LADE Seminar

Abstract

Diffusion Models (DMs) have become a central paradigm in generative modeling. They have attained unprecedented performances on multiple complex generative tasks such as image denoising, conditional image synthesis or temporal data modeling, surpassing the long-time prevailing generative adversarial networks (GANs) both in ease-of-training and diversity of generated samples. For these reasons, DMs are now being increasingly applied to the task of molecular conformational prediction and sampling. In this presentation we aim to provide an overview of the recent advances in applying DMs to molecular conformational sampling with a particular emphasis on RNA. Compared to the well-studied proteins, RNA presents unique challenges when it comes to structural prediction, including scarcity of data and a highly dynamic conformational landscape. We will present some of our thoughts on designing a diffusion-based approach to the task of RNA conformational ensemble prediction using a combination of coarse-graining and torsional diffusion, a variant of diffusion adapted to internal molecular coordinates. Our goal will be to elicit an open discussion about the promises and challenges of DMs applied to the modeling of RNA dynamics.