Gustave Roussy spin-out Orakl Oncology raises a $34M Series A, co-led by Norrsken VC and Singular, to predict cancer-drug trials from patients' tumour organoids
What happened: Orakl Oncology raised an oversubscribed $34M Series A co-led by Norrsken VC and Singular, taking its total to $50M. New investors are InfraVia Growth, Serena, Bpifrance's Patient Autonome Fund, teampact.ventures and Plug and Play, with Verve Ventures, Speedinvest and HCVC following on and angels including Thomas Wolf. Co-founders Fanny Jaulin (CEO), Gustave Ronteix (CTO) and Diane-Laure Pagès (COO) grow patients' tumours as organoids, screen drugs on them and link each result to the patient's records.
Why it matters: About three quarters of cancer candidates fail in Phase 2, the release says, and Orakl sells drugmakers a reading of which drugs, and which patients, will respond before that stage. The money is for new models, five times the data, lung and gastric cancer added to colorectal and pancreatic, US and UK teams for more than 10 drug programmes, and finding new co-targets. Philip Tagari, chief scientific officer of insitro, who helped bring Amgen's KRAS inhibitor Lumakras to patients, joins as an adviser.
Evidence and limits: The 90% accuracy in predicting trial outcomes and the 18 of the 20 largest drugmakers in its pipeline are the company's own figures; the release names no paper or trial set behind them. The round was read in participant InfraVia's release, which gives no valuation.
Qingkui Zhizao, a Beijing maker of robotic chemistry labs, raises an A round of about 100M yuan to build a cloud lab that runs customers' experiments
What happened: Chuangdongfang Investment, the thermal-materials maker Feirongda and Jinhui Big Data took part; no lead was named. Chairman Nanjiang Chen, a Beihang University PhD who worked at its School of Mechanical Engineering, founded it in 2021. It sells modular chemistry workstations (NexoLab), software that runs instruments as agents (AI-LabOS) and models trained on its lab data (MoleMind). The money is for R&D, new industries and the cloud lab. Customers it names include the Chinese Academy of Sciences, Boehringer Ingelheim, Hengrui and Gotion.
Why it matters: Chen's case is that public chemistry data leaves out full reaction conditions and failed runs, so the company built its own robot lab this year to supply data for its models, and has begun trying out experiments as a service from it, with customers watching the data as it comes in. Feirongda, which makes thermal-management materials and liquid cooling, says AI computing infrastructure is raising the bar for cooling and thermal materials, and it wants automated testing for materials formulation: a chip-cooling supplier paying for automated materials R&D.
Evidence and limits: The tens-fold gain in experiment speed and the good repeat-order rate are the company's own words, with no figure or study behind them. The amount is the round figure of 100M yuan, no valuation was given, and the reports name no lead. Chen says AI-designed experiments still need a chemist's sign-off.
Google DeepMind's AlphaGenome 2 predicts what DNA variants do cell type by cell type, and in immune cells finds about twice as many likely causal variants as its fine-tuned predecessor
What happened: Google DeepMind released AlphaGenome 2, which reads a million letters of DNA with one chosen tissue, cell type or state and predicts expression, chromatin and, newly, DNA methylation there. Training included over 107 million single-cell profiles. On immune-cell variants from the TenK10K cohort, which it was not trained on, it found 30% of likely causal variants at 90% precision, against 16% for the first model fine-tuned on single cells. In CRISPRi experiments it matched distant DNA switches to their genes better than AlphaGenome 1 (0.78 vs 0.73, out of 1).
Why it matters: Most disease-linked variants sit in non-coding DNA and act only in some cells; telling which cell is the step a drug programme needs to go from a genetic association to a gene. The model reproduced a 2024 finding that a distant variant drives IBD by boosting ETS2 in gut immune cells, and James Lee of the Francis Crick Institute, who led that study, told Nature it got in 30 minutes what took his team a year and a half of experiments. Peter Donnelly, chief executive of Genomics, says it looks forward to adding the model to its Mystra platform.
Evidence and limits: Every figure is DeepMind's own, from a paper on its own site with no DOI and no peer review. The paper says that at high precision it still misses most likely causal variants, and does worse outside open DNA regions. Peter Koo of Cold Spring Harbor Laboratory told Nature the ideas are not novel but the engineering and performance are impressive; Lee says it does not yet remove the need for experiments. It is not validated for clinical use.
Ruhr University Bochum gives a Claude Code agent its crystal database, models and reaction graph, and the agent proposes candidate compounds with synthesis plans
What happened: Miguel Marques and Silvana Botti's group connected Claude Code, running unattended, to Alexandria (about 15 million DFT calculations), its structure generator, potential and property models, and a graph of 17.4 billion reactions between 193,416 stable compounds. Asked for an untried solar absorber, the agent dropped a candidate whose best reaction releases only 6 meV/atom and proposed Sr2ZnP2 with crucible, atmosphere and heating steps. In the Ba-Zr-S system it found an unreported Ba4ZrS6. With web search alone, it proposed a compound the literature already discusses.
Why it matters: Most materials agents so far give a model ways to run calculations, which a frontier coding agent can already do itself. Here, in one run per setup, the difference tracked a group's existing data and a way to reason about synthesis: with the database and models but no reaction graph, the agent still found candidates but took its recipes from the literature. For computational groups that suggests a role serving data and models to agents that experimental groups query in plain words, with tool descriptions that say when each number is wrong.
Evidence and limits: Every candidate is computed and none was made. Each example is one run, which the authors call examples, not statistics. A later hybrid-functional calculation gave the agent's air-stable pick, Sr2ZrCu2S5, a 0.98 eV gap, too small for a solar absorber; the backup polymorph it offered came out at 1.63 eV. The group tested its own tools.
Cornell engineers let LLM agents write turbulence models and train them inside a differentiable flow solver, and the agents' models score better on their target than the standard ones
What happened: Jian-Xun Wang's group gave GPT-5-mini agents a GPU flow solver that returns gradients through chaotic simulations. The agents propose a closure, a formula or small network standing in for turbulence too fine for the grid, build its training program and choose whether to run, probe or train it. In channel flow, scored against a fine reference simulation, the agents' best model of the drag at the wall had an error of 16.5 against 55.9 for the standard equilibrium model. Each run had 30 experiments.
Why it matters: Turbulence models are still mostly written by hand, and engineering and natural-flow simulations depend on them. Earlier LLM searches for such models scored each proposal by running it; here the agent can also fit a proposal's parameters by gradient. With the formula fixed, that recovered the reference coefficients, which forward runs alone did not reach in the same budget, and the gradients the solver reported changed the agent's own choice of settings. The language model picks the form; the differentiable solver is what makes it trainable.
Evidence and limits: The authors call the agents' wall models comparable to the standard one, which still matches the reference velocity profile better in the log region. All three are textbook flows on coarse grids, each model fitted to one statistic; the channel case was one 82-hour run. The models largely rebuild known forms, and the spectral score could not tell structurally different subgrid models apart. A preprint, testing the group's own system.
AuthorsYi Liu, Xiantao Fan, Mingqi Xu, Jian-Xun Wang
Google puts Gemini 4 Carbon into Antigravity for Ultra subscribers, opens applications for free access for scientists, and previews agent teams
What happened: Google rolled Gemini 4 Carbon into Google Antigravity for Google AI Ultra subscribers, and scientists and researchers can apply for a chance at six months of free access to both Carbon and Argon at its top usage tier; enterprises get Carbon through Gemini for business. Alongside it, /swarm (preview) splits one prompt among a coordinator, parallel workers in separate git worktrees and adversarial verifiers that run tests and challenge each change, all posting to a shared board. Each agent can be pinned to a different model.
Why it matters: When Google announced Carbon on 28 October it went to cyber defenders first, with no date for anyone else; two weeks later scientists can apply for free access to it. Carbon is the model whose agents wrote the proofs in DeepMind's Signal verification, published the same day, and Antigravity is where DeepMind has shown its own agents hunting antimicrobial peptides and redesigning a drug. /swarm turns many agents with separate checkers into a preview feature in the Antigravity app and command line.
Evidence and limits: This is Google's own product post. The CursorBench 4.0 and WANDR rankings it cites come with no figures there, and /swarm's case rests on five weeks of Google's internal use, with no numbers. Google warns that swarms use up quota much faster than a single chat. Free access for scientists is by application, with no number of places given in the post.
Google DeepMind's AlphaProof Code proved key security properties of libsignal's compiled code, in about 6.5 million lines of Lean
What happened: Google DeepMind's AlphaProof Code, agents that wrote proofs on Gemini 4 Carbon, proved forward secrecy, authentication, post-compromise security and handshake secrecy for libsignal, Signal's reference implementation, in its compiled ARM machine code. The trusted part people must check is about 26,000 lines of specification and machine semantics; the agents wrote about 6.5 million lines of proof, each checked by Lean. A team of five did it in four weeks. With two old, fixed bugs put back, the system proved both builds insecure.
Why it matters: Formal proofs of real software have cost years of expert work: the seL4 kernel took twenty person-years, and it was written to be verified. If agents write the proofs and people check only what is being proved, the cost moves to writing the specification, which changes who can afford verified code. DeepMind says the agents sent only 1% of the code back for a person to look at again; DeepMind names medical-device firmware and power-grid software. The proofs are in Lean, the language mathematicians use to check proofs by machine, and DeepMind has opened them.
Evidence and limits: This is DeepMind's own paper, not peer reviewed. The symbolic proof treats the cryptography as perfect, with two honest parties, bounded inputs and no side channels; the computational one rests on the primitives' hardness and does not cover quantum attackers. Both cover DeepMind's own command-line build, not Signal's apps, and rest on 26,000 trusted lines and on Lean. Signal's part is its CTO's endorsement and feedback on the method, not an independent audit.
Sahir Raoof, formerly of SignalFire, raises a $27.5 million debut fund, Proximal Ventures, for start-ups bringing AI to the tools and services behind drug development
What happened: Proximal Ventures will write pre-seed and seed cheques of $500,000 to $1.5 million and has already backed seven start-ups, TechCrunch reported on 10 November. Its backers include the venture firms SignalFire and Kleiner Perkins, Bertelsmann Healthcare Investments, Nan Li and Zavain Dar, co-founders of Dimension, and Amy Abernethy, a former principal deputy commissioner of the FDA. Raoof spent two years as a principal at SignalFire and earlier led business development at Iterative Health and partnerships at Flatiron Health.
Why it matters: The fund bets on the service side of drug development. Drugmakers have long paid contract firms to make thousands of compounds, work out how to manufacture a drug at scale and analyse trial data, and Raoof argues AI can deliver that same work cheaper and faster, as it has begun to in legal and financial services. So the buyer he is counting on is pharma's outsourcing budget, which pays for services, rather than software, which pharma has not been thought a good buyer of.
Evidence and limits: Fund size, cheque size and strategy are as TechCrunch reported them, with Raoof quoted; no announcement or filing from the fund was read. Of the seven start-ups only ChronicleBio is named, and the report gives no customer or result for any of them.
Lila Sciences' open-endedness team releases four preprints on getting AI models to generate new ideas and build on them
What happened: Kenneth Stanley, who leads the team, published an essay naming four things an open-ended discovery system needs: generating ideas, judging which are interesting, an internal representation suited to insight, and managing many research lines. Four arXiv preprints with Joel Lehman, Alon Albalak, David B. D'Ambrosio and Thomas X. Sheppard cover two of them. In one, small models made to forget a topic propose more novel ideas; in another, a training add-on cuts the steps a network needs to learn a modular arithmetic task over twelvefold in one setting.
Why it matters: Lila describes its goal as scientific superintelligence, and its automated lab is meant to be fed by machine-generated hypotheses. These papers show what the team Stanley leads at Lila is working on for that hypothesis side. They argue that scaling models up will not by itself make them creative, which runs against the common view that scaling will fix it.
Evidence and limits: Method papers on language models and small networks: 1-billion and 7-billion-parameter models, an image task and modular arithmetic. None involves a laboratory experiment or a scientific result, and none is peer reviewed.
AuthorsKenneth O. Stanley, Joel Lehman, Alon Albalak, David B. D'Ambrosio
IBM and Mitsubishi Chemical train a circuit-writing transformer on a 70-qubit IBM processor and run it on models of tin photoresists for EUV lithography
What happened: Researchers at IBM Japan, Mitsubishi Chemical, Keio and NVIDIA trained transformers to write circuits for sample-based quantum diagonalization, on a simulator, then on IBM's Heron processor ibm_kobe at 46 and 70 qubits. Applied without retraining to methyltin trihydroxide and its hydrates, minimal models of tin photoresists for extreme-ultraviolet lithography, 27 of 28 runs at 46 to 70 qubits came within 15 mHa of a classical DMRG reference, median 4.5. At 56 to 70 qubits the chosen circuits used 34 to 72 times fewer two-qubit gates than an LUCJ circuit.
Why it matters: Tin compounds are a leading candidate for the resists that EUV lithography patterns chips with, and how their tin-carbon bond breaks under light needs multireference chemistry, the strongly correlated kind quantum computers are meant to help with. A chemicals maker co-authoring hardware runs on these molecules shows where industrial quantum chemistry is pointed. The method's point is that the processor itself gives the training signal at sizes where exact classical simulation is out of reach, from about 46 qubits.
Evidence and limits: The reference is a classical DMRG calculation, so the runs match classical answers rather than beat them, and no resist was made or measured. Each circuit was sampled once and each setting trained with one seed; the authors call the evidence of transfer to new chemistry qualitative. The gate savings are for the best of ten circuits, at about 11 times the processor time per job.
Tushen Zhihe, an AI protein-design company, raises several hundred million yuan over two rounds, its Pre-A led by Shanghai Kechuang Group
What happened: Tushen Zhihe, founded in Wuxi in late 2023 by Yuguang Wang, an associate professor at Shanghai Jiao Tong University's Institute of Natural Sciences, closed two rounds totalling several hundred million yuan. Shanghai Kechuang Group led the Pre-A, joined by Zhongsheng Yinling, Zhoushan Tuyi and others; Jinding Capital, MegaRobo, Zero2IPO Holdings and others invested in the Pre-A+. The money is to take several of its own programmes toward IND filings, upgrade its ToursynClaw design platform and hire. Its previous round, in September 2025, was an angel+ of around ten million yuan.
Why it matters: The money goes first to the company's own drug candidates, and its own candidates are where an AI protein-design company has to show that its designs work as medicines. MegaRobo, a maker of lab automation, invested to link its automated labs to the company's design-and-test loop, and the Pre-A came from Shanghai's state-backed venture group, which calls AI for science in biomedicine one of its main investment directions.
Evidence and limits: The results are the company's own, in the round's announcement: more than 100 antibody-engineering projects checked in real experiments, one design-and-test round in 14 days, more than 5,000 experimental data points. No paper or partner-reported result was found, its programmes and targets are not named, and neither the split between the two rounds nor a valuation is disclosed.
Google adds two TPU systems to its eighth generation: a dual-chip TPU 8i Max and a TPU 8tx with hardware confidential computing
What happened: Google announced TPU 8i Max, which puts two TPU 8i chips on one package to double compute, on-chip memory and memory bandwidth in the same rack space, and TPU 8tx, which pairs TPU 8t silicon with AMD EPYC hosts so data and model weights stay protected and isolated even while in use. With the training TPU 8t and inference TPU 8i from April, that makes four TPU systems in a year. Amin Vahdat, Google's chief technologist for AI and infrastructure, says TPUs now have nearly 100 customers, universities among them.
Why it matters: Neither system is a new chip: 8i Max packages two existing 8i chips together, and 8tx reuses 8t silicon. Google says that is the point, since a new system takes months and a new chip years, so new configurations can follow what customers ask for. Two inputs change. What can be bought: four systems in a year, one for fine-tuning on private data that stays protected while the chip works on it. Chip design: Google says AI now works on its own hardware, AlphaChip in TPU design and Gemini producing layouts in hours rather than weeks.
Evidence and limits: Every figure is Google's own, measured against its other TPUs, with no independent benchmark. The post gives no absolute specs, prices or availability dates for the two new systems, names no science customer, and shows no example of an AI-made layout.
Anthropic released Claude Fable 5.2 with two physics preprints in which, it says, a Claude model solved two open problems about black holes
What happened: Anthropic released Claude Fable 5.2, two days after posting two Hexagon preprints in which a Claude model works out open problems in holographic gravity. One finishes a calculation set out in 2003: at weak coupling, the transition that corresponds to black-hole formation in gravity stays abrupt. The other pins a limit used to rule out candidate theories of quantum gravity at about 9.1266, where it was known only to lie between 8.503 and 12. It costs $8 and $40 per million input and output tokens, down from $10 and $50.
Why it matters: Anthropic says the first calculation sat undone for two decades because it was too laborious; the second gives an exact value where the field had a range. Both went on Hexagon, a new mathematics repository that accepts AI-written papers, labelled primarily AI-generated and contributed by Anthropic. For buyers, the release cuts the price by a fifth. Anthropic says its biology safeguards block about 40% fewer requests than Fable 5.1's in early testing, though most research-biology tasks still go to an older model, and the unrestricted Mythos goes only to vetted organisations.
Evidence and limits: Both are Anthropic's own preprints, unreviewed, with no named author. One says a physicist checked every main result; the other, that a physicist checked it against known special cases and a second Claude computation agreed. Both credit an internal Claude research model, not Fable 5.2, and one says the public Opus 5.5 later reached the same value. The launch's other science claims, such as 28 of 34 virus shells modelled accurately, are Anthropic's own tests.
Anthropic had Claude design three electronic devices, and Diode built the boards from its files with no round of changes
What happened: Claude Fable 5.2 was asked, with the board maker Diode Computers, to design a USB speakerphone, a voice-assistant hub and a handheld virtual pet. From short prompts it wrote the specifications and schematics, laid out and routed each board, modelled a 3D-printable case and wrote the firmware. Diode built the boards straight from the files, and all three ran with no fix to their circuitry. It scored 69% on Diode's in-house PCB Gym, against 25% for Fable 5, and 83.5% on the public BenchCAD 2.1, nearly double Fable 5.1.
Why it matters: Models that touch hardware usually stop at chip design or a benchmark score. Here the output is a file a factory built from, and the test is whether the boards worked. For a lab that needs a custom sensor board or instrument controller, layout, routing and the manufacturing checks are the weeks of an electrical engineer's time that stand between a design and a working device; Diode, which does this work for a living, says it did not have to go through the usual rounds of adjustments with the files Claude produced.
Evidence and limits: This is Anthropic's own post with Diode, a commercial partner, and comes with no paper or dataset; PCB Gym is Diode's own unpublished benchmark, and both scores are compared only with earlier Claude models. All three are prototypes: no electrical validation, antenna tuning, thermal testing or certification, and none designed for volume production. Diode's engineers built and tested everything, and the prompting was minimal rather than absent.
Pierre Fabre signs XtalPi to design small-molecule drug candidates across several cancer programmes
What happened: Laboratoires Pierre Fabre, a drugmaker based in France, will have XtalPi design and deliver new small molecules against undisclosed targets with its AI and robotics platform, in several research programmes in oncology and possibly dermatology and rare diseases. Pierre Fabre picks which compounds advance and runs clinical and commercial development. XtalPi gets an upfront payment, milestones and a downstream share; no figures were given. Francesco Hofmann leads Pierre Fabre's pharma R&D; Raymond Zheng leads XtalPi's integrated drug discovery.
Why it matters: Pierre Fabre bought outside AI design earlier this year too: in January it signed Iktos for one oncology target. A drugmaker that keeps the disease biology and the clinic in house, and buys the molecules from outside on several programmes at once, is paying AI and robotic chemistry for the step between a target and a candidate. The release says the programmes go after targets that have been hard to drug.
Evidence and limits: The targets, the number of programmes and every payment are undisclosed, and no compound or result is reported yet. It shows what one buyer pays for, not that the design works.
argenx agrees to buy Unnatural Products, which designs ring-shaped peptide drugs with machine learning and parallel chemistry, for $550 million upfront and up to about $1.15 billion
What happened: argenx, which already held a minority stake, will buy the rest of Unnatural Products (UNP) for $550 million in cash and up to $595 million more on development and regulatory milestones. UNP designs macrocyclic peptides, ring-shaped molecules meant to reach hard targets and be taken as pills, by pairing parallel synthesis and screening with machine learning. Cameron Pye and Joshua Schwochert founded it in 2017 out of R. Scott Lokey's lab at UC Santa Cruz. The team stays in Santa Cruz; closing is expected by year end.
Why it matters: argenx calls what it is buying a peptide discovery engine: in March UNP's own pipeline was still described as nearing the clinic, and the company had signed discovery deals with argenx, Novartis, Merck and BridgeBio. For investors it is an exit for a company that pairs parallel chemistry with models, eight months after its $45 million Series B. For argenx, an antibody company, CEO Karen Massey presents it as a route to oral and long-acting medicines.
Evidence and limits: The release shows no data on any UNP molecule and does not say how much of the platform's output comes from its models rather than its parallel chemistry. The $595 million is contingent, and the programmes and milestones it is tied to are not disclosed.
Shanghai Jiao Tong and NTU researchers test eleven AI models on alloy lab reasoning: calculations from given data score above 92%, estimates of a material's state 23% to 31%
What happened: Xuyu Dong and colleagues at Shanghai Jiao Tong University and Nanyang Technological University built MatScope from published aluminium and high-entropy alloy experiments: 200 tasks, an agent test, and 20 multistage cases where measurements arrive one by one. Across eleven deployments, including GPT-6 Astra, Claude Fable 5.1 and Opus 5.5, calculations from given inputs scored 92% to 99% and predictions for a specific material state 23% to 31%, with estimates typically off by a factor of 1.6 to 2.1. Only 13.5% to 19% of tasks were fully solved.
Why it matters: An agent that plans the next heat treatment has to say how much a grain size or a strength will change, not only explain the mechanism. MatScope separates the two and finds models weakest at comparing nearby states and choosing under numerical constraints. New measurements helped unevenly: of the forecasts that started in the top score band, Astra kept 4 of 7 to the end and Fable 1 of 3. Two of the corresponding authors are at MATARIA, a materials-model startup: Zheng Liu founded it and Bo An leads training of its materials model.
Evidence and limits: The tasks cover mainly two alloy families and were built by the authors; explanations were graded by another model, Gemini 3.1 Flash Lite, and numbers by code. The agent test allowed three searches and eight actions per question; on 20 failed cases Astra gained 1.6 points with an interval spanning zero and Fable 4.2. None of the model makers is among the authors.
Carnegie Mellon's AI agents propose a 13-parameter density functional that lowers ωB97M-V's held-out error by 11.7% in self-consistent calculations
What happened: Achuth Chandrasekhar and Amir Barati Farimani connected teams of AI agents to a GPU service that fit each proposed formula on training reactions, scored it on validation ones and ran four physical and numerical checks. The formulas are exchange-correlation functionals, the approximate term in density functional theory, chemistry's standard method for molecular energies. Their pick, O159, run self-consistently, cut ωB97M-V's error on 1,152 held-out MGCDB84 reactions by 11.7%. Claude Opus 5.5 agents first passed the screening gate on about 40 times fewer tokens than GPT-6 Astra agents.
Why it matters: Exchange-correlation functionals decide how far DFT can be trusted in chemistry and materials, and improving one has meant long manual or evolutionary searches. Here agents proposed the formulas, the evaluator fit them, and the agents' own controls turned the new term into a design rule: scale down the whole same-spin factor, by 15% to 30% at its peak, at intermediate density. Against SyFES evolution at about 400 fits each, the agents' forms reached lower validation error by adding an explicit density scale that SyFES's representation lacks.
Evidence and limits: GAS22, from a 2022 evolutionary search, still has the lower held-out error (3.59 against 3.81 kcal/mol). On 36 outside GSCDB137 datasets O159 matches ωB97M-V (1.139 against 1.144, not significant). The agents' margin over SyFES is close to the spread of repeated fits, and within the representation the two share neither beat GAS22. The paper does not explain why the window helps heavy-atom energies.
Pushmeet Kohli says Google DeepMind is seeing a phase transition in the pace of scientific discovery, and that its releases over the next few weeks will give a glimpse of it
What happened: On 8 November Kohli posted that across Google DeepMind's projects "we are witnessing a genuine phase transition in the pace of scientific discovery", and that its releases over the next few weeks would give a glimpse of it. He leads the lab's AI for Science work and is chief scientist at Google Cloud. A second post the same day said that as the acceleration unfolds, the focus will be on helping the wider scientific community use these tools responsibly.
Why it matters: This is not a general claim that AI will speed up science. It is a judgment on his own lab's work by the person who runs its science, with specific releases staked on a few weeks. What to watch is whether it reaches the research workflow rather than the model: how much faster a hypothesis is proposed, an experiment designed, a result checked. Only a discovery cycle that is measurably shorter would carry the word. His own hedge is in the sentence: the releases give a glimpse.
Evidence and limits: Kohli names no release, no result and no number, and does not say what would count as a phase transition or against what earlier pace it is measured. The releases he points to had not appeared when this was recorded, so there is nothing here to check yet.
Qingxi Zhiyan, a Beijing start-up putting AI into materials instruments, raises nearly 100M yuan in an angel round backed by GL Ventures and AgiBot
What happened: Qingxi Zhiyan, founded in Beijing in May 2026 by Siming Dong, a Stanford materials science PhD with more than 15 years in semiconductor R&D, with Zhenlin Zhao as COO and Huan Wang as chief scientist, announced an angel round of nearly 100 million yuan from Zhongxin Juyuan, GL Ventures, AgiBot, Fengwu Capital, Orient Jiafu and industry funds, with no lead named. Dingfeng Kechuang invested in July. It builds models that read materials measurements such as X-ray diffraction, puts them into instruments, and takes contract materials R&D.
Why it matters: Two backers work in industries the company aims to sell materials to. AgiBot makes humanoid robots, and the company says it is developing permanent-magnet materials for robot joint motors with robot makers; Zhongxin Juyuan invests in semiconductors, where the company lists photoresist and memory materials among its programmes. Its route to training data runs through the instruments: AI on characterisation and process equipment, so that each measurement it runs is stored in a form a model can learn from.
Evidence and limits: Every figure is the company's own, given to Z Potentials and Wallstreetcn: contracts worth tens of millions of yuan in six months, tens of millions of experimental records, and a rare-earth magnet about 60% cheaper and 30% more heat-resistant than the previous generation, developed in five months, with mass production targeted for 2027. No paper or outside test was found, and Wallstreetcn notes the magnet figures have no independent source. Valuation not disclosed.
Lawrence Livermore's new atomic force model for fusion wall alloys comes close to the accuracy of models ten times its size and runs tens of times faster
What happened: Flynn Walsh and colleagues fit a machine-learned interatomic potential, the function giving each atom's force so a simulation skips the quantum calculation, for the tungsten-tantalum-hafnium-chromium-vanadium-titanium alloys with helium that are candidate fusion wall materials. Rather than share one set of distance functions across the expansion, it learns one per term, which buys accuracy for almost no compute. On the HEA25S benchmark: 4.8 meV per atom and 112.9 meV per angstrom, against 4.1 and 102.6 for a fine-tuned foundation model, at 0.36 against 10 microseconds per atom-step on one GPU.
Why it matters: How far a simulation of radiation damage can go is set by cost per atom per step: the defects that degrade a reactor wall develop across millions of atoms and nanoseconds, and the accurate models have been too slow to follow them. This says the choice between a fast model and an accurate one is not fixed by size. As a demonstration the authors drive 2 MeV collision cascades through a billion atoms of pure tungsten and of a tungsten-tantalum-vanadium-titanium alloy, for 40 picoseconds and hundreds of thousands of timesteps.
Evidence and limits: Everything here is computed: errors against DFT calculations, speeds from running code, nothing made or measured in a lab. On their own set the model reaches GRACE 1L's energy error while the larger models stay slightly better on forces. The HEA25S comparison figures are the GRACE paper's, and this paper's own table note says that validation set is less strictly separated; the cost there was measured on 2,000 atoms, too small to fill the GPU. An arXiv preprint, not peer-reviewed.
University of Delaware's two-agent setup rewrites chemistry codes and cuts the time PySCF spends on the excited-state step by 84 to 94%
What happened: Mathew Chow and Tao E. Li had one agent (OpenAI's GPT-5.6-sol) propose code changes on training cases while a second accepted or rejected each on separate test cases, by speed and accuracy. On ten other molecules the excited-state (TDDFT) step in PySCF took 84-94% less time, with nearly identical energies; a separate job cut the ground-state calculation 5-28%. LAMMPS water simulations took 19-22% less time with identical trajectories, and a serial molecular dynamics code became CPU and GPU versions two to three orders of magnitude faster.
Why it matters: Computational chemists often write their own code, and the first version is usually serial, as the authors note; speeding it up is skilled work that takes time from the science. Here each change had to pass a second agent on cases the first had not seen, and outputs were compared with the original code. In TDDFT the gain was in the method, not only low-level tuning: on one molecule the solver needed 9 iterations instead of 45. Faster codes mean more simulation for the same compute.
Evidence and limits: One lab's preprint, scored by its authors. Repeat runs of the same model ended at different scores; the LAMMPS gains did not hold in pressure-controlled simulations, where one part got slower; the GPU code ignored the request for mixed precision; and the agent-built SCF strategy did not reliably reach the lowest energy or the shortest time. The paper does not say any change was merged into LAMMPS or PySCF. OpenAI supported the work through its academic researcher programme.
Cancer Grand Challenges funds Harvard's AURORA AI co-scientist, and three of its teams start testing it in their own research
What happened: Cancer Grand Challenges, founded in 2020 by Cancer Research UK and the US National Cancer Institute, gave Marinka Zitnik's group at Harvard Medical School nearly $7M to build AURORA, an agent that reads papers and databases, calls tools such as AlphaFold and proposes hypotheses and experiments. Three teams get about $1.3M each to use it: PROMINENT on why some mutated cells turn cancerous, NexTGen on CAR-T for children's solid tumours, MATCHMAKERS on how T cells recognise cancer.
Why it matters: A research funder is paying outside teams to put an AI co-scientist into research already under way, and the evaluation is part of the grant: as the teams test its hypotheses in the lab, Harvard will compare the system's reasoning traces with their results, which would show how often its suggestions hold up. Harvard says it is the first Cancer Grand Challenges project on AI and the first to span several funded teams.
Evidence and limits: Nothing has been tested yet: the project launched on 5 November, and neither release gives an end date, a measure of success or a plan to publish the comparison. Neither names the language models AURORA runs on. Harvard gives the money as nearly $7M plus $4M for the teams; Cancer Research UK puts the project at £8m.
Kos Biotechnology Partners joins Fathom Therapeutics' Series A, which the company now puts at $55M, up from the $47M it announced in April
What happened: Kos, a life sciences investor based in New York and Athens, said it invested through its first fund, Kos BioVentures, in Fathom's oversubscribed $55M Series A led by Sutter Hill Ventures; co-founders Simos Simeonidis and Alex Tzoukas spoke for the firm. Fathom's own release of 29 October gives the same $55M, dates the round to May and names Lilly Ventures among the participants. It also named Bret Williams, from Magnet Biomedicine and before that Relay Therapeutics, as chief scientific officer.
Why it matters: Since April the stated size of the round has grown by $8M, and the October release names Lilly Ventures, the venture arm of a large drugmaker, among its backers. Fathom says it is now turning molecules designed by its protein-motion simulations into drug programmes, and hired a chief scientific officer who has led target and drug discovery to do it.
Evidence and limits: Neither release says when the extra $8M came in or who put it in, or whether Lilly Ventures was among the unnamed investors in April. The claim that the platform already yields potent, selective candidates against targets once thought undruggable comes with no data.
Cornell physicists turn a pretrained atomic model into a crystal designer that changes elements and structure together, and 39 of 45 new superconductor designs picked for full checks compute above 10 K
What happened: Cristóbal Méndez, Justin Tahmassebpur and Tomás A. Arias at Cornell University let a pretrained EquiformerV3 model read blends of elements and partly filled sites, so one gradient search changes a crystal's elements, atom count, positions and cell together. Of 45 new superconductor designs picked for high predicted critical temperature, phonon calculations put 39 above 10 K, the best ZrTiNb2CN3 at 24.0 K; H4LiPd3 (23.2 K) computes as stable. Ten solar-absorber designs landed within 20% of the band gap each was aimed at and within 70 meV per atom of stable.
Why it matters: Generative models and screening propose crystals in bulk and score them afterwards, and a score does not say what to change. Here the choice of elements becomes something a gradient can push cheaply, using weights already trained on public data; the authors say a new target needs only a fine-tuned property model. Two findings matter to anyone running such a pipeline: designs inherit the band-gap error of the calculation method the model learned from unless the target corrects for it, and cheap finite-temperature checks with machine-learned potentials change which candidates look worth making.
Evidence and limits: Every check is the authors' own calculation; nothing was made or measured. The critical temperatures are simplified estimates with the electron repulsion fixed at 0.10, and the four leading carbonitrides fall 56-87 meV per atom short of stable. After the empirical correction, the five designs aimed at the raw computed gap would sit at 1.49-2.17 eV, above the target. Only 10 of the 45 superconductor designs allowed the finite-temperature check. Code and structures come on publication; a preprint.
AuthorsCristóbal Méndez, Justin Tahmassebpur, Tomás A. Arias
South China University of Technology trains one model on 78,889 glass compositions, failures included, to predict which form glass across oxide, chalcogenide, halide and metallic glasses
What happened: Yao Ji and Qinyuan Zhang of South China University of Technology read 78,889 compositions off published glass-forming-region diagrams, 62,133 of them ones that did not form glass, and trained an XGBoost classifier on element-property statistics. In ten-fold cross-validation it was right 89.6% of the time, against 78.8% for always guessing no glass. On two lead-bearing systems kept out of training, it recovered 74.2% and 84.2% of the compositions reported to form glass.
Why it matters: The authors say earlier glass models mostly covered one family or learned mainly from compositions that did form glass; this one also learns from those that did not, and spans four families with one set of element descriptors. It needs no simulation per candidate, so the authors offer it for ranking candidate compositions for functional glasses before anyone melts them. Its explanation step suggests that mixing in more varied elements helps glass form only up to a limit each family sets, and that the limit is tighter in less rigid networks such as halides than in oxides.
Evidence and limits: Every test is the authors' own, against literature records; no new glass was made, and the four-component check reuses their earlier experiments. Labels ignore cooling rate: one report of glass makes a composition glass. With whole chemical systems held out, the ranking score (AUC; 0.5 is chance) fell from 0.950 to 0.904. On 42 metallic alloys the all-family model was right 60.0% of the time, the metallic-only model 81.0%. The authors call the mechanism findings hypotheses.
Third Rock Ventures signs Funnel Biosciences to design antibodies with AI against targets Third Rock picks, for up to about $2.16 billion
What happened: Funnel Biosciences, a Beijing company founded in February 2026, will use its epitope-guided AI antibody platform to discover biologics, antibodies included, against several targets Third Rock Ventures nominates. Funnel gets an undisclosed platform access fee and up to about $2.16 billion in total if conditions in the agreement are met. The platform starts from a chosen spot on the target, generates antibodies with AI and refines the binding surface with physics-based energy functions. Its co-founders are Mingchen Chen, a computational biophysicist, and Jintan Liu, formerly at MSD.
Why it matters: Third Rock usually builds biotech companies itself; here a venture firm is buying antibody design from a Chinese start-up founded this February, target by target, the way a drugmaker would. pharmaphorum counts it as Third Rock's second publicly disclosed deal of this kind, after a peptide programme with Shanghai's Perpetual Medicine. The release aims the platform at hard targets such as proteins that cross the cell membrane several times, where animal immunisation and phage display struggle to keep the target's natural shape.
Evidence and limits: The $2.16 billion is a ceiling on conditional payments; the access fee, the targets and their number are not disclosed. Zero-shot picomolar binding without lab affinity maturation, good developability and low immunogenicity are the company's own description, with no data shown.
OpenAI and Ligo's BioBasicsBench scores AI agents on 100 tasks built from real wet-lab data, and GPT-6 Astra earns 66.7% of the points
What happened: OpenAI and Ligo, an enzyme-design startup, released BioBasicsBench: Ligo scientists turned data from their own experiments into 100 tasks with scoring rubrics (the examples shown include a gel, colony plates and a handwritten plate label), each reviewed by another Ligo scientist, an outside PhD-level wet-lab scientist and a final internal check. At xhigh, the highest setting run on all seven models, GPT-6 Astra earned 66.7% of available points, GPT-6.1 Sol 62.0% and GPT-5.6 Sol 43.8%; 37 model and setting combinations were run, five attempts per task.
Why it matters: Reading which colonies carry the right plasmid, or whether a control plate is valid, is routine work an agent beside a biologist would do, and tasks taken from a lab's own runs keep the ambiguous images that figures chosen for a paper leave out. The failures shown are of that kind: Astra sorted colony PCR lanes into the right groups, then took the larger product as success and picked colonies still carrying the old plasmid.
Evidence and limits: OpenAI built it with Ligo, scored it and compared only its own models; OpenAI says higher scores do not yet mean models can make reliable judgments in the lab. No release of the tasks is reported, and it tests reading data, not doing bench work.
George Church resigns from Radical Numerics after its AI-designed phage announcement did not name Tabula, the partner that ran the lab work
What happened: The New York Times reports that Radical Numerics has acknowledged it did not work alone on the 20 October announcement: Tabula carried out much of the research, including making the phage cocktail the patient received. Eric Nguyen, its chief executive, wrote to the paper that the two worked closely for seven months and that the contract required both to keep the collaboration secret. Tabula says it led the phage therapy. Both agree Radical designed some of the cocktail's phages and Tabula supplied others.
Why it matters: Radical Numerics did not name Tabula when it announced the treatment on 20 October, and Tabula did much of the lab work. So who made an AI-designed therapy is contested by the two companies that made it. George Church held roles at both; he left Radical's on 5 November and told Nguyen that the argument from confidentiality was very weak on common sense and basic ethics. The Times sets it beside two autumn disputes over whether AI companies used other researchers' work, on an OpenAI proof and on Anthropic's enzymes; both companies denied it.
Evidence and limits: The Times's report was read through a licensed translation; nytimes.com refuses automated readers here. It says Church left the board; where he sat is the one thing the accounts differ on. Neither company has published the contract, and the Times writes that the research was done and the treatment given. Tabula says it will give its own account later; the patient's outcome is still undisclosed.
Anthropic's economics team sorts 207,000 research sessions on Claude: scientists use it most to analyse data and write, least to gather data and form ideas
What happened: A language model labelled 207,049 research sessions on Claude's consumer plans, from 2 August to 30 September 2026, by field and task. Data analysis took 33.7% of Claude sessions and 48.1% on Claude Code, against an estimated 27.6% of researchers' time; data gathering about 12% against 28.6%; ideas 4.2% against 7.1%. In a survey of 528 scientists, mostly in the life and health sciences, those two lagging tasks ranked as the most important. Johannes Hermle, Ben Jones and Lukas Althoff led it.
Why it matters: Like Google's study in September, it puts the gain in work done at a screen, not at the bench. Physical steps take 38-41% of researcher time in chemistry, materials and the biomedical sciences, against 4% in the median field, so the report's model caps their output at three to four times what it would be without AI even if AI did all other work, against up to 15 times in computer science and statistics. On Claude Science, the version wired to research databases, idea generation takes about twice its share on Claude.
Evidence and limits: Only consumer-plan sessions count. Models, not people, assigned each session's field and task, estimated researchers' time shares, and scored each task's AI capability by judging its steps in pairs; that score correlates 0.78 with a reinforcement-learning suitability index also rated by models. The survey reached active AI users. The speed-ups are model scenarios that count research output, not the quality of ideas, in Anthropic's report on its own products.
Discovered Materials offers up to $1M in prizes for five unsolved problems in materials physics and chemistry
What happened: The startup, which uses AI agents to look for chip materials, posted five problems written with outside physicists and chemists, Stanford's Felipe Jornada among them: a predictive theory of cuprate superconductors (up to $400,000); cheap but accurate reaction energies, and how chip interfaces get their electrical properties from processing ($200,000 each); what a laser pulse does to a material, and how heat crosses solid interfaces ($100,000 each). Four are scored on published results; the chemistry one on a hidden computed benchmark.
Why it matters: Mathematics has prize problems that show whether AI is closing in; TechCrunch notes materials science has no such set. These come with fixed numbers to hit, from superconductors to the interfaces inside chips. Co-founder Advaith Sridhar told TechCrunch the aim is goalposts that show whether models are improving in this field. On the company's own benchmark, models proposed over 500 new materials but a plausible way to make only one.
Evidence and limits: That frontier models given days on each problem made little headway is the company's account, with no runs published. The company alone judges a solution, may solve a problem itself and close it unpaid, and takes a perpetual, non-exclusive commercial licence to every submission, winning or not.
Google Research's model reads a Pixel Watch's pulse and motion signals to flag hypertension, finding 66% of cases in a 196-person study checked against 24-hour cuff monitoring
What happened: Paolo Di Achille, Ming-Zher Poh and Google Research colleagues, with Stephen Juraschek of Beth Israel Deaconess, pretrained a network on about 492,000 hours of unlabelled watch data, then trained a classifier on 663 people with 24-hour blood pressure monitoring. In a prospective study at five US sites, 196 adults with no known heart disease or blood pressure drugs wore a Pixel Watch 3 for a week. Against the 24-hour monitor it found 65.8% of those with hypertension at 90.0% specificity; office cuff readings at the first visit found 55.3%.
Why it matters: More than half of adults with hypertension do not know they have it, and a watch many already wear could tell them to get checked. The model needs no cuff to calibrate it, and the study scores it against the 24-hour monitor that guidelines treat as the reference, where earlier wearable studies often used home cuff readings. The authors frame it as a prompt toward a cuff-based diagnosis, not a diagnosis.
Evidence and limits: One study of 196 people, funded by Alphabet, with every author but Juraschek a Google employee. It used one watch, left out people already treated or with heart disease, and missed a third of cases. The authors say randomised trials are still needed to show it finds people who would otherwise go undiagnosed and improves outcomes.
Potential Sciences, founded by Hatem Helal in Cambridge, comes out of stealth with an eight-figure seed backed by the UK's Sovereign AI fund and Twin Path Ventures
What happened: Founded in October 2025 by Hatem Helal, the company says it rebuilt gold-standard quantum chemistry, CCSD(T), to run on GPUs at a hundredth of the usual cost, and uses it to compute training data for models of how atoms interact, starting with small-molecule drugs. The UK government says its Sovereign AI fund invested and adds time on national supercomputers, Isambard-AI among them; Joséphine Kant spoke for the fund. Twin Path Ventures, Nauta Capital, Odyssey and four other firms joined, with angels Patrick Collison and Chris Gibson. No amount was given.
Why it matters: Models that predict how a drug binds learn from fewer than 300,000 measured protein structures, and the company bets that accurate computed data can add to them, including the failed cases labs rarely publish. Meta's OMol25 dataset had to use a cheaper, less accurate method to reach its scale, so cost at CCSD(T) accuracy is the bottleneck the company says it has removed. The fund pays in compute as well as cash; Sky News reported the talks in September.
Evidence and limits: The hundredfold saving and the claim to beat published results on reaction barriers and binding affinity are the company's own, with results promised in a later release. 'Eight-figure' is all the sources give on size, and the company counts compute in it; the fund does not disclose amounts. Pharmaphorum reports Sovereign AI and Twin Path co-led; the company calls them key investors.
Chai Discovery signs Roche to use its protein design models in a number of Roche's drug discovery programmes
What happened: Chai Discovery announced a collaboration under which Roche will use its AI models and platform in a number of drug discovery programmes across Roche's therapeutic areas. Joshua Meier, Chai's co-founder and chief executive, and Boris L. Zaïtra, Roche's head of corporate business development, spoke for the two companies. The release names no targets, no number of programmes and no financial terms. It says Chai-3, the latest model, reaches high target success rates and binding affinities, with no figures.
Why it matters: Roche is the seventh drug company this year to sign for access to Chai's models, after Lilly, Pfizer, Novartis, argenx, Bristol Myers Squibb and GSK. Like the others, the release describes use of Chai's general design models across many programmes rather than molecules for named targets, so another large maker of antibodies and other protein drugs is taking in outside design models beside its own research.
Evidence and limits: Announced without terms, targets or numbers: it shows Roche taking up Chai's models, not that Chai's designs work in Roche's programmes. Chai-3's success rates are the company's own description.
Ginkgo Bioworks reports its first sale of an autonomous lab to a large drugmaker, Novo, at Novo's new Waltham research site
What happened: In its third-quarter results Ginkgo said it sells two things built on one robotic platform: lab systems made of its reconfigurable automation carts and Catalyst software, bought by Novo, Pacific Northwest National Laboratory, the NSF and ARPA-H, and experiments run in Nebula, its own lab of over 100 carts. Novo's lab, announced in September, is the first such sale to a large drugmaker; a subcontract under ARPA-H's GIVE programme, worth up to $17.5 million, puts the carts into RNA-drug manufacturing. Revenue was $28 million, down 6% on a year earlier.
Why it matters: It shows who is paying for labs that AI can drive: a drugmaker buying robots for its own scientists, government programmes building shared cloud labs, and AI companies buying experimental data. Ginkgo says frontier labs such as OpenAI use its cloud lab to generate training data, and that 22 customers signed for its drug-property testing in the quarter, 20 of them new. Jason Kelly, Ginkgo's co-founder and chief executive, said the US government now treats autonomous labs as national infrastructure.
Evidence and limits: The figures and customer counts are Ginkgo's own; the price of Novo's lab and the size of the AI-data work are not given. A sale shows demand for the equipment, not that it speeds up Novo's research.
Astellas signs on to use Iktos's molecule design and synthesis-planning software, Makya and Spaya
What happened: Iktos announced that research teams at Astellas Pharma will use two of its products. Makya generates molecules only from purchasable building blocks and known reactions, so every design comes with a way to make it; Spaya works backwards from a molecule to ranked synthesis routes (retrosynthesis). Yann Gaston-Mathé, Iktos's CEO and co-founder, said most of the chemical space worth exploring has never been made.
Why it matters: This is a software licence rather than a discovery deal: Astellas's own chemists run the tools, while in Iktos's deals with Servier and Pierre Fabre earlier this year Iktos designs and makes the molecules for the partner. The pitch is that a design arrives with a route to make it, which is the step where generated molecules have to meet a real lab.
Evidence and limits: Iktos announced it and Astellas has made no statement of its own. The release gives no money, duration, number of users or targets, so it shows Astellas taking up the tools, not that designs made with them work.
Inductive Bio's models rank first in OpenADMET's blind test of predicting how strongly 750 compounds block four drug-metabolising enzymes
What happened: OpenADMET, an open-science effort on drug-safety prediction, measured how strongly compounds inhibit four cytochrome P450 enzymes, which break down most drugs, and withheld the values for 750 test compounds until the challenge closed on 3 November. In the direct-inhibition track, 271 participants took part and 103 entries made the final board. Inductive's entry ranked first, an unnamed entry close behind sat in the same statistical tier, and NVIDIA's entry came third. Inductive, led by co-founders Josh Haimson and Ben Birnbaum, calls it its fourth OpenADMET win in a row.
Why it matters: A drug that blocks these enzymes can raise other drugs to dangerous levels, so chemists want to flag the risk before a compound is made. Here models were scored on measurements produced for the test and hidden from every team, which is a harder check than a public benchmark. Inductive, whose models are trained on an ADMET dataset pooled from industry partners, ranked first again, in a field of 271 participants.
Evidence and limits: The organiser puts the top two entries in one tier, not statistically distinct, so first place is a rank within a tie. Inductive's own method report says the 750 test compounds are close analogues of 75 training compounds, an easier test than new chemistry. It scores predictions against measured data; no drug programme was run on them.
University of Maryland's PolyGPT, told to make its polymers electrolyte-like, returns under half as many distinct ones that hit a temperature target
What happened: Tanmay Sarkar Akash and Siddhartha Das of the University of Maryland report that PolyGPT, their 38.3-million-parameter model that writes polymer structures on request, repeats itself when asked for electrolyte-like polymers: distinct outputs fell from 99.99% to 51.74% of valid ones. With a 246 K target for glass-transition temperature (where a polymer turns glassy), adding that request trimmed the average miss by a property predictor's estimate from 26.97 to 26.07 K, but distinct polymers within 50 K of the target fell from 82.35% to 32.31% of attempts.
Why it matters: Generative models for materials are usually judged by whether outputs are valid and how close they land to the requested property. Here the setting that scored slightly better on that error left a screening step with well under half as many distinct candidates, and fewer that were absent from the training data (19.71% of attempts against 59.90%). The authors' point for anyone building or comparing such generators: count, per attempt, the distinct and new structures that meet every request. Combined requests also lost control: asking for a chemical group plus electrolyte-likeness cut compliant outputs, on average, from 94.27% to 31.27%.
Evidence and limits: Every property is a predictor's estimate, and some of those predictors also supplied the training labels, so hitting a target means agreeing with them; no polymer was simulated, made or measured, and ion conductivity was not assessed. The electrolyte request means only resembling the lithium-electrolyte polymers in the training data. Results come from one training run, evaluated by the two authors; code and weights are available only on request.
A team from USTC and Hefei's Institute of Artificial Intelligence built DP-Learner, which predicts materials' electronic Hamiltonians with 33-41% less error than the team's earlier NextHAM
What happened: Shi Yin, Zujian Dai, Xinyang Pan and Lixin He built DP-Learner, a Transformer that predicts a material's electronic Hamiltonian, the matrix its electronic bands are computed from, keeping each block in the matrix's own layout instead of converting it to the spherical components most symmetry-aware models use. Against NextHAM, the group's previous model, average error fell from 1.417 to 0.841 meV and from 2.394 to 1.595 meV on two datasets, with about a tenth of the inference time and under a tenth of the GPU memory.
Why it matters: A density functional theory (DFT) calculation gets a material's Hamiltonian by repeating a step whose cost grows with the cube of the system's size; models like this one start from a cheap first guess and predict the converged answer, which is what screening many materials for conductivity or optical behaviour needs. Beyond accuracy, the gain is in cost: one GPU and under 1 GiB per structure, where NextHAM spread its four sub-models over four GPUs. The team also built MP-HAM-110K, 113,800 Materials Project structures over 68 elements, six times NextHAM's dataset, computed with the DFT code ABACUS.
Evidence and limits: Every number is the authors' own, against one baseline, their own NextHAM, on that paper's dataset and the new one, where they retrained it; no other Hamiltonian model, such as the DeepH family, is compared. Timing leaves out building the first guess and is not set against running DFT. The new dataset has no spin-orbit coupling or magnetism. Code, weights and data are offered only through anonymous review links with access codes. The paper is under review at ICLR 2027.
Haixu Wu's TensorWorld raises a $105M seed to build a physics foundation model
What happened: TensorWorld (Beijing Wuyuan Technology), founded in July by Haixu Wu, announced on X a $105M seed from Cathay Capital, HSG, Ant Group, MSA Capital, Meituan, IDG Capital, CDH VGC, Link-X Capital, Fortune Venture Capital and NIO Capital; no lead is named. Wu did his PhD at Tsinghua with Mingsheng Long and a postdoc at MIT CSAIL with Wojciech Matusik. The team named includes Junsheng Zhou (Uni3D), Weirui Ye (EfficientZero) and Wen Zheng.
Why it matters: Wu's research is learning physics from data: Transolver is a neural solver that predicts how physical systems behave on irregular meshes, and GeoPT pretrains such solvers on geometry without labelled simulations. TensorWorld says it will build on that toward a model that predicts how objects move, interact and respond to an action. It is a scientific machine-learning postdoc who founded a company in July and raised a seed this size four months later, from venture firms and from Ant Group, Meituan and NIO's fund.
Evidence and limits: The company has shown no model or result yet. The round was read through Zhidongxi's report of the company's X post, not on X. The $600M valuation is media-reported (Zhangtongshe, The Crane); QbitAI reported in October that the company had closed several rounds, which this announcement does not split.
Anthropic's biology lab builds vault shells from phage and bacterial proteins in E. coli and maps one at 3.2 angstroms
What happened: Vaults are large hollow protein barrels in cells across the eukaryotic tree, and nobody knows what they do. The team searched about 15 billion protein sequences, found relatives of the human vault protein in 29 bacterial and 4 archaeal phyla, and put 24 into E. coli. Seven made vault-like particles under a first microscope and cryo-electron microscopy confirmed four, one read as a ring on that pass. L05, from a phage, builds a shell of 96 copies at 3.2 angstroms, enclosing 25% more volume than the human vault.
Why it matters: This is structural biology out of Anthropic's own biology lab, not a design campaign sent out to be assayed, with the microscopy run at Thermo Fisher's cryo-EM centre and Berkeley's EM lab. The loop is what to read: people chose the proteins and the analysis calls, and Claude searched the sequence databases, wrote each protein into DNA for the host cell, recommended the column and the shortcut of loading ruptured cells straight onto it, and ran the cryo-EM processing. Isolation went from about three days to 90 minutes, a first model 34 hours into processing.
Evidence and limits: What vaults do is still unknown, and the authors say their work deepens the question. The particles were made only in E. coli, not shown to form in the phages and bacteria they come from. A small molecule sits in every wall copy and is unidentified: coenzyme A is a shape match. The report was not yet posted on bioRxiv when Anthropic published, the time savings are the team's own estimates against no baseline, and a provisional patent is filed.
Isomorphic Labs is in talks to raise about $5B at a valuation of about $50B, Bloomberg reports
What happened: People familiar with the matter told Bloomberg that Alphabet's drug-design company is seeking roughly $5B in new money at about $50B. The financing has not closed, no investor is named, and Isomorphic did not respond. Bloomberg's report of 8 October gave at least $40B, possibly as high as $50B, and no amount. Max Jaderberg, its president, said at Bloomberg Tech in London on 2 November that its work is still preclinical; Demis Hassabis, its CEO, said in January that first trials should start by the end of 2026.
Why it matters: At about $5B, the round would be some 2.4 times the $2.1B Series B of May, for a company whose own drugs have not yet been tested in people. Investors would be pricing the drug-design engine and its pharma deals before any clinical result, and the size shows how much private money one AI drug-discovery company may now draw ahead of its first trial.
Evidence and limits: The only source is unnamed people speaking to Bloomberg: Isomorphic has confirmed neither the amount nor the valuation, no investor is named, and the round may change or fall away. Jaderberg's own word is that the work is still preclinical.
Axiom Math turns the whole Stacks Project, the open reference for modern algebraic geometry, into one Lean library of about 16 million lines
What happened: Johan de Jong of Columbia started the Stacks Project in 2005, and hundreds of mathematicians have grown it to about 7,500 pages. Axiom Math's coordinated agents translated all of it into Lean, linked each of its permanent tags to the matching formal declarations, and released the library open source with a side-by-side reader. The company says the work turned up dozens of statements that lacked a hypothesis or did not hold as stated, and records counterexamples to them.
Why it matters: Axiom says modern algebraic geometry has been far harder to formalize than fields with mature Lean libraries: one modern proof leans on long chains of commutative algebra, schemes and cohomology that had no Lean counterpart. With those foundations importable, a person or an agent checking a new result can start from verified definitions instead of rebuilding them, which matters as AI writes more proofs than people can read. Where the group classification a week earlier was one theorem, this is a library meant for reuse.
Evidence and limits: Size, faithfulness and the errors found are Axiom's account: its own agents audited whether each formal statement matches the text, with targeted audits of key definitions and theorems, and no outside review is reported. De Jong asked on 21 October that AI-generated contributions not be sent to the project itself, so the fixes stay in Axiom's repository.
Shanghai Jiao Tong University's AutoDNA has AI agents run a DNA lab of more than 20 instrument types, with people only restocking reagents and servicing hardware
What happened: In Nature, Mingyu Wu, Zhaoguo Wang, Chunhai Fan, Haibo Chen and colleagues describe a lab built around AI agents from the start. From a user's request, the agents plan and run DNA and RNA work (synthesis, transcription, amplification, sequencing) across more than 20 instrument types and over 11,000 steps. In four tests it matched expert-tuned yields for enzymatic DNA synthesis, ran concurrent requests 3.2 times faster, stored a text in DNA and read it back exactly, and found and confirmed a KRAS G13D cancer mutation that clinical testing had detected.
Why it matters: The authors say existing self-driving labs sit mostly in chemistry and materials, chase one goal through a fixed workflow on a few instruments, and use AI for one step. Molecular biology needs many different machines and long chains of steps. If agents can plan and schedule a whole nucleic-acid lab, it can take requests from several users at once, which the authors call science as a service. Milad Abolhasani of North Carolina State University, who refereed the paper, told Nature's news team the system stands out for its autonomy and for how readily it takes in different equipment.
Evidence and limits: All four tests were set and run by the authors in their own lab. The yield comparison is against expert-optimized synthesis; the 3.2-fold gain is for running requests concurrently, not for any one experiment. The mutation test confirms one known mutation, a demonstration rather than a diagnostic study.
A USTC team's open RNA model, structRFM, learns from 21 million sequence-structure pairs and predicts RNA 3D structure on par with AlphaFold3 in CASP tests
What happened: In Nature, Heqin Zhu, Peng Xiong, S. Kevin Zhou and colleagues at the University of Science and Technology of China train a language model on RNA sequences together with how their bases pair, with the pairing taken from several structure predictors at once so no single one's bias dominates. Its 3D version, Zfold, performs comparably to AlphaFold3 on CASP15, CASP16 and RNA-Puzzles. The model also sets new results on 2D structure benchmarks and carries over to newly released RNA families. Dataset, weights and code are open.
Why it matters: What an RNA molecule does follows from how it folds, and far fewer RNA structures have been solved than protein ones, so models trained on sequence alone learn little about folding. Building base pairing into pre-training is one way around that shortage. With the 21 million pairs and the weights released, other groups can retrain the model, test it on their own RNAs or build on it.
Evidence and limits: The results the abstract reports are scores the authors computed on existing benchmarks; it reports no new experiment testing a structure the model predicted. The pairing it learns from is itself predicted, not measured. On par with AlphaFold3 is the paper's summary across three test sets, not a claim of beating it.
UCB signs Cradle to use its protein design platform in selected antibody discovery projects, from first hits to preclinical candidates
What happened: UCB, the Brussels drugmaker, announced a strategic collaboration under which it will apply Cradle's AI protein design platform within selected discovery projects, from hit generation through lead optimisation to preclinical readiness. Results from UCB's experiments feed back into models built for each programme. Alistair Henry, UCB's chief scientific officer, and Stef van Grieken, Cradle's co-founder and chief executive, spoke for the two companies. No financial terms, targets or number of projects were disclosed.
Why it matters: It is Cradle's fourth drug-company platform deal this year, after Bayer in January, Lundbeck in June and Hansa Biopharma in October, and it came the day before Roche signed with Chai Discovery. UCB presents it as one part of its own digital antibody programme, beside its in-house discovery platform and a network of technology and research partners: an outside design model added to the drugmaker's own antibody work, not in place of it.
Evidence and limits: Announced without terms, targets or a count of projects: it shows UCB taking up Cradle's platform, not that Cradle-designed antibodies do better in UCB's programmes.
Speciale Invest leads a $2.7M pre-seed in Venomics AI, an IISc spin-out sequencing Indian venoms and screening their peptides against human targets
What happened: Venomics AI, founded in 2026 by Kartik Sunagar, a professor at the Indian Institute of Science in Bengaluru, and Vinay Prabhu, its chief technology officer, raised $2.7 million. Speciale Invest says it is leading the round; Inc42 and BioSpectrum India call it co-led with South Park Commons, with Operators Studio taking part. The company sequences venoms, measures what their peptides do against human drug targets in its own lab, and keeps the results as the Venom Atlas, which models built on ESM and AlphaFold are trained on to rank candidates.
Why it matters: It is a bet on data that is hard to get rather than on a bigger model: Speciale says fewer than 0.1% of an estimated 20 million venom peptides have been studied and the public venom databases hold 40,000 to 60,000 entries between them, and India's amended Biological Diversity Act requires prior approval for foreign-controlled entities to use Indian biological resources. Most of the targets are ion channels, and Speciale says small molecules bind deep inside them, where subtypes look nearly identical, so they often hit the wrong one, while venom peptides act on the outside, where subtypes differ.
Evidence and limits: A pre-seed with no valuation and no cumulative total disclosed, and the company has published no result: its pipeline pages name targets and no data. Venomics AI's own site dates the $2.7 million to April 2026, while Speciale Invest's post announcing it is of 4 November, so the round closed months before it was made public. The two public accounts of who led it differ.
UC San Diego's RobPicker beats the top entries of a Kaggle cryo-ET challenge at picking small, rare proteins out of cryo-electron tomograms
What happened: In Nature Methods, a team led by Pengtao Xie at UC San Diego, with James Zou of Stanford among the authors, describes RobPicker, which uses a small set of cleaner labels to correct wrong or missing labels in the training data and to give rare protein species more weight. On five real cryo-ET datasets and the CZII benchmark it scored higher than three top-ranked Kaggle pipelines; averaging its beta-galactosidase picks gave a 5.2 Å map, against 7.8 Å from the ground-truth picks.
Why it matters: Cryo-electron tomography images proteins inside intact cells, but each particle must first be found, using labels that are slow to make by hand and often wrong when made automatically; small, scarce complexes are the ones models miss. RobPicker asks a lab to curate a small, cleaner validation set instead of re-annotating everything, and once trained, its inference was about 160 times faster than the authors' estimate for PyTom template matching on the CZII test set. For groups with many tomograms, the hand work moves from cleaning every training label to curating that subset; the rest can stay noisy.
Evidence and limits: Every comparison is the authors' own, and two of the five datasets are their unpublished data; no outside group has reproduced it. The sharper map held for one protein: thyroglobulin maps came out about equal (8.3 Å against 8.5 Å). The authors say it does not model differences between labs and acquisition settings, and the cross-organism test used one bacterial species, judged by map resolution rather than scored picks.
NVIDIA releases ALCHEMI Toolkit 0.3, which can distill large chemistry and materials models into small ones
What happened: ALCHEMI Toolkit 0.3.0, released by NVIDIA on 4 November, adds distillation, which trains a small student model on a large teacher model's predictions, to its open-source (Apache 2.0) Python framework for running AI atomic simulations in batches on NVIDIA GPUs. It also adds LoRA fine-tuning, enhanced sampling for rare events, nudged elastic band searches for reaction paths and tools for predicting molecular crystal structures. It runs MACE, AIMNet2 and UMA models, or a user's own.
Why it matters: Machine-learned interatomic potentials let AI stand in for slow quantum calculations in materials and chemistry simulation. An NVIDIA scientist working on AI for science writes that the most accurate of them are often too large to run at the scale many simulations demand, and that distilling one into a compact model makes those workflows practical on far more systems and longer timescales. NVIDIA writes this layer itself and gives it away, and its GPU code is built on CUDA and NVIDIA's own Warp library: groups that adopt it train, distil and sample these models on NVIDIA GPUs.
Evidence and limits: A software release with no results: the changelog and README report no accuracy or speed for any distilled model, though the release adds tools to measure a student model's accuracy and stability. That large models are too big to run at scale, and that distillation fixes it, is one NVIDIA scientist's account in a LinkedIn post.
Tencent's Hy team releases HyCell for single-cell data and HyPDE for physics simulations as open models, alongside Hy4
What happened: Tencent's Hy team released HyCell on 2 November and HyPDE on 4 November, the day it launched its flagship Hy4, with weights and code under Apache 2.0; Hy4's own weights are promised "in the coming days". HyCell, 16.1 million parameters, turns a human cell's gene counts into an embedding and scored 0.735 on seven held-out atlases, level with the 218-million-parameter Stack. HyPDE, 323 million parameters, forecasts simulated 2D physical fields and beat Walrus one step ahead on 9 of 13 datasets.
Why it matters: Single-cell foundation models have grown to hundreds of millions of parameters; in Tencent's comparison of eleven, HyCell scores level with the top one, Stack, at a fourteenth of its size, embedding about 1,500 cells a second on one GPU and fine-tuning into a cell-type annotator in one to three minutes. HyPDE covers 13 physical systems, from acoustics to reaction-diffusion, with one set of weights. Apache 2.0 lets companies build on both. Tencent says both were developed with Hy4 and Hyra, its research agent: by its account, a general-model lab is using its own agent to make tools for scientists.
Evidence and limits: Every score is Tencent's own, with rival models rerun by its team; no outside group has reproduced them. HyPDE was tested on held-out simulations from the 13 datasets it trained on, and loses to Walrus on four. Hy4's claimed lead on TenureTrail (123 tasks, seven disciplines) comes with no scores. Its lower hallucination (12.95% to 7.2%) and unsupported-assumption rates (23.1% to 5.6%) are internal, and its biosafety review rates its open-ended experimental design below frontier models'.
Edison Scientific reports Kosmos matched 83% of the follow-up experiments Incyte's review committee asked for on six drug targets
What happened: Edison Scientific and Incyte built Kosmos Target Assessment, a version of Kosmos that assembles the evidence package a committee reads before deciding whether to pursue a drug target. Edison's case study, by Lydia Auch, Fabian E. Ortega and Angela Yiu, says it assessed a target across all categories in under one workday. On six targets with known committee decisions, its proposed de-risking experiments were a full or conceptual match to the committee's 83% of the time, a partial match 16%, and missed 1%.
Why it matters: Choosing a target commits a drug programme's later spending, and the study says packages are often strong on biology and thin on safety, competition and portfolio fit, which adds weeks or months of specialist follow-up. Its scoring rubric was built from Incyte's own past decisions and review-meeting transcripts, and two named Incyte computational biologists are quoted on using it.
Evidence and limits: Edison's own retrospective study: six targets, with the committee's decisions already known. The claim that Kosmos beat human-prepared packages on a shared rubric comes with no scores. "What would have taken me 3 months" in a morning is one Incyte director's remark. No target decision is attributed to the tool.
Ghent's test of eight AI models that read peptides straight from mass spectra finds most of their errors come from spectra missing key fragments, whichever model reads them
What happened: Sam van Puyenbroeck, Tim Van Den Bossche and Tine Claeys at VIB-UGent and colleagues ran eight deep-learning de novo sequencing models, among them Casanovo, InstaNovo and π-PrimeNovo, on three public proteomics datasets, with database-search identifications as the answer key. On the Orbitrap dataset, 64.2% of spectra were read right by at least one model, the best alone 50.4%. Three rescoring and refinement tools gave little or no gain. Most errors did not depend on the model and traced to spectra missing fragment ions; the models' wrong answers themselves rarely agreed.
Why it matters: These models are how proteomics would find peptides no database lists. If their errors come from what a spectrum lacks rather than from the model, a better architecture alone will not fix them; the authors point instead to information outside the fragments, such as retention time and ion mobility, and to better fragmentation in the experiment.
Evidence and limits: The answer key comes from database search, so the test covers only peptides a database already holds. Performance fell sharply on timsTOF data, because the models were trained on Orbitrap spectra.
Entegris buys a diamond-abrasives maker and partners with De Beers' Element Six to make materials that pull heat out of stacked AI chips
What happened: Entegris, which supplies chemicals and materials to chip factories, has completed its purchase of Advanced Abrasives, a New Jersey maker of diamond consumables for polishing hard substrates, and signed a long-term development partnership with Element Six, the synthetic-diamond maker in De Beers Group. The aim is silicon carbide and diamond materials that carry heat out of logic and memory chips stacked in 3-D packages. Entegris plans a manufacturing line in Taiwan and keeps research on next-generation thermal interface materials in the US. No price was disclosed.
Why it matters: Stacking chips for AI packs more heat into less room, and the layers between chip and cooler set how much of it gets out. So a materials supplier is now paying to work on heat flow across those layers: which silicon carbide and diamond layers to use, how to polish them and build them into the package, and how to cut the resistance to heat where two materials meet. Those are materials and process questions, the kind AI for materials sets out to answer, with a buyer attached.
Evidence and limits: The release gives no price, capacity, customer or performance figure, and says nothing about using AI in the work. Entegris's Capital Markets Day on 9 November added none of those figures: its slides list thermal interface and substrate materials as a new market in advanced packaging.
Google DeepMind adds chemistry skills built on RDKit to Antigravity, and shows an agent redesigning a lung-cancer drug
What happened: A post by Zoi Tsangalidou, Ruijie (Kelly) Zhang, Sukhdeep Singh and Michael V. LeVine releases Cheminformatics Skills for Antigravity, installed through its Science plugin and linked to ChEMBL and PubChem. In the walkthrough an agent pulls thousands of compounds related to crizotinib and lorlatinib from the literature, compares pairs that differ only by a closed ring (matched molecular pair analysis), trains potency models, and proposes ring-closed designs with synthesis routes, in about an hour.
Why it matters: It splits the work between the language model, which picks the steps, reads papers and argues the chemistry, and fixed chemistry code built on RDKit, the open-source toolkit, which does the molecular calculations, so a molecule's properties come from established code rather than the model's guess. For a medicinal chemist it puts data gathering, pair analysis and a first model, which the post says took several days per round, into one session.
Evidence and limits: The post gives no model scores, no count of designs and no test of any design; the authors say each is a hypothesis that must be made and measured. The one-hour figure and the several-day comparison are the team's own, from one worked example on a drug pair whose history is already known.
AuthorsZoi Tsangalidou, Ruijie (Kelly) Zhang, Sukhdeep Singh, Michael V. LeVine
Scinetics writes protein structures as 8,192 discrete tokens and reports a language model above AlphaFold 3's published protein-protein score
What happened: A research note and preprint give every residue, ligand atom and ion one identifier from a vocabulary of 8,192, learned by an encoder that sees the whole complex, so a symbol carries its contacts. Decoded back, known structures reach all-atom lDDT of 0.835 on CASP15. AtomLM-9b, built on Qwen3.5-9B and trained on 64 H100 GPUs for a week, predicts them from sequence and chemistry alone. On FoldBench: 75.01% success on protein-protein interfaces against AlphaFold 3's published 72.93%, and 67.45% against 64.90% on protein-ligand; AlphaFold 3 leads on protein monomers.
Why it matters: Structure prediction today runs on models built for geometry, and a lab that wants structure in the same system as its text and reasoning has to bolt two models together. If geometry can be written as tokens a language model predicts like words, structure becomes one more thing one model does, and the company's own science model is where this is heading. The tokenizer is the piece released, as code that trains from scratch rather than a trained model.
Evidence and limits: The AlphaFold 3 numbers are copied from the FoldBench paper, not rerun, and each model used its own inference settings: a comparison with a published reference. AlphaFold 3 is also ahead on antibody-antigen and protein-DNA interfaces, 47.90% against 44.00% on the first. AtomLM-9b's numbers are the company's own, in-house and not peer-reviewed. Each structure is the best of 25 candidates its confidence model picked; with one, protein-ligand success is 62.10%. Only the tokenizer code is out, not the weights.
OpenAI says GPT-6 Astra designed a small device's enclosure, circuit board and firmware, and a contract manufacturer built prototypes
What happened: OpenAI says two staff, as a side project, had GPT-6 Astra set the requirements for a small device for talking to Codex and design its enclosure, circuit board and firmware in under a week, setting out to make no direct human edits; a contract manufacturer built prototypes within a month. The same post reports an internal model scoring 0.875 on BenchCAD 2.0, OpenAI's test of rebuilding 3D parts and reading circuit boards, against 0.594 for Astra, and a new research collaboration with Mazda.
Why it matters: Drawings, simulations and part design take much of an engineering team's hours, and OpenAI is now training models on that work and asking engineering firms and software vendors to bring real problems. Mazda's aerodynamics and crash-safety engineers will use the models to explore vehicle shapes in simulation, and Chip Ganassi Racing tests turning 3D scans of parts into editable CAD and production drawings.
Evidence and limits: All OpenAI's own account. BenchCAD 2.0 is an internal test with no public tasks, and the stronger model is unnamed. OpenAI says Astra did not consistently respect the manufacturer's limits, tolerances or assembly details, so production took several rounds of review. Mazda's results are promised for the coming months.
Peking University's FoldScape generates protein shapes at a chosen stage of folding, and its sets span about an order of magnitude more of the folding path than AlphaFold 3 run with different seeds
What happened: Lisheng Zhang, Zhengwei Xie and colleagues at Peking University trained FoldScape on about a million simulated conformations of 10,000 proteins. Given a sequence and a folding fraction Q, the share of native contacts formed, it draws backbone shapes at that stage. On 20 human proteins of 36 to 857 residues, with 100 structures from each method, its sets covered a median 72% of the Q range the two methods spanned together, against 5% for AlphaFold 3 run with 20 seeds, every protein the same way.
Why it matters: Drug and protein designers want to know how a protein moves, not only where it ends up: a pocket that opens for a moment, a half-folded state that aggregates. AlphaFold 3 returns near-native shapes whatever the seed, and its authors say seed variation is not meant to sample the solution ensemble. FoldScape takes the folding stage as an input, so a team could ask for shapes between unfolded and folded without running a new molecular dynamics simulation for each protein.
Evidence and limits: Coverage measures spread, not whether each shape is right, and part of it comes from values the user sets. Accuracy along the path was shown on chains not identified as held-out, the 20 test proteins were not checked against training, and only seed-varied AlphaFold 3 was compared, not ensemble generators such as BioEmu. The paths come from fast simulated pulling; nothing was measured in a lab.
10x Genomics releases Sentira Single Cell, an agent built on Claude that analyses Chromium single-cell data from a question in plain language
What happened: 10x Genomics, which makes the Chromium single-cell sequencing instruments, launched Sentira Single Cell on its 10x Cloud. A researcher asks a question, or lets the agent propose one from the data; the agent drafts an analysis plan for the researcher to review, writes and runs the code, and keeps the methods and code in an editable notebook. It runs on Anthropic's Claude models. It is paid in credits, and 10x told GEN a complete analysis takes about 60.
Why it matters: Single-cell work usually stalls at analysis: each lab's question needs its own code, and many biologists who produce the data cannot write it. Putting an agent inside the instrument maker's own cloud is meant to let Chromium users take that step without a bioinformatician. Michael Schnall-Levin, 10x's chief technology officer, told GEN that analyses that took weeks could take hours. It extends the Sentira platform 10x announced at the end of September, whose spatial version is in early access.
Evidence and limits: The release shows no test of the agent's analyses against a bioinformatician's, and 'weeks to hours' is the company's estimate. Schnall-Levin told GEN the agent is not perfect and its output needs review before it goes into a paper.
MIT's Hardness Group treats ETH as refuted and moves lower bounds onto the 1990 Satisfiability Hypothesis, which it proves equal to Weak ETH
What happened: Erik Demaine and six coauthors, writing as the MIT Hardness Group, note that the OpenAI and Anthropic algorithms break ETH but stay consistent with Richard Stearns and Harry Hunt's 1990 Satisfiability Hypothesis (SH). They prove SH equivalent to Weak ETH, which rules out 3-SAT in 2^(n^(1−ε)) time, and give a framework under which most ETH-based lower bounds survive under SH, losing an arbitrarily small constant in the exponent.
Why it matters: Decades of conditional lower bounds rest on ETH. If the refutation stands, this paper says most of them can be kept by swapping the assumption, at a small loss, rather than thrown out; the authors set that against a WIRED headline calling the research irrelevant. This paper, like notes it cites by Michael Lampis and UC San Diego's ETH Group, already builds on the claim being true before anyone outside has confirmed it.
Evidence and limits: The paper assumes the refutation is correct and does not check it. Its results are human proofs in an unreviewed preprint, and the authors say OpenAI's GPT 6.1 Sol helped with the writing.
Johns Hopkins researchers find language models recover more materials they were never shown when more of their training facts appear only once
What happened: Arda Uzunoglu, Eric Nalisnick and colleagues at Johns Hopkins, with Mehrdad Farajtabar of Apple, extend a theory of why models hallucinate to discovery: both draw on facts seen once in training. They fine-tuned gpt-oss-20b and gemma-3-12b on sets whose share of such facts they set from 0 to 100%. One task asks for a formula with given properties, scored on materials new in the Materials Project's September 2025 snapshot; the other edits a DNA sequence to a target activity. In both, exact matches rose steadily with that share.
Why it matters: It points to how training data are sampled, not only model size or reasoning, as a lever on how often a science model proposes something true and new. Showing the same examples several times, which keeps diversity fixed, made the trend disappear, so diversity alone does not explain it; more reasoning raised hit rates without changing the trend. The same lever raises false answers too: on held-out data from the training distribution, discoveries and hallucinations rose together.
Evidence and limits: A discovery here is exact recovery of a known answer held out by date or at random, scored in software; nothing was made or measured. Two small open models and two tasks; the theory's bounds rest on parameters the authors say are hard to estimate. An unreviewed preprint.
Nanyang Technological University has AI agents redo 59 published results in computational quantum science, and Claude Code with Opus 5.5 reproduces about 93%
What happened: Yuxuan Du, Dacheng Tao and colleagues at Nanyang Technological University turned 60 published studies in quantum simulation, algorithms and measurement into tasks: rebuild the method from the paper and reproduce one stated result. Checking agents' attempts against the released code, with experts confirming, found problems in 18 of the 60, from missing details to one study whose paper and code could not be reconciled, which was dropped. On the other 59, Claude Code with Opus 5.5 succeeded about 93% of the time; six third-party agent frameworks scored 19 to 48%.
Why it matters: It separates two reasons agents fail to reproduce published science: the agent, and the paper. Once the gaps between papers and their code were repaired, the strongest systems reproduced most results, and spending more on inference did not reliably buy more success. The authors' own agent, which stores fixes experts verified and reuses them, lifted the low-cost DeepSeek V4.1 Flash from 68% to 85%.
Evidence and limits: Computational results rerun in software and judged against each paper; nothing was measured on quantum hardware. The authors chose the 60 studies and built the agent they compare; its stored fixes came from 5 of the 59 tasks, which its 85% includes. One field, and an unreviewed preprint.
Deep Principle releases Hydrogen-1.0, an open-weight 753-billion-parameter model for chemistry, materials and life sciences, post-trained from GLM-5.3
What happened: Deep Principle, the Hangzhou company behind the Mira research agent, further trained Z.ai's open model GLM-5.3 on chemistry, materials and life-science data, feedback from its own scientists, its lab notebook records and the step-by-step histories of its own materials searches, failed attempts included. The weights are on Hugging Face under the MIT licence, behind a sign-up gate. In the company's tests it beats GPT-6 Astra on the FrontierScience-Research benchmark and on two of three in-house research tasks.
Why it matters: The discovery results reported this year by frontier labs came from closed models; this one can be downloaded, inspected and adapted, and it was trained on how a lab's research decisions turned out, not only on published knowledge. The test closest to discovery is a search: choosing new metal complexes over five rounds, seeing only the results of its own picks, the model's best energy gap, averaged over runs, reached 4.318 eV, against 3.689 for GPT-6 Astra and 3.353 for Kimi-K3.
Evidence and limits: Every score is the company's own, on benchmarks it chose and internal tasks it wrote, and the public-benchmark numbers appear only in a chart. The post reports the complex search as evaluations, with no complex made in a lab, and it covers one property. GPT-6 Astra still leads on assigning structures from NMR spectra, by 3.33 points.
UCL spin-out Octopath AI raises a £1 million seed led by Twin Path Ventures, plus a matched £500,000 Innovate UK grant
What happened: Octopath AI raised £1.5 million on 2 November: a seed round led by Twin Path Ventures and a £500,000 Innovate UK grant the investment unlocked on a matched basis for clinical evaluation and regulatory work. Its models read routine cancer slides cell by cell, Kraken on H&E stains and Nemo on immunohistochemistry, trained on more than 14 million labelled cells across 14 cancer types. The founders are Esther Bär (chief executive), Zhuoyan Shen (chief scientist) and Charles-Antoine Collins-Fekete (chief technology officer), all from UCL medical physics.
Why it matters: Pathology is a bottleneck in cancer care: the announcement cites more than 55,000 people in England waiting past the NHS four-week standard in May 2026, and 3% of histopathology departments reporting enough staff. Octopath sells counts and scores taken from the slides hospitals already make, to labs, to drug developers running trials and to contract research organisations. Twin Path Ventures also backed Potential Sciences, the Cambridge quantum-chemistry start-up that came out of stealth on 5 November.
Evidence and limits: The figures and the deployment at five sites, including about 200 slides a day at University College London Hospitals, are the company's own; the site marks its products for research use only, not diagnosis. The private seed is the £1.5 million total less the £500,000 grant; valuation and close date are not disclosed.
Isomorphic's president Max Jaderberg says it does not sell access to its models: it earns from pharma deals and funds its own drugs
What happened: Speaking at Bloomberg Tech in London on 2 November, Max Jaderberg, president of Isomorphic Labs, said it does not sell access to its models, tokens or API calls, and described its own programmes in oncology, immunology and inflammation as "deep in pre-clinical development", R&D World reported. In January Demis Hassabis, Isomorphic's CEO, said at Davos that it expected its first clinical trials by the end of 2026, Reuters reported; R&D World counts about eight weeks left. Jaderberg also predicted that AI will eventually design every drug.
Why it matters: Jaderberg set out how Isomorphic earns money: pharma partners bring the biology and ideas for medicines, and those deals can bring upfront, milestone and royalty payments; Isomorphic funds its own programmes through discovery and clinical testing. Bloomberg reported on 4 November that it is in talks to raise about $5B at about $50B, before any of its own molecules has been tested in people.
Evidence and limits: Jaderberg's words are as R&D World reports them from Bloomberg Tech. The year-end target is Hassabis's from January, as Reuters reported it, and the eight weeks are R&D World's count; Jaderberg gave no new date and named no programme or biological target. That AI will design every drug is a forecast with no date attached.
Recursion reports REC-4881 cut polyp burden by a median 47% in 15 patients with an inherited polyp disease, and 11 of 13 stayed below baseline 12 weeks after stopping
What happened: Recursion gave new data from TUPELO, its open-label Phase 1b/2 trial of REC-4881 in familial adenomatous polyposis (FAP), an inherited disease that fills the gut with precancerous polyps; Lisa Boardman of Mayo Clinic presented it at the CGA-IGC meeting. After 12 weeks at 4 mg daily, 13 of 15 adults had less polyp burden. In the upper gut 12 of 13 improved, two by 98% or more. Six of eight with polyps over 5 mm had fewer. Four of 26 had a grade 3 drug-related side effect, none higher.
Why it matters: FAP has no approved drug: patients are managed with repeated endoscopy and surgery, and Recursion puts the market at $15B by its own estimate. The company says a phenotypic screen run with its AI foundation models found that blocking MEK1/2 could reverse the cell effects of losing APC, the gene behind FAP, so the trial tests whether a link its platform found in cells holds in patients. It plans a registrational trial in the second half of 2027, pending talks with the FDA, and has not yet given its design.
Evidence and limits: A single-arm trial with 15 evaluable patients, known only from the company's releases; the presentation was not read. For context Recursion cites an Amsterdam registry in which polyp burden rose in 47 matched untreated patients, which is not a control arm. Follow-up off the drug is 12 weeks, and four patients stopped because of side effects. Fewer polyps is not yet fewer surgeries or cancers. The releases credit the platform with finding the mechanism, not with designing the molecule.
Joseph Shunia and GPT-6 Astra build a deterministic primality test on a theorem that extends OpenAI's own zero-free result
What happened: Joseph Shunia's test decides primality with one main congruence, (1+X)ⁿ = 1+Xⁿ in a cyclotomic quotient ring, in L³² bit operations for an L-bit number. Choosing its parameter needs a theorem by D. Cureton and J. Thorner that no Dedekind zeta function vanishes past real part 31/32, which extends OpenAI's 4 October zero-free theorem for Hecke L-functions and was checked in Lean with AxiomProver. The paper says GPT-6 Astra developed and wrote the proofs, from Shunia's conjectured test, notes and direction.
Why it matters: OpenAI's theorem was extended by D. Cureton and J. Thorner, checked in Lean with another company's prover, and then used by GPT-6 Astra to build an algorithm. An AI-produced result is serving here as an input to new work, two steps on, not only as something to check or rewrite.
Evidence and limits: It is not faster: Lenstra and Pomerance's unconditional test already runs in L⁶, and this one reaches L⁶ only with a stronger zero-free bound or GRH. No outside check of the paper's own proofs, which the model wrote, is reported; the Lean check covers only the Cureton and Thorner theorem it rests on.
