Why it matters: biology evals are starting to look more like real lab work instead of tidy exam questions. Anthropic says Claude tackled 99 bioinformatics problems and solved roughly 30% of the 23 that stumped expert panels.
Anthropic put hard numbers on Claude's biology capability claims instead of vague lab hype. In 99 real-data bioinformatics problems, the company says experts were stumped on 23 and recent Claude models solved roughly 30% of that hardest slice.
A Hacker News post highlighted OpenMed’s end-to-end protein AI pipeline, from ESMFold and ProteinMPNN to CodonRoBERTa, plus a 25-species model suite trained in 55 GPU-hours. The thread praised the engineering detail but also raised the usual questions about biological validation.
Google DeepMind said on X that it is expanding AlphaFold Database with millions of AI-predicted protein complex structures in collaboration with EMBL-EBI, NVIDIA, and Seoul National University. The release pushes AlphaFold beyond single-protein structure prediction toward a broader public resource for studying how proteins interact.