Anthropic Reveals Technology to Accelerate Over 30 Biomolecular Models by 4x on Average Using Claude and Predict Large-Scale Systems

Anthropic announced that it optimized over 30 open-source biomolecular models used by scientists in approximately 4 weeks using Claude. This resulted in an average speed improvement of approximately 4x and improved memory efficiency. In particular, the introduction of low-memory mode enabled accurate predictions for biomolecular systems with over 10,000 tokens on a single NVIDIA GPU node.

Key Technologies and Performance Improvements

During the optimization process, a custom kernel called FlashPairformer was developed to resolve bottlenecks in the latest models such as AlphaFold3 and OpenFold3. Compared to existing standards, triangle attention improved by an average of 2.7 to 2.9x, and triangle multiplication improved by 1.7 to 3.2x. Through individual model optimization, structure prediction models were accelerated by an average of 4x without precision loss, and speed improved by approximately 1.6x for the same output.

Large-Scale System Prediction and Limitations

The 'Big' mode developed by Claude enables accurate modeling of systems with over 10,000 tokens on a single NVIDIA B300 node. It successfully folded complex molecular machines such as human mitochondrial complex I and the TRiC chaperone complex, which closely match experimentally determined structures. However, a limitation was identified where accuracy decreased for systems sized 31,000 to 70,000 tokens, which are more than one order of magnitude larger than the training context.

Community Support

Anthropic is collaborating with Adaptyv Bio, Modal, and Twist Bioscience to hold a protein design competition, supporting a total of $1.25 million worth of credits ($1 million in Claude credits and $250,000 in Modal compute credits) and experimental validation for over 5,000 designs.