Cronin's group, whose founder leads Chemify, makes kinase inhibitors on a Chemputer and finds one that beats regorafenib in colorectal organoids
What happened: A robot synthesised a 4-anilinoquinazoline library, screened in a KRAS-mutant colorectal cancer fly model. Hit AP2-83 improved fly survival; the optimised AP4-43 beat regorafenib in a mammalian organoid growth assay.
Why it matters: A robot-made molecule got through two rounds of design and testing. Evidence stops at flies and organoids, one lab's own work, and read from the abstract only.
Imaging liver cells picks up compound activity at lower doses than standard toxicity assays, but does not predict cell-free assays
What happened: Axiom Bio, the Broad Institute's Imaging Platform and co-authors at the EPA and Bayer ran the Cell Painting imaging assay and two cytotoxicity assays on primary human liver cells against 1,085 compounds at eight concentrations. Among the 131 compounds active in every readout, the images flagged activity at 2.5 times lower concentrations than the metabolic assay, 7.9 than counting nuclei and 15.8 than the membrane-damage one. Predicting 412 curated ToxCast endpoints, they worked on cytotoxicity and cell-based assays and no better than shuffled labels on cell-free ones.
Why it matters: Axiom Bio sells liver-toxicity prediction from exactly this kind of imaging, and this is the public paper behind it: the images carry what happens inside a cell and nothing about a protein in a tube, so what the method can replace is bounded by biology, not by model size. A pretrained vision transformer and a convolutional network scored about the same as hand-engineered image features, so the limit here is the assay, not the architecture. The cells are one pooled five-donor hepatocyte lot, so donor variation is untested, and the ToxCast comparison uses other labs' assay designs.
Sakana's AI Scientist, in Nature: one of three fully AI-written papers passed first-round review at an ICLR workshop, after people picked which outputs to submit
What happened: Chris Lu, David Ha, Jeff Clune and colleagues at Sakana AI, Oxford and the University of British Columbia describe a system that proposes machine-learning ideas, writes and runs the code, writes the paper and reviews it. Three manuscripts went, with the organisers' consent, to the ICLR 2025 ICBINB workshop; one scored 6, 7 and 6, above the average acceptance bar, and was withdrawn as planned. An automated reviewer they built agrees with past ICLR decisions about as well as human reviewers agree with each other.
Why it matters: It is a fully AI-written paper getting through blind human peer review, now on record in Nature. The bar was low: a workshop that accepts 70%, one paper in three, and the team filtered the most promising outputs at each stage. The authors list naive ideas, wrong implementations and hallucinated content as common failures. All of it is machine-learning research run in software.
A Fukuoka lab makes MnFeCo4Si2, a GNoME prediction, in the predicted structure: a soft magnet with a Curie temperature of 1039 K
What happened: Shuhei Naganuma and Jiro Kitagawa (Fukuoka Institute of Technology) picked MnFeCo4Si2 from the GNoME entries in the Materials Project, arc-melted about 2.5 g and annealed it at 800 C for four days. Lab X-ray diffraction fitted the predicted rhombohedral structure, with lattice parameters slightly larger than predicted; electron microscopy found no impurity phase. Magnetisation measurements show a soft ferromagnet with a Curie temperature of 1039 K and a saturation moment close to their own calculation.
Why it matters: It is the first case on this record of an outside lab making a GNoME prediction because GNoME listed it, and getting the predicted crystal and a working magnet. The GNoME paper's 736 "independently realized" crystals were made by others before anyone looked at its list.
Evidence and limits: One compound, one lab, a preprint. Mn, Fe and Co scatter X-rays almost alike, so a lab XRD fit cannot show whether they sit on separate sites as predicted or share them.
Xaira's X-Cell, trained on 25.6 million perturbed cells, beats earlier virtual-cell models up to five-fold on its own held-out genes, but its test loss barely improves with scale
What happened: Xaira's bioRxiv preprint pairs X-Atlas/Pisces, a CRISPRi Perturb-seq set of 25.6 million cells across 16 contexts, with X-Cell, a diffusion model that predicts how knocking down a gene changes expression. On 200 held-out genes in two cell lines it scores Pearson deltas of 0.51 and 0.48 against 0.10 for STATE, the next best; on Tahoe-100M drug data the lead is 0.31 to 0.22. Scaled to 4.9B parameters, training loss falls as a power law (exponent 0.32), test loss far more slowly (0.03). Ci Chu and Bo Wang are senior authors.
Why it matters: A virtual cell is a bet that a model can stand in for some perturbation experiments, and this tests it on genes and cell types it never saw, not only new cells. The five-fold lead is on Xaira's own held-out set, the baselines were not retrained at its scale, and the flat test loss says more parameters alone will not buy accuracy: the authors say data is the limit. No prediction was tested in a new experiment. The 55M model is released under a non-commercial licence; the 4.9B one is not.
RxnCaption and RxnID parse reaction diagrams by locating molecules first, then letting a VLM read the relationships
What happened: A detector numbers each molecule so the VLM only relates indices; RxnID uses the paper's labels, like 1a. Zero-shot, a frontier VLM leads the lenient metric; trained models lead the strict one except on scans.
Why it matters: Splitting localisation from reasoning beats asking one large model to do both; but per-diagram metrics, even the strict one, do not measure correction-free records per paper.
BioMap, backed by Baidu founder Robin Li, has reportedly filed confidentially for a Hong Kong IPO
What happened: Bloomberg, citing people familiar with the matter, says BioMap could raise hundreds of millions of dollars, with CICC, Morgan Stanley and UBS. The company has not confirmed it.
Why it matters: If it lists, a life-science foundation-model company faces public investors; until a prospectus is public, even the filing is only reported.
One model from Anew Labs and Tsinghua designs small molecules, peptides and nanobodies against a KRAS pocket, and peptides and small molecules against PCSK9
What happened: AnewOmni, from Xiangzhe Kong, Kai Liu, Jianzhu Ma and colleagues at Anew Labs and Tsinghua University, trains on more than 5 million complexes and assembles designs atom by atom. Against the KRAS G12D switch II pocket, 7 of 30 linear peptides, 7 of 20 cyclic ones and 2 of 3 small molecules inhibited below 100 micromolar, and 4 of 7 nanobodies bound. For PCSK9, with no binding site given, 4 of 7 peptides and 3 allosteric compounds bound below 10 micromolar, one co-crystallised 0.92 angstrom from its design.
Why it matters: Design tools are usually built one per modality: a model for small molecules, another for peptides, another for antibodies. Here one model covers all three on the same pocket, and the designs went to assay a few dozen at a time rather than through a modality-specific screen. The co-crystal is the firmest part: the pose the model drew for a PCSK9 compound sits 0.92 angstrom from the measured structure, so the model put the molecule in roughly the place it actually binds.
Evidence and limits: A preprint, not peer reviewed, and every measurement is the team's own. The potencies are micromolar, a starting point rather than a drug, and the hit rates rest on small batches: 76 molecules in all, and the abstract's 75% is three nanobodies out of four. The paper says nine allosteric designs were synthesised, then that three of ten bound. The authors say the filtering is computationally heavy and the model cannot be told what affinity to aim for.
BigHat made and measured 289 antibody designs to compare how protein language models are sampled, and the sampling method changed the hit rate
What happened: Calvin McCarter, Nick Bhattacharya, Sebastian Ober and Hunter Elliott score a whole sequence at once rather than resampling masked positions, which lets a beam search steer on scores that need a complete sequence, such as humanness or charge. In a live antibody campaign they made and measured 289 designs across 14 model and sampling combinations. Their beam search beat Gibbs sampling on all three models where both were tried, and a classifier trained on the campaign's own 729 earlier samples did better still.
Why it matters: Attention in this field goes to which protein language model to use, and the paper's point is that the sampler on top of it is a consequential choice alongside the model, tested on antibodies made and measured rather than scored. For a team running a design loop the practical reading is to spend the first lab budget on labels: ranking with a classifier trained on the campaign's own early results raised the success rate considerably, and steering generation with it raised it further. The company published its internal head-to-head, not only the method it kept.
Evidence and limits: The authors state the limit: one campaign, one starting antibody, so the ranking may not carry to another antibody, let alone another protein. Success means a design could be made and gave a detectable binding signal, not that it bound better than the parent, and the measurements are BigHat's own. The 100% rests on at least 21 designs in that arm, and guidance made the antibodies less human, a risk the paper flags and does not solve.
Unreasonable Labs launches from stealth with a $13.5 million round
What happened: Playground Global led, with AIX Ventures, E14 Fund and MS&AD Ventures. Yuan Cao, formerly of Google DeepMind, is CEO and MIT professor Markus Buehler is CTO. Pilots in energy, materials and pharma name no partner.
Why it matters: A DeepMind researcher and an MIT professor raising seed-sized money for a discovery engine built on concept-linking rather than lab automation. The release shows no result; what the money tests is unproven.
Yann LeCun's AMI Labs raises a $1.03B seed round to build world models, reported at a $3.5B valuation
What happened: The Paris lab, led by CEO Alexandre LeBrun with Yann LeCun as executive chair, raised $1.03B in a seed round co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions, Tech.eu reports, quoting LeCun. Its site lists the founding team as LeBrun, Laurent Solly, Mike Rabbat, Min Lin, Pascale Fung, Saining Xie and LeCun, with teams in Paris, New York, Montreal and Singapore.
Why it matters: It is a large bet against language models as the route to understanding the physical world: AMI builds models that predict in a learned representation of sensor data rather than generating it, and names industrial process control, robotics and healthcare as uses. For science it matters because lab automation and physical experiments need models of how the world behaves.
Evidence and limits: Only Tech.eu's report of the round was read; the company's own release was not found, and the valuation's basis is not stated. The company has no published result yet.
Cradle's CRADLE-1 preprint claims protein lead optimisation 4-7 times faster than rational design, counted in wet-lab rounds
What happened: Protein language models are tuned on sequences related to the target protein and then on data generated in the lab-in-the-loop. The abstract reports in-vitro validation across nanobodies, scFvs, IgGs, peptides, enzymes, CRISPR systems and vaccines, with single and polyspecific binding down to picomolar.
Why it matters: Optimising a protein someone already has is what most lab work actually is, and the claim is counted in rounds of experiments rather than scores. What the 4-7 times is measured against is not in the abstract, and the campaigns behind it are the company's own work with its customers.
Imperial College's closed-loop optimiser makes 46 amine resins for capturing CO2 from air and finds a straight-line trade-off between capacity and speed
What happened: Tristan Spreng, Ronny Pini, Camille Petit and colleagues at Imperial College London and the University of Vienna let a Bayesian optimiser pick recipes for amine-grafted porous polymers, made and measured on an automated platform: six rounds of eight recipes, 144 samples, 46 stable materials. At 400 ppm CO2, uptake and adsorption speed traded off along a straight line; some adsorbed about ten times faster than benchmark sorbents at half the uptake. The highest-uptake material came within three rounds; a full grid of the space would need about 16.
Why it matters: A direct air capture plant pays for how much CO2 a bed of sorbent moves per day, and the commercial benchmark resin, Lewatit VP OC 1065, was not designed for it and runs cycles of about three hours. Optimising capacity alone, the usual target, would miss the best point: in the authors' process model a material from the middle of that line gave 194 kg of CO2 per cubic metre per day against 95 for Lewatit at similar energy, and grafting a larger amine (PEI) doubled one material's uptake at unchanged speed.
Evidence and limits: Uptake and speed were measured in dry gas at 303 K; humid air, repeated cycling and oxidative or thermal ageing, which set a sorbent's life in a plant, were not tested. The productivity figures come from the authors' process model, and the PEI materials were modelled assuming only uptake and speed changed. A preprint, not peer reviewed.
Evo 2 brings million-base genomic context into an open foundation model
What happened: Published in Nature, Evo 2 is trained on roughly nine trillion DNA base pairs across all domains of life with a one-million-token context window.
Why it matters: Genome modeling moves beyond local sequence effects toward chromosome-scale context, while reusable models and resources turn the paper into scientific infrastructure.
RealPXRD-Solver, led by DP Technology, reads crystal structures off measured powder diffraction patterns, and solves 39 database entries whose structures had never been reported
What happened: The model was trained on 6,250,238 theoretical crystal structures, deliberately roughened to look like real measurements, and reads a pattern as a fingerprint of peak spacings and intensities, so simulated and measured data share one representation. Given the unit cell, it puts the right structure among its 20 best candidates for 91.9% of the measured CNRS set and 92.9% of RRUFF, and ranks it first for 77.9% and 78.8%. Without the unit cell, the same two CNRS figures are 79.4% and 52.4%.
Why it matters: Powder diffraction is what a chemist gets when a new material will not grow a single crystal, and earlier AI solvers worked on simulated patterns and fell over on measured ones, where peaks overlap, crystals line up one way and impurity phases show up. Coupled to an automated refinement pipeline, the model determined the structures of 39 materials in the Powder Diffraction File that had carried a unit cell but no atomic positions since they were first reported, work that normally takes an expert weeks an entry.
Evidence and limits: Every number is the authors' own, on public reference sets they chose; the 39 new structures were solved and judged by their own pipeline on standard refinement agreement factors, with no independent redetermination, and nothing was synthesised. The subsets held back from training are small, 23 and 14 entries, and their first-choice rates fall to 47.8% and 57.1%. The diffraction database's licence barred testing on all of it. A preprint, not peer reviewed.
