AI FOR SCIENCE · TIMELINE
The 13 entries the timeline holds for January 2026, newest first. Each was read at source. The full timeline, with filters →
What happened: The LaTeX workspace combines drafting, citations and model assistance and is made free to ChatGPT users.
Why it matters: The consequential move is into the daily research surface, where assistance can shape how evidence is written, cited and shared.
What happened: Anna Goldie (CEO) and Azalia Mirhoseini (CTO), who built AlphaChip at Google, founded the Palo Alto lab to speed up chip design with AI. Lightspeed leads at a $4B post-money valuation, less than two months after launch; DST Global, NVentures, Felicis, 49 Palms, Radical Ventures and Sequoia join. Guru Chahal of Lightspeed speaks for the lead. The money is for hiring and compute.
Why it matters: Chip design is an engineering science where AI already has a deployed track record: the company says AlphaChip was used in four generations of Google's TPU and by outside chipmakers. The round puts frontier-lab money behind using AI to design the hardware AI runs on, a loop several 2026 startups now chase.
Evidence and limits: A round shows where money goes, not that the method works: the company has published no result of its own. How widely AlphaChip is used is the company's own account.
What happened: Auditing 35 tokenizers, Michigan and Carnegie Mellon find closed atom vocabularies turn valid but rare bracket atoms into [UNK]; Smirk's 165 base tokens cover all of OpenSMILES by construction.
Why it matters: A model's ceiling starts with what its input preserves; the gains are the same team's computation, not isolated to the tokenizer.
What happened: Under the deal Isomorphic makes the predictions and designs, and Johnson & Johnson, through Janssen Biotech, runs the experimental assays and takes programmes forward. Isomorphic says its design engine makes candidates against hard targets as small molecules, antibodies, peptides and molecular glues, and that this deal uses that range, large-molecule and biologics design included. Sarah Skerratt, Isomorphic's chief research officer, is quoted in the release.
Why it matters: Isomorphic's 2024 deals with Lilly and Novartis were for small molecules. This one takes its large-molecule and biologics design to a third large drugmaker, under a split that leaves every wet-lab test to the partner, so J&J's assays are where the engine's designs get checked.
Evidence and limits: No financial terms, targets, number of programmes or timeline were given, and no design from the deal has been reported since.
What happened: The authors of the 2023 A-Lab paper, including Ekin Dogus Cubuk and Amil Merchant of Google DeepMind, concede that its novelty claims "were subject to misinterpretation": the materials were "new to the prediction platform, not necessarily new to science". The title drops "novel". A manual re-analysis of the diffraction data, peer-reviewed after publication, confirms 36 of 40 reported successes; four are now counted as inconclusive, and Zn2Cr3FeO8, which was in the training data, is removed.
Why it matters: A-Lab and GNoME were published together as the case that AI could predict new materials and a robot could make them. The correction narrows the A-Lab half to what the critics said in 2024: the lab made compounds its own prediction platform had not listed, most of them already known to chemistry.
Evidence and limits: Being made in the lab still stands for 36 compounds. The correction does not say which of them are new to science, and powder X-ray diffraction alone cannot rule out the disordered forms the critics pointed to.
What happened: Ge Lei, Ferran Brosa Planella, Sterling G. Baird and Samuel J. Cooper had GPT-5 propose charging policies as code from a plain-language list of limits, scored in PyBaMM battery simulations. In one route (P2P) the model writes the whole current profile; in the other (P2O) it designs a small neural network whose parameters a Bayesian optimizer tunes. After 100 accelerated-ageing cycles of 30-minute charges to 90%, the best policies ended at 92.17% (P2O) and 92.10% (P2P) state of health, against 87.93% for a three-step constant-current protocol.
Why it matters: Testing a charging protocol on real cells is slow and costly, so optimizers are usually confined to a few fixed shapes such as three current steps. Here the language model widens what can be searched while safety limits are stated in words: the best P2P policy charges hard while voltage is low and backs off near 4.2 V, and P2P got its result within a matched budget of 210 evaluations. Giving the baseline 35 steps did not help within the same budget: it became hard to tune and produced jagged results.
Evidence and limits: Everything is simulated: one PyBaMM cell model with ageing sped up (the SEI reaction rate scaled fivefold) and an added voltage penalty for side reactions the model lacks. No policy was run on a real cell; the authors say that would need calibration, safety screening and cell tests. P2O's best result used evaluations beyond the matched budget.
What happened: Bessemer Venture Partners led the oversubscribed round, with TLV Partners, Vintage Investment Partners, Saras Capital and unnamed executives from Meta, OpenAI and Wiz; the company says it takes its total raised to $30M, after a $5.5M seed led by TLV Partners in November 2024. Dov Gertz, chief executive, Oded Kalev and Iddo Weiner founded it in 2024; it is based in Boston and Tel Aviv. It trains models on DNA, RNA and protein sequences and sells them as finished systems that plug into a customer's workflow.
Why it matters: Converge sells AI to drug developers rather than developing its own drugs, and its pitch is the distance between a model that scores well on a benchmark and a system a biologist can use. The release says it ran over 40 programmes for more than a dozen customers in a year, finding antibodies that bind at single-digit nanomolar strength and raising protein yields 4- to 7-fold. Those are the company's own figures, from its release, not from a paper. No valuation was disclosed.
What happened: EDEN, from Basecamp Research, is a family of DNA models including one of 28 billion parameters, trained on 9.7 trillion nucleotide tokens of the company's own BaseData. Prompted with 30 bases of a target site, it designed large serine recombinases: the paper reports active ones at all ten disease-linked genes and four safe-harbour sites it tested, a 63.2% hit rate on prompts outside the training data, and half the designs active in human cells, inserting a CAR into primary T cells.
Why it matters: Putting a whole gene into a genome usually means a virus or a cut in the DNA, which caps the payload and brings safety risk. A recombinase swaps DNA in at a site it recognises instead, but each site has needed its own enzyme, and finding one was a project in itself. Here 30 bases of a target were enough to get working enzymes at every site the paper reports testing. If that holds up outside the company, the enzyme is designed for the site rather than searched for in nature.
Evidence and limits: A preprint, not peer reviewed. The experiments are the authors' own, Basecamp with its co-authors (NVIDIA, Microsoft, and the University of Pennsylvania on the peptides), with no outside evaluator. "Active in human cells" is cell culture, not an animal or clinical result, and BaseData is proprietary, so others cannot rerun the benchmark comparisons. GEN's report of the same work carries figures the abstract does not (insertion at over 10,000 locations, over 90% tumour-cell clearance); the full text was not read.
What happened: In Science, Yinjun Jia, Yanyan Lan and colleagues at Tsinghua's Institute for AI Industry Research and other labs train a contrastive model that puts protein pockets and molecules in one space, so screening becomes a lookup, up to 10 million times faster than docking. In the lab its picks hit 15% of the time on the norepinephrine transporter, with two inhibitors solved in complex with it, and 17.5% on TRIP12, a target with no known small-molecule binders, using only AlphaFold2 structures.
Why it matters: Docking is too slow to run against every human protein; this makes it cheap, and the precomputed results are open as GenomeScreenDB. The hits were measured in the lab, but by the authors themselves, and the hit rates come without a figure for how potent the hits are. It is the method behind AIPher, the Tsinghua AIR spin-out that raised an angel round in September.
What happened: Noetik announced a five-year, non-exclusive licence giving GSK's AI and therapeutics teams direct access to its OCTO-VC models of lung and colorectal tumours, trained on Noetik's own spatial profiling of human tumour samples. The deal carries $50M in upfront capital and near-term milestones plus annual licence fees, and the two will generate new spatial datasets together. Kim Branson, GSK's head of AI and machine learning, said foundation models are only as good as their training data.
Why it matters: Noetik presents it as a move from paying for AI services to licensing a model by subscription, and its chief business officer, Shafique Virani, called it among the first and largest such deals; that ranking is the company's own. It shows what GSK will pay for tumour data built into a model, not that the model finds drugs: the release reports no result from using it.
What happened: Bayer's antibody scientists get full access to Cradle's platform for finding and optimising leads, after a proof of concept in which Bayer chose Cradle over several vendors. The two will also run a joint machine-learning research project. Anastasia Hager, Bayer's global head of drug discovery sciences, and Stef van Grieken, Cradle's co-founder and chief executive, spoke for the companies. No financial terms were given.
Why it matters: Bayer is buying software its own scientists run, not handing molecules to Cradle for milestone payments, the asset-based deal van Grieken sets this against. Choosing it after a proof of concept says the platform passed Bayer's own trial; the release gives no result from that trial.
What happened: The company mined human developmental cell atlases for candidate molecules (24 for cortex, 17 for heart), grew 3D tissue from stem cells, imaged it over time and let Bayesian optimisation pick each next round. Over three rounds, the share of constructs beating same-plate controls on both alignment measures rose from 28.5% to 45.6% in cortical tissue and from 38% to 56.6% in cardiac, and by the third round 94% and 68% fewer conditions were being tested, with cost per experiment down by about 70%.
Why it matters: Growing human tissue has meant years of hand-tuned recipes for each tissue type. Here the same loop moved from neurons to heart cells with little retuning, over a smaller search that took five weeks. Tissue whose structure can be set on demand is what Polyphron wants as a test bed for AI drug predictions, and this is the first data it published on making it.
Evidence and limits: A company blog post, not peer reviewed, reporting hit rates and medians without statistical tests. It optimises one structural feature, cell alignment, and shows no tissue function, such as contraction or signalling, and no drug response.
What happened: The problem, from a 1975 paper on binomial coefficients by Erdős and three co-authors, had been restated by the erdosproblems.com community after trivial solutions turned up. After feedback on a first attempt, ChatGPT proved the intended version. The Lean tool Aristotle repaired its minor errors and produced a Lean proof, and participants used ChatGPT to turn that into a write-up whose writing Terence Tao judged within reach of a research paper's standard.
Why it matters: Tao called it a milestone: earlier AI solutions to Erdős problems had mostly turned out to be in the literature, and no prior solution to this one was found. He found a second change more interesting: with Lean checking each version, a proof can be rewritten many times, at whatever level of detail a reader needs.
Evidence and limits: Within days the method turned out to be very close to a 2014 paper by Carl Pomerance, who then showed in a short note that his method also solves the problem. Tao adds that the problem lacked prior literature partly because its original wording was flawed.