DSAI seminar with Juan Viguera Diez

Abstract

Efficiently sampling state distributions of chemical and physical multi-body systems is a major outstanding challenge in the computational sciences, with significant implications for various applications, including drug design, quantifying binding affinity to targets, and assessing drug solubility. While machine learning approaches have shown promise to tackle this issue, they often fall short in capturing crucial entropic contributions to the equilibrium ensemble, critical for downstream applications.

In this talk, I will present a simple method aimed at generating equilibrium ensembles of non-cyclic molecules, seeking to overcome the aforementioned challenge. Our approach uses surrogate models for the torsional degrees of freedom within a molecule, which require long simulation times for representative sampling. By leveraging these surrogate models, we effectively re-weight and conduct short Molecular Dynamics (MD) simulations.

Through extensive evaluations involving geometrical and thermodynamical observables, our results demonstrate the effectiveness of our method in generating diverse and physically realistic conformations. Notably, on average, our method generates lower energy ensembles than the MD baseline, which served as the source of training data, and matches the energy of long Replica Exchange simulations with high accuracy.

About the speaker

Juan Viguera Diez is a WASP industrial PhD student at AstraZeneca (AZ) and Chalmers University of technology. His research focuses on deep generative model for molecules. His thesis work is supervised by Simon Olsson (DSAI, Chalmers) and Ola Engkvist (DSAI, Chalmers and Molecular AI, AZ).

Juan holds a BSc. degree in Physics from University of Zaragoza and a MSc. in Complex Adaptive Systems from Chalmers University of Technology.

This is a seminar from the DSAI seminars series usually held every Monday at 14:00 by the Data Science and AI division. The seminars are usually hybrid. No registration is required.