4 October endorsed a one-page declaration led by Michael Kratsios, director of the White House Office of Science and Technology Policy. They "endeavor" to widen researchers' access to what the text calls "super intelligence" tools, scientific data, computing and experimental facilities, and encourage AI that drives "autonomous, closed-loop discovery", alongside metascience and new ways to fund research. Signers include Japan, Germany, the UK, Italy, Korea, Singapore and the UAE.
Why it matters: Seventeen governments now name AI-run, closed-loop discovery as something they mean to support, the direction the US already pays for through the Genesis Mission. But the text commits no money, sets no targets and binds no one: it says "endeavor" where the White House release says "pledged". China and France are not on the list.
Zhao Song and Microsoft Research's Xinzhi Zhang post a proof of the 1995 KLS conjecture, working with AI tools they directed
What happened: Independent researcher Zhao Song and Microsoft Research's Xinzhi Zhang posted a 140-page arXiv paper on 4 October claiming a proof of the Kannan-Lovász-Simonovits conjecture: in any dimension, a flat cut through a log-concave distribution is within a fixed factor of the best cut of any shape. The paper says that with GPT-6 Astra, GPT-5.6 Sol and Claude Fable 5 and 5.1 they explored more than 100 approaches on the way to an intermediate bound, choosing which to pursue, and it describes the final steps as the authors' own ideas.
Why it matters: Results in sampling, optimisation and high-dimensional statistics have assumed this conjecture; if the proof holds, they stand without it. It also shows a way of working: Microsoft's post says unguided runs of several leading models on about the same token budget fell short, while agents exploring in parallel, with a researcher pruning each round, got there. Pierre Bizeul, Boaz Klartag and Joseph Lehec posted a 32-page proof the same evening, built on the criterion in Song and Zhang's first version, and write that ChatGPT found most of its proofs and ideas.
Evidence and limits: No journal has reviewed either paper. arXiv shows the Bizeul, Klartag and Lehec proof posted about two hours before Song and Zhang's complete version. A public Lean 4 formalization (the KLS500 repository) checks the theorem with constant 500, chiefly along the second group's route, not the 140-page argument. Microsoft says a shortened proof was verified in Lean and is not yet released.
Xaira says X-Design Vega took its XA-1 nanobody from DNA synthesis to a lead in 7 weeks, and made an antagonist for a GPCR that library screening and llama immunisation failed on
What happened: On its own site, Xaira describes two internal programmes run with X-Design Vega. XA-1: 182 designed nanobodies, 18 confirmed binders, the best at 17 nM KD to the human target and passing every developability test; 7 weeks from DNA synthesis to a lead. XA-4, a GPCR where nanobody library screening and llama immunisation had both failed: about 60,000 designed, 92 tested, 7 bound and 6 of those inhibited, best IC50 183 nM.
Why it matters: It tests two questions for de novo design: can it shorten hit-to-lead (XA-1 reached a selective, developable lead), and can it open targets where other methods failed (XA-4 gave a working antagonist, not just a binder). Limits: targets undisclosed, so outsiders cannot judge how hard they are; in-house assays, no independent replication; no paper, so method and benchmark go unreviewed; XA-1's epitope was chosen in advance, so it shows design and optimisation, not finding a binding site.
XtalPi moves its own AI- and robot-designed hair-loss drug into IND-enabling studies after a mouse result it compares with minoxidil
What happened: The company says pharmacology, safety and manufacturing studies are under way toward an IND. In a mouse model of androgenetic alopecia, a candidate reached a hair-growth score comparable to 5% topical minoxidil at one-hundredth of its concentration. It was optimised by a loop of AI-guided design, free-energy calculations and robotic synthesis and testing; XtalPi already sells two AI-developed scalp-care cosmetic ingredients.
Why it matters: An AI-and-robotics company taking a molecule of its own toward the clinic, not only running projects for partners, is the step investors in such platforms wait for. The evidence is one mouse model in a press release, with no data, paper or named scientist, and IND-enabling is the stage before any human dose.
Ai2 opens the weights and training data of AstaBrief 8B, the model that writes cited literature reports in its Asta tool
What happened: AstaBrief is Qwen3-8B post-trained on 47,000 reports generated for real researcher queries, plus 6,000 preference pairs, and it writes a report in one pass instead of section by section. In Asta it runs as Fast mode at 51.1 seconds a report against 178.5 for the Claude-backed Thinking mode, about 3.5 times faster. Ai2 says the clearest gain came from one simple filter: dropping training reports that left claims uncited.
Why it matters: An institution that cannot send unpublished questions to a hosted model can now run the report writer itself, which is the point of releasing the weights. Ai2 is plain that most of the evaluation was done in 2025 against the frontier models of that time and has not been rerun, and that in its 14-question human study of three researchers, the open DR Tulu was preferred overall. Two of three preferred AstaBrief only on citation accuracy.
ImmuneAgent ranks antibody candidates out of human B-cell sequencing: 60 of 110 cloned neutralised a virus, 12 of them broadly
What happened: The system reads single-cell B-cell receptor data from people who were vaccinated or infected, ranks candidates on cell type, sequence, predicted structure and a trained predictor together, and feeds each round of lab results back into the model. Of 110 cloned antibodies (60 from an antigen-sorted SARS-CoV-2 dataset, 50 from an unsorted influenza one), 60 neutralised and 12 were broadly neutralising. Five protected all six mice each against a lethal flu dose, as the clinical-stage antibody MEDI8852 did.
Why it matters: Antibody discovery usually starts by sorting cells against a known antigen. Only the influenza arm skipped that step, and it is the one the mouse protection came from, which is what would matter for a virus nobody has sorted cells for yet; the larger SARS-CoV-2 half was sorted as usual. What was tested is prophylaxis in mice, six per group, and the authors say the design cannot establish equivalence to MEDI8852. One group's preprint, baselines retrained by the same team, and their own audit found 19% of the SARS-CoV-2 test sequences within 90% identity of training data.
Meta posts six maths papers done with Muse Spark in its ordinary chat app; it says five answer open questions, and for three of them other teams posted answers first
What happened: Mathematicians used Muse Spark 1.1 and 1.2 in Thinking Mode on meta.ai, with no research scaffold. The model wrote the GAP search that found a 384-element counterexample to a 2024 conjecture that semiabelian groups are monomial, produced a three-dimensional counterexample in evolution algebras, and drafted three core technical sections of a number-theory paper. Each paper marks in its margin which passages people drafted and which the AI did.
Why it matters: This shows a public chatbot doing real work on problems mathematicians picked themselves: people chose every question and checked every proof. The papers sit on Meta's own site, reviewed by mathematicians inside the collaboration rather than by a journal, and Meta does not say any proof was machine-checked. On three of the five, others had already posted solutions by other routes, among them the AI agent Nilradical, whose counterexample came with a Lean proof.
AuthorsAykut Arslan, Leonard Dinh, Joseph Phillip Brennan, Andres Barei
Baker lab designs two working enzymes with complex active sites, one with a 2.7 Å crystal structure
What happened: CaGE, from David Baker's lab, fades a structure predictor's geometric constraints out over repeated design cycles, so the final sequence has to hold the active site on its own. Two designs worked in the lab: an iron enzyme giving 83% enantiomeric excess, and a PLP retro-aldolase whose 2.7 Å crystal structure matches the design model to 0.702 Å on the backbone.
Why it matters: An active site that couples several residues to a cofactor and a substrate has been the hard part of designing enzymes from scratch. On the PLP side none of the 30,293 usual-pipeline sequences passed the filter; after CaGE cycling 96 did, and one of those turned out active.
Evidence and limits: No rate constants: the paper gives endpoint yields over 16 hours, not catalytic efficiency, so there is nothing to compare against a natural enzyme. The crystal structure is of the empty protein, so it shows the fold, not the active site. Six of 71 iron designs gave a signal above control. No outside lab tested it. A preprint.
MIT's VISTA harness lets Claude Opus 5.0 finish all 25 public ARC-AGI-3 games from screen images, with no new model trained
What happened: Kaiming He's group at MIT wraps existing multimodal models in a harness that keeps every game frame and lets the model zoom into or replay any of them. Claude Opus 5.0 scored 100 and GPT-5.6 Sol 99.00, against 40.68 and 13.33 under the organizers' official setup, which runs through the API with tighter limits and, for Opus, one step lower reasoning effort. Opus used 57.4% fewer actions than first-time human players.
Why it matters: The authors' own step-by-step test shows where the gain comes from. Images instead of text grids and looser action and time limits took GPT-5.6 Sol from 13.33 to 51.66; keeping every frame to zoom into and replay, plus exact pixel reads, took it from 70.05 to 99.00. For agents watching experiments, that argues for keeping raw observations rather than summaries. Only the 25 public games were tested, and the models may have seen them in training.
arXiv caps each submitter at two papers a month after a record 40,363 submissions in September, a rise it puts down to AI tools
What happened: From 1 October each submitter may send two papers a calendar month, in every category; rejections count, co-authors don't, and the 2024 cap of three active submissions stays. September set a record: 40,363 submissions, 51% over a year earlier, and almost 9,000 support tickets.
Why it matters: arXiv blames AI tools: thin papers, one study split into several, dense AI-written ones; cs.AI is up over sixfold in two years. The cap is a stopgap to share volunteer moderators' time, not a quality test.
Imperial's GEODE makes 11.9% metastable, unique and novel crystals on MP-20, against 7.7% for the next symmetry-aware model
What happened: Yuchen Lou and Alex Ganose's model samples a symmetry template, then diffuses lattice, coordinates and elements in Cartesian space. Filtering templates lifts the rate to 18.2% without retraining, and a permittivity-guided demo is shown.
Why it matters: Symmetry-aware generation now beats earlier symmetry-aware models. All of it is computed: metastable is an energy check, not a made material, and rates differ between papers' set-ups.
AI Theorist builds a model of a quantum material's optical spectra from measurements nobody had published, and the experiment that would settle it has not been run
What happened: Human scientists measured reflection and photocurrent spectra on a thin a-RuCl3 crystal at 8 K. Agents (Claude Opus 4.6 for hypotheses, critique and synthesis, a fine-tuned Qwen3.5-27B for choosing calculations) then ran Quantum ESPRESSO and BerkeleyGW themselves, rejected two explanations and assigned the three measured peaks to bound electron-hole pairs at 1.077, 1.179 and 1.291 eV. The run took 154 million tokens; the Claude share of that came to about $335 at list prices, with 69 hours of simulation unpriced.
Why it matters: This is a model fitted to measurements that already existed, not a prediction anyone has checked: the test that would separate it from the alternatives (rotating the light's polarisation from zigzag to armchair) is written down as a prediction. The authors call the peak assignments candidates, say one calculated spacing is still too small and that the screening updates are not strictly converged. An unnamed human theorist went over the methods. One group's preprint; code and data are promised, not posted.
MeteoSwiss's 1-km AI weather model Varda-single roughly matches its own 1-km physics model for 33 hours over the Alps, but underestimates some local wind peaks
What happened: Varda-single is a graph-transformer forecaster built in ECMWF's Anemoi framework, trained on ERA5, then a 20-year 1-km Swiss reanalysis, then operational analyses. Checked from April 2025 to March 2026 against analyses and weather stations, it broadly matches the 1-km ICON-CH1-EPS control up to +33 h and generally beats the 2-km ICON-CH2-EPS control up to +120 h. It runs quasi-operationally every 3 hours and its weights are on Hugging Face. Alberto Pennino and Francesco Zanetta lead it; Oliver Fuhrer is last author.
Why it matters: Global AI models already rival physics forecasts; the open question was whether that holds at the kilometre scale, where mountains steer wind and rain. A national weather service now says yes on most headline scores, on its own verification. It is deterministic only, smooths convective rain, and underestimates some local wind maxima, which are much of what a mountain forecast is for.
AuthorsAlberto Pennino, Francesco Zanetta, Oliver Fuhrer
Shift Bioscience says the metrics behind 'perturbation models do not beat baselines' are miscalibrated, and that newer models win once better-calibrated metrics are used
What happened: Across 14 genetic perturbation datasets and 18 metrics, the paper adds a positive control built from duplicated data and a measure of how much of a metric's range a model can use. Mean squared error and control-referenced correlation often fail that test. On one Perturb-seq dataset at least one deep-learning model beats every baseline on every metric. On the Norman19 combinatorial dataset an additive baseline still captures about 88% of the gap to ideal performance on the best-calibrated metrics.
Why it matters: If the measure is wrong, the field's reading of a negative result is wrong too, and this argues the earlier benchmarks hid the signal rather than finding none. It is a re-analysis of public data with no new experiment, and every author has a commercial interest: five work for a company selling perturbation modelling and the sixth is a chief AI scientist at Xaira.
AuthorsHenry E. Miller, Lucas Paulo de Lima Camillo
