Research in Progress | Rising Star Seminar

Taming the Biomolecular Structure Revolution: Representation Learning for dynamics, RNA, and substructures

Abstract: The post-AlphaFold biomolecular structure revolution has set off a wave of structure-aware machine learning, but most of that progress still rests on several convenient assumptions, such as that one static structure is enough to describe a protein, that proteins are the workhorses of molecular function, and that the rigid substructure units are a sufficient unit of analysis. Each is now under pressure. Protein dynamics, the ensembles of conformations a protein samples to function, break the single-structure assumption behind virtually every current encoder. RNA 3D structure, governed by distinct biophysics and tightly coupled to function, remains underserved by tools built for proteins, despite the therapeutic appeal of RNA targeting. Finally, much biomolecular function lives in cohesive substructures, including catalytic triads, binding pockets, and base-pairing networks, that traditional encoders tend to dilute.

I will discuss how our group has approached these challenges with a unifying recipe: building biology-aware benchmarks that expose where current models break, designing representations that respect the geometry, sparsity, and modularity of the data, and reasoning at the substructure level as one important lens on where function actually lives. Taming this revolution, I argue, will hinge less on bigger models than on representation learning that embraces richer modalities and biologically meaningful units of analysis.

Carlos Oliver, PhD is an Assistant Professor at Vanderbilt University, with a joint appointment in the Departments of Molecular Physiology & Biophysics and Computer Science, and a core member of the Vanderbilt Center for AI in Protein Dynamics. His research group develops geometric deep learning methods for structural biology, with a focus on RNA 3D structure, protein conformational dynamics, and functional substructure discovery in biomolecules.