Google's GenFocal turns coarse climate projections into local weather, and cuts the error in US heatwave risk against the method the National Climate Assessment uses
What happened: Published in Nature Machine Intelligence, the model learns from 1980-1999 without needing simulated and observed days paired up, then is scored on 2010-2019. Downscaling a 100-member climate ensemble over the contiguous US, it cut bias in the 99th-percentile summer heat index by over 35% against two statistical methods, and bias in 5-day heat streaks by 44% and 57% against BCSD and STAR-ESDM.
Why it matters: Cities and insurers plan for local extremes that global climate models are too coarse to show, and this is checked against the method a US national assessment recommends, not a forecast score. The check is against reanalysis for the past decade, in one region, by the team that built it.
AuthorsZhong Yi Wan, Fei Sha, Leonardo Zepeda-Núñez
Cradle's redesigned cetuximab fragment binds EGFR at 1.2 nM, the best of 400 designs Adaptyv Bio tested
What happened: In Adaptyv Bio's second binder-design competition, 130 designers sent in EGFR binders and Adaptyv made and measured 400 in its own lab; 53 bound. Cradle's entry, cetuximab's scFv with 10 framework mutations its protein language models chose, bound most tightly, with a KD of 1.2 nM. Cradle puts the unmodified scFv at 9.94 nM and the next team at 5.18 nM.
Why it matters: An outside lab measured it, so the number is not Cradle's own. It improves an approved antibody rather than making a new binder: the loops that bind EGFR were kept. One design won, and Cradle says Adaptyv's computed ranking had placed its entries between 322nd and 623rd.
The Chemistry Nobel recognizes protein design and AlphaFold
What happened: David Baker receives half the prize for computational protein design; Demis Hassabis and John Jumper share the other half for structure prediction.
Why it matters: AI-enabled prediction and design receive the strongest institutional recognition in science.
A patient-tissue drug screen plus machine learning finds an antidepressant that boosts glioblastoma chemotherapy in animals
What happened: Nature Medicine: 132 drugs tested on surgery tissue from 27 glioblastoma patients; machine learning on drug-target networks then screened over 1 million compounds in software. Vortioxetine boosted chemotherapy in vivo.
Why it matters: Per Prevision's team page, co-founder Snijder's lab built this screening method, which Prevision sells. The drug benefit shown is in animals, not patients.
Chai Discovery releases Chai-1, a structure-prediction model with open weights that companies may use
What happened: Chai-1 predicts the structures of proteins together with small molecules, DNA, RNA and chemical modifications. Weights and inference code are released under Apache 2.0, with a free web interface, commercial drug discovery included. Chai reports 77% on the PoseBusters ligand-docking benchmark against 76% for AlphaFold 3, and says it keeps most of its accuracy without multiple sequence alignments.
Why it matters: At release AlphaFold 3 was only a web server, so this gave companies a model of that class to run and build on. The scores are on public benchmarks Chai ran itself: they show the model computes well, not that its predictions hold up in a lab.
Leash Bio's BELKA competition: about 2,000 teams, 300 million training examples, and no one predicted binders in a new chemical library better than chance
What happened: Leash gave Kaggle teams about 300 million examples, aggregated from 4 billion DNA-encoded-library measurements of molecules binding three proteins, all from one triazine-core library. The test had three parts: molecules sharing building blocks with the training set, molecules that did not, and a different library, kinase0. The winner did well on shared blocks, worse on unshared ones and could not predict kinase0. Taking the 100 most confident kinase0 picks, the winner had two true binders for one protein and none for the other two, where chance gives about one.
Why it matters: It showed, before any Leash model, that a good score on one library can come from memorising building blocks, and that more data of the same kind did not carry over to a different library. Leash's later tests, a separate chemical library and protein mutants, are built on that result.
Evidence and limits: Leash designed the data, the splits and the scoring, and its own post lists reasons the gap with published hit rates could have, including that its data might be poor; it promised to release the raw replicates for others to check. Only three proteins and one library design were tested, and the practical test was a simulated catalogue pick, not a drug programme.
Google researchers report that self-copying programs arise on their own in a soup of random code, with no goal or fitness score
What happened: Blaise Agüera y Arcas and seven co-authors at Google's Paradigms of Intelligence team and the University of Chicago filled a soup with 131,072 random 64-byte programs in a self-modifying variant of the Brainfuck language (BFF), then kept running random pairs joined end to end. Self-copying programs took over in 40% of 1,000 runs within 16,000 rounds, and about as often with no random mutation. Forth and the Z80 and 8080 instruction sets did the same; SUBLEQ, whose shortest hand-written self-copier is 60 bytes, never did.
Why it matters: The best-known artificial-life experiments, Tierra and Avida, start from a replicator someone wrote by hand. Here replicators appear from programs rewriting each other with no fitness target, and the paper gives a rate, a way to spot the moment it happens (a compression-based complexity measure) and open code (cubff) that others can rerun. It turns the idea that the first replicators came from interaction rather than luck into an experiment with numbers.
Evidence and limits: Computed by the authors in a preprint; no journal version was found. Takeover is detected with the complexity measure and token tracing, not by checking every program. Random starts run for only 128 rounds were taken over in 3 of 1,000 runs, and a planted hand-written copier took over in 22%, so most single replicators die out. What happens after takeover is shown through examples only.
EvolutionaryScale's ESM3 protein model writes a green fluorescent protein only 58% identical to any known one, and in the lab it glows about as brightly as natural GFP
What happened: Thomas Hayes, Alexander Rives and colleagues trained ESM3 on protein sequence, structure and function together, at sizes up to 98 billion parameters. Prompted with a few key residues at GFP's light-making site, the 7-billion-parameter version produced designs the team made in E. coli. Of 88 in a first plate, several close to natural GFPs glowed as brightly; one far from them glowed, 50 times dimmer. A second plate grown from it gave esmGFP, 58% identical to the nearest known fluorescent protein and 36% to jellyfish GFP.
Why it matters: Engineered GFPs usually carry 5 to 15 mutations, and at most 40 to 50 with machine learning; esmGFP is 96 of 229 residues from its closest known relative, a gap the authors estimate evolution takes over 500 million years to open. So a working function was made far from known sequences and shown in the lab, not only predicted. The test is narrow: one protein family, the authors' own assays, two plates of designs, fluorescence measured in bacteria, no solved structure. A 1.4-billion-parameter model was released openly; Science published the paper in January 2025.
AlphaFold 3 predicts interactions across biomolecules
What happened: The model extends beyond proteins to complexes containing DNA, RNA, ligands and modified residues.
Why it matters: Structure prediction moves closer to the interaction problems central to drug discovery, while restricted access triggers an openness dispute.
Leash Bio, a Salt Lake City start-up of former Recursion scientists that collects its own protein-binding data, raises a $9.3M seed round led by Springtide Ventures
What happened: MetaPlanet, Top Harvest Capital, Mitsui Global Investment, MFV Partners and Recursion's Chris Gibson and Blake Borgeson joined. CEO and co-founder Ian Quigley's company said it had measured over 17 billion protein-chemical interactions and planned to screen 500+ protein targets by 2025. The money is for more data collection and computing, a general machine-learning model of medicinal chemistry, and moving internal programmes toward animal studies. Five of its six employees came from Recursion; Claire Smith of Springtide spoke for the lead.
Why it matters: The money went to a pitch built on data rather than a model: the company's own measurements of proteins binding chemicals, collected first. A six-person team raising $9.3M is a small bet on that approach; the BELKA competition launched with it, and the models and agent environments Leash later published rest on the same screens.
Evidence and limits: The 17 billion measurements and the 500-target plan are the company's own statements in its release. The date is the release's dateline, as reproduced by The AI Insider; the release itself (GlobeNewswire) was not opened, and a $7.9M figure attributed to Endpoints was seen only in a search snippet.
A Liverpool lab steers its search with a model trained on known crystals and finds Li7Si2S7I, a new lithium conductor at about 10 mS/cm
What happened: Guopeng Han, John B. Claridge, Matthew J. Rosseinsky and colleagues at the University of Liverpool narrowed 296 lithium four-element systems, let a variational autoencoder rank them (Li-Si-S-I came 9th), chose Li-Si-S-I for its two very different anion sizes, and aimed synthesis with structure prediction. The new phase Li7Si2S7I has a total conductivity of 1.01 x 10^-2 S/cm at 303 K and ran 240 hours of lithium plating and stripping.
Why it matters: This is the lab Matter Machine came out of, and the result shows its method reaching a material worth testing, not only a new structure. The model only ranked chemistries; the design idea, packing two anions like an alloy, was the chemists'. The same team made and measured it, the material decomposes above 2.5 V, and a phase change below room temperature cuts its conductivity.
