AI FOR SCIENCE · TIMELINE
The 55 entries the timeline holds for 2025, newest first. Each was read at source. The full timeline, with filters →
What happened: Insilico sold 94.69M shares at HK$24.05 and began trading the same day as 3696.HK. The Hong Kong tranche was 1,427 times subscribed.
Why it matters: Public investors now price an AI drug-discovery company directly; the trials of its fibrosis drug, not the platform, will move that price.
What happened: Academic groups and Deep Principle, including its co-founders Chenru Duan and Haojun Jia, built SAGA: LLM agents read each round's results, propose new objectives and write them as scoring code that an optimiser then runs. Of 28 antibiotic designs made and tested, four cut E. coli growth by over 80% alongside a membrane-opening helper; one showed minimal toxicity to two human cell lines at its MIC, 128 µg/mL. Three of 24 PD-L1 nanobodies bound, at 300-400 nM.
Why it matters: The paper's claim is that design agents fall short at choosing what to optimise, so the agent should rewrite its targets as it goes. The lab results are first hits from the authors' own assays: the antibiotic screen used a helper that opens the bacterial outer membrane, and one weak hit is far from a drug. The materials and chemical-process designs were checked only by calculation.
What happened: Latent-X2 is given a target structure and an epitope and writes both the sequence and the all-atom structure of a VHH or scFv antibody. Testing at most 24 designs per target per antibody format, Latent Labs got binders against 9 of 18 targets at picomolar to nanomolar affinity, with expression, aggregation and stability matching approved antibodies and no optimisation step after generation.
Why it matters: Immunogenicity is one of the reasons antibodies fail in the clinic, and the authors say nobody had measured it on an AI-designed antibody before. Four VHHs designed against one target, TNFL9, were put on immune cells from ten healthy donors: no detectable T-cell proliferation and no detectable cytokine release, a profile the paper calls comparable to caplacizumab, an approved VHH drug. Donor cells in a dish are a proxy; no animal or human data exists yet.
What happened: Linfeng Zhang of DP Technology and co-authors from Shanghai Jiao Tong University, Peking University and other institutions describe Bohrium as a platform that turns scientific tools, models, databases and laboratory systems into services an agent can call, with each run traced, and SciMaster as the layer that runs long multi-step workflows on it. They list eleven agents deployed or prototyped, and report cycle times falling from months to days in numerical physics and from weeks to hours in literature work.
Why it matters: DP Technology sets out here what its platform is: not one model but the plumbing that lets agents use many tools and keep a record of what they ran. For a lab or company choosing where to run agent workflows, that traceable record bears on whether results can be checked.
Evidence and limits: A paper by the platform's builders. The time savings are their own figures for chosen cases, with no outside measurement, and some of the eleven agents are prototypes.
What happened: Each question sits inside a research scenario rather than standing alone. Where open datasets allowed, questions were templated from them; the experiment-related ones were drafted by a domain expert, and every one went through panel review. Best scores: 0.71 biology (Claude Opus 4.1), 0.60 chemistry (Claude Sonnet 4.5), 0.75 materials and 0.60 physics (GPT-5). Turning GPT-5's reasoning effort from medium to high moved three domains barely at all (0.70 to 0.69 biology, 0.74 to 0.75 materials, 0.58 to 0.60 physics) and chemistry by 0.07.
Why it matters: Exam-style benchmarks put these models near the top on science questions; put a question back in a researcher's context and a quarter to two-fifths of it goes wrong. More thinking time mostly does not close that, which points at what the models lack rather than how long they get. The scores were computed by the team that assembled the questions, and the company that led the work sells scientific agents. The paper declares no competing interests.
What happened: OpenAI reports strong results on Olympiad-style questions but much lower scores on open-ended research tasks.
Why it matters: The useful result is the gap itself: solving bounded problems does not yet transfer cleanly to open-ended research.
What happened: In Nature Communications, METiS TechBio reports MTS105: a lipid nanoparticle carrying mRNA that makes the liver itself produce a T cell engager against the liver-cancer marker GPC3. The company designed the mRNA sequence with its own MrNatty algorithm, which it says pairs RNA structure prediction with natural-language-processing methods. On three weekly doses of 0.5 micrograms a mouse, liver tumours disappeared completely and had not returned after 84 days; monkeys dosed to 45 micrograms per kilogram showed no severe adverse effects. A first-in-human trial, NCT06689540, is recruiting.
Why it matters: METiS listed in Hong Kong as a company that designs drug delivery with AI, and this is the first paper on this record where that design appears as a drug tested in animals rather than as a platform description. What the algorithm designed is the mRNA sequence; the lipid the paper says was designed and made in-house, with no method given for choosing it. The tumour result is the company's own mouse experiment, and the antibody drug it is compared against was picked by the company. The human trial has reported nothing.
What happened: In Biotechnology and Bioengineering, DataHow and Merck Serono, with DataHow co-founders Michael Sokolov, Fabian Feidl and Alessandro Butté among the authors, link Bayesian experimental design and a hybrid process model to 24 parallel mini-bioreactors. After a 24-run training experiment, the software set temperature, stirring, perfusion and pyruvate once a day from day 9. One bioreactor reached a 30% viable-cell-volume target; standard runs average under 20%. For a new clone, aiming at 98% viability, steered runs kept viability above unsteered ones.
Why it matters: Self-driving labs usually make molecules or materials; this one steers living cell cultures for weeks, in the step of making a drug once it is found. Each goal was tried in one or two bioreactors at one company, against three unsteered ones, with results shown as plots, not effect sizes. For the new clone the model got it right only after retraining on half the run's data, and product quality was not controlled.
What happened: Part of a wider UK government partnership. Robots directed by Gemini are to make and measure hundreds of materials a day; DeepMind names superconductors, batteries, solar cells and chips as targets.
Why it matters: DeepMind's materials work moves from predicting crystals to running its own experiments. Artem Mishchenko has led building the lab since November 2025.
What happened: Mehran Karimzadeh, Hani Goodarzi and colleagues at Exai Bio, UCSF and the Arc Institute trained a 3.6-million-parameter transformer on small RNA fragments circulating in 8,339 plasma and serum samples, out of a corpus of about 13,000. It fills in hidden measurements with an R-squared of 0.89, against 0.57 for an averaging baseline. Built on its compressed representation, a classifier for ovarian cancer, a disease left out of pretraining, trained on 48 cases scored 0.1 higher in area under the curve than one built on the raw measurements.
Why it matters: A blood test for a rare cancer usually fails for lack of patient samples to train on. A model pretrained on many people's blood that lets a classifier learn from a few dozen cases is what would make such tests buildable. The same representation also let a classifier trained on plasma keep working on serum, where the raw-feature classifier dropped from 0.74 to 0.56. Network Bio, which Goodarzi co-founded, now presents Exai-1 as its first model.
Evidence and limits: All samples came from the company's own assay and pipeline, and every test is on its own held-out data; the authors call validation on independent datasets the next step. Network Bio's later blog claims on head and neck cancer and osteosarcoma are not in the paper.
What happened: RlapsRisk BC reads a routine H&E slide and a few clinical facts. At the San Antonio Breast Cancer Symposium 2025, Owkin and co-authors at Gustave Roussy and MD Anderson compared it with EndoPredict in 381 Gustave Roussy patients (7 distant recurrences) and with Oncotype DX in 154 node-negative MD Anderson patients (42), all followed at least five years. Areas under the curve for five-year distant recurrence-free survival: 0.73 against EndoPredict's 0.57, and 0.78 against Oncotype DX's 0.58. Its high-risk group held 6 of 7 recurrences and 35 of 42.
Why it matters: Oncotype DX decides whether a woman gets chemotherapy, so a model that reads the slide the hospital already has would be cheaper and need no tissue sent away. This is not yet evidence it can take that decision: the analysis is the company's own, looking back at archived slides, and 7 recurrences in one cohort cannot separate two tests. The abstract itself says the test 'may support' de-escalation.
What happened: The Shanghai-founded company announced the round and the lab the same day. Founder and CEO Sean Lin says the plant runs a 24/7 automated synthesis line that plans, runs and records chemistry, and that the money goes on engineers and chemists in Singapore; Granite Asia partner Yinghui Kuang spoke for the lead. ChemLex says it has served more than 70 customers in three and a half years, six of them among the ten largest drugmakers, and signed a memorandum with Singapore's drug development centre, EDDC.
Why it matters: ChemLex puts its global headquarters and its automated lab in Singapore, with the state investment board at the launch, which is where the money for this kind of work is going this year. What is on the record is a lab and a customer count the company states; no yield, cycle time or comparison with a conventional CRO is published, so “months of synthesis and optimisation into weeks or even days” stays the chief executive's own claim.
What happened: Justin Grace, Gino van Heeke and colleagues at LabGenius searched 44,160 HER2 x CD3 designs (valency, topology, linker, affinity) with multi-objective Bayesian optimisation, building and assaying 980 of them over five cycles under six weeks each. Several leads killed the cancer cell model below 10 pM with no measurable killing of the healthy model up to 100 nM, which the paper puts as selectivity of 10,000-fold or more, over 400 times the 22.9-fold of runimotamab, a clinical-stage benchmark, in the same assay.
Why it matters: On-target damage to healthy tissue is what has kept T-cell engagers out of solid tumours, so a loop optimising the gap between two cell lines aims at the real obstacle. The selectivity figure is a floor, not a measurement: it comes from seeing nothing at the highest dose tested. Everything is cells in a dish, run in-house with two donors per lead, and the paper says these readouts may not match a body. The healthy model is one cell line, T47D, which the paper says carries about as much HER2 as cultured heart muscle cells.
What happened: Shahana Chatterjee, Félix Therrien and colleagues at Mila, four Canadian universities and Entalpic ran their own DFT checks and put the error of a single adsorption energy at about 0.3 eV on metal surfaces and 0.5 eV on oxides. With that error, the thermodynamic overpotential for hydrogen and oxygen evolution puts 37% of OC20's surfaces and 34% of OC22's (47% after a quality filter) level with platinum or iridium oxide, though the materials were drawn at random. A stability test keeps fewer than 5% of the OC22 oxides.
Why it matters: Machine-learning catalyst work, including models trained on Open Catalyst data, commonly ranks candidates by this overpotential, and generative models could optimise for it. If a third or more of random materials tie with the best within error, the score cannot pick winners, even though it still finds platinum. The authors suggest screening first on stability, cost and ease of synthesis. It is a short arXiv preprint built mostly on the datasets' own calculations, with no laboratory experiments.
What happened: Entalpic, with Hugging Face, Mila and university co-authors, released an open evaluation suite and a public Hugging Face leaderboard for models that generate inorganic crystals. Across 12 recent models, stability went with lower novelty and diversity on average, and no model led on every measure.
Why it matters: A common yardstick for generated crystals makes claims of new materials comparable across labs. Every score is computed: no structure was made, and stability here is a calculated estimate, not a measurement.
What happened: GB-0895, an anti-TSLP antibody, enters two placebo-controlled trials in about 1,600 people with severe asthma, dosed every six months. By August 2026 both were recruiting, approved in 33 countries.
Why it matters: Phase 3 tests the drug, not the design method. No efficacy data yet, and the AI role is the company's own account.
What happened: Rather than one reaction type per robot run, Iktos grouped reactions whose chemist-set temperature and time windows overlap, so different chemistries share a reactor. Its retrosynthesis tool Spaya picked the routes, chemists assigned conditions, and a scheduler packed 135 one-step reactions of 27 types into six clusters over three campaigns on a Chemspeed robot. LC-MS found the product in 86 (64%); rerunning 14 failed reductive aminations under new conditions raised that to 97 (72%). Of 12 sent out for purification, 10 came back above 90% pure.
Why it matters: Generative models propose lists of unrelated molecules, and robots built for one-chemistry libraries stall on exactly that, so this works on the make step where AI design meets the bench. The bar for success is the product seen at 5% of the UV trace, not isolated yield, and purified yields ran 4 to 48%. Chemists still chose the targets, the conditions and the reruns: this is a scheduling method for a robot, not an autonomous lab.
What happened: An executive order directs the Department of Energy to connect federal scientific data, compute and facilities around AI-driven research challenges.
Why it matters: AI4S becomes a federal programme that ties national-lab data, compute and facilities together, rather than only a lab or company strategy.
What happened: Mission BioCapital and Santé Ventures led the Boston company's round, with existing investors Lightchain Capital and Cobro Ventures. Founder and CEO Chris Bahl designed miniproteins as a postdoc with David Baker and was founding faculty at the Institute for Protein Innovation; the company says it has made molecules against more than 150 targets. It also moved to a hub-and-spoke model, with programs as LLC subsidiaries. Steve Tregay of Mission BioCapital and Omar Khalil of Santé spoke for the leads.
Why it matters: It is money for the miniprotein route out of the Baker lab: small proteins designed from scratch, made and tested at high throughput, as an alternative to antibodies. What backs it is the company's own count of 150 targets and 'multiple' programs with animal proof of concept, with no data in the release, plus a 2024 BMS option deal worth up to $400M in milestones. The subsidiary model is a bet that the platform outlasts any one drug.
What happened: Chai moved Chai-2 from antibody fragments to full-length IgGs. Of 88 designed IgGs against 28 targets, 86% raised at most one flag on four developability tests, and 24 targets gave at least one clean design. Cryo-EM of five complexes matched the predicted binding sites. Testing 10 to 73 designs per target found a binder for each of six GPCRs, two of them agonists, and two binders to a KRAS G12V peptide-MHC that spared G12D and wild-type.
Why it matters: Binding alone does not make a drug; this tests whether the designs also look manufacturable and hit the intended spot. Every assay was Chai's own, in a preprint, and developability flags are screening thresholds, not evidence a molecule will work in people.
What happened: Choosing 25 compounds a round from a semi-synthetic EGFR dataset, a binding-strength model that says how unsure it is about whole batches of molecules at once reached the same best potency in about 5 times fewer rounds, and the same top-ten average in 7 times fewer, than ranking by predicted score. Terray's researchers built it on the company's COATI chemistry model, partly pretrained on synthetic data. The paper reports up to 10 times fewer rounds on 50,000 blinded compounds from Terray's own screen.
Why it matters: Making and testing molecules a batch at a time is the slow, costly step in drug discovery, so choosing each batch well saves rounds of chemistry. Picking only the top predicted compounds tends to fill a batch with near copies; a model that knows how its guesses move together can spread the batch out. Terray runs a large in-house binding screen, which is where a method like this would be used, and the code is public.
Evidence and limits: Retrospective only: every result replays compounds already measured, with their values precomputed, on a semi-synthetic benchmark and on one undisclosed target from the company's own screen, with no new molecule made. All authors are at Terray, and the paper is a preprint.
What happened: The team joins Biohub, the Chan Zuckerberg nonprofit, and co-founder Alex Rives becomes its head of science. Biohub is growing its compute tenfold, to 10,000 GPUs by 2028. No terms were disclosed.
Why it matters: A startup that raised a seed of more than $142M in 2024 now works inside a nonprofit: its next protein models are paid for by philanthropy, not by customers.
What happened: Takeda and Flying Fish Partners co-led the seed, joined by Columbus Venture Partners, Cercano Capital, Washington Research Foundation, Alexandria Investments, Pack Ventures and Argonautic Ventures. Matthew Bick (CEO), Javier Castellanos (CTO) and Hector Rincon (CSO) founded the Seattle company. Pfizer signed a research and licence deal with an upfront payment and over $330M in possible milestones; Kite Pharma signed a multitarget deal for molecules used in cell therapies.
Why it matters: Antibodies and multispecific drugs mostly block pathways. Accipiter designs proteins from scratch, including ones that switch pathways on, several on the same cell in one molecule, in place of a combination of drugs that may not reach the cell together. The founders come from Seagen, Merck, Neoleukin and the Institute for Protein Design, and worked on the first fully de novo protein to enter clinical trials. Heather Gorham of Flying Fish sits on its board.
Evidence and limits: The release shows no data on any designed molecule, and the under-two-month cycle from design to lab is the company's own figure.
What happened: A for-profit company is created to commercialize scientific agents while FutureHouse remains a nonprofit research lab.
Why it matters: The field experiments with separating public-interest research from commercial deployment.
What happened: Given an objective and a dataset, Kosmos shares a structured world model between a data-analysis agent and a literature agent over about 200 agent rollouts; a run executes 42,000 lines of code and reads 1,500 papers on average. Scientists checked 102 statements from three reports: 85.5% of data-analysis statements reproduced, 82.1% of literature statements held, and 57.9% of interpretive ones. Of seven highlighted findings, three reproduce unpublished or preprinted results it had not seen.
Why it matters: The test is whether an agent can stay on one question long enough to reach a finding, and here outside data and unpublished results are the check. But the checkers and the groups who estimated 6.14 months of work per run are the paper's co-authors and collaborators, the weakest statements are the interpretations, and the new findings were checked against other existing datasets, not new experiments; one wet-lab test is under way.
What happened: At SITC 2025, Yubin Xie, Jérémie Decalf, Daniel Bear and colleagues at Noetik showed OCTO-vc, a model trained on spatial gene-expression maps of over 40 million cells from 1,399 non-small cell lung cancer resections; it predicts a cell's genes from its neighbours. They set a virtual macrophage among 8 copies of a patient's tumour cells or fibroblasts, 500 neighbourhoods per patient, 873,000 virtual macrophages in all. Low-oxygen tumour programs went with the immunosuppressive SPP1 state, and fibroblasts linked to it showed higher matrix-gene expression in KRAS-STK11 co-mutant tumours.
Why it matters: This is the model GSK licensed in January 2026, and these simulations are the use Noetik sells it for: asking how a cell would behave in a neighbourhood no lab assembled. The KRAS-STK11 finding gives a candidate reason why those lung cancers respond poorly to immunotherapy, a hypothesis a lab can test.
Evidence and limits: A company poster, every author at Noetik. The findings are associations inside the model's own simulations; no lab experiment or patient outcome was checked, and the link to immunotherapy resistance is the poster's suggestion.
What happened: Genesis's preprint reports Pearl beats AlphaFold 3 on the public Runs N' Poses and PoseBusters tests, by 14.5% and 14.2% at placing a molecule within 2 Å with a valid pose.
Why it matters: Genesis ran the comparison itself. The 3.6x gain is on its own private targets, and Pearl runs only inside Genesis.
What happened: The randomised, double-blind, placebo-controlled trial enrolled 264 patients whose pseudobulbar affect followed ALS, multiple sclerosis, stroke, head injury or Parkinson's disease, at 48 hospitals over 9.5 months, led by Dongsheng Fan of Peking University Third Hospital. The company reports statistically significant gains over placebo on the CNS-LS lability score, on clinician and patient impression of change, on the SF-36 mental-health index and on bulbar function, with benefit from week 2 across 12 weeks of treatment. It expects to file for approval in China in 2026.
Why it matters: The AI here designed the formulation, not the molecule: METiS's AiTEM tool predicted how a drug and its excipients behave together, to make a tablet that dissolves in the mouth for patients who often have trouble swallowing. The company says that took formulation work from the industry's usual one to two years down to under three months. What the trial tested is the tablet, in patients: whether it relieves the symptom. Whether the tool shortened the work is a separate claim, and the trial does not speak to it.
Evidence and limits: No numbers are public: the release does not name the primary measure and gives no effect size, p-value or adverse-event rate, and no paper or registry record was read. The three-month figure is the company's, against an industry average it states itself. MTS-004 is a Category 2.2 improved new drug in China, a new formulation of an active ingredient already known, which the release does not name, so the trial tests the formulation and the indication, not a new mechanism.
What happened: Konstantin Jakob, Aron Walsh, Karsten Reuter and Johannes Margraf (Fritz Haber Institute, Imperial College London, University of Bayreuth) trained models on the ICSD crystal database to predict from a composition alone whether a compound shows substitutional disorder, where atoms of different elements share a site. The best reached 90% accuracy on a held-out test set. Applied to GNoME's 554,054 public structures, it predicts 80-84% likely disordered, against 39-47% for the Materials Project.
Why it matters: GNoME, like most computed databases, lists each compound as a perfectly ordered crystal. If most of its compositions would form disordered in a real furnace, the predicted structures and properties describe a material that would not actually be made, and many "new" entries would be known disordered compounds written in ordered form. A composition-only check like this can be run before anyone tries a synthesis.
Evidence and limits: The estimate is itself a prediction, and the authors call GNoME a "sizable covariate shift" from the classifier's training data. They add that ordered predictions remain useful; no GNoME compound was made to test it.
What happened: Lila closes its Series A at $350 million: Braidwell and Collective Global co-led the first $235 million in September, and NVentures, Analog Devices and IQT joined the second close. Total raised reaches $550 million, for more AI Science Factories and a first cohort of customers. Geoffrey von Maltzahn, co-founder and chief executive, announced it.
Why it matters: Investors paid $350 million before Lila had published any result, for the plan of turning a lab into a data source for models. Whether customers use the platform is the next thing to check.
What happened: Former OpenAI and DeepMind researchers form a company focused on AI-driven physical experimentation and materials discovery.
Why it matters: A record-sized seed round signals investor belief that experimental infrastructure can be a standalone AI platform.
What happened: Yuge Ji, Fabian Theis and colleagues pretrained a transformer on 4.7 million conditions from public screens (26,000 genetic and chemical perturbations, 1,550 cell states) to predict gene expression, viability and cell shape for perturbations and cells it had not seen together. It beat linear, random-forest and neural-network baselines. Screening all 1.9 million ChEMBL compounds across 1,406 cell lines, 2.7 billion predicted experiments, it picked compounds for melanoma; four of the five tested reduced viability in two melanoma lines but not in HeLa cells.
Why it matters: Measuring every compound in every cell type is out of reach, so a model that carries what it learned in one assay or cell line to another could choose which experiments are worth running. Prophet's claim is that training across unrelated screens makes its predictions better than a model trained on one, and the melanoma screen tests it on compounds the training data had not covered. It is the basis of Reflector Bio, which the first author founded.
Evidence and limits: The lab check is five compounds in three cell lines, run by the authors, and the paper gives no doses for the five; cell death by apoptosis was confirmed for two. The compounds were chosen because their predicted effects resembled approved drugs, so the screen points to known biology rather than a new mechanism. It is a preprint, not peer reviewed.
What happened: Meta's FAIR team with Georgia Tech, Carnegie Mellon, Oak Ridge, Rice and five CuspAI authors report nearly 60 million DFT single-point calculations of CO2, H2O, N2 and O2 adsorption in 15,000 metal-organic frameworks. CuspAI's transformer generated 53,000 candidates, 4,000 survived filtering and 460 entered the dataset. Of the machine-learned potentials released with it, eSEN trained on the full set reports an adsorption-energy error of 0.078 eV.
Why it matters: A startup's generated structures go into a public dataset others will train on, so its model's output is there for anyone to test against. The released potentials are not scored against DFT alone: their Henry's coefficients are set beside experimental values from the literature, where the journal version has eSEN trained on the full set reaching 0.0085 eV for nitrogen against a conventional force field's 0.0190, while on CO2 the larger UMA-Medium leads instead.
Evidence and limits: CuspAI co-wrote the paper and nobody made a sorbent for it. The experimental check is three MOFs and two gases, against literature coefficients that are scarce: where duplicates exist their spread is 0.016 eV for CO2 and 0.009 for nitrogen, about the size of the best model's error. That nitrogen figure also moved between versions, from 0.0433 in the preprint, worse than the force field, to 0.0085 here.
What happened: Ali Madani's team trained language models on over 1 million CRISPR operons and generated editors about 400 mutations from SpCas9. Several matched or beat it on activity and specificity in the paper's tests, and OpenCRISPR-1 also works with base editing.
Why it matters: A generated protein was made and run in human cells, past a score on a screen. Profluent wrote and ran the tests, and the abstract reports no test in animals or patients.
What happened: Latent-X generates a binder's all-atom structure and sequence together with the target, for a chosen epitope. Of the top 30 macrocycle designs submitted per target, 11 to 17 were made and measured, and 91-100% bound MDM2, MCL-1 or PD-L1; on five mini-binder targets, among them IL-7R-alpha, TrkA and the SARS-CoV-2 receptor-binding domain, 100 designs each gave hit rates of 10-64%, with the strongest reported below the assay's 0.01 nM floor. Seven targets in all, each one a benchmark earlier methods had used.
Why it matters: A hit rate that high changes what a design round costs: tens of molecules to make and assay instead of thousands, which is the step where a generative model meets a bench budget. Latent Labs chose the designs and set the thresholds; GenScript ran the binding assays, and the specificity check was in-house. Its comparison takes the rivals' hit rates from their papers and remeasures only their best binders, where its own replication disagrees sharply with the published numbers, and Latent-X does not lead on every target. No peer review and no outside lab has repeated it.
What happened: OpenAI and Google DeepMind report natural-language solutions reaching the IMO gold threshold.
Why it matters: Mathematical reasoning advances beyond specialized geometry and formal-only pipelines into broader problem solving.
What happened: X-Master is a reasoning agent that writes and runs code and calls tools while it thinks; X-Masters runs several of them in what the authors call a scattered-and-stacked workflow. The authors, from Shanghai Jiao Tong University and DP Technology, report 32.1% on Humanity's Last Exam, a test of expert-level questions, against 26.6% and 26.9% for OpenAI's and Google's Deep Research, and say it is the first score above 30%. The code is open source. It is Part I of the SciMaster series.
Why it matters: SciMaster is the agent layer of DP Technology's Bohrium platform, and this is a public score behind it. The test is general question answering, so the score says the agent's tool use works on hard questions, not that it does science.
Evidence and limits: The authors' own run, on the leaderboard as it stood in July 2025. The figures for the other systems are as the paper quotes them.
What happened: Up to 20 designs per target, 52 targets with no antibody in the PDB. One round gave at least one hit for 50% of targets; miniproteins hit 68%.
Why it matters: If it holds, antibody design needs a few dozen wet-lab tests, not thousands. The tests were Chai's own, on a preprint, and the entry is read from the abstract only.
What happened: DeepMind’s model predicts regulatory activity and variant effects directly from DNA sequence, pushing foundation modeling beyond structure toward noncoding function.
Why it matters: The harder question is no longer only whether sequence can be modeled at scale, but whether those predictions remain reliable across biological contexts and prospective use.
What happened: Dylan M. Anstine, Olexandr Isayev and colleagues at Carnegie Mellon, with co-authors at Purdue, Notre Dame, Michigan State and Deep Principle, trained it on about 4.7 million density functional theory calculations of neutral, closed-shell reactions of carbon, hydrogen, nitrogen and oxygen. On 1,650 reactions its barrier heights were off by 2.0 kcal/mol (RMSD), and its transition states sat 0.07 Å from DFT's on average. A batched pathway search on one GPU converged over a million 10-atom reactions a day, against 5 for the reference DFT.
Why it matters: General-purpose machine-learned potentials are trained mostly on stable molecules, and the authors show two of them missing barriers by 15 to 25 kcal/mol on a standard reaction benchmark, too far off for kinetics. Training on reacting structures closes most of that gap, at a speed that makes searching millions of pathways practical. Every check is against DFT, the model's own teacher, including an 11-step glucose pyrolysis pathway taken from an earlier computational study that ends at the product experiments observe. Charged, open-shell and metal-containing reactions are outside its scope.
What happened: Inside an alloy system researchers bounded, active learning started from 140 data points and picked eight new compositions over four rounds, each made and tensile-tested, reaching HEA05 (Fe35Ni29Co21Al12Ta3): 1.75 GPa yield at 25% uniform elongation.
Why it matters: Small data carried a real alloy design, not a large screen: 140 points and eight new compositions, each physically tested; but strength and elongation still trade off across processing routes.
What happened: The bioRxiv preprint says two genome-wide screens targeting all human protein-coding genes give eight million cells, each sequenced to over 16,000 UMIs, and calls it the largest public Perturb-seq atlas. Ci Chu is the corresponding author.
Why it matters: Xaira makes its own perturbation data and shares part of it. A dataset and method paper: no drug target, no working model shown.
What happened: The system produces 50 possible cyclone trajectories up to about two weeks ahead with input from operational forecasters.
Why it matters: AI forecasting moves from retrospective papers toward a live interface shaped by real forecasting institutions.
What happened: A randomized, double-blind Phase 2a trial reports safety and efficacy results for rentosertib, a generative-AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis.
Why it matters: Once an AI-discovered molecule reaches controlled human trials, clinical safety and efficacy become more important evidence than the sophistication of the discovery model.
What happened: Gemini proposes programs while automated evaluators select and evolve them across algorithms, chip design and data-center scheduling.
Why it matters: A reusable discovery pattern emerges wherever outcomes can be measured cheaply and automatically.
What happened: Specialized agents for literature search and scientific reasoning move from internal research into a public platform.
Why it matters: Scientific-agent capability becomes a usable service rather than only a paper result.
What happened: Basecamp Research and co-authors published HiFi-NN in iScience: it retrains ESM-2 protein embeddings so that enzymes sharing more levels of their Enzyme Commission (EC) number sit closer, then labels a new sequence by its 20 nearest neighbours. On the 149-enzyme Price set its F1 score was 0.62, against 0.49 for CLEAN, the strongest deep-learning method compared, and 0.39 for BLASTp. It also withheld the recorded label from 86% and 90% of two BRENDA sets in a class an earlier lab study found largely mislabelled.
Why it matters: At least a third of microbial proteins cannot be given a function by matching them to known ones, and a function label is what lets someone looking for a catalyst find one. On the Price test, labels improved when the authors added two million sequences from Basecamp's own collected metagenomic database to the lookup set; on a second, time-based test they made no difference.
Evidence and limits: All scores are the authors' own computer benchmarks; the paper says no lab test checked the cases where its labels differ from other tools'. The BRENDA check shows the method withheld a suspect label, not that the label it gave instead is right.
What happened: The paper charts the trade-off between accuracy, speed and memory. It finds that models that are neither rotation-equivariant nor energy-conserving can still predict physical properties, including ones that need higher derivatives of the energy.
Why it matters: Speed and size are why a potential can run whole-system simulations. The accuracy comparison uses the authors' chosen evaluations, and nothing was made or measured in a lab.
What happened: Bo Chen, Jie Tang, Le Song and colleagues at BioMap and Tsinghua University trained one protein language model with 100 billion parameters on 1 trillion tokens, combining the two usual ways of training: filling in hidden letters, which suits reading a sequence, and writing the next letter, which suits making one. They report it beats strong baselines on 18 protein benchmarks in four categories, underpins a structure predictor ahead of other language-model tools, and writes new sequences that resemble natural ones. A preprint came out in July 2023.
Why it matters: Most protein language models were trained either to fill in sequences, which suits reading them, or to write them letter by letter, which suits making them, and the authors say that left them struggling to do both. This paper shows one model doing both, at a size the authors call unprecedented, and it is BioMap's main peer-reviewed model paper. The weights and the downstream datasets are public on Hugging Face.
Evidence and limits: The abstract reports only computed scores on existing datasets and sequences judged by software, with no protein made and measured, and gives no margins over the baselines.
What happened: Its first external financing is directed toward internal oncology and immunology programs and partnered drug discovery.
Why it matters: The AlphaFold-to-drug thesis attracts large-scale outside capital before clinical proof.
What happened: Zechen Wang, Liangzhen Zheng, Weifeng Li and colleagues at Shandong University, the Shenzhen Institute of Advanced Technology and Zelixir, its founder Sheng Wang among them, trained a model on BRENDA and SABIO-RK data to predict kcat, Km and kcat/Km; it beat DLKcat and UniKP, most clearly on kcat, on test sets that keep similar enzymes apart. For a vanillin-making step it shortlisted 150 enzymes; five were tested. One, from Sphingobium, had 19.53 times the reference's activity, and two mutants it helped pick were 3.2 and 3.3 times as active again.
Why it matters: Enzyme speed is a large part of whether a biomanufacturing route pays, and here a prediction led to enzymes made and measured. It is one reaction, five candidates and relative activity measured by a Zelixir author, and the model only narrowed a list that sequence and structure filtering also shaped. On TurNuP's test set its fit on the least familiar enzymes was weak (R squared 0.13 against 0.33) until retrained on TurNuP's data. The paper does not say whether this vanillin work is the flavour Zelixir now makes.
What happened: Flagship Pioneering unveils Lila, founded in its labs in 2023, with $200 million of committed seed capital to build scientific models and autonomous labs for life science, chemistry and materials. Geoffrey von Maltzahn is co-founder and chief executive; Noubar Afeyan, Flagship's founder, was co-founder and chairman at launch.
Why it matters: Here the money pays for the experiments themselves, as a source of data for models, not only for software or a drug pipeline.
What happened: A multi-agent Gemini system generates, critiques, ranks and evolves hypotheses; collaborators test selected biomedical proposals.
Why it matters: Hypothesis generation becomes an explicit search process rather than a single language-model response.
Evidence and limits: One showcase result, on how a family of bacterial gene islands spreads between species, came from José Penadés's lab. Its bioRxiv write-up shows the system's background cited the lab's 2023 paper: the islands build their own capsids and need only phage tails. The top hypothesis, that they take tails from other species' phages, went one step past that paper, on a question the lab had already answered in a paper then under review.
What happened: Benedikt Alkin, Johannes Brandstetter and colleagues at Emmi AI and JKU Linz built AB-UPT, a neural network that stands in for a computational fluid dynamics (CFD) solver on car aerodynamics. On DrivAerML, high-fidelity simulations of about 140 million cells each, its error on surface pressure was 3.82%, against 4.81% for Transolver and 4.35% for a plain transformer, and lower on airflow speed and swirl too. It can also predict from the car's CAD shape alone, without building the costly simulation mesh.
Why it matters: High-fidelity car simulations are expensive (the DrivAerML set cost about $1M to compute), so engineers can only try a few shapes. A model that returns the pressure and flow fields in seconds lets them try many more, and AB-UPT computes drag and lift from its own predicted fields. The paper, published in Transactions on Machine Learning Research, is the measured accuracy behind the car-aerodynamics models Emmi was selling when Mistral AI bought it in May 2026.
Evidence and limits: Every comparison is the authors' own, on public car datasets, with baselines they retrained; Transolver++ was left out because both the authors' version and its public code scored worse than plain Transolver. Drag and lift are checked against simulations from the same CFD dataset, not against a wind tunnel or a new design.
What happened: Microsoft's diffusion model proposes stable inorganic materials and includes physical synthesis of a generated candidate.
Why it matters: Materials AI shifts from screening a fixed list to inverse design: asking what material should exist.