From AlphaFold to MMseqs2-GPU: How AI is Accelerating Protein Science - Ep. 273

What It Covers

Chris Delago from NVIDIA and Martin Steinegger from Seoul National University discuss GPU-accelerated protein structure prediction tools, including MMseqs2-GPU's acceptance to Nature Methods, which reduces homology search time from 80% to 20% of total AlphaFold computation.

Key Questions Answered

• •Homology Search Acceleration: MMseqs2-GPU inverts AlphaFold's computational bottleneck by reducing homology retrieval from 80% to 20% of total execution time, enabling the machine learning inference step to become the primary focus for further optimization and allowing structure prediction on standard gaming GPUs.
• •Protein Interaction Prediction: Multimer structure prediction remains significantly less accurate than monomer prediction, representing the next frontier. Solving protein-protein interactions enables reasoning about cellular pathways and drug targets, though combinatorial complexity creates massive computational scaling challenges requiring efficient search methods.
• •Data Explosion Management: Pre-AlphaFold databases contained 200,000 structures; post-AlphaFold databases contain hundreds of millions. Every existing computational biology tool must be redesigned to handle this thousand-fold increase, requiring new approaches like FoldSeek for rapid structural comparison and FoldDisco for identifying functional motifs at scale.
• •Open Source Collaboration Model: NVIDIA's digital biology strategy focuses on accelerating community tools through partnerships rather than proprietary development. The MMseqs2-GPU collaboration required patent-free, fully open-source code from inception, enabling startups to secure funding rounds by unblocking computational bottlenecks in their drug discovery pipelines.