An AI company starts doing biology
-
Nº CXXXI
- Date
- 29 Oct 2026
- Issue
- 131
- Stories
- Six
- Editor
- ARC
Anthropic reports an enzyme find from its own lab, while today's preprints ask how anyone checks such claims.
Claude agents find an enzyme system of unknown function
Claude agents found an enzyme system whose function is still unknown, Anthropic reported on its research blog. Agents here means AI software that carries out multi-step work on its own, in this case built on Anthropic's Claude models. The finding is one of the early results from Anthropic's new life sciences research lab. Since nobody yet knows what the enzymes do, the agents have so far only spotted a new system, and its biology still has to be worked out. Readers will want to check the company's own write-up for how the agents flagged it and what experimental evidence supports the call.
Preprint fixes how automated cell labs pick experiments
Fix the measurement model first, argues a bioRxiv preprint on self-driving laboratories that experiment on living cells. A self-driving lab pairs automated experiments with an optimiser, the algorithm that reads each result and picks what to run next. The preprint proposes fixes to that optimiser, and its title says which comes first: correct how the system models its own measurements before tuning the rules it uses to choose experiments. Readouts from living cells are noisy, so the way an optimiser interprets each measurement affects every choice after it.
Anthropic adds 20 partners to its AI for Science program
Anthropic expanded its AI for Science program and highlighted 20 scientific foundations, institutions and research groups from around the world that it now partners with. The expansion comes alongside its in-house life sciences lab, so Anthropic is doing research itself while also adding outside partners. Because the list includes foundations and institutions as well as research groups, the program now reaches the organisations that fund and host science, not only the people doing it.
Preprint aims to make AI-run science checkable
Verifiable agent-run science on public infrastructure is the subject of a new bioRxiv preprint. Whether results produced by AI agents can be checked independently is one of the field's open questions.
One learned force field models biomolecular complexes
One neural network potential handles the dynamics of biomolecular complexes at coarse-grained resolution, a new arXiv preprint reports. These potentials are machine-learned stand-ins for the physics force fields that drive molecular dynamics simulations.
HGPRGDel ranks gene deletions for growth-coupled design
HGPRGDel, a new method described in an arXiv preprint, learns and ranks gene deletions for growth coupling. Its model takes into account gene-protein-reaction (GPR) rules, the logic linking genes to the metabolic reactions they enable.
Reply with your discoveries. A human reads them. Forward freely.
|