How Claude is uplifting biomolecular modeling

Anthropic researchers shared results showing how Claude optimized more than 30 open-source deep learning models trained for biological tasks, including structure prediction, protein design, genomics, and protein language models. Working within Claude Science under the supervision of two staff members without prior inference optimization experience, Claude completed the work in under four weeks. On average, Claude sped up the tasks roughly 4x with minimal loss in precision, and nearly 2x with identical outputs, while also improving memory utilization. To achieve these optimizations for structure prediction models, Claude helped develop FlashPairformer, a set of custom kernels for triangle attention and multiplication that outperform field standards. Additionally, Claude created a low-memory mode that enables accurate modeling of systems larger than 10,000 tokens and inference on systems larger than 70,000 tokens using a single NVIDIA GPU node, successfully folding large molecular machines like human mitochondrial complex I and bacterial ribosomes. Anthropic is open-sourcing all of the optimized code and technical reports. To support the broader community, Anthropic is also co-sponsoring a protein design competition with Adaptyv Bio, backed by up to $1 million in Claude credits, $250,000 in Modal compute credits, and wet lab validation for over 5,000 designs, with contributions from Modal and Twist Bioscience.