Bonjour à tous,

Vous êtes cordialement invité.e.s à l'évaluation du Predoc III de Jiarui Lu, le 27 août à 13h30 (mode hybride).

Jury

Abstract

The dynamic behavior of biomolecules, such as proteins, plays a crucial role in their biological functions and properties. Traditionally, studying these dynamics has however relied on time-consuming molecular dynamics simulations, which often struggle with high energy barriers that prevent thermodynamically favored transitions within a feasible number of simulation steps.

This proposal explores the application of modern generative models, for example the score-based models on Riemannian manifolds, to capture and sample the complicated distributions of protein systems. Following a comprehensive review of the relevant techniques and literature, we propose a simple yet effective framework that leverages diffusion models with appropriate geometric modeling for zero-shot protein conformation sampling. This data-driven, transferable generative sampler achieves state-of-the-art performance in efficiently sampling unseen protein systems and generating conformation ensembles that align with long MD simulations.

Additionally, we introduce structure language models that generate conformation ensembles through a quantized latent space of structures. By leveraging vector quantized auto-encoders for representation learning, these models effectively capture the distribution of conformations within the discrete latent space. This approach not only demonstrates successful modeling of structural dynamics across different proteins but also presents a novel scheme for protein conformation sampling.

The proposal further investigates the ongoing work and future directions on this topic, such as the adaptive integration of generative models with feedback from physical simulators, the development of universal samplers for multiple biomolecules and complexes.