Tuesday, November 17, 2026 · 12:00 PM
Add to calendarShriram Center · Room 262
Title: Generative Models for Protein Design
Abstract: Generative models are flexible tools with which we can model complex data distributions. When applied to proteins, we aim to build a model capable of generalizing the patterns observed in natural protein sequence/structure/function to the design of novel proteins conditioned on arbitrary desired functions. Samples from a generative model are hypotheses of the physical world. I show that generative models can expand functional space beyond what nature has already explored. First, I propose the SHAPES framework to benchmark how well generative models cover the distribution of native protein structures. I present the perspective of an oracle trap, in which filtering designs by structure prediction models leads to oversampling of idealized helices and undersampling of loops. Second, I present Protpardelle-1c, a collection of generative models of protein structure with improved conditional generation capabilities. I focus on motif scaffolding, in which we can condition on function by identifying a substructure/subsequence pair which geometrically defines the desired function. Protpardelle-1c excels on MotifBench, a standardized benchmark for motif scaffolding, proposing designable scaffolds to problems previously unsolved by other models. Finally, I apply generative models to the design of lanthanide binding proteins. I demonstrate how partial diffusion with Protpardelle-1c allows access to local perturbations of the lanmodulin fold to unearth a previously unexplored sequence space which has inverted rare earth selectivity, favoring heavy rare earths over light rare earths. Given a feedstock initially enriched in La/Ce (70%), the designed protein removes ~99% of lanthanum and cerium to enrich for more valuable heavy rare earths. I also validate motif scaffolding experimentally, showing that EF hands hosted in novel non-lanmodulin scaffolds recapitulate their intended selectivity profiles, and use this to probe the modularity of selectivity and the interplay between motif and scaffold.
Shriram Center 443 Via Ortega, Stanford, CA 94305 Room 262
Tuesday, November 17, 2026 · 12:00 PM