Anthropic Research has published findings detailing how Claude, an internal general-purpose research model, optimized open-source models used in biomolecular modeling. Over a period of just under four weeks, Claude processed more than 30 such models, enhancing their speed and memory efficiency.
Key Points
- Claude optimized over 30 open-source biomolecular models in under four weeks.
- The optimizations resulted in an average speedup of approximately 4x for these models.
- Claude also created a low-memory mode, allowing accurate prediction of biomolecular systems larger than 10,000 tokens on a single NVIDIA GPU node.
- All optimized code is being open-sourced by Anthropic.
- A protein design competition, co-sponsored with Adaptyv Bio, offers up to $1 million in Claude credits and wet lab validation for over 5,000 designs.
- Claude Mythos 5.1 previously accelerated seven open-source biology models, making them run up to 2.5 times faster.
- An earlier demonstration of Claude's protein binder design capabilities required up to $10,000 per target on Modal, equivalent to approximately 2,500 NVIDIA H100 GPU hours.
Context
According to Anthropic, the optimizations focused on inference to make biomolecular research more accessible. The work involved developing FlashPairformer, a set of custom kernels for triangle attention and multiplication operations, which are computationally intensive in modern structure prediction models like AlphaFold3, OpenFold3, and Boltz-2. FlashPairformer outperformed existing field standards, achieving an average speedup of 2.7-2.9x on triangle attention and 1.7-3.2x on triangle multiplication. Beyond these transferable kernels, Claude also applied specific optimizations to individual models, such as caching redundant computations and simplifying dead code branches. These efforts were supervised by two Anthropic technical staff members experienced in biomolecular modeling, who had no prior experience in inference optimization or kernel development.
Why It Matters
This work demonstrates how large language models can contribute to fundamental scientific research by automating complex code optimization tasks. For builders and researchers in biomolecular fields, the open-sourced code and improved efficiency could reduce computational barriers, accelerating drug discovery and development, and enabling the study of larger, more complex biomolecular systems.
What To Do
- Review the open-sourced optimized code for the biomolecular models to understand the specific changes implemented by Claude.
- Investigate the technical report for detailed information on the optimization methods and results.
- Consider participating in the protein design competition co-sponsored with Adaptyv Bio to leverage Claude credits and potential wet lab validation.
- Note the resource requirements for previous agentic protein design work (up to $10,000 per target or 2,500 NVIDIA H100 GPU hours) when planning similar projects.