Learning Thermodynamics from Sequence: Predicting and Designing Complex Biomolecular Mixtures
SLAM Seminar
Biological and bio-inspired materials derive function from sequence-encoded interactions in multicomponent mixtures. Predicting how biomolecular sequences determine thermodynamic behavior and material properties is therefore a central challenge across biophysics and molecular engineering. In this talk, I introduce a machine-learning framework that maps sequences directly to context-dependent chemical potentials—without requiring direct free-energy calculations. The approach learns low-dimensional, physically interpretable representations in which distances encode differences in thermodynamic behavior across mixtures. Using intrinsically disordered proteins as a model system, I demonstrate how this framework enables quantitative prediction of partitioning and multicomponent phase diagrams, while revealing how sequence composition and patterning jointly control mixture thermodynamics. More broadly, this work outlines a path toward quantitative models of complex mixtures that bridge machine learning, biophysics, and molecular engineering.