NVIDIA on AlphaFold, MMseqs2-GPU, and Protein Science Acceleration
NVIDIA's Noah Kravitz interviews Chris Dallago (NVIDIA Research Lead) and Martin Steinegger (Seoul National University, AlphaFold Nobel co-author) about a wave of NVIDIA announcements at the intersection of AI and protein science. The centerpiece is the formal acceptance of MMseqs2-GPU in Nature Methods, a GPU-accelerated version of Steinegger's widely used homology search tool that inverts AlphaFold's compute profile: the homology retrieval step drops from ~80% of total inference time to ~20%, leaving the deep-learning stage as the new bottleneck that NVIDIA's other tools can then attack. NVIDIA also released La Proteina, an open-source generative model for protein design trained on AlphaFold-predicted structures and built on scaled transformer architectures, alongside targeted accelerations for attention operations and NIM inference microservices for AlphaFold, OpenFold, and Boltz, plus new Blackwell GPU compatibility. The guests frame NVIDIA's strategy as contributing accelerated infrastructure rather than competing as a drug-discovery shop, and emphasize fully open-source distribution that has already helped biotech startups secure funding. Looking ahead, both guests identify multimeric and protein-protein interaction prediction as the next major frontier, along with extending folding to DNA/RNA/small-molecule complexes (à la AlphaFold 3 and Boltz), molecular dynamics at scale, and eventually reconstructing cellular systems. They highlight combinatorial search-space blowup as the core scaling challenge and argue that GPU acceleration of search and inference tools is essential to making this tractable.
Original sources
Podcast NVIDIA AI Podcast
Episode:From AlphaFold to MMseqs2-GPU: How AI is Accelerating Protein Science - Ep. 273 ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.