Harvard Medical School's Walter Lab runs 1.7 million protein interaction predictions using NVIDIA DGX Cloud

Overview

The Walter Lab at Harvard Medical School, with access to NVIDIA DGX Cloud through the National Science Foundation's NAIRR pilot program, completed nearly 1.7 million protein-protein interaction predictions on ColabFold in three months using 32-node DGX clusters with 256 A100 GPUs, a task that previously would have taken years. The team curated 40,000 high-confidence protein interactions among 300 human genome maintenance proteins using two custom ML tools (KIRC and SPOC), publishing results via the Predictomes website.

The challenge

Other AI tools are trained to predict 3D structures of proteins but often do not reliably separate relevant protein-protein interactions (PPIs) from false-positive predictions, and manually determining protein structure was an incredibly time-consuming process that used to require years of research. The Walter Lab sought to expand its success with AlphaFold beyond a single lab's focus to the scale of the entire human proteome — 20,000 proteins interacting in perhaps a million ways.

The solution

The Walter Lab developed a multi-step pipeline: a machine learning tool called KIRC to identify protein pairs worth in-depth analysis by AlphaFold, and a second tool, the Structure Predictions and Omics-Informed Classifier (SPOC), to separate true and false PPI predictions. They first applied SPOC to a matrix of around 300 human genome maintenance proteins, generating 40,000 protein interaction predictions. To scale the project, the team adopted NVIDIA Accelerated Computing infrastructure and NVIDIA DGX Cloud — 32-node DGX clusters with 256 A100 GPUs — made available through NVIDIA's partnership with the National Science Foundation's NAIRR pilot program, running 1.7 million PPI predictions on ColabFold in three months. Results were published via the Predictomes website as an accessible, classifier-curated interactome resource.

Reported business value

The Walter Lab completed nearly 1.7 million protein-protein interaction predictions in three months — a process that would otherwise have taken years, if not more, according to graduate student Ernst Schmid. The classifier generated 40,000 high-confidence protein interaction predictions among 300 human genome maintenance proteins, providing new structural hypotheses for disease-relevant pathways.