Timeline

What happened: Flagship Pioneering combined the platforms of three of its companies, which BioPharma Dive says had raised $374M between them, with its in-house AI unit, Pioneering Intelligence. Adrian Rawcliffe, formerly CEO of Adaptimmune, is CEO; Flagship senior partner Alec Reynolds chairs the board, and Noubar Afeyan is a co-founder. It starts with programmes in blood disorders such as myelofibrosis, in metabolism, including partnered MASH work, and in immunology. Cellarity's sickle cell drug CLY-124 goes to a separate company.

Why it matters: Each company held one kind of evidence: Cellarity read the state of whole cells, Etiome followed how a disease changes from stage to stage, and Empress showed what a molecule actually binds. Flagship is betting that the three together, plus data on how cells respond when a gene is silenced or a chemical is added, can tell which targets cause a disease, and so which drugs will work in patients before a trial. Reynolds framed the target as the roughly half of drugs that fail in Phase 2.

Evidence and limits: No new funding was announced. The 150-million-cell foundation model, which the release calls the industry's largest, and the more than 30 prediction models are the company's description; no result from them is shown.

What happened: Physicists at National Taiwan University, Shanghai Jiao Tong University, SLAC, the University of Washington and others pretrained a model of whole collision events on 500 million simulated events, then tuned it for four analyses. In simulated searches for new heavy particles and exotic Higgs decays it beat XGBoost, TabPFN and SPANet, most clearly with little data, and in simulation it measured top-quark entanglement 37% more precisely than a published benchmark. In real 2016 CMS data it found the known Upsilon meson at 7.6 sigma, against 6.4 for the published method.

Why it matters: Collider physicists today train a separate model for each analysis. EVENET tests whether one pretrained model can serve many: one large pretraining run, on 512 GPUs at NERSC, then cheap fine-tuning for each new question. The gain is largest in searches for rare signals, where each new-particle hypothesis comes with only thousands of simulated signal events. The pretraining data and three of the four downstream datasets are public on Hugging Face under the MIT licence.

Evidence and limits: Only one test uses real collisions, and it finds a particle already known; the other three run on simulated events. The authors say the model does not yet carry detector and theory uncertainties through, and that use in a live LHC analysis is the decisive test.

By Ting-Hsiang Hsu, Vinicius Mikuni, Yuan-Tang Chou, Yulei Zhang

What happened: Fanxu Meng, Kedar Hippalgaonkar and colleagues at UC Berkeley's Singapore centre, Nanyang Technological University, the National University of Singapore, Monash University and Lawrence Berkeley National Laboratory built SKIER: AI models call analysis procedures the researchers wrote, and plain code does the calculations. On 49 records from the group's own published study of methanol oxidation on platinum, all 11 AI configurations finished the data processing with the shared procedures and none without. The models then proposed three mechanisms and tests to tell them apart; the team ran the tests.

Why it matters: AI models are increasingly asked to interpret experiments and propose mechanisms, and their claims are often hard to trace to the measurements behind them. SKIER splits the work: researchers fix how each measurement is analysed, the model chooses and reasons, and each conclusion points back to its data. In the case study the loop closed once: isotope-labelling, residence-time and oxygen-exchange tests argued against two of the three mechanisms and placed C-O bond cleavage before formic acid forms.

Evidence and limits: One reaction system, from the authors' own earlier study. The 100 single-choice questions on which GPT-5.6-sol scored 100% were expert-designed, and the paper does not say by whom. The authors note that intermediates below the detection limit and pathways they did not consider are not ruled out. A preprint.

By Fanxu Meng, Kedar Hippalgaonkar

What happened: The New York company buys or licenses Phase 1 and Phase 2 drugs from other companies and runs their development itself. Its new Drug Development Accelerator Fund has $600M committed by research foundations, endowments, hospitals and philanthropies; Formation Bio is the general partner through a subsidiary and puts in money of its own. Each drug sits in its own company. CEO and co-founder Benjamine Liu, chief business officer David Steinberg and chief scientific officer Michael Ehlers spoke in the release.

Why it matters: It is another way for an AI company to pay for drug development: Formation Bio raises money against individual drugs, and investors buy into the drugs, not the platform. Liu said it opens paths to liquidity without selling the parent company and should lower its cost of capital over time. The claim to test is that one shared, AI-supported development team takes each drug to a clinical result more cheaply, within 2 to 4 years of buying it.

Evidence and limits: This is a fund for drug programmes, not an equity round in Formation Bio. No limited partner and no drug the fund will buy is named, and the release shows no case where the platform shortened a programme.

What happened: BigHat Biosciences and Anthropic ran two lead-optimisation campaigns on BigHat's Milliner platform, about 1,010 designs over seven rounds in all, four on one molecule and three on the other, with Claude in the loop. Reading the raw instrument curves, they found a thermal measure on each scaffold did not measure what its name implied: on the nanobody, the clumping-onset temperature. They fitted a physical model of the curves, chose new measures and changed the protocol. Both leads beat their parents on binding, heat stability and clumping.

Why it matters: A closed-loop lab optimises the number it is given, so a measure that stands in badly for what a drug needs cannot be rescued by a better model or more data. The gain here came from people and a model reading the instrument signal and changing the experiment, not from more rounds of search. Published lab-in-the-loop work mostly takes the reported assay value as a correct label; this campaign did not.

Evidence and limits: Both companies' own report, not peer reviewed, every measurement in BigHat's lab: two molecules on one platform. The report says each lead's values rest on one or two replicates and that confirmation runs come first, and the new measures were chosen inside the campaign that produced them. The leads are antibody fragments, not drugs, and nothing is tested in animals or people.

By Samuel Stanton, Peyton Greenside

What happened: Over three rounds of design and measurement with AI agents, the company built a 45-amino-acid signalling tail for a CAR, less than a third the length of the 4-1BB-CD3zeta domain in approved therapies. Against that domain it raised receptor levels on the cell surface and antigen-driven IFN-gamma release, with none detected when antigen was absent; it kept killing going longer after mRNA delivery, carried four receptors on one mRNA dose, and depleted more B cells in donors' blood cells after lentiviral and lipid-nanoparticle delivery.

Why it matters: Shortening the tail was expected to cost function, because more activation motifs were read as stronger signalling. The short tail made more receptor per unit of mRNA, which is why the team could put four receptors on one dose, and the report suggests compact tails may lower the dose needed to make CAR T cells inside the body. It is also a worked case of agent-guided design: the agents proposed designs and revised them as measurements came in.

Evidence and limits: The company's own technical report, not peer reviewed, and all of it in cells: human blood cells from a few donors, no animal test. The report says how compact tails perform in a living animal, and how many cells carry all four receptors, remain unknown; exhaustion was not measured, and whether agent-guided design beats expert design, or needs fewer rounds, is untested against a matched baseline.

By Liam Bai, Wujie Wang, Jason Yi

What happened: Enveda tested CDP, its Claude-based checker for drug manufacturing batch records, on four executed records from one clinical campaign: 1,061 largely handwritten pages. Human reviewers had raised 11 actionable issues; CDP raised nine as direct matches and two judged intended matches, across its runs rather than in one pass. Single runs on Claude Opus 4.7 added 7 actionable comments no reviewer had raised. A review holding CDP's comments took 9.4 times less time.

Why it matters: Batch-record review is required before a batch of a drug is released, and Anthropic's post says Enveda's manufacturing chemists had been spending up to 75% of their time on paperwork. This is the regulated-paperwork end of drug development, which rarely gets measured at all, and the yardstick here is what human reviewers caught on real records rather than a score on a benchmark.

Evidence and limits: The company's own report, not peer reviewed: four records from one campaign, checked against reviewers' comments, two of the four reviewer arms being outside consultants. No CDP comment cited content the records do not contain, and each run also posted one misread at Major, and the human reviewers could consult earlier-step records CDP was not given. Whether the 7 extra comments matter is Enveda's judgment. The time figure comes from a review holding CDP's comments in advance.

By Andrew Martin, Brandy Schenck, Daniel H. Wee, Viswa Colluru

What happened: On its Science page Anthropic said it has acquired Mithril Medicine, which was developing treatments for two rare childhood muscular dystrophies, and that its founder, Cody Tranbarger, has joined the company to continue that work and start further programmes. The same post reports research partnerships with six organisations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate:Biomedicines and The Column Group, which is setting up a company, Pleiades, to run drug programmes with Claude.

Why it matters: A model supplier has become a drug developer. The partnerships put Anthropic's own scientists inside other companies' programmes, and the acquisition gives it programmes it owns, so it now carries the risk and would hold the result. The post also says Anthropic plans to scale this work up substantially.

Evidence and limits: No terms were disclosed for the acquisition. Anthropic says that if the candidates continue to hold up it expects a drug to be ready to start human trials as soon as late 2028, which is its own forecast, and the mouse results behind it are in a technical report it published itself, not peer reviewed; that report studies two tool molecules in a pilot study, not a candidate. Pleiades has no disclosed programme yet.

What happened: Omalizumab, the only approved biologic for food allergy, neutralises circulating IgE but does not effectively remove the kind bound to FcεRI on mast cells. Guided by its research team, Claude ran 22 phases toward a two-armed antibody that would prise it off. Of 91 disruptor designs screened at one concentration, 20 beat every negative control; eight were purified and titrated, three of them removing more than omalizumab from 25 to 400 nM. Three of 57 anchor designs bound IgE, at apparent affinities of 3.4 to 235 nM.

Why it matters: Pulling bound IgE off the receptor needs a precise geometry, which display and immunisation libraries cannot specify, so designing from a structure is the point here rather than a convenience. Activity survived as many as eight substitutions from the parent, and hits came from every design scope and method the campaign tried. Divergence says the two arms came in weeks rather than the months screening takes.

Evidence and limits: Divergence's own report, not peer reviewed, all biophysics on a sensor chip: no cells, no animals, the two-armed molecule not yet made. In the titration the report says the variants did not appear to improve on the parent sdAb026, the best of them matching it at most concentrations, and Anthropic's post summarising the work does not mention the parent. The anchor affinities may be confounded by avidity, and the method needed an existing disruptor's sequence and solved structure.

By Ezra Erives, Emilio Ferrara, Bowen Jing

What happened: The model generates video together with music, ambient sound and interactive dialogue, which DeepMind calls its first real-time world model to generate audio natively, and it takes natural-language edits while a simulation runs, to change the weather, move the camera or add a person. It builds on the Gemini family, with Gemini Omni anchoring real-world appearance, and every output will carry a SynthID watermark.

Why it matters: DeepMind's stated purposes include developing training environments for robots and agents, and testing AI systems in different environments before taking them to the real world. One clip is generated from the prompt "Universal Robots UR5 holds a pipette and dispenses liquid into a lab plate, triggering tiny electric sparks", the only lab scene in a post whose testers are creative industries, journalism, education, gaming and engineering.

Evidence and limits: No benchmark, score or comparison with Genie 3 appears: it is clips, captions and an internal safety test with no result given. The two-minute consistency figure in press coverage is in neither post; neither gives a session length, and the post says object permanence can drift the longer a simulation runs. Trusted testers only, no public access.

By Raia Hadsell, Diego Rivas, Hang Qi

What happened: In a bioRxiv preprint, Paride Antinucci, Dominic Hall and colleagues at Biographica, a London crop-AI startup, feed a model a gene's DNA plus the measured levels of transcription factors, the proteins that switch genes on, in each tissue or condition. On held-out data from six plant species its median per-gene correlation across contexts is 0.831, against 0.486 for the same model without that input, and 57-246% above three fine-tuned genomic foundation models. It then designed edits of up to 20 DNA letters in soybean and tomato.

Why it matters: Many crop genes affect more than one trait, so an edit that helps in one tissue or condition can cause unwanted changes in others. A model that predicts expression per context lets a breeder rank candidate edits by the change wanted in one place and the change everywhere else, before any plant is edited. Biographica sells gene-editing targets to seed companies.

Evidence and limits: The edits exist only as predictions: none has been made in a plant and measured. The test data come from the public PlantExp collection and were scored by the company's own team, and the preprint is not peer reviewed.

By Paride Antinucci, Dominic Hall

What happened: Deputy Prime Minister Gan Kim Yong announced MARVEL-AI at SWITCH 2026: a platform led by A*STAR, starting with S$150 million of government money, in which AI models pick candidate materials and design experiments and self-driving labs make and test them. It is anchored at A*STAR's Institute of Materials Research and Engineering, whose executive director, Xian Jun Loh, also leads MARVEL-AI. Named partners include ST Engineering, Heraeus Electronics and Foxconn Industrial Internet on materials, and CuspAI and Google Cloud on the AI side.

Why it matters: The speech names the gap it is meant to close: models such as Google's GNoME and Microsoft's MatterGen propose materials, but a material still has to be made, tested and manufactured before industry can use it. MARVEL-AI puts the models and the labs in one national programme and opens its equipment, testing and researchers to companies, startups and SMEs included. CuspAI, which designs materials with AI, will connect its system to the platform's automated labs.

Evidence and limits: This is an announcement: the speech and reports give no split of the S$150 million, no timeline and no material yet. The figure of up to 470 alloy samples a day, more than 400 times conventional throughput, describes an existing lab process Enterprise Singapore cited, not a MARVEL-AI result.

What happened: Adam Lahouari, Jutta Rogal and colleagues at New York University and the Flatiron Institute studied acridine, a small molecule with at least eight crystal forms. Melt experiments solved the structure of form VIII for the first time. Their CrystalMath method kept 15 of 341,683 generated structures without computing any energy, and the 15 include all seven forms with known structures, form VIII among them. A MACE potential fine-tuned on acridine, with full free-energy calculations, reproduced the switch from form II to form III as temperature rises that experiments show.

Why it matters: One molecule can crystallize in several forms that behave differently, which matters for drugs, organic electronics and agrochemicals, and crystal-structure prediction with machine-learned potentials is how the field hopes to rank the forms before making them. Here the ranking step needed tuning: four off-the-shelf foundation potentials put the forms in the wrong order, and simpler free-energy approximations missed the observed crossovers. The study also grows acridine from the melt with 18 additives and finds the two strongest effects match computed defect energies.

Evidence and limits: One molecule. The predicted crossover between the two most stable forms sits near 290 K, against about 318 K measured, some observed transformations disagree with the free-energy curves, and form IV was left out of them. The additive link rests on two of the 18 additives. A preprint.

By Adam Lahouari, Jutta Rogal

What happened: The S-1, filed on 26 October, shows $16.3bn of revenue for the first half of 2026 against $4.6bn for all of 2025, and an operating loss narrowed to $693m from $2.1bn. The $80.7bn net loss is almost all a non-cash $72.4bn remeasurement of convertible notes, derivatives and SAFEs. Compute commitments include at least $114.5bn with Google, $110bn with Amazon and $44.9bn with Microsoft, owed regardless of usage.

Why it matters: For science, the filing says frontier AI 'may extend' to robotics and to scientific research, including drug discovery and materials science, and that Claude with the Model Hardware Standard can 'now' operate microscopes, liquid handlers and robotic arms. Its vision of 'Powerful AI' adds tasks running for weeks and 'compressing decades of scientific progress into years'. In the accounts, though, there is one segment, and the life sciences lab appears among the risk factors.

Evidence and limits: The June figures are unaudited interim ones. The three cloud commitments are three of seven named, with $125bn to $150bn more TPU compute estimated outside the table, and the aggregate line is blank. The top five customers were about a quarter of revenue. A Founder LLC's new share class carries 50.1% of the vote, and the safety trust's own class elects a majority of the board.

What happened: Jared Lewis's group at Indiana University, with David Baker, tested 96 designs; 63 expressed, and one scaffold called DE3 bound one mirror image of an iridium photocatalyst at 94% enantiomeric excess. As designed it gave low selectivity in catalysis. Screening mutations reached a 3:97 enantiomeric ratio, with turnovers above 50,000 in air and the protein reusable over three cycles.

Why it matters: Generative design supplied a pocket that holds a cofactor nature has no protein for, which is what the field could not get from existing folds. On this reaction the enantioselectivity came from screening mutations afterwards; a second design campaign later reached 90:10 on a different reaction with no evolution at all. Metal polypyridine photocatalysts had resisted both routes until now.

Evidence and limits: One scaffold carried through, out of five folds tested for binding. The pocket was designed around a ruthenium cofactor and bound only iridium analogues, and no cofactor-bound structure was obtained. Other reactions in the scope are only moderately selective. Designed and evaluated by the same groups. An accelerated preview.

By Sandip Mishra, David Baker, Jared C. Lewis

What happened: OpenAI's manuscript, dated 25 October and added to its public repository the next day with a Lean formalization, gives a deterministic algorithm that counts the satisfying assignments of fixed-width CNF formulas in subexponential time. Anthropic's 49-page paper, dated 26 October and pushed on the 27th, claims the same thing and prints its author as Claude, Anthropic. It dates its model's discovery to 16 October.

Why it matters: The hypothesis has organised decades of lower-bound work: Anthropic's paper says a literature search by Claude found more than 800 papers proving a result of their own only under it or something stronger. If either proof holds, those conditional results have to be read again, and Ryan Williams, who is helping organise the review, says that if true it should move every computer scientist's weight a little toward P=NP.

Evidence and limits: Nobody outside either company has verified either proof. Lean checks that a proof follows from a formal statement, not that the statement means what the paper claims; Anthropic writes that down and adds that at the time of writing no person is known to have checked its Lean. OpenAI's scope note formalizes a weaker bound than its manuscript, on canonically encoded inputs, still subexponential. Anthropic calls it purely asymptotic.

What happened: OpenAI told three complexity theorists it had the result and would release it, and asked them to organise its examination. Russell Impagliazzo and Ramamohan Paturi at UC San Diego and Ryan Williams at MIT wrote to their field the same day: "Our involvement should not be taken as certification of the claim." Williams said on his own account that the group signed no NDAs and takes no pay, and that OpenAI agreed to release its raw data and sponsor a conference.

Why it matters: Who certifies a machine's result is what the last six weeks have been about, and here a lab named its examiners before releasing the claim. The letter is them refusing the role that implies while keeping their leverage: Williams says they asked for the raw data and a sponsored conference and OpenAI agreed, and the letter says the effort should confer ownership on no company and no chosen few, themselves included.

Evidence and limits: OpenAI's announcement page is 403 here, so its own account is not quoted; its history file calls the three an independent working group, naming no others. The physics result Al Jazeera reports as deferred has no public artifact, and OpenAI's 721-manuscript maths catalogue carries no quantum chromodynamics paper. The NDA, payment, data and conference points are Williams's words, not the letter's, which is undated.

What happened: OpenAI's 59-page paper, its author line reading OpenAI, dated 25 October and added to its public repository the next day, proves Sogge's conjecture for every spatial dimension at least two, with arbitrarily small Sobolev loss and none at p=2. From it the paper deduces the Bochner-Riesz conjecture, sharp bounds for Stein's square function, and the Kakeya and Nikodym maximal conjectures. OpenAI says it is releasing the model's output directly.

Why it matters: The published record had the plane from Larry Guth, Hong Wang and Ruixiang Zhang, and three-dimensional Kakeya from Hong Wang and Joshua Zahl last year; this claims every dimension at once, and cites two OpenAI preprints from September that it supersedes. Unlike the complexity result released the same day, the Lean here covers the headline claims: the local smoothing bound in every dimension, and that a set holding a unit segment in every direction has full Hausdorff dimension.

Evidence and limits: Nobody outside OpenAI has checked it, and OpenAI's own repository calls these manuscripts raw research data rather than polished papers. The restriction estimates for the sphere, paraboloid and cone hold only inside the conjectured range, so that conjecture is not settled here. The same repository withdrew three manuscripts on 7 October over a sign error.

What happened: A Roche team at Prescient Design and Large Molecule Research used a language-model pipeline to pull 89,074 antibody-antigen records out of 10,110 US patents, then built 11,247 pairs: an antibody a patent says binds, with no solved structure, against a sequence-similar decoy. Six structure models were scored. Protenix-v2 led at 72% of pairs and 87% on its most confident half, and picked the true binder out of 50 candidates 29% of the time. The best model touched a listed epitope residue in only 62.8% of cases.

Why it matters: Teams screening designed antibodies read a model's folding confidence as a stand-in for binding. This puts a number on what that is worth: much better than picking at random, nowhere near enough to name the one candidate to make. And telling a binder from a decoy does not mean the pose is right, so reading an epitope off these models is a separate claim these numbers do not support.

Evidence and limits: Nothing here was assayed: the binders are what patents claim, the non-binders are repertoire antibodies assumed not to bind, so the accuracy is against a constructed negative set rather than measured affinities. The structures were not in the models' training data; the patent text may have been. The preprint is CC BY and says the dataset is released, without stating its licence. Not peer reviewed, and the full text was not read.

By Daria Tupikina, Andrea Roncoli, Alexander Bujotzek, Brennan Abanades

What happened: Junkil Park and Aron Walsh at Imperial College London gave a composition-only synthesizability model a second task: rank known materials by the year each was first reported in the ICSD crystal database. On about 10,000 held-out reported materials it recovered 52.4%, against 47.0% for the model it extends and 48.4% with the years shuffled, all at one calibrated threshold where chance is 8.1%. Trained only on materials reported up to 2018, it put 7 later-reported compositions in its top 100, against 3 for the baseline.

Why it matters: Generative models now propose candidate materials far faster than labs can try them, so the question is which to make first. Synthesizability models usually treat every reported material as an equal example. This one uses the reporting order as free extra supervision, on the authors' reading that the order carries which precursors, methods and applications made each material reachable. The shuffled-year control is what ties most of the gain to the order itself rather than to the added task.

Evidence and limits: All checks run on existing databases. The authors say learning from history may favour chemistry already studied and score unexplored families lower. It reads composition only. The authors say the largest gaps come in the top 100 at the later cutoffs, where the statistical uncertainty is also larger. A preprint.

By Junkil Park, Aron Walsh

What happened: Sarath Menon, Braulio Sánchez Ortega and Ralf Drautz at ICAMS, Ruhr University Bochum, add a linear correction on top of GRACE, a foundation interatomic potential, fitted to free energies computed in simulation so that phase transitions land on the assessed SGTE values. Across 11 metals and 14 transitions, every corrected temperature falls within the set tolerance: 15 K for melting, 50 K for solid to solid. Gold melted at 844 K in GRACE and at 1346 K after correction. For titanium it took about 55 H100 GPU hours.

Why it matters: Foundation interatomic potentials learn from DFT data, and an error of a few meV per atom in which phase is more stable can move a melting point by hundreds of kelvin. This layer fixes the balance between phases without touching the base model: it needs only the energies and structures the model produces, not its parameters or training data, so in principle it fits any potential, open or closed. Copper's melting curve, fitted only at zero pressure, then follows measurements within 15 K up to 4 GPa.

Evidence and limits: The corrected temperatures are fitted to the SGTE values, so hitting them is calibration, not prediction, as the authors say. The layer can only rebalance phases the base model already describes and cannot add missing physics such as magnetism. Where the correction is large it moves other properties too, mostly towards experiment: liquid gold and silver become 3 to 6% denser, but vacancy energies move away from experiment for vanadium and aluminium.

By Sarath Menon, Braulio Sánchez Ortega, Ralf Drautz

What happened: Ruichen Li, Ji Chen, Weiluo Ren and colleagues at ByteDance Seed and Peking University split a real-space wavefunction into a spatial part, computed by a neural network (LapNet), and a spin part built to carry one exact total spin. Earlier neural network methods added a spin penalty to training, which can leave the wavefunction stuck mixing spins. The method matched experimental singlet-triplet gaps of biradical molecules about as well as the penalty method, gave an antiferromagnetic spin ladder for [Fe2S2] and a spin gap for an [Fe4S4] cluster.

Why it matters: Iron-sulfur clusters are central to biological nitrogen fixation, and their many spin states of nearly equal energy are where standard methods such as DFT, and even CCSD(T), can go wrong. With spin built in by construction, a calculation cannot drift into the wrong spin state, and chemists' own ideas (core and active orbitals, a first guess at how the spins couple) can be written into the network. With 10 active orbitals it reproduced the local spin trend on [Fe2S2(SCH3)4] that DMRG needed 32 to show.

Evidence and limits: Everything is computed. The iron-sulfur results have no experimental reference: they are compared with other calculations (FCIQMC, DMRG), and for [Fe4S4] the authors say the order of the three methods' gaps is not tied to any one cause. The November 2025 preprint called that gap more accurate; the published version drops the claim. Only the biradical gaps are checked against experiment.

By Ruichen Li, Ji Chen, Weiluo Ren

What happened: The Cabinet Office announced the AI Biosecurity Lab, an independent non-profit it calls the world's first dedicated AI biosecurity organisation, meant to detect biological threats earlier and speed up vaccines and other countermeasures. OpenAI is putting in £100m of compute and model credits through its Rosalind Biodefense programme, Anthropic up to £100m in credits, access and support. Google DeepMind and Basecamp Research join as research partners, not funders.

Why it matters: The £300m headline is £100m of public money; the rest is credits and access rather than cash, and the research partners give access, not money. It also puts frontier labs inside a national function: a National AI Biosecurity Initiative puts the lab with UKHSA, APHA and Dstl, and informs the MHRA's regulation of AI-designed medicines.

Evidence and limits: No budget breakdown, no timeline and no named staff; the scientific advisory committee is still to be announced. Anthropic's £100m is an 'up to' figure, and the page gives no period for either.

What happened: A Toronto, NIST and KAUST team compared deep-ensemble uncertainty against loss-landscape estimates: ALIGNN on band gap and formation energy, a multilayer perceptron on solubility. Calibration methods generalised only variably from in-domain to out-of-domain data, and calibrated uncertainties generally failed to cut the data an active-learning campaign needed on out-of-distribution data. Random forests and XGBoost on the same tasks showed the same poor uncertainty.

Why it matters: Active learning picks the next sample by uncertainty, and a campaign aimed at new chemistry is out of distribution by construction. The same failure in tree models as in the network points at the data: at least partly intrinsic to it, the authors say, rather than model capacity alone.

Evidence and limits: All computed, on three public property tasks. The claim is about these architectures and calibration methods on these datasets, not a proof that no calibration can work. Peer reviewed.

By Ashley S. Dale, Jason Hattrick-Simpers

What happened: Bristol Myers Squibb led; the company puts the post-money valuation at about $3B, up from $1.6B at its $365M Series D in May 2025. Pathos, a New York company co-founded in 2022 by Eric Lefkofsky and Ryan Fukushima and run by CEO Iker Huerga, licenses drugs other companies made and uses its agent platform, Foundry, to pick patients and adjust trials. BMS gets non-exclusive access to Foundry and three-year exclusive rights in neuroscience; Tempus will build a neurology patient dataset that Pathos licenses for $55M over five years.

Why it matters: Pathos puts its AI on the decisions that decide whether a trial succeeds: choosing which existing drug to run, in which patients, and when to change the trial. It does not design molecules. A drugmaker leading the round and taking access to the platform shows that one large pharma company will put money behind this use of AI, and the deal carries the method from cancer to Alzheimer's, a field with long trials and many failures. Simon Davies, who heads data and quantitative sciences at BMS, spoke for it in the announcement.

Evidence and limits: The round, the valuation and the deal terms are all from the company's two releases of 23 October; neither names the other investors or says whether BMS pays for its access beyond its stake. That Foundry identified JSKN016, an antibody-drug conjugate licensed from Alphamab, and has improved any trial's odds is the company's own account: no trial result from its approach has been published.

What happened: Bo Wang, Ci Chu and colleagues describe X-Automate in a companion post to X-Scientist: an off-the-shelf six-axis arm with a camera on it, machined grippers and control software written in-house, directed by the same agent system. Its first target is an Agilent HPLC, which has no motorised door and a drawer that shifts when touched. Over seven days in the company's own lab, a mobile robot and the arm loaded and analysed 18 plates, holding the HPLC busy 97.2% of the time against 53.6% in its model of weekday hours.

Why it matters: Protein-design groups can now make thousands of candidates a week, and the measurement that tells them which ones worked is the step that waits for a person to open a door. Xaira's answer is vision on the arm rather than taught positions, which is what lets one arm move between instruments that were never meant to be automated, and it sidesteps the instrument maker's paid interface. The company says adapting the arm to another instrument now takes weeks rather than months.

Evidence and limits: A company blog post, no paper, with the company reporting its own demonstration and calling the week a technical proof of concept. One instrument, one week, and the 80% is against a schedule the company modelled rather than ran. Four interruptions in plate storage upstream needed a person to recover them, with no errors it attributes to the arm or the agents. The plates were analysed but no scientific finding is reported.

By Bo Wang, Ci Chu

What happened: Hongjun Xiang and Xingao Gong's group at Fudan built an equivariant graph network that predicts both the charge density and the projector augmented-wave augmentation occupancies. Earlier universal models gave only a pseudo charge density, which cannot rebuild the electronic structure consistently. A trainable radial basis and a delta-learning step for charge transfer make it 50 to 150 times faster than earlier charge-density models; over 1,000 unseen materials its median band-energy error is 0.045 eV.

Why it matters: Charge density is what density functional theory solves for, so a model that reproduces it at the accuracy real codes use can stand in for the expensive self-consistent step rather than just predicting one property. The augmentation occupancies are what makes that possible here.

Evidence and limits: All computed, by the authors. The universal model covers non-magnetic structures. The speed is measured against two other machine-learning models and plane-wave DFT on one GPU. Peer reviewed.

By Xiwen Li, Hongjun Xiang

What happened: A Hong Kong University of Science and Technology team matched the crystals of 55 databases by structure into one set of 14.6 million crystals and 131.8 million property values, only 4.5% of whose crystal-property pairs carry a value. Given 100 verifications each, MatterGen conditioned on 400 GPa returned 100 stiff hits against screening's 55; but every generated hit contained osmium, 82 repeated a recorded composition, and screening's best hit was stiffer, 433.8 against 375.4 GPa.

Why it matters: A novelty share measures the reference set as much as the generator, and the usual reference is Materials Project alone. Against all 55 sources, GNoME's matches go from one to 18,234, its new share from 100% to 96.7%. The forward screen turned up 53,813 ultra-incompressible candidates with no recorded bulk modulus; 62 have been made, among them ReN2, measured at 428 GPa.

Evidence and limits: All computed, plus a literature audit: ReN2's 428 GPa is a published measurement, not a new experiment here, and the screen had predicted 349 GPa, 18% below it. The MatterGen runs are the authors' own, from released weights. A preprint.

By Zhilong Song, Zijun Pang, Lixue Cheng

What happened: An EPFL and Fribourg team fine-tuned the PET-MAD potential on r2SCAN data and ran large-scale molecular dynamics on Li6PS5Cl, screening carboxylic, sulfonic and phosphonic acid coatings by chain length and coverage. Simulation said lithiated head groups adsorb more strongly and that long-chain lithiated carboxylates form dense brush-like hydrophobic layers, with protection setting in near 8 Å. Infrared, impedance and H2S measurements followed.

Why it matters: The loop half closes: simulation set the chain length, and the H2S measurements run the same way, longer carboxylate chains suppressing it better. It also names the conflict rather than hiding it, since a thicker coating blocks lithium transport, so the best coating is a middle length.

Evidence and limits: One electrolyte, Li6PS5Cl, and one potential fine-tuned by the authors. The 40 to 55% figure is what is left after processing, not a gain. The experiments cannot rank the head groups, and the paper says so. A preprint.

By Zhiyi Wang, Leonie Braks, Hanna Türk, Davide Tisi

What happened: Tetsuya Yamada and Henrik Kaessmann at Heidelberg, with Svante Pääbo at the Max Planck Institute for Evolutionary Anthropology, predicted the regulatory effect of every single-nucleotide substitution on the human lineage across 305 tissues and cell types. Most of the 71,345 gains sit outside the regions evolution researchers had flagged as fast-changing, and more often than not they come from a regulatory element being reused in a new tissue rather than a new one appearing.

Why it matters: The question of which non-coding changes made humans human has been stuck on data nobody can get: great-ape tissue panels are thin, and no functional measurement is possible on an extinct hominin at all. A sequence model sidesteps that, because it scores a reconstructed ancestral allele as readily as the living one. The payoff here is the archaic comparison: nearly every gain now fixed in humans was already there before the split, so the modern-human-specific ones are few, and the paper names which ones move gene expression most.

Evidence and limits: The map is prediction, not measurement. Where it could be checked, against accessibility measured in four great apes' cell lines, AlphaGenome beat Enformer and ChromBPNet and 82.7% of its strongest calls held up, against 31.1% of controls, but it recovers only a quarter to a half of the measured human-specific changes depending on the threshold, and misses clustered gains more often than isolated ones. A preprint.

By Tetsuya Yamada, Henrik Kaessmann

What happened: Jihan Zhou's group at Peking University put a two-stage network after the reconstruction step: a ResUNet-3D variant corrects distortions across the volume, then a set transformer reads the shape of the electron density around each atom to name its element. Trained on 17,340 simulated samples and benchmarked on 42,588, at typical noise it cut the median coordinate error from 0.244 to 0.100 angstrom and raised atom recovery from 93% to 99%, elements called right 99.5% of the time.

Why it matters: Electron dose is what has kept atom-by-atom tomography to materials that survive the beam, which leaves out most of what people want mapped. The gain here is in post-processing rather than in the microscope, so a tilt series already sitting on disk can be re-read: thinned to a sixth of its dose, a published Pd@Pt series matched the full-dose conventional reconstruction on peak sharpness and repeatability.

Evidence and limits: No new experimental data were taken for this paper: the one demonstration on real data thins a published series, halving its projections too. Every other number is simulated, and the claim that this reaches beam-sensitive materials such as perovskites rests on simulation alone. Peer reviewed.

By Yao Zhang, Lanyi Cao, Zhen Sun, Jihan Zhou

What happened: A KAUST and Karolinska team ran image-based, transcript-based and multimodal segmenters, from Watershed to Cellpose, Baysor and three versions of Segment Anything, over 15 datasets from 5 platforms and 12 human and mouse tissues. On their overall score Segment Anything led at 0.58 and Watershed came last at 0.27, but count how many cells each found and SCS leads instead; the ranking also reorders by dataset: SAM did well where cells look distinct, SCS where they look alike.

Why it matters: Segmentation is the step every spatial result stands on, because it decides which transcript belongs to which cell. This benchmark shows the choice carries all the way through to cell-type proportions, how much signal bleeds between neighbours, and inferred ligand and receptor pairs. So a published spatial finding depends on a step with no standard, and the pipeline released here is an attempt to make the comparison repeatable.

Evidence and limits: There is no true cell boundary to score against, so the metrics are biological proxies: expression purity, marker genes, cell-type proportions against a reference. The authors say platform design, image quality and gene panels cannot be fully separated from the algorithms, and that their gene-set sensitivity test ran on one dataset only. Two of the 11, Swin-S and Cascade Mask R-CNN, were left out of the scored comparison. Peer reviewed.

By Xabier Martínez-de-Morentin, David Gomez-Cabrero, Jesper Tegner

What happened: Lightspeed led; Premji Invest, Mayfield, IVP, Lux, Conviction, Menlo Ventures and 12 other firms joined; angels include OpenAI's chief research officer Mark Chen and its CFO Sarah Friar. The South San Francisco company is a contract research organisation: it will make variations of a lead molecule for drug hunters, with chemists guiding the work. Its co-founders are CEO Kevin Weil, chief scientific officer Sol Reisberg (medicinal chemistry, Terremoto Biosciences), CTO Manohar Paluri (Meta Superintelligence Labs, a founding member of FAIR) and COO DJ Patil, the first US chief data scientist.

Why it matters: Weil's case is that AI has sped up software and mathematics but not chemistry, because chemistry moves only when someone makes a compound and tests it. A lab that sells that work to drugmakers also produces the experimental data better models need, and Warp says it will have no drugs of its own. The other half of the pitch is location: the release cites a 2024 industry survey in which 74% of companies had Chinese contracts for preclinical and clinical work, and the BIOSECURE Act signed at the end of 2025.

Evidence and limits: The round is the company's own announcement, which names no stage. The $865M post-money valuation comes from press reports citing the company and Weil, not from the release. Nothing is automated yet on the public record: the release describes labs being built and a first service guided by chemists, and the site names no customer, only a small number of design partners.

What happened: Google DeepMind and Google Research, with the World Resources Institute and IIASA, released Forest Typology 2020, which labels every 10-metre patch of land as one of six types, from primary forest to plantation, tree crops and other land. Built on AlphaEarth Foundations embeddings, it is 90.2% right overall against 8,190 expert-labelled sites, and is open on Earth Engine.

Why it matters: The EU rule turns on what a plot was in 2020, not just whether it had trees, and a tree-cover map cannot tell a natural forest from one logged and replanted as plantation. This map makes that distinction.

Evidence and limits: The 90.2% is carried by the large other-land class; per class, F1 runs from 82.9% for primary forest down to 58.2% for planted forest. Accuracy was checked on the Geo-Wiki platform of IIASA, a co-developer.

By Maxim Neumann, Anton Raichuk

What happened: Harsha Nori, Eric Horvitz and colleagues at Microsoft AI built a benchmark where a case starts from a short presentation and the diagnostician must ask questions and order tests, each with a cost, before naming a diagnosis. Their orchestrator, MAI-DxO, has language-model agents play a panel of doctors. It was right on 83.2% of New England Journal of Medicine clinicopathological cases and 80.3% of MIMIC-IV emergency admissions; generalist physicians scored 36.1% and 46.8%. Mustafa Suleyman, Microsoft AI's CEO and a co-author, called it a step toward medical superintelligence.

Why it matters: The benchmark, first posted as a preprint in June 2025 and now peer reviewed, tests diagnosis the way a doctor works: not picking an answer from a full case write-up, but deciding what to ask and which tests are worth paying for. Scoring cost beside accuracy lets a system be judged on whether it orders fewer tests, not only on whether it gets the answer. The Science version adds emergency-department cases and newer models.

Evidence and limits: The cases are published records run in simulation; no patient was seen. The doctors worked alone on cases usually handled by specialists, who were not compared. TechCrunch reports much of the gain comes from the underlying model: GPT-5 alone scored 78.9% on the hard cases, and on emergency cases the orchestrator added little.

By Harsha Nori, Eric Horvitz

What happened: Carbon-X, built by Hugging Face with Zhongguancun Academy and TIGEM in Naples, starts from Qwen3-14B, is trained on 302 billion more tokens (55% of them raw DNA), then tuned to answer questions about sequences pasted into a chat. On five DNA benchmarks it averages 66.7%, against 67.5% for GPT-6.1 Sol and 60.9% for Opus 5.5, and leads every model tested on coding-variant effects (74.9%) and next-base prediction. The weights are out under Apache 2.0 and run on one 80 GB GPU.

Why it matters: DNA models usually output scores a scientist has to interpret, while a chat model may explain what a gene does yet struggle to judge a change in its sequence. Carbon-X puts both in one open model a lab can run on its own hardware, so unpublished or patient sequences need not go to an outside API. The team estimates that scoring ClinVar's 2.4 million variants of uncertain significance would cost about $300 with Carbon-X, against $32,000 to $4 million with the frontier models it tested.

Evidence and limits: Every score is the team's own, and two of the five benchmarks (FrontierVEP and NextBase) are ones it built. Nothing was measured in a lab or a clinic, and the team says the model is not clinically validated. Frontier models still lead on insertions and deletions and on single-base variants in long sequences, GPT-6.1 Sol also on sequence-feature questions, and Carbon-X can miscount bases or misplace coordinates in long sequences.

By Loubna Ben Allal, Diego di Bernardo

What happened: The company's research note and preprint set out a four-step cycle: train compact experts on separate skills, break each answer into checkable requirements and accept an expert's contribution only where it beats the current one, write the accepted parts into the target model, then specialise again from it. In the 4B and 8B settings it ranks first among the methods compared on all five scientific benchmarks, scientific reasoning at 55.0% against GRPO's 53.0%. With 8B experts and a Qwen3.5-27B target, all seven benchmarks improve, the mean from 87.3% to 94.0%.

Why it matters: Most science models are trained once and then stand still, and a lab that wants a new skill in one trains a new model. The claim here is that a smaller model's strengths can be moved into a larger one, which the 8B-to-27B run is the test of, and that the filtering is what makes it work: without the verification step the same supervision drops scientific reasoning from 55.0% to 50.6%, and scattering it at random costs 6.3 and 12.1 points. The open 8B checkpoint lets someone outside the company check what the recipe produced.

Evidence and limits: Benchmark scores only, all of them the company's own, scored in-house and not peer-reviewed. The note says its own report does not show that repeated cycles keep accumulating gains, and the 27B run also changes model generation, which it reports separately; the model card says the open checkpoint renames and republishes an existing one rather than a fresh training run. The rankings are first among the methods compared, not against the field.

By Xun Xu, Zaixi Zhang

What happened: Michele Ceriotti's group at EPFL applied the Point-Edge Transformer, which drops explicit O(3) algebra for edge-to-edge attention, to molecular dipoles and polarizabilities, NMR chemical shielding tensors and auxiliary-basis expansions of the electron density. A bigger backbone improved accuracy and equivariance error together. It matches or improves on the released models at lower inference cost, the density only when trained on a better-posed target.

Why it matters: Strictly equivariant networks pay for their guarantee in tensor products that grow steeply with angular order, and can fail to represent a target if truncated. Targets like shielding tensors and basis expansions are where that bites. Learning the symmetry instead removes the trade-off, if approximate equivariance is good enough.

Evidence and limits: All computed, by the authors. Equivariance here is approximate rather than guaranteed, and the authors say only that it affects most observables little, not all. Electron density is the weakest case: its equivariance error grows with angular order, and above order two it falls with backbone size no faster than the symmetrized error does. A preprint.

By Joseph W. Abbott, Michele Ceriotti

What happened: Gus Hart at Brigham Young University and Jacob Pursley at the DOE's Kansas City National Security Campus tested 14 universal machine-learned potentials on 1.45 million ordered intermetallics against AFLOW, a database none of them trained on. For half the structures AFLOW, Alexandria and OQMD differ by under 6 to 8 meV per atom, but a minority differ far more, raising the average to 16 to 30. Re-referenced, the best potential agrees with AFLOW as closely as Alexandria does.

Why it matters: Leaderboards rank these potentials by formation-energy error against one database, as if it were ground truth. The disagreement between independent databases sets a floor under any such number, and the top models now sit at it. Screening out broken elemental references aligns the energy and stability rankings, and the leader's 7 meV per atom margin shrinks to 0.4 when both are scored at a common geometry.

Evidence and limits: All computed: this compares databases and models, with no new measurement. Intermetallics only, chosen as a region where DFT is relatively accurate. A preprint.

By Gus L. W. Hart, Jacob N. Pursley

What happened: Statisticians at Karlsruhe and Marburg put Pangu-Weather, GraphCast and Aurora against ECMWF's HRES under one experimental design. HRES came out ahead when scoring was restricted to events that actually happened; the AI models came out ahead under weighted scoring functions. A contingency table shows why: the AI models forecast extremes less often than they occur, so misses dominate their scores, while HRES forecasts them more often than the AI models do, so false alarms dominate its.

Why it matters: Published studies disagree about AI models on extremes, and this says the disagreement is largely in the protocols rather than in the models. The decomposition puts the score gaps down to calibration, not discrimination: every model tells an extreme from a normal day about as well, Pangu-Weather least of them, and the table shows the AI models are simply too conservative about saying so. That is a fixable problem, and a different one from not seeing the event.

Evidence and limits: All computed, against the analysis that initialised all four models, which the authors say slightly favours HRES; station recordings would test it. One physics model and three AI models, on the variables and lead times chosen here. A workshop paper at NeurIPS 2026, so shorter and less reviewed than a journal article.

By Tobias Biegert, Sam Allen, Zhongwei Zhang, Sebastian Lerch

What happened: Chengming Zhang and Luonan Chen, with Kazuyuki Aihara at the University of Tokyo, split the latent space of a hierarchical model into causal and spurious factors and apply do-calculus to the causal half, then recurse to rebuild the regulatory layers. On synthetic data, stem-cell differentiation, a CRISPR activation screen of 31 single and 26 paired conditions with some held out, a type 2 diabetes multiome and gut microbiome data, it beat the methods compared on causal triggers in most settings.

Why it matters: Most regulatory networks read off single-cell data are built from correlation, so a gene that moves with a state change is indistinguishable from one that drives it, and that is the difference between a target worth perturbing and one not worth the bench time. Encoding the intervention itself, rather than the association, is a different route to the problem, and the output is a ranked list of candidates to test.

Evidence and limits: All computed on public datasets, with no new experiment. The authors say the inferred triggers may still contain spurious biological signal, that the in silico perturbations are a qualitative read of direction rather than a predicted magnitude, and that simple additive baselines still do well on part of their own perturbation benchmark. Its results also depend on the regulatory hierarchy assumed. Peer reviewed.

By Chengming Zhang, Deyu Cai, Kazuyuki Aihara, Luonan Chen

What happened: A Nanjing University, CUHK and NUS team split the problem in two: generate an RNA shape in a pretrained structural latent space, then predict the contacts that place it against the protein, rather than generating shape and pose together. On 124 protein-RNA complexes held out by date, it leads ODesign, RFDpoly, RNAFlow and NACraft on complex recovery, with DockQ 0.131 against 0.116 for the next best. Twenty of the 124 clear DockQ 0.23.

Why it matters: RNA that binds a chosen protein is a route to reagents and to drugs, and unlike protein binder design it has stayed out of reach, partly because so few protein-RNA structures have been solved. What this paper pins down is where the state of the art stands when the binding site is withheld: five methods, and the best of them recovers the native pose on one target in six.

Evidence and limits: All computed, no synthesis and no binding measurement, and the authors say structural recovery establishes neither affinity nor activity. The headline ranking uses the native interface after generation to pick which of RiboBridge's own pockets is scored, so it is not a blind result. RiboBridge also clashes more than two of the baselines and does not lead on sequence recovery. A preprint.

By Zhou Zhang, Hanqun Cao, Tianfan Fu

What happened: In INTerpath-001, 1,137 patients with resected stage IIB-IV melanoma were randomised 2:1 to intismeran, an mRNA encoding up to 34 neoantigens chosen from each patient's tumour sequencing, plus KEYTRUDA, or to placebo plus KEYTRUDA. After a median 24.6 months, two-year recurrence-free survival was 82.2% against 70.0% (HR 0.58), and the risk of distant metastasis or death fell 38% (HR 0.62). The data go to ESMO 2026's Presidential Symposium and are accepted by NEJM.

Why it matters: Until now only the topline was public. The two arms part only after six months, once the course of up to nine vaccine doses is given: the hazard ratio is 0.99 in the first six months and 0.42 after, in a prespecified sensitivity analysis. That is the timing expected if the benefit comes from an immune response the vaccine builds against each patient's chosen neoantigens. Merck says it is now talking to regulators.

Evidence and limits: Merck and Moderna released these numbers ahead of the talk and the paper. Overall survival is immature: the release reports a 'favorable trend' and no number. The trial tests the whole product, so it cannot say how much the neoantigen selection adds over another way of choosing antigens. Grade 3-5 treatment-related events were 14.2% against 13.4%.

What happened: RFD4-Proteina, from the University of Washington Institute for Protein Design and NVIDIA (David Baker corresponding), designs proteins, DNA, RNA and enzymes in one shared latent space, using reasoning and guided search rather than a separate model per task. On AlphaFold3-refolded designs it reports a 10.6% pass rate for protein binders against BindCraft's 5.9% at 42 times the throughput, and higher rates than every method tested on small-molecule, enzyme-scaffold and DNA-binding tasks. In the lab, 158 of 188 designed TrkA binders bound below a micromole.

Why it matters: Protein design has used a different specialist tool for each job: backbones, binders, enzymes, DNA. RFD4-Proteina puts them in one model that also covers modified amino acids, sugars and nucleic acids, and still comes out ahead on all four. The part that matters is the wet lab: designs across many target classes were made and measured, including a designed enzyme that converted 79% of an intermediate in making the HIV drug lenacapavir, a de novo catalyst doing a real step of drug chemistry, further than designed enzymes had reached.

Evidence and limits: A preprint, not peer reviewed, and all the wet-lab work is the authors' own, so nothing is independent yet. In silico a 'pass' means an AlphaFold3 refold matches the target (a model checking a model), and the pass rates are low in absolute terms (3.5% for DNA binders). The binder and enzyme successes are counts from the authors' own assays, not an outside lab's, and the enzyme result is a single proof-of-concept.

By Nathaniel Corley, Kieran Didi, David Baker

What happened: A. Ali Heydari, Daniel McDuff, Xuhai Orson Xu and colleagues at Google Research built an orchestrator over three Gemini 2 agents: one writes and runs analysis code on Fitbit and blood-test data, one checks health knowledge against sources, one coaches. The data came from the WEAR-ME study of 1,165 people. Across 11 tasks and 1,350 hours of rating, 45 users put it first 46.7% of the time; three health experts scored its overall quality 72.8%, against 64.3% for the three agents answering separately and 54.4% for one agent.

Why it matters: Questions about your own wearable data need code and statistics, while coaching needs a different way of talking, and the authors found one model struggles to balance both. Splitting the work let each part be tested on its own: the data agent's analysis plans scored 75.6% against 53.7% for plain Gemini, and its code passed 75.5% of expert unit tests on the first try against 58.4%. The framework and its benchmark are open under Apache 2.0, so other groups can swap in their own models.

Evidence and limits: Everything was rated in conversation, not measured in health: the live users played study personas instead of asking about their own data, and no one tested whether behaviour changed. The data are Fitbit only, some scoring used model raters, and a response took 205 seconds and 6.5 model calls, against 36 seconds for one agent. The authors call it a research prototype, not a diagnostic tool.

By A. Ali Heydari, Xin Liu, Daniel McDuff, Xuhai Orson Xu

What happened: Mirendil reports that one researcher built Mirendil-Base from an empty repository: agents wrote over 99% of the code, ran the experiments and diagnosed failures, while the researcher chose what to try and what to trust. It learned only from public code, web pages and datasets, with no distillation, and the cost covers all compute, the agents' tokens included. On seven code repositories published after training it scored 0.222 bits per byte (lower is better), behind Kimi-K3's 0.193 and ahead of GLM-5.3, DeepSeek-V4-Pro-Base and Nemotron-550B-Base.

Why it matters: Pretraining a competitive base model has taken a team of specialists in architecture, data and chips. If one person directing agents can do it, the scarce inputs become compute and judgment, and a group whose expertise is in its own science, such as a robotics lab or a hospital, could train a model on its own data. It is the first result from a company that raised on the pitch that labs using AI in biology or chemistry should not have to become AI labs.

Evidence and limits: Every number is Mirendil's own, in a blog post with no paper, no weights and no named researcher. The only score is how well the model predicts unseen code, not whether it solves coding tasks, and Mirendil notes that Kimi-K3 and GLM-5.3 were scored after post-training, which usually raises this loss slightly. The two larger models are still in training.

What happened: Jeonghyeon Kim and Philip A. Romero connected GPT-5.6 Terra to a standard Opentrons OT-2 pipetting robot and a webcam through PEARL, five model roles sharing one record of the experiment. Given only a goal, it refused an impossible instruction in all ten runs, read from the colours that a human had turned the plate 180 degrees mid-run, and in seven of ten diluted dyes to see their colour first. Asked to characterise an enzyme assay, it set concentrations and timings, discarded a spiked well, and repeated runs missing early readings.

Why it matters: Lab automation runs the procedure it was given, and a moved plate or a saturated reading breaks it unless someone wrote a check for that case first. The paper's own comparison locates the behaviour: in a plain tool loop none of the 80 execution runs passed, one agent keeping a persistent record of the experiment passed 60%, and the five-role version passed all 80. So the judgment is mostly the general model's and the scaffolding buys reliability, which is what would let a lab with an ordinary liquid handler try this rather than build a platform.

Evidence and limits: Eight scenarios on one bench, run by the authors: colour mixing and one horseradish peroxidase assay, whose kinetic constants matched their own human-run reference and previously reported values, not an outside lab's. The scoring of planning, adaptation and execution is theirs, on their own rubric, and every perturbation was one they introduced, so nothing tests what a working lab meets unplanned. Colour was read off a webcam calibrated against a plate reader. It is a preprint.

By Jeonghyeon Kim, Philip A. Romero

What happened: AlphaVariant pairs a sequence generator trained on related natural proteins with activity predictors retrained after each round of results. On savinase, a detergent protease, two rounds produced libraries in which 75.4% and then 83.8% of the variants tested beat wild type, as the positions it could change went from 25 to 67; its best variant was 7.1 times more active than wild type on a synthetic substrate. A third round also scored heat tolerance and patent distance, and of 32 variants built, three beat the commercial protease Uno 100 L.

Why it matters: Directed evolution gets its gains mostly one position at a time, because the combinations explode and the measured data is far too sparse to fit a model over them. Here the useful variants carry changes at eight to sixteen positions at once, and the final round needed only 32 measurements, which is the scale at which a design method is worth a company's screening budget rather than a benchmark table. Putting patent distance in the reward is the unusual part: freedom to sell was optimised alongside performance, and it is a constraint any enzyme has to satisfy.

Evidence and limits: Every assay and the wash test were run by the authors with the industrial partner Bestzyme, which has filed a patent application on the variants; no outside lab has tested them. The wash test is a Chinese standard method on one soiled-cloth type, not a product trial. The 7.1-fold figure is kinetic fitting on a peptide substrate, and round three used casein. On the six public landscapes the clear wins were the three scored by a surrogate. Preprint, CC BY-NC-ND.

By Xiaopeng Xu, Jie Wang, Xin Gao

What happened: Bo Wang, Ci Chu and colleagues describe X-Scientist on the company's blog: a lead agent plans and checks the work of worker agents that call bioinformatics tools and Xaira's own models. It nominated 381 genetic perturbations to undo what TGF-beta does to HepG2 liver cells; the company also tested 321 its own scientists picked and 433 drawn at random, measured all of them by Perturb-seq, and counted a hit at three standard deviations back toward untreated cells. Nominating took 20 minutes against more than three hours by hand.

Why it matters: TGF-beta drives scarring in chronic liver disease, and which genes to perturb against it is the kind of choice a biologist spends a morning on. The measurement here is of the model's picks, not of the model's own score: every nomination was made, perturbed and read out in cells. Three of its hits outside the known TGF-beta pathway beat the median of the ones inside it, and none were in the scientists' list, which is what the loop is for, since a list of known pathway genes is one a scientist already has.

Evidence and limits: A company blog post, no paper and no peer review, and Xaira both chose and scored the comparison. One cell line, one exposure out of the twelve it ran, and the 7.1% against 6.9% gap is well inside what 381 and 321 perturbations can resolve. The benchmark scores beside it compare harnesses on the same models, not biology. Nothing has gone back into the models yet.

By Bo Wang, Ci Chu

What happened: AI Proteins will design and optimise miniproteins from scratch against several targets Astellas picks. Astellas, through its affiliate Astellas Innovation Management, takes an exclusive worldwide licence to any candidate and runs development, regulatory work and sales. Upfront and milestone payments come to more than $400M, plus royalties. Chris Bahl, the Boston company's founder and CEO, and Morten Sogaard of the Astellas Innovation Lab spoke for the deal.

Why it matters: This is AI Proteins' second large-pharma deal, after Bristol Myers Squibb's in late 2024, also valued at up to $400M. Miniproteins are small proteins built from scratch rather than taken from nature. Bristol Myers Squibb took an option on a licence; Astellas takes the exclusive licence outright, so this time the designs are bought as future drugs from the start.

Evidence and limits: The release does not break out the upfront payment or say how many targets are involved, and no designed molecule from the deal is shown. One trade report described the $400M as upfront; the release counts it as upfront and milestones together.

What happened: Fengyuan Dai, Fajie Yuan and colleagues at Westlake University trained a 16-billion-parameter model on 1.7 billion protein-text pairs: a description of the function wanted goes in, a sequence comes out, with less hand-written structural specification. Four tasks were run, a fluorescent protein, a PET hydrolase, an alcohol dehydrogenase and a metabolic H-protein, and each gave the intended function, with catalysis measured for both enzymes. In a reconstituted CO2-fixing reaction a designed H-protein raised product yield 1.7-fold over the natural E. coli one.

Why it matters: Protein design normally starts from a natural protein to optimise, or asks the designer to restate the goal as geometry and sequence rules a model can take. Here the goal is a sentence, and what came out of it catalysed reactions rather than only folding as predicted. The H-protein is the sharper part: a designed carrier protein beat the natural one in a reconstituted CO2-fixation reaction, with no rounds of experimental optimisation in between.

Evidence and limits: The per-task numbers are in the extended data and they are small: 94 PET hydrolase designs assayed out of 2,416 filtered gave one hit, 8 alcohol dehydrogenases tested gave 5, and 20 H-proteins picked from about 1,000 gave 5 active. All the wet-lab work is the authors' own. The inference code is Apache 2.0 and the weights repository states no licence; the full training corpus was withheld for biosecurity and given only to the reviewers.

By Fengyuan Dai, Fajie Yuan

What happened: Fajie Yuan's group at Westlake University, with HKUST (Guangzhou), the Shenzhen Institute of Synthetic Biology and Zelixir, trained Pinal, a 16-billion-parameter model, on 1.7 billion machine-written pairs of a protein and a description. It drafts a structure from the words, then a sequence to fit it. Designs for a fluorescent protein, a PET-degrading enzyme, an alcohol dehydrogenase and an H-protein each showed the function asked for; in a CO2-fixing pathway the designed H-protein gave 1.7 times the product of E. coli's own, with no lab optimisation.

Why it matters: Design models typically start from a natural protein to improve, or need a person to turn the goal into shape and sequence constraints first. Pinal takes the goal as a sentence, which is a different interface: the biologist says what the protein should do and gets candidates to test. The weights are on Hugging Face and the code is Apache-2.0, so the claim can be tried by anyone.

Evidence and limits: Each hit came out of a funnel the headline hides: about 151 designs assayed across the four functions, out of thousands generated, with the PET-degrading enzyme leaving one standout among 94 tested. The measurements are the authors' own, and the training corpus is withheld over dual-use risk. Fajie Yuan told Science and Technology Daily that language-driven protein design is still at the proof-of-concept stage, with finer instructions, such as setting a fluorescent protein's emission band, open for the field.

By Fengyuan Dai, Shiyang You, Tong Si, Fajie Yuan

What happened: OpenAI says it has acquired Orchestra Bio, a platform for coordinating drug development, and is bringing its team and technology into Rosalind. Its own example: a failed manufacturing batch can delay the start of a clinical trial, and Orchestra's agents work out which activities and timelines that hits, gather the relevant information and compare ways to respond, after which scientists and programme leaders decide. The same page names Novo Nordisk and Moderna as Rosalind partners.

Why it matters: Rosalind has been a model and a workbench; this adds the layer that tracks what a programme owes and when. OpenAI puts the case in its own numbers: roughly 10 to 15 years on average from target discovery to approval in the United States, with teams coordinating across internal departments and outside partners on disconnected systems. So what is being sold here is coordination rather than prediction, and the division of labour is stated outright: agents monitor and track, people decide.

Evidence and limits: No price, and OpenAI's page names nobody from Orchestra. The reported funding, about $20 million in total, reaches here through Decrypt citing Caplight rather than from the company, and the seed round's size is not given. Neither company says how many customers Orchestra has, what happens to existing contracts, or when the agents reach Rosalind. Orchestra Bio is not Orchestra BioMed, the unrelated Nasdaq device maker.

What happened: Aniketh Iyengar, Benjamin Siranosian and colleagues report it on the company's blog. Omnii-1 rewrote the DNA for gp17, the tail-fibre protein the T7 phage uses to recognise E. coli. At 11 to 15 amino-acid changes from the natural protein, 69.9% of its designs were recovered after assembly against 4.8% of random controls, and 2,161 of 6,694 variants with measurable activity were enriched above the natural phage. A cocktail containing its designs cleared one patient's resistant bacteria in a dish, then was given to that patient.

Why it matters: A model's design has gone into a person's treatment, not only into a dish: the company says this is the first clinical use of AI-generated phages, under an authorised, physician-led compassionate-use pathway. What the model wrote is one region of a known T7 genome, not a genome. It also ranks which designs to test against a given bacterial strain (on the hardest of three library strains, 24% of its top 100 picks beat the activity threshold against 4.5% of the library), and that ranking is the step a therapy matched to one patient has to get right.

Evidence and limits: A company blog post, no paper and no peer review, and every measurement is the company's own. One patient: the company discloses no clinical course and no outcome, citing privacy, and says a single patient cannot establish safety or efficacy, so nothing here tests whether the treatment worked. "World's first" is the company's own claim. Recovery is a sequencing proxy rather than a plaque count, and activity is enrichment measured against the natural phage.

By Aniketh Iyengar, Benjamin Siranosian, Arpan Sarkar, Avery Noonan

What happened: The team chose five toxins; from one prompt and four follow-ups, an internal research model then worked mostly on its own for 64 hours, picking the sites to aim at and generating 318,484 candidates, 211 of which were made. Adaptyv Bio tested all of them for binding, Twist Bioscience 210 against four toxins. Seventeen met Anthropic's criteria, 15 of them at both labs. The tightest, against the Mexican toxin Cn2, was single-digit nanomolar. All 211 and their data are released.

Why it matters: This is the company's earlier binder work pushed at a target nobody has money to work on (antivenom for a neglected disease), and the loop is what to look at: the model set its own quality thresholds, decided how much to lean on each design tool the prompt offered, and sent sequences to a vendor, with people steering four times. Among designs aimed at a single toxin, 11 of 60 bound, which is what an ordinary protein-design campaign gets, so the claim is about who ran the campaign, not about a better method.

Evidence and limits: Binding is all that was measured, against toxin proteins grown in the lab, not taken from venom; nothing was tested in cells or in an animal, and Anthropic says it has not made an antivenom. None of the 151 designs aimed at more than one toxin met the bar against more than one. The criteria were set after the data came in, and no design baseline was run, so the model's own contribution cannot be separated from its tools'.

By Nathaniel Horwitz, Amir Shanehsazzadeh

What happened: Yazhuo Liu, Ting Zhu and colleagues at Georgia Tech, the University of Osaka and USC build GBClaw: an orchestrator and five agents, routed to two GPT-5.6 models, that set up atomistic shear simulations in LAMMPS and analyse them. Across 14 asymmetric tilt boundaries in nickel the strength at which sliding starts ran from 14 megapascals to 9.3 gigapascals. Four were rebuilt with arrays of secondary misfit dislocations, the defects a real boundary has: one fell from 1,074 to 530 MPa, two by four orders of magnitude, one rose.

Why it matters: Most interfaces inside a real metal are these mismatched boundaries, and they stay understudied because their structure is hard to pin down in both experiment and simulation. What the sweep shows is about the models, not the agent: if the predicted strength depends this much on whether the boundary's own defect structure sits in the cell, then screening boundaries on simplified cells (with a first-principles or machine-learned potential, at real cost) is reading the cell as much as the material.

Evidence and limits: All of it is simulation with one empirical nickel potential: nothing was made or measured, and the authors call their numbers athermal instability thresholds for the cells they modelled, not universal upper bounds. The cells are small and periodic, which suppresses the long-range defect interactions a wide boundary has, and the energy paths carry no temperature. The {111}/{110} anomaly was traced only after a researcher directed the follow-up runs.

By Yazhuo Liu, Ting Zhu

What happened: Miguel Marques and Silvana Botti's group at Ruhr University Bochum ranked about 2.8 million near-stable compounds in the Alexandria database with two machine-learning models and a fitted formula, kept about 1,200, added about 12,800 compounds already computed, and narrowed the combined set to 52 for full electron-phonon transport calculations. Tested on elemental metals, those calculations land mostly within 30% of measured values. The best compounds, Be2IrPt, H3NaNi, LuIr and LiBePt2, come out at about 34 to 38 million S/m, near aluminium's measured 36.59.

Why it matters: Chip wiring and power lines need metals that conduct well, and conductivity is a property materials databases rarely compute because the full calculation is expensive. This shows a workable route: rank millions cheaply, then spend the full calculation on a few dozen. It also explains why silver, copper and gold stay ahead, and points to a mechanism: electrons from light elements can push a compound's high-scattering d-states below the Fermi level, as in LiBePt2.

Evidence and limits: The match with experiment is for elemental metals, not the new compounds, none of which was made or measured here. Even among elements, calcium comes out at 7.98 against a measured 28.99 million S/m. Spin-orbit coupling was left out; adding it cut Be2IrPt and LiBePt2 by 8% and 9%. Several leading candidates contain platinum, iridium or rhodium.

By Thalis H. B. da Silva, Silvana Botti, Miguel A. L. Marques

What happened: A Sydney, Columbia, BAAI and Tsinghua team re-scored 95 predictors over 217 ProteinGym assays and 2.47 million variants by what each would find under a matched budget, on two objectives: finding top variants, and ranking the landscape. Across 20 scores fitted to 20 measurements of the target assay, prior ranking ability explained under a fifth of the differences in what they recovered (r = 0.44). A plain additive model from measured single mutants then overtook them: at 50 to 100 measured singles for top variants, around 500 for ranking.

Why it matters: This changes what a protein-model leaderboard is for. A campaign's question is which variants to measure next, and the paper shows the quantity that answers it is not the one benchmarks report: a score that ranks a landscape well need not direct a budget well. The practical reading is that a model prior is worth most in the first few dozen measurements, after which the lab's own data wins, and that closed-loop design should be scored by what gets discovered at a given budget, not by correlation on a fixed dataset alone.

Evidence and limits: All retrospective, re-scoring measurements that already exist, with no new experiment and no closed loop run. Substitution landscapes only; the authors say insertions, deletions and de novo design would need their own controls and stated measurement costs. The crossings are additivity predicting double mutants from measured singles; on variants with three or more substitutions the models led instead. They come from deep mutational scans, where every variant is already measured, which is not a lab's candidate pool. A preprint.

By Wenze Ding, Haipeng Gong, Zhiyong Wang

What happened: Matthew Carter and colleagues post the numbers on the company's blog. On AlphaGenome's own panel of 21,948 test cases, each pairing a DNA change with a gene and a tissue, half of them changes statistics pinned as the cause and half it ruled out, Omnii-1 scores 0.810 against Borzoi's 0.705 and the 0.713 AlphaGenome's paper reports. On a splicing panel of 14,458 cases it scores 0.814 against 0.721 for Pangolin. It also predicts methylation, a chemical tag on DNA, one site at a time, agreeing with measurements at 0.879.

Why it matters: Most DNA changes tied to disease sit outside the stretches that code for proteins, and working out which of them does anything is the bottleneck. The earlier models read DNA alone. This one also reads lab measurements of how the DNA is packaged and used in a given tissue, and can take those measurements as an input rather than only predict them. Its lead over Borzoi grows the further the change sits from the gene, from 0.05 within 500 bases to 0.19 beyond 50,000.

Evidence and limits: Every score is the company's own, in a blog post with no paper and no peer review. It says it cannot run AlphaGenome itself, so those numbers are read out of the published paper rather than measured side by side, and AlphaGenome's higher supervised score on the same panel, 0.799, is not the one compared against. The labels are statistical inferences from population data, not variants tested in an experiment, and the model is behind a request form.

By Matthew Carter, Armin Thomas

What happened: Max Highsmith and colleagues report it on the company's blog. The campaign began with 1,620 designs from earlier pipelines, 150 of them with measured binding. Since mid-September, Omnii-1 rewrote positions in the best of those, ranked its own proposals, and Aptamino made and measured a selected set, three times over. The designs push out the measured trade-off between how tightly a single-domain antibody binds and how much of it a cell makes; two from the third round bind below the test's floor.

Why it matters: This is a loop where the model's own ranking decides what gets made, and each round's measurements go back into it. Antibody sequences were not a domain of this genome model's original pretraining; they came in afterwards, as public repertoires and the partner's own measured designs. What is being tested is whether a general biological model can be turned to one partner's narrow design problem. Ranked on the same 327 measured designs, what the model has learned predicts expression and potency a little better than protein-specific models, with a plain one-hot encoding of the sequence close behind on potency.

Evidence and limits: A company blog post, no paper and no peer review, and both sides report their own work. One target, single-domain antibodies, three rounds, with the target's identity and some measured properties held back by the partner. Two designs beating the test's floor is a limit of the test, not a measured strength; the antibody post gives no number for that floor, and 0.2 nM comes from the launch post. The gains over the other models sit inside overlapping confidence intervals.

By Max Highsmith, Michael Poli

What happened: Armin Thomas, Gautam Machiraju, Benjamin Siranosian and Christine Parthemore report it on the company's blog. Livermore measured how 609 natural proteins split their binding across the 15 rare earth metals, on a plate assay it built this year. Given two of the metals, Omnii-1 ranked proteins it had not seen by which one they favour, at a median rank agreement of 0.340 over all 105 pairs, against 0.318 for the strongest public model, ESM-C 6B, and 0.254 for the best of three methods that go by sequence similarity.

Why it matters: Telling these metals apart is a refining problem: China refined 91% of the world's supply in 2024 on the International Energy Agency's count, and the elements differ so little in size, 0.013 angstroms between neighbours, that the standard route needs harsh acids and a lot of energy. Proteins can distinguish them, and the question is whether a model can say which protein will, before anyone makes it. This model had no rare earth binding data in its training, and the lab is now testing the candidates the model generated and ranked.

Evidence and limits: Prediction only: the proteins the model designed are at Livermore awaiting testing, and no result from them is reported. The company's post, its own evaluation, no paper and no peer review. On the 17 pairs closest in size, the hardest cases, the company puts the gap to ESM-3 at 0.006 against a paired error of 0.005 and declines to rank the two. AlphaFold2 scored 0.191 here, below the simplest sequence-similarity baseline, 0.198.

By Armin Thomas, Gautam Machiraju, Benjamin Siranosian, Christine Parthemore

What happened: Matt Turk, James Golden and colleagues at Protege DataLab and Heimdall Bio ran 4,249 rollouts of four frontier models, plus 4,486 trials across 26 models. Claude Opus 5 (0.816), GPT-5.6-sol (0.812) and Grok 4.5 (0.785) are not separated by the intervals; Gemini 3.1 Pro Preview scores 0.707 and every gap to it excludes zero. Opus 5 averages 0.907 on metadata lookup and 0.716 on numeric calculation. Numeric tasks are 28 of the 131 but six of the twelve tasks all four models fail.

Why it matters: A researcher picking an agent to measure distances and angles in a structure needs the numeric score, not the overall one, and for Opus 5 the two differ by 0.19. The paper also shows how much of a leaderboard is the setup around the model: on the 99 tasks with eight full runs, one run per task instead of eight puts a different model first in 23.9% of resampled task sets, and swapping each provider's own command-line agent for the reference harness moves scores by -0.071 to +0.086.

Evidence and limits: The four experiments use different task subsets, so model, harness, access and task mix cannot be ranked against one another, and the harness comparison can only detect differences of about 0.12 to 0.20. A task all four fail is not shown to be unsolvable. The 131 tasks lean to biophysics, and the scores describe those model endpoints in the collection windows the appendix gives. A preprint, not peer reviewed.

By Matt Turk, Brandon L. Neel, Diana M. Lopez, James B. Golden

What happened: Manuel Gallone, Shyam Ravidath and Srihith Madireddy built an out-of-distribution harness (two datasets, six dataset-split combinations, scVI, CPA and a Ridge baseline, three uncertainty methods), then retrofitted null baselines it had never included. The harness fed each test cell to the model and scored it on reproducing that same cell, so an identity predictor reaches a perturbation-effect correlation of 1.0 everywhere. Against a constant train-mean predictor, scVI wins on 99.7% of Norman cells and 0.2% of Replogle cells.

Why it matters: Perturbation prediction is one of the directions this record follows most closely, and the audit says metric names there can outrun what is measured. The six failure modes are mechanical rather than statistical (scoring reconstruction as prediction, holding out sequencing batches instead of perturbations, silently swapping an unseen perturbation for a control), and each comes with a diagnostic that needs no retraining, so another harness can be checked largely from what it already stored. The one model never handed the target cell, the linear baseline, is the only one that beats its null on a genuine counterfactual.

Evidence and limits: It is the authors' own harness, not a published benchmark; PerturBench is cited but never audited. The audit needed no retraining, though models were re-run counterfactually, and everything is computed on two datasets with no new experiment. One result survives: split-conformal coverage holds at 0.900 in distribution and falls to 0.847 across cell types, while on Norman that nominal 0.900 covers 0.271 to 0.342 of differentially expressed genes.

By Manuel Gallone, Shyam Ravidath, Srihith Madireddy

What happened: Joseph Hart, Samuel G. H. Brookes, Fabian L. Thiemann and Christoph Schran screen six models on density-functional energies, forces and binding curves, then put four through melt-quench-anneal dynamics at seven densities, five runs each. Every pure-carbon run finishes. Adding oxygen and hydrogen separates them: MACE-MH-1 completes all 35 runs, ORB-v3 33, MACE-MP-0b3 25 and MACE-OMAT-0 18. UMA never reached this stage (its carbon-dimer curve has two spurious minima) despite low errors on flat-carbon environments.

Why it matters: A foundation potential is sold on transferring to systems it never trained on, and the usual evidence is a table of energy and force errors. Here those errors do not say which models break: UMA scores well on flat carbon, then is thrown out for a dimer curve no potential should have, while models sharing a training set share ring statistics across architectures. All four that ran reproduce the main experimental peaks for nanoporous carbon, so the question is where they stop, not whether they work.

Evidence and limits: One element family and one workflow. Stability is scored in 216-atom cells; the structures compared with experiment are larger, at one density, and matched to published measurements of different samples. The first author co-authored both the model that comes out best and the scoring framework the error table uses, and the paper declares no competing interests. Two of the six never reached dynamics, one on cost alone. It is a workshop paper.

By Joseph Hart, Samuel G. H. Brookes, Fabian L. Thiemann, Christoph Schran

What happened: METiS announced a strategic collaboration with Lilly on 19 October to research lipid nanoparticle delivery that reaches one tissue rather than everywhere. METiS runs its own AiLNP platform, which designs the ionizable lipid inside the particle, and Lilly provides its drug discovery, development and commercialisation expertise. METiS takes an upfront payment of over $10M, is eligible for up to about $1.1B in development, regulatory and commercial milestones, and single-digit tiered royalties on net sales.

Why it matters: The record holds METiS's Hong Kong listing, its mouse data for a liver-cancer mRNA drug and a Phase 3 formulation result. This deal is about the platform rather than one drug asset: Lilly is paying for the design step, which is where METiS says its AI sits. Where a nanoparticle ends up in the body decides what an mRNA or gene therapy can treat, and it is still mostly found by trial and error.

Evidence and limits: A deal, not a result: no target, no programme count, no timetable, and nothing public about how well AiLNP picks lipids. All the money beyond the upfront is contingent on programmes advancing, and the upfront is stated only as a range.

What happened: A Zhejiang University team ran DFT calculations, about a million CPU core-hours, on 45,909 frameworks from six databases, compressed the resulting fields about 26-fold and turned them into point clouds. Models reading the fields beat a structure-only model on 12 of 13 properties, all the adsorption and separation ones among them; the structure model kept band gap. Combining the two did better still on every property. Electron density predicted from structure is off by 0.51% on average, and starting DFT from it cut run time by 13.59%.

Why it matters: MOF models mostly learn from atom positions, while gas adsorption depends on how a guest molecule feels the framework's charge. Aligned electronic fields at this scale were scarce because they are costly to compute and store; the authors say the dataset will be released when the paper is published. The results say the fields carry information structure-only models were not using, most of all for adsorption and separation.

Evidence and limits: All computed, by the authors. The structure baseline is one model, CGCNN from 2018, so the comparison with structure rests on it. The electrostatic potential is reconstructed far worse than the density, 8.42% off and worse in large cells. The 13.59% saving leaves out the time spent predicting the density, and is an average: some frameworks ran slower.

By Zhaolin Hu

What happened: James Kirk, Lee von Kraus and colleagues trained a Transformer, without labels, on over 4 million hours of recordings from electrodes in more than 1,400 plants of 100 species, in labs, greenhouses, orchards and forests. A second part, trained on over 41,000 handheld leaf readings, estimates seven stress measures such as how efficiently leaves photosynthesise and how warm they run. On held-out plants of six species under drought, its estimates moved with the readings at correlations of 0.27 to 0.71.

Why it matters: Earlier models of plant electrical signals were each trained for one task, usually on one species. One model that carries across species would let a grower read a crop nobody has labelled. On species left out of training the estimates still moved with the readings, less closely, and on plants stressed by salt or a herbicide, which the model never saw labelled, they shifted the way the literature expects.

Evidence and limits: The data, sensors and evaluation are all the company's own, in a preprint. The correlations follow one plant over time; about half the error is a fixed offset per sensor, and on the error measure, three of the seven estimates did worse than guessing each species' average. One of the six densely labelled species read at or below zero, and nothing yet shows a grower acting on it.

By James Kirk, Piotr Bednarski, Filip Twardy, Justin Mast, Lee von Kraus

What happened: The two endpoints return ranked web results with a relevance score and an authority tier, and a page's content as Markdown. DeepSeek says its in-house engine has long served the search in its web app and API, handles hundreds of millions of requests a day and indexes hundreds of billions of pages, with filters for SEO spam, GEO manipulation and low-quality AI-generated text. Calls made through the server-side tools in DeepSeek Harness, Claude Code and Codex stay free until 31 October.

Why it matters: A literature agent needs a search backend it can call by program, and almost all of them sit on Bing's or Google's index. This is one of the few that does not, from a lab whose models many Chinese AI4S groups already run, priced per call rather than by tier. Whether its index covers journals, preprint servers and Chinese-language sources well enough for science is for anyone building on it to test.

Evidence and limits: The request and index figures are DeepSeek's own and unaudited; the English release note gives the daily requests and the Chinese one the index size. No coverage or freshness comparison against other search engines was published, and the endpoints were not called here.

What happened: Zachary Cutts, Pooja Rao, Annalisa Pawlosky and colleagues gave Gemini-based agents literature search, folding tools and a Python environment, and had them design 1,000 single-stranded DNA aptamers per round against two proteins. Each pool was synthesised, pulled down on beads and sequenced, and the agents read the counts before designing the next. Corrected for bead background, hit rates went 0% to 26% to 42% on alpha-thrombin, or 37% counting only designs free of published motifs, and 10% to 4% to 34% on myoglobin, at about $3 a candidate.

Why it matters: Earlier studies of reasoning agents in molecular design tested their designs once, at the end. Here the measurements came back three times, and what the agents did with them is on the record: after a counter-screen showed that 97% of the cross-reactive sequences were G-quadruplexes, which stick to charged surfaces whatever the target, they moved the myoglobin designs to pseudoknots and hairpins, and specificity there went from 19% back to 72%. The authors picked aptamers because the designed sequence is the molecule, so one cycle is 10 days instead of the 3-6 weeks a protein needs.

Evidence and limits: Binding is measured in the authors' own pooled assay. Single-clone titrations covered eight thrombin designs only, did not reach saturation, and the authors call those affinities preliminary: nanomolar for reworked literature motifs, above a micromolar for de novo ones. Round 1's 0% on thrombin was a construct artifact the agents then fixed, the harness changed after it, and the authors say control arms with no feedback would be needed to credit the loop itself. A preprint, not peer reviewed.

By Zachary Cutts, Pooja Rao, Annalisa Pawlosky

What happened: Tatsuki Hori, Yoshitaka Moriwaki and Ryuichiro Ishitani at Institute of Science Tokyo and the University of Tokyo release RGI-Toolkit, which corrects the denoiser output at each reverse-diffusion step with a differentiable restraint loss, as a backend-agnostic library wrapping six predictors: Boltz-2, Protenix-v2, AlphaFold3, Chai-1, OpenFold3 and ESMFold2. On 209 protein-ligand complexes, conformer restraints lifted chirality agreement from about 50% to 100% and cis/trans from 70-80% to 100% for all six, with lower bond-length and bond-angle errors. No retraining. MIT-licensed.

Why it matters: Chemically invalid ligand geometry is a known weak spot of AlphaFold3 and its kin, and this fixes it from outside the model, at inference, without retraining it or changing a weight. It does not touch where the ligand sits: the authors report pose accuracy is neither helped nor hurt, so the gain is a clean stereochemistry patch, not a better binder prediction. Geometric and custom restraints can also push sampling toward a chosen conformational state.

Evidence and limits: The chirality and cis/trans figures are computed agreement rates against each ligand's ideal conformer, on 209 complexes with one seed each, not a physical measurement. The restraints only work as well as the reference conformer: a bad RDKit embedding misguides them. The two earlier RGI papers this builds on were in Boltz; the new work is extending that to six predictors and packaging it. A preprint, not peer reviewed.

By Tatsuki Hori, Yoshitaka Moriwaki, Ryuichiro Ishitani

What happened: Scripps Research, OpenAI and PerturbAI gave one agent to each of 1,946 perturbations in an in vivo whole-brain CRISPR Perturb-seq atlas, with the source manuscript withheld and published after the model's stated knowledge cutoff. Of 55 findings derivable from the expression data alone, the agents matched or refined 48. They adjudicated 1,519 high-confidence findings against the literature, reading 39,185 papers and keeping 16,445 citations. The run took 32 minutes.

Why it matters: The bet is coverage through many parallel agents rather than one stronger agent, and the paper measures it directly: at 893 perturbations a single agent cited 11 references while the distributed run kept 711 times more. Blinded experts preferred the distributed readings. This is a throughput claim about interpretation, the step that now limits what a large screen yields, not a claim that the agents found new biology.

Evidence and limits: All computed, on one atlas the same group produced, and every corresponding author has a stake: one at OpenAI, whose model this evaluates, two who co-founded PerturbAI, which hosts the data. Recovery is scored by the source study's own authors, 14 of 71 findings uncovered. Held-out predictions reach 0.705 balanced accuracy: 82% of the silent perturbations called, only 59% with a real response. The expert preference ranks readings, not their correctness. A preprint, not peer reviewed.

By Alexander Sun, Xin Jin

What happened: Andi Gu, an Anthropic STEM Fellow, with Alek Kemeny of Anthropic and Harvard's Susanne Yelin and Mikhail Lukin, had Claude Mythos 5 write circuits in Python, each checked and scored by an evaluator. In six tasks the circuits cut both Toffoli and qubit counts against the best published design; a seventh had no published reversible design to beat. In the eighth, a new reversible modular inversion, its key step formalised in Lean, gives one design 17 times fewer gates than the smallest previous one, for 12% more qubits.

Why it matters: Most of the fall in what useful quantum chemistry is expected to cost has come from redesigning circuits, not hardware: the FeMoco estimate dropped about 150,000-fold over eight years of human work. A loop where the model writes circuits as code and a program checks every one speeds that redesign up: Anthropic sets the elliptic-curve design, from a single 53-hour run, against eight years of field-wide work. The same loop cut the costliest step of the DQI optimisation algorithm, its decoder, about 12-fold.

Evidence and limits: All of it is computed: the circuits have not run on any quantum computer, the counts come from the team's own evaluator, and the arXiv preprint is not peer reviewed. People wrote the briefs and at times told Claude to search more broadly.

By Andi Gu, Alek H. Kemeny, Susanne F. Yelin, Mikhail D. Lukin

What happened: A model could be blocked only on a foreseeable, imminent risk of over $1 billion in damages or of killing or seriously injuring more than 50 people, from a biological weapon, a model evading its developer's controls, or a cyberattack. Commerce would ask a federal judge in Washington to order it; no other agency could. Evaluators would register with Commerce, may be lab-paid if the fee does not turn on their findings, and must report an imminent risk within 72 hours. Some state AI laws would be preempted.

Why it matters: A frontier lab has written the text that would govern it. Politico notes its threshold language mirrors California's SB 53, and Anthropic calls the bill its Advanced AI Framework, pushed since June. For science the category that matters is biological weapons, one of three grounds on which a model could be stopped: an evaluator's report to Commerce would be the first step toward a court order there.

Evidence and limits: The draft is confidential and Politico did not publish it, so every provision here is Politico's account of a text it obtained; three people with knowledge are its sources for the pitch itself. The article does not name which lawmakers received it. An unnamed Anthropic spokesperson endorsed the bill as the company's framework without publishing the text.

What happened: Google DeepMind fitted its own capability index, GCI v1, to 123 models on 92 benchmarks. Public Gemini models gain about 0.1 points a month: 0.091 before 1 April 2026, 0.118 after, a difference that is not statistically significant. Most production code is now over 75% AI-written, yet fully delegated pull requests rose only from 7.5% in April to 11.9% in September; about 88% still need multi-turn human steering or human-written code.

Why it matters: Whether AI is speeding up its own development now has a frontier lab's internal measurements behind one answer, set against the September working paper co-signed by Geoffrey Hinton, Yoshua Bengio and OpenAI's chief scientist, which estimated that fully automated research could make progress tenfold faster. A survey of 12 senior researchers puts automation highest on well-specified, verifiable tasks and lowest on open-ended scientific ideas and research taste. Argon sits above the trend, which the report suspects is due to more compute and organisational investment.

Evidence and limits: The index is DeepMind's own and built on benchmarks: a steady slope means measured capability grew steadily, not that no research step got faster, as the report itself notes. The post-April slope rests on few models (0.053 to 0.171 at 95%), so it cannot rule out a speed-up either. The survey and the agent-use sample, graded by Gemini Flash 3.8, are internal, and the finding supports the report's own conclusion that the risk is low.

What happened: Vinay Saji Mathew and colleagues at Penn State and the University at Buffalo distilled the diffusion stage of AlphaFold 3 and Boltz-2, the part that turns a predicted layout into atoms, into an eight-step student. On about 2,300 Boltz-2 benchmark targets, the AlphaFold 3 student scores 0.8635 complex lDDT against its teacher's 0.8615, and predicted ligands passing PoseBusters' validity checks fall from 45.0% to 43.0%. For Boltz-2 the cost is larger: ligand validity falls from 60.9% to 52.0%, with structures carrying any geometry violation rising from 30.3% to 41.9%.

Why it matters: Protein design and screening pipelines call a structure predictor tens of thousands of times, so the sampling stage's cost adds up. The paper also argues that speed-up claims have been scored too loosely: earlier few-step samplers were judged on similarity scores with the chemistry checks switched off, which stay high while bonds, angles and clashes go wrong. Blitz reports both, under one protocol for every method, and shows a one-step sampler losing about 32 points of lDDT once the checks are on.

Evidence and limits: Computed, by the authors, on public benchmarks. Only the sampling stage gets faster: the paper says the trunk, the network that runs before sampling, is unchanged and still costly, so eight steps is not a 25-fold faster prediction. No end-to-end timing is reported.

By Vinay Saji Mathew, Soundar Kumara, Gretta Kellogg, William K. M. Lai

What happened: Jiaru Bai, Alán Aspuru-Guzik and colleagues at the University of Toronto describe agents that start from the equipment in a room, rebuild it as a simulated twin, learn missing robot skills by trial, first in simulation, and write the workflow. In the one physical test, after a human approved the plan, the system wrote a 236-command liquid-handler protocol, had the arm carry the plate to a sealer and then a plate reader, and an 8-hour cadmium sulfide quantum-dot screen found the ligand and pH range where emitting dots formed.

Why it matters: A self-driving lab today runs only the procedures it was built for, and adding a new one takes specialist engineering. If agents can build the missing pieces, what a lab can try is limited by its equipment rather than its code, and researchers set goals instead of programming robots. The authors also turn each solved skill into a script that runs without the language model, because the agent loop was too slow for the chemistry.

Evidence and limits: One physical workflow on known chemistry, run by the authors; coordination across several robots was shown in simulation. The paper gives no count of human interventions or of engineering time saved, and says people still had to settle disagreements between agents and supply physical facts the instruments could not sense. It is a preprint.

By Jiaru Bai, Yang Cao, Varinia Bernales, Alán Aspuru-Guzik

What happened: Sifted reported on 15 October that Accel is set to lead a $60M round for Revolution Labs, founded by Nando de Freitas after about a decade at DeepMind and a stint at Microsoft. Dealroom's summary of the report adds that GV and NEA are said to have joined. The company builds diffusion models, extending them from images and audio to text. In September Bloomberg had reported talks to raise at least $100M initially.

Why it matters: The September figure was a target in talks; this is the first report of a round sized and led. At $60M, the first close is smaller than the $100M+ floor reported then, though Bloomberg also said the raise would likely come in several tranches. It is one more DeepMind alumnus raising for a bet on a different kind of model than today's language models.

Evidence and limits: Media-reported and not closed: Sifted's headline says Accel is set to lead, and Sifted's text could not be read here. GV and NEA come only from Dealroom's summary. No stage, valuation or company statement; no model has been published.

What happened: insitro will use its Virtual Human platform to find disease-causing biology in both diseases and take targets AstraZeneca nominates from discovery to a development candidate. The collaboration is worldwide and exclusive, and AstraZeneca runs clinical development, regulatory work and sales. Beyond the $40M upfront, insitro can earn up to about $1.46B in discovery, development and commercial milestones, plus single-digit royalties. Daphne Koller, insitro's founder and CEO, announced the deal.

Why it matters: A large drugmaker is paying for an AI company's reading of disease biology at the target stage, the earliest and riskiest decision in drug R&D, and not only for molecule design: insitro's platform looks for what causes the disease, then takes targets AstraZeneca picks through to a drug candidate. insitro already works with Bristol Myers Squibb, Lilly and Gilead; AstraZeneca is the fourth.

Evidence and limits: The $1.46B is a ceiling that depends on milestones being met; only the $40M is certain. The release names no target, no number of programmes and no timeline, and gives no data from the platform.

What happened: The two plan a long-term partnership on foundation models for science, starting with high-energy physics. The goal is a single pretrained Large Physics Model that could reconstruct collisions, identify particles, flag unexpected signals and help evaluate future detector designs: jobs now spread over many separate tools; reconstruction alone runs through a chain of specialised algorithms. Enrica Porcari, CERN's chief information officer, and Johannes Brandstetter, Mistral's vice-president of AI for engineering, spoke for the two.

Why it matters: Particle physics has used machine learning for years, but one model per task. The plan is for a frontier model company to learn across the whole reconstruction chain, at the lab holding about an exabyte of LHC data, before the High-Luminosity LHC raises the collision count tenfold around 2030. Brandstetter co-founded Emmi AI, the physics-simulation start-up Mistral bought in May, and worked at CERN before that.

Evidence and limits: A statement of intent, not a contract: CERN's release gives no budget, timeline, compute figure or terms on who owns or can use the model. Nothing has been trained yet, and 'the first foundation model for collider physics' is the release's own description.

What happened: OpenAI's terms: up to 30 ChatGPT Business Standard seats at no charge for up to twelve months, for eligible labs at US universities, with research data on those accounts kept out of model training by default. A seat does not unlock GPT-Rosalind. That needs a separate trusted-access review of the lab head's identity, employment and authority, the research purpose, and a named institutional safety contact. Subscription seats only, so no API, and a university holding an Enterprise agreement applies through the school.

Why it matters: Until now the model went to organisations through trusted access, with Amgen, Moderna, the Allen Institute and Thermo Fisher as its named partners. This is the first way in for a university lab on its own, and it says where OpenAI expects life-science demand to come from next. Without the API it stays something a researcher types into rather than something a lab wires into its own pipeline, which is the difference between a tool and infrastructure.

Evidence and limits: What counts as an eligible lab, how long the trusted-access review takes, how many labs will be admitted and what happens to access after twelve months are all unpublished. Benchmark scores for GPT-Rosalind are circulating from a single secondary source and are not on OpenAI's page, so they are left out here.

What happened: Gideon Shaked and Siddharth Nagisetty, both at the University of Michigan and AfterQuery, build each task as a directed acyclic graph of subtasks over a real biological dataset, posed one per conversational turn and scored against an answer key by 58 graders that match entities through reference databases, not by string. All twelve models ran on one harness, two on two more. Means ran 0.43 to 0.57 with no feedback and 0.54 to 0.67 when a failing answer could be retried; the slowest agent averaged 3.6 hours a task.

Why it matters: Most bioinformatics benchmarks ask one question in one prompt, which cannot show whether an agent carries intermediate results through a real analysis. Here every combination scores lower after its first turn, and lower again on subtasks that build on its own earlier answers, with feedback winning back only part of it. That shape of failure is exactly what a single-prompt score cannot see.

Evidence and limits: Scores measure agreement with an answer key, not whether the analysis is biologically sound. The authors call the retry mode open to reward hacking by design and do not quantify how much of its gain that explains. 94 of the 97 tasks are private, so the numbers cannot be reproduced from outside, and harness effects cannot be separated from model effects. The failure labels come from Opus 5, one of the models judged. A preprint, not peer reviewed.

By Gideon Shaked, Siddharth Nagisetty

What happened: Zone II and CC.VC backed it, with angels the release places in finance and Fortune in defence and finance. Fortune reports the round is all equity and values Holyvolt at $245M, and later struck Fifth Wall from its own list of this round's investors after what it calls a disclosure error by the company; the release names it as an existing investor. The lab is Wildcat Discovery Technologies, bought in 2026: 18 years of runs, more than 500,000 cells cycled. Mathias Ingvarsson is CEO, Steven Prast, formerly of Tesla, COO.

Why it matters: Most AI-for-science companies build a model and then look for data to feed it. Holyvolt bought the lab and the 18 years of cycling data first, then trained on them, and its models now pick what the lab makes and cycles next. The timing comes from procurement: from 1 October 2027 the Pentagon cannot buy batteries from CATL, BYD and four other named Chinese makers, and Holyvolt plans to sell its DRX cathode into drones first.

Evidence and limits: Nothing here is a result. The ten-times figure comes from one customer case study, and the hundred-times figure for the closed loop is the company's own. DRX data has been shown at scientific conferences, the International Battery Conference in March 2026 among them, and the suppliers are undisclosed. The $245M valuation rests on one person Fortune does not name.

What happened: Ibuki Okuda, Izumi Takahara and Teruyasu Mizoguchi extend Microsoft's MatterGen so one network denoises the crystal and, for every atom, either its Bader charge (how much charge it carries) or its atomic density of states (which energies its electrons occupy). Of 1,024 structures per setting, counting only stable, unique, novel ones, 1.1% of the baseline's landed inside both a 2 eV band gap and a -4 eV per atom formation energy, against 3.5% of the density-of-states model's. Single targets gain more in points and less in proportion.

Why it matters: Asking a generator for a target property costs structural quality, and the electronic variable makes it cost less: at that two-target setting the share of stable, unique, novel structures falls from 33.3% to 21.8% with the descriptor and from 33.6% to 15.4% without it. The control experiments are the part to carry away. Corrupting the charges, or attaching a random number that means nothing, also raised the unique-and-novel share, so a novelty gain on its own shows only that something random was added, and it takes stability, validity and the target hit rate to tell physics from noise.

Evidence and limits: All of it is computed: structures are generated and scored by DFT, and nothing was synthesised or measured. A hit means landing within 0.5 eV, or 0.5 eV per atom, of the target. The validity filter is the authors' own caveat (about 30% of their training structures fail its charge-neutrality check), so both the stricter and the looser counts are reported. Several single-property gains are a few points, and some do not reach significance at 1,024 samples.

By Ibuki Okuda, Izumi Takahara, Teruyasu Mizoguchi

What happened: James Weber (University of Reading) and Makoto Kelp (University of Utah), equal authors, fine-tune Microsoft's Aurora on the UKCA chemistry-climate model, in three variants. The first reproduces 23 of 27 evaluated fields at R-squared of at least 0.85 in troposphere and stratosphere, with annual mean biases under 4% for ozone, CO and temperature; a simulated year takes 5.5 minutes on one GPU against 9.5 hours on 576 CPU cores. A second, trained on paired emissions scenarios, reproduces ozone's response to NO and CO. A third is trained on satellite ozone.

Why it matters: Chemistry is the most expensive part of a climate model, so it is usually simplified or fixed in advance, and trying many emissions scenarios needs it fast. The third variant is the interesting one: trained on Aura MLS satellite ozone and tested on years held out, it cut the error in ozone averaged around each latitude band by 23%, from 0.37 to 0.28 ppm, roughly 10 to 25 km up. That correction sits in the weights, so it carries to runs where no observation exists: an emulator can be better than the model it learned from.

Evidence and limits: Most of what is measured is agreement with UKCA, so a well-reproduced field inherits UKCA's own errors, and the species predicted are a subset the authors chose of UKCA's 100-plus. The satellite gain holds only in the latitude-band average: the same model is 9.6% worse at individual retrievals. The emission-response variant has its winds and temperature prescribed from the control run, and the 50-year run recycles one year of forcing.

By James Weber, Makoto Kelp

What happened: Hojae Choi, Samuel Pellock and David Baker at the University of Washington's Institute for Protein Design, with Donald Hilvert at ETH Zurich, report enzymes that use an activated cysteine to hydrolyse the polypeptide backbone in a sequence-dependent way. 13 of 69 designs worked, 19%, with rate enhancements over the uncatalysed reaction of up to 3x10^7. The folds are unlike natural proteases (TM score below 0.50), six crystal structures sit within 1.2 A of the design models, and the designs cleave their targets in mammalian cells.

Why it matters: A peptide bond in neutral water lasts hundreds of years, and de novo enzyme design had mostly managed reactions with low barriers. Cutting one where the designer chooses turns protein design into a way to make programmable scissors: pick the sequence, get an enzyme that cuts it. That is a different kind of tool from a binder, which is what designed proteins have mostly been.

Evidence and limits: Activity is still short of natural cysteine proteases: the best design, after optimisation, reaches 413 M-1s-1 against 350,000 for the Ulp1 the motif came from, which the paper puts down to imperfect active-site preorganisation and substrate positioning. In cells the proteases cut a substrate the authors designed and transfected in with them, not an endogenous protein, and the work is the designing lab's own.

By Hojae Choi, Samuel J. Pellock, Donald Hilvert, David Baker

What happened: Anqi Chen, Kejia Wu and David Baker at the same institute, again with Donald Hilvert at ETH Zurich, published the same day, used RFdiffusion to put a water-activating zinc site next to a peptide-binding pocket. Of 135 designs tested, 36% cut their target bond; the best beat the uncatalysed reaction by more than 10^8-fold, and four substitutions took it past 10^10. Designed proteases selectively cut the disease proteins TDP-43, amyloid-beta and serum amyloid A, up to 10^9-fold, and switch on caged cytokines.

Why it matters: The disease targets are the point: TDP-43 and amyloid-beta are proteins drug makers have tried to clear, and cutting one is a different kind of intervention from binding it. With the cysteine-protease paper beside it, the lab now reports two independent chemistries for the same hard reaction, one metal-based and one not.

Evidence and limits: Rate enhancements are measured against the uncatalysed reaction, not against a natural protease: out of the computer the designs run 10^3 to 10^4 below the best natural ones, two orders after four substitutions. The disease proteins were cut as purified protein in vitro, and nothing reaches a cell interior, an animal or a patient. The designers have filed two provisional patents on the methods and designs.

By Anqi Chen, Kejia Wu, Donald Hilvert, David Baker

What happened: The system pairs ACE, which runs the atmosphere at 100 km, with a new downscaler, HiROv2, that sharpens it to 3 km and now predicts several fields at once rather than rainfall alone. Given coarsened X-SHiELD data it matches that simulation's high-resolution statistics to within 3%, even for events that occur once in a million data points, and estimates tropical-cyclone energy for storms it never trained on. Over a 10-year run the hurricane energy is comparable to X-SHiELD's. The models, notebooks and a comparison demo are public.

Why it matters: Work on extreme-event risk needs many years of storm-resolving data, and Ai2 puts generating it with X-SHiELD at roughly 300 times the cost. At about a three-hundredth of the cost, a decade of global 3 km rain and wind becomes something a research group can generate itself, several times over, which is what estimating rare events takes.

Evidence and limits: Every number is measured against X-SHiELD, the simulation HiRO learned from, not against observations. Ai2 says HiRO only works with the ACE version tuned to X-SHiELD, so it cannot yet produce a warmer climate or reanalysis like ERA5, and winds and pressure around hurricane eyes stay more lopsided than in X-SHiELD.

What happened: A team led by Frances E. Reuland and Adam R. Brandt at Stanford released metered methane in Arizona as satellites passed overhead: 481 releases of 9.5 to 1,184 kg an hour and 265 blank controls, with the teams not told which was which. Fifteen teams from nine countries made more than 2,250 estimates from 12 government, commercial and non-profit satellites. Instruments built for methane detected 92-100% of releases with 0-11% false positives; repurposed ones 62-88%. Teams applying AI methods to Sentinel-2 and Landsat images made 79% of all false positives.

Why it matters: Satellite plume reports already feed leak alerts and national methane rules, and the EU's proposed import standard for gas leans on them, yet Carbon Mapper's Tanager-1, EMIT, GHOSt and the AI detection methods had had little or no blind testing. This one gives each a measured detection rate and error against known releases, and leaves an open dataset for training the next detectors.

Evidence and limits: Teams using AI scored 0.62-0.77 accuracy, better than chance, but made most of the false alarms, and the authors say AI methods need more work before routine use. Releases were at a known site: the authors' conclusion holds when the source location is roughly known, and cloud cover over 20% degraded every system.

By Frances E. Reuland, Adam R. Brandt

What happened: Iambic sold 9,375,000 new shares at $16 for $150M gross, and the underwriters can buy 1,406,250 more within 30 days. The shares started trading on 15 October as IAM, opened at $16.15 and closed at $18.91, 18% above the offer price; on 16 October they closed at $23.81. J.P. Morgan, Jefferies, BofA Securities and Citigroup ran the books. Tom Miller, co-founder and CEO, and Fred Manby, co-founder and CTO, lead the company.

Why it matters: Public investors have now put a price on an AI drug-discovery company whose most advanced drug, the HER2 inhibitor IAM1363, is still in Phase 1. The prospectus puts the money, with cash on hand, toward finishing that trial and starting Phase 2 and 3 trials, taking IAM217 and IAM-C1 through Phase 1 dose escalation, and the AI platform, and says it lasts into the fourth quarter of 2028. By the prospectus's share count, $16 a share values Iambic at about $760M.

Evidence and limits: $150M is before fees; the offering was due to close on 16 October and the extra-share option is unexercised. The price says what investors will pay, not whether Iambic's AI-designed drugs work.

What happened: Minos turns Leash's DNA-encoded-library screens, millions of molecules measured against many proteins, into environments where an agent chooses molecules within a budget and is told what was measured. The flagship task gives no starting molecule and a budget of 100 molecules over 20 rounds, across 33 targets with about 6 million molecules each. Leash reports about 100,000 agent trajectories in six weeks, releases a few environments and thousands of reasoning traces, and finds that stronger models do better and weaker ones stay too long in poor regions of chemical space.

Why it matters: Because every molecule was measured in advance, an agent's choices cost compute rather than lab weeks, and the results of the molecules it did not pick are known too, so its decisions can be judged, not only its final answer. Leash CEO Ian Quigley's argument is that data collected this way is reusable experience, in the way game moves are for a game-playing model.

Evidence and limits: All results are Leash's own; agent performance is shown in figures this page did not read values from, and the preprint was not opened. The environments cover binding to one target among molecules already screened, with no pharmacokinetics or toxicity, and the published ones are a small sample of what Leash holds.

What happened: Probabilistic bias correction (PBC) learns from a forecast system's past predictions and the weather that followed how to shift its probabilities. Over 2016-2024 it improved ECMWF's already debiased dynamical model on 91% of pressure, 92% of temperature and 98% of precipitation targets; on the 2025 test year it doubled the skill of ECMWF's AI system. Entered as MicroDuet in ECMWF's AI Weather Quest, it scored best for every variable and both lead times in the September-November 2025 season, and best again in the winter season that followed.

Why it matters: Two to six weeks ahead is where farmers, grid operators and flood planners need a forecast and where models do worst, often no better than the long-term average for the date. PBC is a layer that works on top of any model, and its contest scores come from forecasts issued before the weather happened, set against operational centres' own systems. Hannah Guan and Soukayna Mouatadid led it, with Judah Cohen of MIT and Lester Mackey of Microsoft Research; Cohen says Kenya's meteorological office has asked about using it.

Evidence and limits: The contest covered two seasons, and ECMWF, which runs it, has two authors on the paper. For precipitation, all six centres' debiased forecasts scored below the climatological average, so the gain starts from a low base. Better weekly-average scores do not yet show a decision made better, and the paper reports no operational use yet.

By Hannah Guan, Soukayna Mouatadid, Judah Cohen, Lester Mackey

What happened: Danaher's release says the two are working to combine its instruments, automation and lab workflow software with NVIDIA's computing: edge hardware such as DGX Spark, RAPIDS for analytics, and real-time inference and foundation models including the BioNeMo Open Models. Kimberly Powell of NVIDIA states the loop the release describes, experiments improving models and models designing experiments; Julie Sawyer Montgomery of Danaher says AI is reshaping how science is done. Initial efforts cover a select group of priority instruments.

Why it matters: Danaher sells instruments that sit in a great many labs, so where its boxes compute matters: an autonomous lab needs the measurement and the inference close together, and edge hardware at the instrument is one way to get there. The release names the parts rather than a product. It stays an intention until something ships or a customer says what changed.

Evidence and limits: No money, no equity, no timeline and no product. Which instruments are in the select group is not named, the wider portfolio is only a possibility, and the Cosmos world models and Omniverse digital twins are described as being explored. NVIDIA has issued no release of its own. Danaher's separately announced autonomous antibody lab is still at the plans stage, expected to run at scale in early 2027.

What happened: Across 15 targets and 3,666 designs, folding only the contact shell plus a margin reproduced the full target's interface to a median 0.37 angstrom, and cut GPU hours by 88% on the 719-residue outer part of fibroblast activation protein alpha. With a researcher setting the strategy and a Gemini 3.1 Pro agent choosing the crops and parameters, the in-software hit rate rose from 5.06% to 9.33% while saving a further 42%. Yeast display then gave four clones at apparent 3.0 to 9.1 nM.

Why it matters: AlphaFold3 charges by token bucket, so most of the compute in screening designs goes on target residues far from the binding site, and a big outer domain is what makes a target impractical to design against. The paper treats cropping as something to test rather than a free saving: whether the interface geometry survives, whether the ranking that separates binders from non-binders survives, and whether the saving crosses a bucket boundary. Left to pick its own strategy from a map of the whole target, a fully autonomous arm converged on the same catalytic-rim approach the researchers had specified.

Evidence and limits: Cropping cost ranking: mean interface ROC-AUC fell from 0.777 to 0.729, and on one target the ranking collapsed, recovered only by dropping the target's alignment, which helped deep alignments and hurt a shallow one. Seven of the 15 targets were small enough to gain nothing. Affinities are apparent values on the yeast surface, not purified protein; one target carried both the wet lab and the agent arms, nobody outside repeated it, and this is a technical report.

By Jun W. Kim, Annalisa Pawlosky

What happened: Generate takes $30M upfront, up to $35M in early preclinical milestones and milestones of nearly $1.2B in all, plus tiered royalties, for two initial programmes with an option on a third. Generate runs research through lead-candidate selection; Zealand takes clinical development, regulatory filings and worldwide commercialisation. Utpal Singh, Zealand's chief scientific officer, says proteins complement its small molecules and peptides; Mike Nally, Generate's chief executive, says the teams will define the target profile first.

Why it matters: Obesity is where the metabolic money is, and Zealand is buying protein design rather than building it: its own 25 years are in peptides. For Generate it is the third pharma collaboration after Amgen in 2022 and Novartis in 2024, and the first since it listed on Nasdaq in February, which is the kind of revenue a newly public platform company needs.

Evidence and limits: The $1.2B is what could be paid if programmes reach development, regulatory and commercial milestones; $30M is what is paid now. No target, candidate or data is named, and the work starts at design.

What happened: New investors Delta-v Capital, Socium Ventures, HighSage Ventures, Coral Capital and Legion VC joined, and Coatue, J.P. Morgan Growth Equity Partners, Index Ventures and Applied Ventures followed on. CEO and co-founder Nick Talken says the money expands Albert's chemistry AI; Morgan Gerlak of TCV spoke for the lead. The partners the release names include AkzoNobel, BASF, Henkel, Kenvue and Clorox.

Why it matters: Albert sells to the makers of coatings, adhesives and consumer goods: AI that searches their own formulation records, suggests candidate formulations and designs experiments, rather than discovering new molecules. The evidence that its model helps is Albert's own: a white paper says that on the hardest set of a benchmark Albert built, its retrieval model Valence put the right passage from about a million drawn from US patents in its top 100 results 81.8% of the time, against 2-4.5% for OpenAI and Voyage models. No valuation was given.

What happened: Pawel Stankiewicz's lab at Baylor College of Medicine gave Gamow Labs de-identified whole genomes from 26 infants whose lung disease first-line labs had not diagnosed; the lab had solved 19. Gamow's George-0.1 pipeline, built on frontier models including Gemini 4 Argon, matched all 19 and flagged candidates in 2 of the other 7. In one, read evidence pointed to a balanced translocation separating FOXF1 from its enhancer, consistent with alveolar capillary dysplasia; Sanger sequencing in Stankiewicz's lab confirmed it.

Why it matters: The step forward is that a candidate the AI proposed went on to be confirmed in a lab: Baylor's own sequencing verified a chromosomal rearrangement that had gone undiagnosed for three years. That is the step from flagging a variant to a diagnosis a family can act on.

Evidence and limits: It is a small study, 26 cases, disclosed by the company in a post rather than a paper, and the outcome of the second candidate is not given. Argon is one of several models in the pipeline, so the result belongs to the pipeline, not to Argon alone.

By Daniel McKinnon

What happened: On 13 October Google opened Gemini 4 Argon to developers, enterprises and eligible subscribers, and invited 10,000 scientists to apply for six months of free access to Antigravity's top usage tier. Its science case studies cover enzyme design, quantum computing, fusion research and rare-disease genome analysis.

Why it matters: What stands out is less the model's scores than frontier AI beginning to leave traceable results in different research workflows. The fusion work has a public preprint, and the speed-up of the quantum chemistry package xTB has code submitted to its repository; Gamow Labs reports that its AI pipeline found a previously undiagnosed chromosomal rearrangement, which a partner lab confirmed by sequencing.

Evidence and limits: The cases rest on different kinds of evidence. The fusion paper is still a preprint; the three xTB pull requests are open, not merged, and a submission is not an independent check of the speed-up; Gamow's study is the company's own account, and Argon is one model in a multi-model pipeline. The enzyme activity and some quantum results lack public material an outside group could check.

What happened: Xiyao Yu of Nanyang Technological University shows that the usual test, force error on equilibrium snapshots, barely sees the barrier region that sets a rate, so worst-case rate error grows exponentially with barrier height. Weighting energy errors by a 'leverage density' concentrated on reaction paths instead gives an estimated interval for the rate. In alanine dipeptide, five designed perturbations of matched force error changed the rate 0.49 to 3.5 times; in Li3OCl the intervals ordered six models' ion-hop rates, but five measured rates sat about 7% above the upper end.

Why it matters: Machine-learned potentials are judged by force error but used to compute diffusion and reaction rates. This offers a check costing a few hundred reference energy calculations, reported as two numbers beside force error, and picking training data by it cut rate errors even as force errors rose. The Li3OCl agreement holds only once hops are sorted by mechanism; on the pooled set the intervals do not track. Every rate tested against is the author's own simulation, not experiment, in an unreviewed single-author preprint.

By Xiyao Yu

What happened: Fan Nie, Haotian Ye and colleagues split verification into proposing candidate errors and then judging whether each is real, with the threshold set so that almost no valid proof is wrongly rejected. Across three maths benchmarks that cuts the invalid solutions a reviewer must read by 46% on average against a voting baseline at the same false-rejection rate, and by 57% once the validator is trained. Pointed at Xinwen Zhu's 2017 Annals of Mathematics paper, it reported 45 issues where the same model reviewing directly found 11.

Why it matters: Expert review becomes the scarce resource once a solver produces more candidate proofs than anyone can read, and this treats that capacity as a budget: reject as much as possible while letting through almost every valid proof. The authors of the papers it checked confirmed the errors and revised. Paired with a solving agent it also helped settle three open conjectures in relative Langlands duality, which experts then confirmed.

Evidence and limits: The 11 preprints' main conclusions hold after the corrections. The confirmations come from the papers' own authors, and for the preprint credited with 88 issues one of them is Shurui Liu, an author of this paper; he also wrote up the three Langlands proofs, co-wrote the Manin constant paper and co-wrote the solver OSVerify is paired with. No formal proof checker is used anywhere.

By Fan Nie, Haotian Ye

What happened: Bloomberg, citing people familiar with the matter, says the round is worth tens of millions of dollars and would value Winverse at about $400M after the money, with Meituan's Long-Z Investments also committed. The deal is not final and its terms could change. Han Hu founded Winverse this year after leading Tencent's vision-language and embodied foundation model teams; he was earlier a principal researcher at Microsoft Research Asia. The company is still in stealth.

Why it matters: It shows a model lab putting money into the AI that lets robots act, not only into robot hardware: DeepSeek already took a stake in Unitree when the humanoid maker listed in Shanghai. Bloomberg's point is that China builds robot bodies well, while the models that turn instructions into reliable actions lag. Nothing has closed, none of the parties commented, and Winverse has shown no product or result.

What happened: Matt Rosseinsky of the University of Liverpool's chemistry department founded the company. It sits in the university's Materials Innovation Factory and pairs AI prediction with experimental and characterisation expertise, aimed at inorganic crystals for batteries and catalysts. IP Group led the oversubscribed round, with Northern Gritstone, the University of Liverpool Enterprise Fund, Norrsken Evolve, LYVA Labs and the Parkwalk-managed Northern Universities Venture Fund alongside. The money is for the team, the platform and commercial partners.

Why it matters: Matter Machine represents a different commercial route for materials AI: pairing model predictions with a university's existing synthesis and characterisation capability. Its founder's Liverpool lab has repeatedly made new materials that computation pointed to, and the round adds an executive the release says helped build Oxford Nanopore.

Evidence and limits: The round funds the team and the platform. The company has not yet published material data of its own from which its technology could be judged.

What happened: Four departments and agencies say pilots have helped identify and fix over 1,000 potential vulnerabilities in government systems, and the programme will grow. The government will explore using the Cyber Security and Resilience Bill to have serious AI incidents and near misses identified and reported, and to cover loss-of-control risk, where autonomous systems evade human oversight. The AI Security Institute and the National Cyber Security Centre are already looking at how to respond to incidents involving AI agents. AI Minister Kanishka Narayan announced it.

Why it matters: AI safety governance is extending from testing models before release to reporting incidents and responding after deployment. That matters for the science agents the record follows, which are starting to run code and lab equipment on their own.

Evidence and limits: The vulnerability pilots are already running, and the 1,000 figure is the government's own, with no method or model named. Mandatory reporting of serious incidents is still being explored, with measures to be worked through with parliamentarians, technical experts and operators.

What happened: Zhuoran Yang, Vladimir V. Gusev, Matthew J. Rosseinsky and colleagues at Liverpool write it as a mixed-integer optimisation under crystallographic constraints: fit each pattern from a library of known phases, then rank samples by the signal left unexplained. Hide one of four known phases and the samples holding it rank near the top: all ten for SnO2, 11 of 14, 12 of 16, 16 of 18 elsewhere. On 31 patterns from a past Ba-Y-Si-Ru-O campaign, the leftover signal reproduced the main reflections of the two phases that campaign found.

Why it matters: High-throughput and autonomous synthesis produce more diffraction patterns than experts can read. AriadneX does not identify new crystals; it picks out, from many patterns, the samples most worth an expert's look, and shows what their unexplained peaks look like.

Evidence and limits: The results come from retrospective tests on data whose answers were known; the open-world test is one campaign of 31 patterns, assigned afterwards by the same group's experts. The next step is a new synthesis campaign showing it actually cuts expert analysis time. The code is MIT but needs Gurobi and licensed structure databases.

By Zhuoran Yang, Vladimir V. Gusev, Matthew J. Rosseinsky

What happened: Jihoon Han and Seth L. Shipman at the Gladstone Institutes and UCSF tiled six RNA targets at single-nucleotide steps, measured all 7,383 sensors, and trained a model that compares two candidate sites rather than scoring one. Held-out transcripts: its top sites were enriched for high-performing sensors. Frozen and run on the SARS-CoV-2 N gene in E. coli, with that transcript supplied alongside, 4 of its ten top sites gave working sensors against 0 of 20 random ones by qPCR, 5 against 2 of 20 pooled.

Why it matters: Toehold switches can be programmed for any RNA, and most work has gone into the switch itself; where in the transcript to aim has had far less, though the authors find neighbouring sites differ nearly as much as random ones. The gain comes from a base-pairing rule in the sensor stem that transfers between transcripts. Numbers this small (4 of 10, one transcript, one lab) measure enrichment, not a design rule. Toehold-VISTA, which works in its own setting, did not rank these sensors; the authors leave open whether the architecture, the readout or both explain that.

By Jihoon Han, Seth L. Shipman

What happened: Ivan Zlobin and colleagues at the Russian Academy of Sciences' Kurnakov Institute predicted which of two metals a ligand binds more strongly, from 88,525 selectivity values the authors derived from the Critical Materials Institute's compilation, holding out whole ligands rather than rows. Keeping the safest half of predictions, the model's own predictive uncertainty cut mean-of-pairs error by 65%, the best descriptor-space distance by 32% and the best ligand-only distance by 18%, while fingerprint nearest-neighbour and learned-embedding distances ordered error no better than chance.

Why it matters: An applicability domain is usually asserted: a prediction is called reliable because the molecule resembles the training set. Here that assumption is tested against held-out error, and similarity mostly fails it while the model's own uncertainty holds up. The same split reframes the published model for this compilation: its reported 0.599 mean absolute error on binding strength reproduces, but that test set shares 94% of its ligands with training, and a ligand-disjoint retrain gives 1.484. One task, one compilation, and the selection exercise replays fixed data rather than running an experiment.

By Ivan Zlobin, Alexander Pavlov

What happened: Qingran Cai and five colleagues at the company, with founder Zaixi Zhang at HKUST corresponding, pair a reasoning language model with an all-atom protein generator. From the reaction the enzyme should run, the reasoner writes four rounds (the amino acids that do the chemistry, the pocket around them, the chain's shape, then sequence and structure), and each round steers the generator, whose output becomes the next round's context. The preprint reports it ahead of the baselines compared on five of six measures.

Why it matters: This is the company's first public method since its first round in September, and it ships with the weights and the scripts that produced them, so the procedure can be inspected rather than taken on trust. Its design choice is to make the model think in structures: each step has to be built as a structure before the next one is written, instead of a chain of reasoning that stays in text.

Evidence and limits: Everything here is computed, and the headline numbers are narrower than they read. The 57.0% is a predictor's enzyme-function code for the design overlapping the target's at three levels, on the 2,435 test proteins shared with the baselines; the authors' own script says all 2,801 score about 8 points lower. The 0.729 TM-score refolds the design's sequence against the model's own structure, not a real enzyme. Nothing was made or assayed. The weights are Apache-2.0, the code carries no licence.

By Qingran Cai, Zaixi Zhang

What happened: Andreessen Horowitz led; Sequoia Capital and existing investor DCVC joined, and Martin Casado of a16z joins the board. CEO Diogo Almeida worked on RLHF and InstructGPT at OpenAI and was at Google Brain; his co-founders are Sasha Sheng, a former research engineer at Meta FAIR, and Erik Gafni, who founded Ravel, an AI company for DNA sequencing. TypeSafe came out of stealth on 15 September with a $40M seed led by DCVC.

Why it matters: Jev writes no text: it answers yes or no, picks an option or gives a score, each with a probability, and by TypeSafe's and a16z's own figures is one to two orders of magnitude faster and cheaper than frontier chat models at this. a16z calls it the first of a new class of models built to be called by code, not talked to, a bet that decisions in software reach further than chat. Neither names a science use; ranking candidates or checking an agent's work is where it could reach a lab, and nobody has shown it.

What happened: Chai Discovery announced a collaboration giving GSK access to its protein folding and design models for discovery across GSK's pipeline. It follows an evaluation in which Chai's models designed molecules for GSK targets without target-specific training, and GSK's own wet lab found binders to all the targets it tested. Christopher Austin, GSK's head of R&D technologies, said the models complement GSK's internal ones, which it retrains on its own lab data. No financial terms were given.

Why it matters: It is Chai's sixth drugmaker deal on this record this year, after Lilly, Pfizer, Novartis, argenx and BMS, and the release says GSK ran the wet-lab test itself before widening the deal. That makes the check external, but it gives neither the number of targets nor any binding figure: the designs bound, and GSK calls the affinities strong without a number. In January GSK also licensed Noetik's tumour models, so it is buying outside models while training its own.

What happened: Hedda Oschinski, Christian Carbogno, Karsten Reuter and colleagues at the Fritz Haber Institute had GPT-5.4, with no fine-tuning, propose elpasolite compositions one at a time and fed back each one's formation energy from a precomputed set of machine-learned predictions. Targeting below -2.26 eV per atom, a class holding under 0.2% of the space, it found 3,577 of 3,740 targets in 5,000 proposals across three runs. With the same 5,000 attempts, the GAN, VAE and reinforcement-learning models it is compared against found 1,726, 1,511 and 1,569.

Why it matters: The gain comes from the loop, not the chemistry knowledge: asked for all 5,000 at once with no feedback, the same model did far worse. It reads its own growing history of hits and misses, which the one-shot generators it beats cannot do, and the authors say that asymmetry should be weighed when comparing. Everything happened in software against a lookup table. The paper is explicit that the figure is recovery of a region an oracle defined, not a count of stable or makeable materials, and that 262 of the target compositions contain noble gases.

By Hedda Oschinski, Maximilian L. Ach, Konstantin S. Jakob, Christian Carbogno

What happened: Xiaona Yang, Hui Shen and colleagues at Inner Mongolia University, the Qingdao Institute of Bioenergy and Bioprocess Technology and the Nagoya Institute of Technology made the copper-hydride cluster Cu13H6 in one pot at 21.37% yield on copper, and resolved the hydride positions with stochastic surface walking global optimisation driven by a neural-network potential, in the LASP code. The cluster oxidised alkyl aromatics and alcohols to ketones under mild light, and also drove defluorination, deuteriation, cyclopropane ring-opening and dehydrogenative coupling.

Why it matters: Hydrogen is nearly invisible to the X-ray diffraction these clusters are normally solved by, and pinning it down usually needs neutron diffraction and a large enough crystal. The paper's abstract twice says the computation only tentatively identifies the positions, so the structure stands as a joint computational and experimental inference, not a direct determination; the count of six hydrides, though, is settled by mass spectrometry. The chemistry either side of it is physical: the cluster was synthesised and the reactions run and measured.

By Xiaona Yang, Yunzi Xin, Xiaoyan Sun, Hui Shen

What happened: Shaohui Zhang, Jing Huang and colleagues at Pennsylvania, Lehigh, Stanford and Children's Hospital of Philadelphia scored ten configurations on 161 cases drawn from the MedXpertQA benchmark, each with a narrative, findings, at least one image and a reference diagnosis. Accuracy rose with reasoning strength, 32.9% to 46.0%. Web search cut the routing mode from 38.5% to 26.7% while the reasoning modes each gained a little. Agent and deep-research modes scored 36.6% and 29.8%, below the plain model, and deep research cited nothing in 29 cases.

Why it matters: It is the rare benchmark where added machinery is measured and found to cost accuracy, which is the opposite of how agent systems are usually reported. The failure labels come from GPT-5 reading each trace, with physicians confirming 827 of 928 labels on 51 cases, a check the paper says measured precision, not recall: images misread, evidence not integrated, no reasoning to audit. The cases come from an existing benchmark rather than a clinic, nobody was treated, and the top score is under half.

By Shaohui Zhang, Jiarong Qian, Kai Zhang, Jing Huang

What happened: Daniil Poletaev and Artem Oganov at Skoltech searched for LaH10's structure at 150 GPa and 300 K with USPEX, ranking candidates by a free energy that includes quantum anharmonic effects (SSCHA), which DFT alone makes far too slow. Potentials trained on the fly, the universal potential MatterSim-5m and DPA3 effective potentials trained on SSCHA data each found the cubic phase seen in experiments; MTP-based effective potentials did not. Of seven universal potentials, only MatterSim-5m relaxed a test structure correctly, and it needed fine-tuning to approach DFT accuracy.

Why it matters: Standard structure search ranks candidates at 0 K, and for LaH10 that picks the wrong answer: the same search with 0 K DFT chose lower-symmetry structures over the cubic phase at 11, 44 and 88 atoms per cell. The effective potentials made the 88-atom search affordable. The test is one compound whose answer experiment already gave, and direct SSCHA search still costs about a hundred times a 0 K DFT search, even with a universal potential.

By Daniil Poletaev, Artem R. Oganov

What happened: From its first advisory on 6 October, when Isaias was still an unnamed depression, NHC forecast a Category 2 hurricane, shaping its forecast to match DeepMind's model, whose newest version is WeatherNext 3, CNN reported; many physics-based models then showed a weaker storm on a different track. That first discussion put the NHC track in best agreement with the DeepMind and ECMWF AI aids. Isaias peaked near 100 knots offshore, then landed near Destin, Florida, late on 9 October at 90 knots, east of where that first track pointed.

Why it matters: This is the model used in practice, not scored afterwards: NHC director Michael Brennan says 2026 is the first season WeatherNext is fully part of the centre's forecasting aids, after it guided last year's Hurricane Melissa call, and that NHC helped train it on its own best-track data. One storm is not a verification. The formal storm report comes months later, and the verdict that DeepMind's track was best so far comes from reporters and a TV hurricane specialist, not from NHC.

What happened: Daniel Widdowson, Olga Anosova and Vitaliy Kurlin (University of Liverpool) and Tatiana Kurlina (UCL) compared every crystal in two public GNoME versions (384,938 entries in 2023, 554,054 in 2024) by continuous geometric invariants. Matching atom positions only, ignoring elements, 63-78% of entries lie within 0.05 Å of a crystal already in the ICSD or Materials Project, depending on database and metric. About a fifth contain radioactive elements. Of 1,133 groups of exact duplicates in 2023, 1,010 were deleted whole in 2024.

Why it matters: The audit reopens a basic question: in what sense is an AI-predicted material new? GNoME's 2023 Nature paper is the record's reference for how many new stable materials AI has found, and the audit argues a crystal within 0.05 Å of a known geometry should not count. Similar geometry does not mean the same composition or properties, so structural novelty, compositional novelty and whether a material can be made need to be judged separately.

Evidence and limits: The analysis is computational and not yet peer reviewed. Most near-matches carry different elements, so whether a new composition on a known framework counts as a new material is the point in dispute, not a settled finding.

By Daniel Widdowson, Tatiana Kurlina, Olga Anosova, Vitaliy Kurlin

What happened: A University of Washington, Ai2 and NUS team, with David Baker among the authors, text-conditioned the Complexa structure generator on 1.8 million structures paired with 52.4 million captions, written by a language model from database annotations. It leads the baselines on 15 of 18 structural tasks and on at least 13 of 20 binder targets, raising a computed interface-confidence score 14% over the strongest. Against GM2A, which has no reported de novo binders, Adaptyv Bio assayed 89 designs picked from a thousand: nine bound, six below a micromolar.

Why it matters: A designer can now ask for a fold, a disulfide count and an interface size in one written sentence, instead of a separate input format per property. That matters most for the agent loop the paper tests: a language model researches a target, writes the specification in words, reads the scores and rewrites it, with no translation step into a model's own labels. The GM2A binders are what separate this from a benchmark result.

Evidence and limits: Binding is not function: Adaptyv Bio, commissioned by the authors, measured apparent dissociation constants by surface plasmon resonance, fitting eight of the nine, and nothing shows they do anything to GM2A. The paper's 6.7% hit rate counts the six sub-micromolar binders against 89 designs the authors picked from a thousand. The 14% and the task counts are the authors' own runs of the baselines, and the preprint is not peer reviewed.

By Joseph Lee, Pang Wei Koh

What happened: Yiwei Chen and colleagues at UC San Diego and Amazon put each task in a natural-language scientific scenario backed by a hidden probabilistic graphical model: a language model writes the readable surface, while code builds the causal graph, fixes the intervention semantics and computes the oracle answer, so runs are graded without human labelling. Agents get four turns to request observational or interventional samples. The main board is 180 tasks, with a separate 150-question yes/no check; seven current models, two older ones and three active agent designs were run.

Why it matters: Most science benchmarks hand a model the evidence and ask what it concludes. This one makes it decide what to collect. The paper's evidence ledger then says where a run broke: it tested the wrong intervention, lost evidence it had already won, or kept querying after the answer was supported. Those are the parts of an investigation that a final-accuracy score hides.

Evidence and limits: Every world is synthetic, with fully observed categorical variables and known intervention semantics, so no latent confounding, continuous variables, measurement error or open-ended experiment design. The authors say success here should not be read as readiness for real scientific or policy settings. Claude Opus 4.6 is both one of the evaluated models and the fixed parser between every agent and the simulator.

By Yiwei Chen, Rose Yu

What happened: Mason Minot, Gisbert Schneider and colleagues at ETH Zurich and Roche had a tree-based model pick which of 22,253 Roche compounds to try in C-H borylation, a reaction chemists use to modify an already finished drug molecule. Three rounds tested 50 new substrates, taking the dataset to 6,865 reactions on 568 substrates, with site labels for 812. Graph neural networks trained on it then predicted the borylation site for ten unseen substrates, and the top-ranked site matched the experiment in all ten.

Why it matters: Reaction models for drug chemistry lack data, above all failed reactions and labels for where a reaction lands; this loop produced both, 106 negatives among them, and the data is public. The ten were run after the predictions, by the team itself, which is what makes it a test. The site model reads no reaction conditions, though ligand and solvent can move the site, and on unseen scaffolds the graph models scored only 0.43 to 0.50 MCC, where 1 is perfect, at telling whether a reaction works. The authors ask for more drug-like molecules: accuracy fell on the hardest.

By Mason Minot, David F. Nippa, Kenneth Atz, Gisbert Schneider

What happened: Shaon Basu, Kevin Moreau and colleagues at AstraZeneca, Charité Berlin and the Francis Crick Institute measured the proteomes of liver cells exposed to a library of 204 androgen-receptor degraders; the cells lack the receptor, so every response was off-target. A gradient-boosted model trained on them predicted toxicity (holdout AUPRC 0.78) and pointed to respiratory complex I. It scored AstraZeneca's compound 1 at 91 of 100 and two analogs differing only in the linker region at 47 and 35; those barely touched complex I and spared primary human liver cells.

Why it matters: Degraders join two drug parts, and the joined molecule can be toxic where neither part is; this catches that in cells before animals. The call was then tested with new experiments: cell respiration, an enzyme assay on heart mitochondria, and primary human liver cells: all cells and isolated enzymes, the team's own work, from one cell line and one target class. A 28-day mouse study, seven treated animals, showed the safer analog still shrank tumours, but had no liver endpoint, so it tests the drug and not the toxicity call. The analogs were already in the library.

By Shaon Basu, Kevin Moreau

What happened: People familiar with the effort told Bloomberg the round could value Alphabet's drug-design company as high as $50B. The amount, the lead and the terms are not reported, the financing has not closed, and Isomorphic did not comment. Demis Hassabis, who in August moved from running Google DeepMind to chairing it and became Alphabet's chief scientist, remains Isomorphic's CEO. The May Series B gave no valuation.

Why it matters: It shows what investors may pay for a drug-design engine before its own drugs have results: the record holds Isomorphic's pharma deals and models, and no clinical readout. It is talks, reported by unnamed people, so the figure can move or fall away; nothing here says the engine works.

What happened: Justinas Slepavicius and colleagues at FTMC and Vilnius University, who run the Crystallography Open Database, relaxed its entries with universal potentials to test physical plausibility, a check that had relied on expert judgment. Of 522,086 entries, 317,457 passed a quality filter and were relaxed; 99.0% of those converged without a large volume change. Hand-checking 327 structures showed no single threshold on energy, force or displacement separates valid from invalid. On 524 hard cases, M3GNet distorts cyclopentadienyl rings and PET models do not.

Why it matters: The models disagree in ways that depend on chemistry, not on quality: each has its own blind spots, so a large change after relaxation can be the potential's failure rather than the structure's. The paper says its criteria are not calibrated against ground truth, so how many bad entries they miss cannot be estimated. Their own conclusion is to combine several models rather than trust one. Two coordinate errors they found, in CIFs rather than the original papers, are now corrected, and the relaxed structures are deposited so a reader can see whether relaxation confirms a published one.

By Justinas Slepavicius, Saulius Grazulis, Linas Vilciauskas

What happened: Ziming Pan, Kun Li and colleagues at Tsinghua's Institute for AI Industry Research train on simulator trajectory segments: kinetic Monte Carlo for three reactor pressure vessel steels, molecular dynamics for a Cu-Zr glass and a Li3N electrolyte. Per macro step, the model predicts a sparse set of structural edits plus the physical time elapsed. Edits are restricted to sites the teacher's rules can reach and species counts projected to stay conserved. Macro-step inference runs 1,000 to 10,000 times faster than event-by-event simulation.

Why it matters: The constraints are the result: on held-out steel the model violates reachability and conservation in no case and scores 0.91 on predicting which sites changed, while the same architecture without them keeps only 45.2% of edits reachable and 59.2% of counts conserved, and its clock drifts to 0.35 of the true elapsed time. But the ablation drops all three constraints at once, so which one earns what is unknown, and the teacher is a simulator rather than a measurement. A separate billion-atom comparison runs the baseline on a Sunway supercomputer and this model on A100s.

By Ziming Pan, Ruge Zhang, Haozhi Han, Kun Li

What happened: Anthropic says the money will make Claude available to more than 15 agencies in the federal programme, naming NASA, the National Institutes of Health and the National Science Foundation. Over three years it will give Claude, Claude Code and API credits to several hundred Genesis Mission projects, partner with agencies and national labs on fusion energy and quantum computing, and train scientists on those projects. A White House fact sheet the same day lists it beside NVIDIA at $1B, AMD at $500M, OpenAI at $200M and Google at $150M.

Why it matters: The named deliverables are product access, credits and training, and Anthropic never says the $150M is cash; the White House totals the eleven pledges as science tools and compute credits, so the figure is a valuation of what is given, by whoever gives it. No agency has confirmed a share, Anthropic does not say which 15 it will serve, and the Department of Energy's own Genesis pages did not mention it the next day. Fusion and quantum are named as priorities to partner on, with no project, lab or date attached.

What happened: Miroslav Lebeda and colleagues at the Czech Technical University in Prague ran five binary and two high-entropy refractory carbides through stress-strain molecular dynamics with nine pretrained potentials, none fine-tuned, and compared them against published ab initio molecular dynamics from another group. MACE-MH-1 had the lowest mean absolute percentage error, 5.7%. Most models underestimate stiffness, with equilibrium volumes larger than the reference's. Errors track the metal, not the chemical complexity: 6.6% for ZrC against 20.1% for TaC, and the high-entropy carbides match their constituent metals.

Why it matters: Absolute stiffness and the rate of softening with heat are separate skills, and the paper reports them separately on purpose: a model can get the softening trend right while sitting at the wrong stiffness, so one averaged score hides which. The reference is another group's ab initio calculation, not a high-temperature measurement, so this says how well the potentials reproduce that calculation. The data is available from the authors on request rather than deposited.

By Miroslav Lebeda, Jan Drahokoupil, Simon Svoboda, Petr Vlcak

What happened: Where model output moves hardware without human approval, the policy now requires a qualified individual able to watch the equipment and stop it at any time, equipment that stops or holds a safe state if that person intervenes or the connection drops, and operating limits (speed, force, temperature, pressure, dose) enforced by the equipment or a controller independent of the model. It lists hazardous energy and materials and acting on the human body among the triggers, and excludes laboratory automation handling non-hazardous materials inside enclosed instruments.

Why it matters: The company's Model Hardware Standard, on this record since August, set out how Claude may drive lab equipment; this turns that into a rule with a date. The exclusion marks the boundary: an enclosed instrument with harmless reagents is out, a bench that heats, pressurises or dispenses hazardous chemicals at levels that could injure is in, and its limits must be enforced somewhere the model cannot reach. The policy sets no rule about how long an agent runs, despite how the update was framed.

What happened: The department announced 12 Phase II awards totalling $159 million, taking Phase II to 14 projects; Darío Gil, its Under Secretary for Science, is the official quoted. Commonwealth Fusion Systems leads a digital twin of its SPARC device; UC San Diego will expand an RNA structure database five-fold for model training; the University of Washington works on sequence-structure ensemble modelling for enzyme design; Berkeley Lab's MOAT-Core is to run and design particle accelerators autonomously. MIT, Fermilab, Oak Ridge, Harvard and Argonne lead five of the rest.

Why it matters: Phase I was selection; this is money, and the list says which bets the programme is paying to build. Three of the twelve are the site's own beats (enzyme design, autonomous accelerator operation, inverse design of magnets), and the fusion digital twin puts a company's own machine at the centre of a federal project. None of it is a result: the release states no target any of these projects has to hit, which the earlier PNNL awards did.

What happened: Jensen Huang announced it at the Science: A New Golden Age event in Washington on 8 October. NVIDIA describes the $1 billion as commitments over five years for super intelligence for science and for quantum computing, through higher-education research institutions, quantum investments and cloud providers serving government missions, and says it collaborates on several Phase II awards announced the same day. The release restates its existing DOE work: the department's largest scientific supercomputer at Argonne, and seven new systems across Argonne and Los Alamos.

Why it matters: The $2.4B of industry pledges to the Genesis Mission consortium is already on this record, and this is the biggest single piece of it. What the release does not say is what the $1 billion is: it is "commitments valued at" that sum, spread over five years and across three channels, with no figure for any of them and no cash changing hands on the day. Read it as a statement of where the compute will be pointed, not as a grant.

What happened: Kun Chen of the Institute of Theoretical Physics, Youjin Deng of USTC and colleagues collected 46 mathematics and 36 physics problems still open in the literature, with context and prior progress but no reference answers. Seven models worked offline, up to 12 hours each; four of them also judged every submission. GPT-6 Astra's mean judged solve rate was 14.0%, every other configuration under 8%. All four judges agreed on eight Astra solutions, seven of them physics problems; many physics items ask to extend a given construction.

Why it matters: A score on problems nobody has solved measures something benchmarks with known answers cannot, but every verdict here is a model's: the authors say expert review is still needed to show any 'solved' answer is correct or new, and that the judges agreeing does not estimate how often they are wrong. So 14% is a model-judged figure, not a count of new results, and the eight unanimous answers are where to check. HorizonMath and FrontierMath: Open Problems check answers with computational or problem-specific verifiers instead.

By Zhiyi Li, Kun Chen

What happened: Kuen-Phon Wu and Chuan-Kai Chueh of Academia Sinica and National Taiwan University had Claude Code with Claude Opus 5.5 run Phenix, Coot, ModelAngelo and other established programs, read the validation and pick the next step, in about 36 hours on one workstation. On 15 deposited cryo-EM models the median MolProbity score went from 2.62 to 1.76 and clashscore from about 21 to 6. Three crystal structures built from diffraction data reached Rfree 0.182, 0.207 and 0.238, against 0.173, 0.197 and 0.229 published.

Why it matters: This tests an agent as a supervisor of existing programs, judged against the data each structure was measured with, not as a new refinement method. It did not win everywhere: CaBLAM outliers rose in seven of the 15 cases, one ligand fit its map worse, one in-house model was kept as deposited, and the crystal structures came close to the published ones without beating them. The investigator steered throughout, ligands and ions had to be supplied, and nothing was compared with human experts or other agents.

By Kuen-Phon Wu, Chuan-Kai Chueh

What happened: Werner Müller, Benoit Creton and Carlos Nieto-Draghi of IFP Energies nouvelles and Erich Müller of Imperial College gave Claude Opus 5 three papers and three prompts. Over three sessions it derived SAFT-VR IMP, wrote about 600 lines in the Julia package Clapeyron.jl with 97 verification tests, and fitted it to DIPPR data for six solvents. For the ethers and the sulfide, errors on the fitted data stayed within about 1%. The authors say none of its errors was caught by the researchers; its own tests caught them.

Why it matters: It is a derivation-to-fit chain a reader can check end to end, and the checks are real: exact reduction to the parent model, analytic limits, derivatives and the Clapeyron equation. But the 97 tests were written by the same model as the code, so they cannot catch a misreading of the physics that both share; the authors say auditing this much output is the real limit. And for the three ketones the data rejected the measured dipole: the fits only worked once the dipole was refitted to near zero, which removes the physics the model was built to add.

By Werner A. Müller, Erich A. Müller

What happened: Artur Lyssenko, KyuJung Jun and Rafael Gómez-Bombarelli at MIT make the Drude relaxation differentiable, so bonded, electrostatic, polarization and screening parameters are fitted together against DFT energies, forces and dipoles, with no empirical tuning, at 1.20 times the cost of a fixed-charge simulation. On PEO-LiTFSI the dipole-constrained model matches the measured ionic conductivity about as well as the best charge-scaled model, 17% against 16% mean absolute percentage error.

Why it matters: Charge scaling is the standard way to fake the missing electrons, and on conductivity alone it still wins narrowly. The case for real polarization is what it gets that scaling cannot: the cation transference number in the dilute limit, its rise with salt concentration, and the absence of lithium-TFSI contact pairs at low concentration. A second result cuts against anyone fitting potentials: the variant that matched the DFT data more closely underestimated conductivity by up to an order of magnitude.

Evidence and limits: One polymer and one salt. Conductivity, the cation transference number and ion-pairing trends are all checked against published experiment, but the model reproduces only the dilute-limit transference and its rising trend, not its size: it flattens near 0.17 above 1.7 molal where experiment reaches about 0.5. Training is 5,942 DFT configurations of seven species, 10% held out, no separate validation set. A preprint, not peer reviewed.

By Artur Lyssenko, KyuJung Jun, Rafael Gómez-Bombarelli

What happened: Kieran Bozier and Chris J. Pickard at Cambridge compute the Coulomb parameter in the random-phase approximation for six hydrides and for the ambient-pressure metallic elements. For Mg2IrH6 they get 0.62, and a pseudopotential near 0.39, far above the 0.1 to 0.15 that Tc predictions conventionally assume; using it costs about 40 K against the standard 0.16. The parameter closely tracks the density of states at the Fermi energy, and its energy dependence accounts for a factor-of-two gap between two earlier Tc estimates for that material.

Why it matters: Hydride searches rank candidates by predicted Tc with this repulsion held at a number borrowed from elemental superconductors. If it instead rises with the density of states at the Fermi level, the ranking is skewed exactly where the field looks hardest, because a high density of states is also what makes a candidate look promising. The paper's answer is to compute the parameter, or to correct for the density of states, rather than keep one value for everything.

Evidence and limits: Six hydrides and one set of elements, all computed. Mg2IrH6, YSbH6 and LaSc2H24 are predicted materials with no measurement to check against. For H3S, where Tc has been measured, including the computed repulsion moves the prediction toward experiment, which the paper reports as a trend rather than a number. A preprint, not peer reviewed.

By Kieran Bozier, Chris J. Pickard

What happened: Kyuho Lee, Joseph G. Checkelsky, Heather J. Kulik and colleagues at MIT stack two metastable methods, diffusion and epitaxy, into diffusion-activated epitaxy: a substrate sits in a lower-melting flux and reacts at its surface while its bulk stays inert. In a narrow synthesis window they stabilised four distinct phases on oxide substrates, confirmed single-crystalline by X-ray diffraction and dark-field STEM. Three could not be matched to any structure in the experimental and computational inorganic-structure databases they searched.

Why it matters: Closed-loop materials discovery, from GNoME to DeepMind's robotic synthesis, mostly predicts stable structures first and then tries to make them. These three phases were made but sit outside the computed databases those predictions are drawn from, which is evidence that synthesizable space runs ahead of what the databases cover, so a predict-then-make loop misses metastable phases by construction. The paper's suggestion is a branch that makes first and solves the structure after.

Evidence and limits: The structures of the three unmatched phases are not solved: the claim is that no database structure fits their diffraction and microscopy, not that they are new, which would need a solved structure. One narrow window of one flux system, four phases. No AI method is used in the work; the bearing on predict-then-make pipelines is the reason it is recorded here. A preprint, not peer reviewed.

By Kyuho Lee, Joseph G. Checkelsky

What happened: Brice Ménard, an astrophysicist at Johns Hopkins and Anthropic, had agents in Claude Science gather and cross-calibrate the public UV surveys, among them GALEX's 38,000 observations, and fill the third of the sky never observed in UV by learning how UV brightness tracks visible, infrared and radio maps, then add over 100 million Gaia stars. On measured patches hidden from it, the model came within about 10% of the real UV. Each pixel is marked measured or predicted, with an uncertainty.

Why it matters: A complete UV sky was missing because filling it takes weeks of pixel-level calibration that nobody puts first; here it took a few days of a scientist steering agents between his own projects. The map is a teaching tool, and the post's point is that most fields have a backlog of such resources that is now much easier to clear.

Evidence and limits: The predicted third is a statistical estimate, checked only against held-out patches that were measured; no new UV observation tests it. Ménard caught a calibration artefact that two rounds of agent review had missed, and the post cites no paper or outside review of the map.

By Brice Ménard

What happened: Biohub's Virtual Biology Initiative, launched in April with $500M of its own, adds three blocks: DOE invests more than $500M over five years through the Genesis Mission, NIH brings datasets and repositories built with more than $500M of past federal money, and Google DeepMind, Isomorphic Labs and Meta put in $300M together. The goal is open, standardised data on how cells respond to interventions, across far more cell types. Alex Rives of Biohub, Max Jaderberg of Isomorphic and Pushmeet Kohli of Google DeepMind speak for it.

Why it matters: It bets that data is what holds virtual-cell models back, the same limit Xaira's X-Cell paper reports. The $1.8B is not all new money: NIH's share is past spending on data that already exists, and DOE's is spread over five years. No first dataset, date or access terms for the company-funded data are given. It shows where public and big-tech money is going, not that a virtual cell will work.

What happened: Nextech Invest and Norwest co-led; Invus, BB Biotech, T1D Fund and BOLD Longevity Growth are new, and Access Biotechnology, Atlas Venture, Sanofi Ventures and Digitalis Ventures followed on. Kanishka Pothula of Nextech and Brian Matesic of Norwest join the board; Andre Turenne is CEO. Nathanael Gray and Tinghu Zhang of Stanford, Jianwei Che of Dana-Farber and Edward Chouchani of Harvard Medical School and Dana-Farber founded the company. No valuation was given.

Why it matters: Matchpoint finds covalent binders with chemoproteomics and uses machine learning to choose targets and guide compound design. The money takes its first drug, an oral T-cell modulator for autoimmune disease, to clinical proof of concept, with trials to start in 2027. No clinical data exist yet, and the release does not say how much of MPT-062's design came from the models. Novartis already funds one of its programs under a July 2025 option deal worth up to $1B.

What happened: Artem Zholus, Mahmoud Assran and colleagues at FAIR train predictors from 22M to 8B parameters on a frozen V-JEPA 2.1 encoder, from 23 public manipulation datasets across 12 robot platforms. Imagination error is the per-token distance between predicted and encoded features over a nine-step rollout. It follows a second-order power law in compute. Fitted on the 22M to 2B frontier, it matched the held-out 4B and 8B checkpoints more closely than three rival forms. Planning toward one goal image, grasp success rose from 30% at 22M to 67% at 8B.

Why it matters: The law was fitted without the two largest models and still describes them, which is more than a curve drawn through every point can claim. What it does not do is reach past 8B: the authors report an irreducible error the fit is already near, and say more compute alone will not move it without more and more varied interaction data. The robot runs are the authors' own, 30 episodes per model and task, scored by hand, on a platform whose data was in training. The promised code and checkpoints were not public two days on.

By Artem Zholus, Jeannette Bohg, Nicolas Ballas, Mahmoud Assran

What happened: Jonas Grandel, Philipp Benner and Janine George at BAM Berlin and Jena fine-tuned MACE-MP-0b3 separately for each of 53 materials, against their own new VASP reference calculations. Regularising towards the pretrained weights while training only the node embedding and first interaction layer cut the median phonon error to 0.05 THz against the foundation model's 0.27. On recovering the phase a soft mode leads to, across the nine materials DFT finds unstable, multi-head scored 0.43 and ordinary transfer learning 0.56, both below the untouched foundation model's 0.63.

Why it matters: The finding that matters is the gap between the two scores: getting harmonic frequencies right, and even classifying which structures are unstable, does not mean the model has the energy surface right along the unstable direction. The authors say as much in their introduction and conclusion. So a recipe picked on phonon error can quietly be worse at the thing phonons are usually computed to predict. Everything is model against DFT, in software; the best recipe's margin over doing nothing is 0.66 to 0.63.

By Jonas Grandel, Philipp Benner, Janine George

What happened: The lab uses AI to propose affinity reagents such as antibodies, builds and tests each one, and feeds every result back into the model. It strings together tools from Danaher's own companies (Beckman Coulter Life Sciences, Cytiva, Genedata, Integrated DNA Technologies, Molecular Devices), with robotics built with Automata. CEO Julie Sawyer Montgomery spoke for the company; chief science officer Jose-Carlos Gutierrez-Ramos says it is designed for up to 8x faster discovery and up to a tenfold rise in output, from tens to hundreds of reagents a year.

Why it matters: Danaher is building a closed design-make-test loop out of instruments its own companies already sell, as the first step of a stated push toward smart instruments and autonomous labs. It is a plan: the release says 'plans to launch', though a Simply Wall St piece, also run on Yahoo Finance, said 'launched', and the 8x and tenfold figures are design targets with no data behind them yet. What it would prove is whether such a loop speeds up reagent discovery inside a working business.

What happened: Simon Batzner, who built the NequIP interatomic potentials and was a co-author of GNoME and GenMS at Google DeepMind, announced on X that he has joined Anthropic as a Member of Technical Staff, thanking colleagues at DeepMind and Google Brain.

Why it matters: With him, four of GNoME's six authors now work elsewhere; three of them co-founded Periodic Labs. The team that produced DeepMind's best-known materials result has largely dispersed, while DeepMind builds a materials lab in London with different people. It follows John Jumper's move to Anthropic in June.

Evidence and limits: His post does not say what he will work on at Anthropic. His own website still described him as at DeepMind on 2026-10-22.

What happened: Chris Elliott of Resolution, a lab that studies how AI models generalize, worked with agents on a 112-page step-by-step plan, then had an orchestrator agent, mostly Claude Fable 5.1, carry it out. The proof of Hironaka's 1964 theorem and its modern forms took about a week and some 30 interventions from him: about 100 billion input and 300 million output tokens and 600,000 lines of Lean. Billy Snikkers and Adam Newgas then spent two weeks refactoring it.

Why it matters: It gives formal mathematics a price tag a reader can cite, reached outside a frontier lab with little scaffolding beyond the plan and some agent skills from earlier efforts. Resolution needs the theorem for singular learning theory, which ties the shape of a model's loss landscape to how it generalizes, and says the library already lets it automate more of that theory.

Evidence and limits: The figures, including the token cost, which is an estimate, are Resolution's own. The proof is public with a checker, Comparator, that confirms it proves exactly the stated theorems, but the post reports no outside check. It formalizes known mathematics; the learning-theory results it is meant to enable are promised for the coming months.

By Billy Snikkers, Adam Newgas, Jesse Hoogland, Chris Elliott

What happened: Advent, Temasek and Xora Innovation co-led the Series B; AMD Ventures and Madrona joined, and Eclipse and Khosla Ventures followed on. The $1.5B valuation is the company's own. Vinci says the money takes it from thermal, thermo-mechanical and convective-fluid work in semiconductors to memory, advanced computing, vehicles, aircraft and satellites, on a physics model it says needs no per-design training and reads designs of hundreds of millions to more than 15 billion degrees of freedom in minutes. Hardik Kabaria is co-founder and CEO.

Why it matters: This is the simulation bet that materials and chemistry have been making with learned potentials, moved to engineering: if a model can stand in for a finite-element run and stay accurate, physics stops being something you check at a few points and becomes something you ask while the design moves. AMD is now in through AMD Ventures. The published comparison with conventional solvers is Vinci's own, at EPTC 2025 and on its site; the customers it says benchmarked it are unnamed, and none of their results is public.

What happened: Following Morgan and Tian's account of Perelman's proof, mathematicians wrote a blueprint and designed about 90 milestones; GPT-6-Astra put the milestone statements into Lean and, through coding agents working in parallel, built most of the proofs, with Fable-5.1 helping review statements. The result proves both the topological and smooth theorems, passes a full Lean build and a check against target statements written only with Mathlib, and is rechecked by an independent kernel. Bin Dong and Chunlei Liu are corresponding authors.

Why it matters: After Fermat's Last Theorem, another famous proof is machine-checked, and the paper says where the work went: preparation began in May, and people chose definitions, reviewed the milestone statements and unblocked the agents. The kernel settles that each proof step is valid; whether the statements say what Perelman proved still rests on human review. An independent formalization by Ziyang Qin, Bennett Chow and others was announced at the same time.

By Zhiyuan Zhang, Axel Delaval, Chunlei Liu, Bin Dong