AI FOR SCIENCE · TIMELINE
The 11 entries the timeline holds for February 2026, newest first. Each was read at source. The full timeline, with filters →
What happened: The Flagship company sold 25M shares at $16 and began trading as GENB on 27 February. Goldman Sachs and Morgan Stanley led the books.
Why it matters: Its $370M Series B in 2021 was a private bet on AI-designed protein drugs; the IPO hands that bet to public investors while the drugs are still in trials.
What happened: Google reports gold-medal-level results on the written sections of the 2025 International Physics and Chemistry Olympiads, alongside stronger general reasoning scores.
Why it matters: The breadth matters, but so does the boundary: Olympiad success is evidence of reasoning, not yet evidence of autonomous research.
What happened: Hermes is a deliberately small model trained on Leash's DNA-encoded-library screens (the July 2025 version). Leash tested it on four sets (about 13.8 million examples): new proteins (164 targets), a new chemical library (59), a public ChEMBL-derived set (403) and MF-PCBA (26). It beat the XGBoost baseline on three of the four, was slightly ahead of Boltz-2 on Leash chemistry and behind it on MF-PCBA. Both did poorly on the new library, except BRD4, where Boltz-2 scored near-perfectly, which Leash puts down to PDB structures of molecules from that library.
Why it matters: Binding models are usually tested on molecules made by the same chemists they trained on, so a good score may be recognising their style. Leash trained on chemistry nobody else made and tested on much larger sets than usual, two of them public downloads, which makes a good score less likely to be that.
Evidence and limits: Leash ran every comparison itself, and the paper says Boltz-2 may have seen part of the MF-PCBA data in training. Boltz-2 was run on samples of about 50,000 per set. XGBoost beat Hermes on the new-protein set, where Leash found a leak, and both models did poorly on the new chemical library. The post does not claim Hermes is the better model.
What happened: Huan Souza and Pankaj Mehta reran TranscriptFormer, Geneformer, UCE, scVI and ESM2-CE on the same splits as their own normalized, linear pipelines. Labelling testis cells across eight mammals, their method scTOP averaged a macro F1 of 0.679 against 0.639 for Geneformer and 0.515 for TranscriptFormer. On human cell types in 26 tissues their linear pipeline scored 0.887 against TranscriptFormer's 0.911, and another came close to TranscriptFormer at telling infected from uninfected cells.
Why it matters: TranscriptFormer's Science paper offered its cross-species results as evidence that pretraining across evolution improves biological generalization, and such results are why cell foundation models are seen as a base for virtual-cell work. If a parameter-free projection beats them at cross-species transfer and short linear pipelines come close elsewhere, these benchmarks cannot show what pretraining adds, and anyone funding or building such models needs a linear baseline in every comparison and harder tasks than cell labelling.
Evidence and limits: The pipelines are the authors' own, and those for human cell types and infection were tuned per task; and they still trail TranscriptFormer on human cell types. The cross-species test covers one testis dataset restricted to orthologous genes. The authors say their result is limited to current atlases and benchmarks, and that tasks such as developmental time courses or multimodal data may need larger models. A preprint.
What happened: On the Runs N' Poses cases least like its training data, IsoDDE posed 30 of 60 ligands correctly, 17 of them where AlphaFold 3 failed. On 334 antibody-antigen pairs unlike its training set, 39% of its predictions were high-accuracy, 2.3 times AlphaFold 3 and 19.8 times Boltz-2. It also predicts binding strength, beating free-energy physics methods on public sets, and finds pockets 1.5 times better than P2Rank by one measure. The report shows results, not the model, and nothing is released.
Why it matters: AlphaFold-style models mostly fail on pockets and antibodies unlike what they were trained on, and that is where new drugs are. This report says Isomorphic has closed much of that gap, and the engine is part of what its $2.1B Series B pays for. But every number is computed and scored by Isomorphic itself, the method is withheld, and no outside team can check it. No molecule designed with it has public results.
What happened: The arXiv preprint drops all-atom diffusion and uses only protein C-beta atoms and ligand heavy atoms. It reports about 20% higher Pearson correlation than Boltz-2 on CASP16 and on 18 of Terray's own assays.
Why it matters: Screening millions of molecules needs a cheap model. These are software scores, and the 18-assay set is proprietary, so nobody outside can rerun it.
What happened: Fan Fan, Bo Huang and colleagues at Beijing StoneWise Technology, with Peking University co-authors, publish HEAD (High-Energy Atom Detector) in Nature Communications: a two-stage, open-source check for 3D molecules from generative models. An AI-derived force field flags atoms made implausibly high in energy by their surroundings; a network trained to density-functional accuracy flags rotatable bonds whose torsional energy passes 2 kcal/mol. On the LigBoundConf and CSD sets of real, measured conformations, HEAD keeps above 95% of valid structures, like the geometry-based PoseBusters, but catches high-energy anomalies PoseBusters misses.
Why it matters: As generative models pour out candidate 3D molecules, telling a physically sensible one from a distorted one is the bottleneck, and the usual checks are either fast-but-crude geometry rules or slow, inaccurate mechanics energies. HEAD aims for quantum-chemistry-level judgement at a speed that can screen large batches, and it separates two faults a single score blurs together: an atom sitting in an impossible spot, and a bond twisted into a high-energy pose. Being open source, it lets other groups hold generated molecules to the same bar.
Evidence and limits: The checker is validated by how well it recovers known-good conformations and flags perturbed ones, not by any downstream gain in a real design campaign. 'Valid' and 'rational' are set by the authors' own energy thresholds (2 kcal/mol for torsions). The comparison is against PoseBusters alone, and the five generative models it was run over are named but not re-trained or improved here.
What happened: Srijit Seal, then at Cambridge and the Broad Institute, with Anne E. Carpenter, Andreas Bender, Ola Spjuth and four Axiom Bio scientists including co-founder Alex Beatson, reran four published benchmarks using one feature: how many cells survived. On a 209-assay benchmark, cell count alone scored a mean AUC of 0.68, a network on full Cell Painting profiles 0.67 and an image model 0.73. Of the 66 assays cell count predicted well, 64 were tumour-cell growth tests.
Why it matters: Cell images and gene-expression profiles are offered to drug discovery as extra information about what a compound does. These benchmarks could not show it, because a compound that kills cells looks active in many of their assays. The paper's fix is a cell-count baseline and dropping viability assays; on a cleaned set of 24 protein-target assays, Cell Painting models reached a balanced accuracy of 0.67 against 0.52 for cell count.
Evidence and limits: A reanalysis of public benchmarks; the only new data are dose-response images of six compounds, shown as an illustration. The cleaned benchmark is small: 24 assays across 269 compounds. Four authors work at Axiom Bio, which sells drug-toxicity prediction, and the paper declares that Carpenter and Shantanu Singh advise companies using Cell Painting.
What happened: The trillion-parameter mixture-of-experts scientific model activates 22 billion parameters per query and targets scientific reasoning.
Why it matters: It makes scientific foundation models an open ecosystem contest, allowing capability claims to be inspected beyond a single lab.
What happened: In six steps over six months GPT-5 ran over 580 plates and 36,000 reaction compositions in Ginkgo's cloud lab. For the model protein, cost per gram fell from $698 to $422 and titer rose 27%.
Why it matters: A real hardware loop with a cost number. But it ran in the companies' own lab, the benchmark is another group's preprint, and 6 of 12 proteins showed on a gel.
What happened: Mohammed Abouzaid, Martin Hairer, Nikhil Srivastava, Lauren Williams, Rachel Ward and six colleagues posted ten questions from their own research, each solved by its author in five pages or less and kept off the internet, with answers released on 13 February. Their own tests with Gemini 3.0 Deep Think and ChatGPT 5.2 Pro in early February, posted with the answers, found most of the best attempts flawed. Two were essentially correct, one improving on its author's method.
Why it matters: Most maths benchmarks use contest problems. These come from the last step of real research, with answers that were not online, so a model could not simply look them up. The authors' notes say how the attempts failed: a sketch on a homepage passed off as a proof, invented lemmas, an easier problem solved instead. A second batch of ten followed in June, with referee reports on the AI solutions.
Evidence and limits: Ten questions graded by their own authors are not a benchmark, as the authors say themselves. Unpublished did not mean nothing related was online: Hairer's question had a short sketch on his homepage, a 2024 paper was close to Question 9, and the key idea of the best answer to Question 10 was already in a 2016 arXiv paper.