In its 21 October announcement, Nvidia said it and the UW Medicine Institute for Protein Design had released RFD4-Proteina at GTC Berlin. The open BioNeMo model combines protein design with a way to search possible designs: researchers can give it molecular inputs, constraints and scoring objectives, then generate and refine candidates for laboratory testing.
Checking for signs of disease could benefit from proteins designed to detect them, one potential use Nvidia identifies. A candidate chosen through the model’s search would still need laboratory testing to establish how it performs under tested conditions.
In Nvidia’s account of tests set out in a research draft, RFD4-Proteina surpassed the baselines assessed on both the rate of successful computational designs and the rate of design generation, across four benchmark suites covering DNA binding, enzyme scaffolding, small-molecule binding and protein binding. RFD4-Proteina is available through Amazon Bio Discovery, Anthropic’s Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI and Tamarind Bio, Nvidia says.
