Anthropic Announces Breakthrough in AI-Powered Protein Design
Anthropic reports that its Claude models designed working protein binders, outperforming human experts in some cases. The company shared two experiments on its research blog, highlighting significant achievements in drug development.
Key Findings:
- Protein Binder Design: Claude designed protein binders against 15 targets, succeeding in 14. It achieved binding rates of 22% to 35%, far surpassing the typical 10% to 15% success rate in the field.
- External Validation: Two independent labs, Adaptyv Bio and Twist Bioscience, confirmed the designs, demonstrating the physical validity of Claude’s work, which typically takes weeks.
- Computational Power: Anthropic utilized Opus 4.8 and a preview of its Mythos model, providing them with substantial computational resources, including up to 12,500 Nvidia H100 hours.
- Success Stories: The Mythos preview excelled against RBX1, with a 40% success rate, surpassing human competition. Claude also designed binders for TNFα that worked across human, monkey, and mouse versions.
- Challenges: However, the model faced difficulties with maltose-binding protein and BBF-14, indicating room for improvement.
Experiment Details:
In the first experiment, Anthropic used Claude Science, its research workbench, to generate protein binders using two models. They utilized open-source design and folding models, screening candidates for binding.
The second experiment involved chemical analysis using Claude Opus 5, demonstrating the model’s capabilities in routine compound identification and purity checking.