Accelerating Open-Source Biomolecular Models with Claude

Inference optimizations for 36 implementations of more than 30 open biomolecular models, produced in just under four weeks by an internal general-purpose Anthropic research model working in Claude Science. Two scientists experienced in biomolecular modeling, but not in inference optimization or kernel engineering, supervised it. The repository holds one drop-in kit per upstream tool, covering structure prediction and cofolding (Boltz-2, Chai-1, OpenFold3, Protenix, RoseTTAFold3, ESMFold2, ColabFold, and AlphaFold 3 inference code run on OpenFold3 weights), binder and sequence design (RFdiffusion and RFdiffusion3, BoltzGen, BindCraft, PXDesign, ProteinMPNN), and protein and genomic language models (ESM C, ProGen2, Evo 2, Enformer, Borzoi). Each kit wraps a pinned upstream release with its original weights and switches on under a named mode: exact (outputs identical to the unmodified tool), fast (small numeric differences within the tool's own seed-to-seed variation) or big (lowest peak GPU memory, optionally split across the GPUs of one node). A shared core includes FlashPairformer, new GPU kernels for the triangle attention and triangle multiplication at the heart of Pairformer-style structure models.

On NVIDIA H100s, exact mode ran the forward pass of 14 structure prediction models 1.6× faster on average, and fast mode ran 13 of them 4.1× faster. The share of acceptably predicted interfaces (DockQ ≥ 0.23), pooled over 13 model configurations and 1,925 model–target pairs, moved by less than one percentage point in every mode. FlashPairformer's triangle attention is on average 2.7× faster than the field-standard kernels at the pair width most of these models use. Big mode predicted complexes of more than 10,000 tokens accurately on a single 8-GPU node, including human mitochondrial complex I and the E. coli 70S ribosome, and ran inference on assemblies of up to 70,320 residues on eight B300s, though those largest predictions were not accurate. In a demonstration, a single Claude model with one H200, 24 hours, a prompt of about 1,100 words and no sub-agents designed protein binders whose in silico ipSAE scores matched earlier Claude Mythos 5.1 campaigns that used about 100 times the GPU hours; Anthropic puts the combined GPU and token spend at about $150.

The kits are Apache 2.0 for Anthropic's code, with each upstream tool under its own license, and are a reference release that Anthropic does not plan to update and that accepts no contributions (316 stars at filing). The 140-page technical report lists "Claude Science" as co-first author with Richard Shuai, has Amir Shanehsazzadeh as corresponding author, and includes co-authors from Biohub and Columbia University. Alongside it, Anthropic and Adaptyv Bio launched a protein-design competition on five problems, with up to $1 million in Claude credits, $250,000 in Modal compute credits and wet-lab validation for more than 5,000 designs.