AI FOR SCIENCE · TIMELINE
The 23 entries the timeline holds for July 2026, newest first. Each was read at source. The full timeline, with filters →
What happened: Epoch AI expands the benchmark beyond private-answer questions to problems with no known solution at release.
Why it matters: It changes what evaluation asks: not whether a model can uncover a hidden answer, but whether it can advance a problem with no answer yet.
What happened: GSK pays up to $110M, upfront and milestones, for Relation to run automated cell-perturbation experiments and train foundation models on them. Its model MORGAN has no benchmark, size or paper yet.
Why it matters: A pharma partner pays for data made to train a model, not for a target list. The upfront share is not stated and MORGAN has shown no results.
What happened: The Financial Times reported that Google DeepMind has broken up the team that built AlphaFold. DeepMind confirmed internal moves, to Gemini projects and, per the FT, to enzyme design, nuclear fusion and genomics; others went to Isomorphic Labs, and a spokesperson told Scientific American that many of the moves happened more than a year ago. John Jumper announced in June that he was leaving for Anthropic. The AlphaFold database and prediction server remain available.
Why it matters: AlphaFold was how DeepMind set itself one hard scientific problem and stayed on it for years. Its people now work on Gemini-focused projects and on other sciences. Pushmeet Kohli, who leads the AI for Science team, wrote on the day that the team had expanded its work rather than pivoted. John Moult and Debora Marks told Scientific American they were not surprised: the task AlphaFold was built for was largely done. Moult added that outside groups' own versions now carry the work on.
Evidence and limits: Most details are the Financial Times' reporting, among them its own count that nearly a quarter of the original papers' DeepMind authors have left. That Jonas Adler and Alexander Pritzel followed Jumper to Anthropic was reported by Bloomberg, citing people familiar with the matter.
What happened: Duke's Raygun shrinks or grows proteins using language-model embeddings. In cells, 6 of 8 shortened fluorescent proteins fluoresced, 2 of 11 TurboID variants worked, and two EGF variants bound EGFR better than wild type.
Why it matters: Function survived big edits only in some cases: fluorescence was dim and the half-size TurboID was inactive. Not independently replicated.
What happened: Bunsen takes a scientific question in plain language, picks the computational method, sets up the multi-step workflow, launches the calculations on Schrödinger's own validated tools and interprets the output. Selected customers got early access on 27 July, with NVIDIA and Google Cloud supplying compute; full release is planned by the end of 2026. On 5 August Bristol Myers Squibb, a long-time user of the platform, signed to deploy it at large scale and to co-develop features with Schrödinger's chemical-space search and RetroSynth synthesis planner.
Why it matters: Schrödinger contrasts Bunsen with general-purpose AI that only retrieves and summarises information: it is built to drive one vendor's established physics-based methods, which drug hunters already use to rank molecules before making them. Schrödinger aims it at drug hunters who are not computational chemists as well as at computational chemists, and Ramy Farid, its president and CEO, says licences priced on how much is computed will capture the extra use. A drugmaker deploying it across its research organisation is one test of whether that happens.
Evidence and limits: Neither release gives a measurement: no task success rate, no comparison with chemists working without it, no customer result. The BMS agreement has no stated terms or length.
What happened: In an open-label Phase 1b of ENV-294, nine adults took it daily for 28 days. Mean EASI score fell 68% by day 28, 85% by day 42. One adverse event, a headache.
Why it matters: Nine patients, no placebo, company-reported: not proof it works. The release does not say how AI shaped this molecule.
What happened: DOE picked 278 projects for award negotiations; selection is not a commitment to fund. Brookhaven leads 7 and Stony Brook 3, putting AI into reasoning, instruments, design and data infrastructure.
Why it matters: What DOE chose to test is AI as parts of workflows, not one universal AI scientist, and every capability named is still a Phase I plan.
What happened: Berkeley Lab leads 13 and Oak Ridge 9: Earth and energy forecasting, materials loops, digital twins, and shared software and networks, all selected for award negotiation.
Why it matters: The portfolio follows the data and facilities the two labs already hold; the selection materials, checked 2026-09-27, report no project result that can be checked.
What happened: PNNL leads 16 Phase I projects and joins 26 more, from subsurface simulation to instruments and manufacturing; targets such as a 100-fold speedup are written down in advance.
Why it matters: Targets stated in advance make these awards testable: a missed target will be as informative as a met one. Selection is not yet a funding commitment.
What happened: PoLARIS is to link more than 20 autonomous stations across five labs at the University of Chicago and Argonne, among them Argonne's Advanced Photon Source and Polybot, and take experiment requests through cloud interfaces. The materials are liquids, polymers, gels, colloids, foams and granular matter. UChicago expects 100 to 300 experiments a day once running, end-to-end autonomous experiments on its own first questions within 18 months, and outside users by the end of the grant's second year. Ian Foster leads it with Stuart Rowan, Jie Xu and Yuxin Chen.
Why it matters: Most autonomous labs serve the group that built them. This one is funded to take work from universities, companies, government labs and institutions without such equipment, so a researcher with a good idea and no robots can have the experiment run. If it works, access to automated experiment becomes something a scientist applies for rather than builds.
Evidence and limits: Nothing has run for an outside user yet: the throughput, the 18-month and two-year dates and the access model are all UChicago's own plan, in its announcement. The NSF award record reads $20M intended over four years, of which $5M is obligated for 2026, as a cooperative agreement under a new Programmable Cloud Labs program, so it is a commitment renewed yearly, not money in hand.
What happened: Pushmeet Kohli and Karthik Narain announced, at the DOE Genesis Mission Summit, $40 million in tokens and cloud credits: Genesis awardees get AlphaEvolve, AlphaFold 3, AlphaGenome, WeatherNext and AlphaEarth Foundations, and DOE lab staff get Gemini for Government for a year. The post quotes Steven Spurgeon, a materials data scientist at the National Laboratory of the Rockies: Gemini in the lab's instruments cut microscope calibration from over 90 minutes to about 13, and the steps to focus an image from as many as 50 to two.
Why it matters: It is the clearest case yet of Google's models running inside a national lab's materials instruments rather than proposing ideas from outside. The credits are a subsidy, but the microscope example shows where DeepMind's tools are reaching the physical lab: through other institutions' equipment, not its own materials lab, which has not opened.
Evidence and limits: The calibration figures are a quote in Google's own post; no paper or lab report behind them was found. The post names no materials result.
What happened: Kleiner Perkins and NEA co-led, with Bezos Expeditions, Lux Capital, AMD Ventures and Britain's Sovereign AI Venture Fund new and Temasek and Prosus following on; Bloomberg reported the $2.6B valuation, up from $520M a year earlier. Co-founder and chief executive Chad Edwards told Bloomberg most of the money goes on labs in Cambridge, Singapore and the Bay Area, and that semiconductors take 80% of this year's research, including designing out ruthenium and iridium. The Foundry pools compute, labs and expertise with 45-plus organisations, NVIDIA and Meta among them.
Why it matters: It is the largest closed round in materials and chemistry on this record, and it buys wet labs and partners' instruments rather than models: the Foundry's 45 members are where a predicted material would get made and measured. Against that, the same report says CuspAI has carried no project to a material ready for commercial development, and says its competitors have not either, so the price is on the search, not on a product. The pivot to semiconductors and to designing out two constrained metals says where the paying problem is.
What happened: Anthropic, the Gladstone Institutes and Cerebral Valley ran it online on July 7-13 for $100K in credits; about 300 projects are posted. Research track: NCypher (Faith Ogundimu) first, ranking 10,869 DNA variants from a nearly always fatal childhood brain cancer for lab follow-up; Extremolith (Jaymin Patel) second, redesigning enzymes for heat, acid or salt. Build track: Lazarus (Dean Sherry) first, reviving abandoned research code; Provinans (Shereen Lee) second, merging cancer records. Gladstone's prize: Trialign (Jules Park, Neil Wang), matching cancer patients to trials.
Why it matters: The winners show Claude reaching into several real research steps, and the gap between finishing a task and a trustworthy scientific result. NCypher narrowed 10,869 DNA variants to lab candidates, but its team calls the effect modest and the result hypothesis-generating. Lazarus got 17 dead research repositories running, yet of 20 sampled, 5 reproduced their paper's numbers: running is not reproducing. Extremolith's enzyme-design demo returns a precomputed result until its builder finds affordable GPUs. None is peer-reviewed or lab-tested, and Anthropic and Gladstone judged. AI is helping with screening, analysis and execution here, not making discoveries.
What happened: The 2023 deal is worth up to $140M in upfront and milestone payments. Aqemia designs small molecules on its physics-based generative AI platform; Sanofi runs the wet lab, development and sales. The release gives no amount for this payment and does not name the target.
Why it matters: A partner adding a target shows the deal is being used, not that the molecules work. Maximilien Levesque, Aqemia's CEO, is the only person quoted; Sanofi said nothing in the release.
What happened: Latent-Y takes a goal in words, or a paper, and runs the whole campaign itself: reading the literature, picking the epitope, generating candidates with Latent-X2 and choosing which sequences go to the bench. Its technical report takes nine targets to the bench and gets lab-confirmed nanobody binders against six, down to single-digit nanomolar; most of those campaigns ran end to end, and on the cross-species one an expert had to step in and correct the agent. Approved researchers get 250 designs a day free, with credits for more.
Why it matters: This is the step from a model a protein designer drives to one that makes the design decisions, and the test is who can use it. The launch release names three outside groups that got lab-validated designs, including a UC Davis lab that blocked a human ion channel with four of five designed nanobodies, and says some had no computational design experience. Those outside results are Latent Labs' own account; the 6-of-9 figure is its own wet lab.
What happened: Index Ventures led the Series C alongside Kleiner Perkins, Sequoia Capital and Dimension; Bain Capital Ventures, Battery Ventures and Sapphire Ventures are among the new investors, and OpenAI, Thrive Capital and Menlo Ventures followed on. Chai says the money goes to compute, data and research for its models and to reaching more scientists. Joshua Meier, co-founder and chief executive, spoke for Chai; Nina Achadjian, Ilya Fushman, Pat Grady and Zavain Dar for the four firms.
Why it matters: Seven months after a $1.3B valuation, investors pay nearly three times as much for a company selling model access to Eli Lilly, Pfizer and Novartis rather than developing drugs of its own. The price is a bet on software licensed to pharma, not evidence that its antibodies work.
What happened: Iktos and the contract research firm Sygnature Discovery designed inhibitors of PolQ, a DNA-repair enzyme pursued as a cancer target, using Iktos's generative software Makya alongside docking. Of 88 compounds made and tested, six inhibited it below 5 micromolar. An oxime compound whose group a chemist repositioned reached pIC50 7.3 and matched the reference inhibitor ART-558 in cells; a second generative round around it reached 8.3 in the biochemical assay, against 7.6 for ART-558.
Why it matters: This is a generative design carried through synthesis and assays rather than a docking score, and the paper credits model and chemists together without separating their shares. The cell assay is where it stops: the 8.3 compound showed no potency against BRCA-deficient cells, which the authors put down to poor permeability, so the biochemical gain did not carry into cells. Every number comes from the two companies whose staff wrote the paper, and selectivity against related enzymes and any animal test are still to come.
What happened: The GBP 25M ($34M) takes Alchemab's Series A to GBP 109M and pays for its clinical pipeline and for growing its antibody database from 500 million to over a billion sequences. Its Resiliome platform uses AI to find naturally occurring antibodies in people unusually resistant to hard-to-treat diseases. Jane Osbourn, co-founder and chief executive, and Kate Bingham, board chair, spoke for the company; Charlotte Lawrence and Carmine Circelli for the bank, whose commercial arm British Patient Capital made the investment.
Why it matters: The release itself offers a deal, not data, as its validation: Eli Lilly licensed Alchemab's first antibody, ATLX-1282 for ALS, in May 2025 for up to $415M before any human had received it, and its Phase 1 began that September. Lilly has also signed a discovery collaboration and invested, so its bet reaches the platform too, but none of the releases reports clinical results or shows the platform picks better antibodies than other methods. The money is the UK government's: the British Business Bank is the state's development bank.
What happened: Agents propose hypotheses and an automated culture and metabolomics lab tests them. In yeast it found glutamate slows spermine-treated cells, and aminoadipate partly rescues formic-acid stress.
Why it matters: A full loop from hypothesis to wet-lab check, but two findings in one yeast, by the builders, read from the abstract.
What happened: The Munich company was founded this year by chief executive Hanh Nguyen, chief technology officer Tiago Cerqueira and chief scientific officer Milan Allan, who chairs experimental physics at LMU Munich and heads its in-house synthesis and characterisation. Its engine sits on Alexandria, the open database of predicted crystal structures Cerqueira co-built, plus in-house property datasets and its own lab. Amanda Birkenholz, a principal at UVC Partners, spoke for the lead. The company says it has a pipeline of rare-earth-free magnet candidates whose predicted properties were checked against experimental data.
Why it matters: The release says roughly 90% of advanced permanent magnets are made in China and that export restrictions have made supply uncertain for makers of electric vehicles and wind turbines; a pre-seed to replace their rare earths follows that supply problem, not a benchmark. Alexandria is a real asset, which the company calls the largest and most widely used database of its kind, but what it says about its own candidates is one sentence, with no compound, no measurement and no paper. The line that the last breakthrough in permanent magnets came more than 40 years ago is its release's too.
What happened: Aureka Biotechnologies released OpenDDE under Apache-2.0 with training code, weights and benchmarks: a model that predicts how proteins, nucleic acids and small molecules fit together, built on the open Protenix and OpenFold3 models. On three antibody-antigen test sets it placed an acceptable interface for 51.0%, 70.0% and 66.4% of complexes, the best of the models compared. The third set, 164 structures newer than the training data, was curated by Aureka; there Protenix-v1 reached 50.0% and Boltz-1 25.0%.
Why it matters: Antibody-antigen interfaces are flexible, hard for co-folding models, and where antibody companies need them most, and the strongest reported system, Isomorphic Labs' IsoDDE, is closed. An open model close to it lets any lab run and retrain the structure step itself. It is also the model under Aureka's agentic design harness, so that harness's lab results rest on it.
Evidence and limits: Every score compares computed structures with deposited ones, in runs by Aureka; no outside group has repeated them. The paper says both that it reaches IsoDDE's reported accuracy and that it only narrows the gap, and IsoDDE's recipe is not public. The authors say it is not tuned for small-molecule docking or affinity. A preprint, not peer reviewed.
What happened: The paper reports state-of-the-art cell-type classification, even for species over 685 million years from the training set, and zero-shot disease-state identification in human cells.
Why it matters: One model comparing cells across the tree of life bets that expression data alone carries shared structure. The abstract reports tests on existing annotated data only, with no new experiment.
What happened: Charlotte Knierim, Luca Versari, Robert Obryk, Blaise Agüera y Arcas and Rif A. Saurous of Google's Paradigms of Intelligence team used a direct detector for self-copying 64-byte programs. Sampling random programs from a tuned byte distribution found one after about 94,000 programs on average, against about 5 million when running BFF, some 25 times fewer. Blocking how programs merge did not stop self-copiers from appearing, only from taking over the soup.
Why it matters: The 2024 paper read the soup's takeover as the sign of replicators and argued that paired interactions are an effective way to find them. Detecting replicators directly separates two things: appearing, which random mutation alone does as well or better, and spreading, which needs interaction. The authors write that compositionality is not needed for discovery in BFF, which changes how the 2024 result reads.
Evidence and limits: Computed by the same team in a short preprint, mainly on one BFF dialect (bff_selfmove), though the authors say other variants gave very similar results. The detector checks whether a program copies most of itself over five rounds, a working definition. The best distributions were designed with knowledge of what BFF writes, so the comparison is about search speed, not about what a soup would do unaided.