Machine learning of biomolecular interactions and the human signaling networks they comprise

My lab will use machine learning to build physically-grounded models of biomolecules and their interactions and apply these models at proteome (genome) scale to address basic questions in the systems biology of human signaling. On the modeling front, our efforts will focus on building computational models of protein-ligand interactions, with a specific emphasis on post-translationally modified ligands that cells widely employ in signaling networks. I hypothesize that a step change in accuracy and generality of protein-ligand interaction models is possible using deep learning advances in protein structure prediction and protein representation learning. My lab has been at the forefront of these advances, having developed the first end-to-end differentiable model of protein structure prediction (RGN); the first protein language model (UniRep), a key technique for learning mathematical representations that capture chemical, structural, and evolutionary properties of proteins; and one of the first deep learning methods for protein-protein interactions (HSM). We will leverage our expertise in these domains to predict protein-ligand interactions based on both sequence and structure information. We will further develop specialized models for predicting protein structures and alternate protein conformations for the purpose of predicting protein-ligand interaction, using these predictions as inputs for our protein-ligand interaction models. On the biological front, we will employ these machine-learned models to assemble person-specific signaling networks to understand how normal allelic variation is manifested at the level of signaling networks, and how these networks are perturbed in human diseases. To study general variation in signaling networks, we will use exome sequences (UK Biobank and NHLBI TOPMed) to build individualized networks that map person-specific protein sequences to protein-ligand affinities. We will quantify how network topology varies among individuals and populations and test whether disease-associated traits correlate with topology. We will also compare networks of healthy and disease-afflicted persons to identify topological differences that predispose individuals to genetic diseases. Ultimately, I expect machine-learned models to be sufficiently predictive of ligand binding that mechanistic understanding of pathway rewiring by mutations is possible. While my focus will be computational, I expect to carry out close collaborations—with the Fordyce Lab (Stanford) to experimentally characterize and validate protein-ligand interactions and the Shen Lab (Columbia) to perform statistical genetic analyses—to exploit synergies at the interface of computation and experimentation.