How Claude is uplifting biomolecular modeling

An unreviewed summary of a Science blog post reporting that Claude, supervised by two scientists without kernel-engineering experience, made more than 30 open-source biomolecular models run several times faster and in less memory, and that the optimized code was open-sourced alongside a protein design competition.

Anthropic (Science team) · 2026-09-17 · original

Not reviewed. This note reports what the post itself says. Unlike the other notes in this collection, its claims have not been checked against the underlying papers, reports or data. The work concerns biological design that could be misused, and in-depth analysis of it triggers the safety safeguards of the AI system used to prepare these notes, so it was left at the level of the post's own summary.

What the post reports

The subject is engineering rather than biology: making the open-source deep-learning models that biologists use for structure prediction, protein design, protein language modeling and genomics run faster and in less memory, so that work previously requiring a large compute budget fits on hardware ordinary labs can reach. The post says an internal general-purpose research model, working within Claude Science, optimized more than 30 such models in just under four weeks, speeding them up roughly 4x on average with minimal loss of precision, and roughly 1.6x to 2x with outputs identical to the originals. A set of custom kernels the post calls FlashPairformer targets the two operations that dominate runtime and memory in modern structure predictors; the post says these operations are cubic in system size, and reports the kernels beating the field standard by about 2.7–2.9x on one and 1.7–3.2x on the other, depending on configuration.

The post also describes a low-memory "Big" mode that, it says, allows accurate prediction of systems larger than 10,000 tokens (tokens here being amino acids, nucleotides and atoms) on a single GPU node, and inference on systems above 70,000 tokens. It names several large cellular machines predicted this way as closely matching their experimentally determined structures, and compares this with a published prediction of 7,663 tokens. Pushed further, to systems of 31,000 to 70,000 tokens, the post says the predictions are not correct and the structures collapse, which it reads as the models failing to generalize far beyond their training context.

Finally, the post reports rerunning an earlier binder-design exercise against 16 targets under much tighter resources (a single model on one GPU for 24 hours, with no sub-agents and no human steering), reaching, by an in silico score, roughly the same values as the earlier campaigns at about two orders of magnitude fewer GPU hours, for a combined spend the post puts at around $150. The optimized code is open-sourced, and the post announces a protein design competition co-sponsored with Adaptyv Bio, with up to $1 million in Claude credits, $250,000 in compute credits from Modal, DNA from Twist Bioscience and wet-lab validation of over 5,000 community-submitted designs.

Who did the work and who checked it

Anthropic's Science team publishes the post; the supervising scientists are described as two members of Anthropic technical staff experienced in biomolecular modeling but with no prior experience in inference optimization or kernel engineering. The detail sits in a technical report the post links, with code on GitHub; the post claims no peer review. The competition partners are named as Adaptyv Bio, Modal and Twist Bioscience, and the post also points to Anthropic's Life Sciences Verification Program, all commercial relationships it discloses in the post itself.

Caveats the post itself gives

The post says the largest capability runs are proof-of-concept and are not predicted correctly, and that the fast modes trade a small amount of precision for speed, with the identical-output mode offered as the conservative alternative; it reports verifying for each model that acceleration did not change downstream performance. The binder-design comparison is explicitly in silico only. A figure note adds that one tool's concurrently released fast kernels are not yet benchmarked in the comparison.

How it connects

References