Anthropic Optimizes Over 30 Biomolecular Models with Claude, Cutting Protein Design GPU Time by 1/100th

Anthropic announced that it optimized over 30 open-source biomolecular models for scientists using Claude in less than four weeks. This achieved an average speedup of approximately 4x while minimizing precision loss, and the development of a low-memory mode enabled the prediction of large biomolecular systems exceeding 10,000 tokens on a single NVIDIA GPU node.

Model Acceleration and Large-Scale System Processing

To address the triangle attention bottleneck in modern models like AlphaFold3, the FlashPairformer kernel was developed. This kernel accelerates triangle attention by 2.7–2.9x and triangle multiplication by 1.7–3.2x compared to previous methods. Additionally, the 'Big' mode now allows for the accurate prediction of complex structures exceeding 10,000 tokens, such as the human mitochondrial complex I, on a single B300 node.

Innovation in Protein Design Cost and Efficiency

Previously, de novo protein binder design cost $10,000 per target. Through this optimization, GPU time was reduced by approximately 100x, lowering the combined GPU and token cost to around $150. Applying three models, including Claude Mythos 5.1, to 16 targets achieved wet-lab binding affinity prediction metrics (ipSAE) similar to previous campaigns, significantly improving resource efficiency.

Ecosystem Support and Competition

Anthropic is hosting a protein design competition in collaboration with Adaptyv Bio. The competition involves experimentally validating over 5,000 community-submitted designs across five challenges, offering up to $1M in Claude credits and $250,000 in Modal compute credits. The optimized code has been released as open source on GitHub.