Passer au contenu

/ Department of Computer Science and Operations Research

Je donne


Chence Shi's Predoc III Presentation

Dear all / Bonjour à tous,

We are happy to invite you to Chence Shi's Predoc III defense on Thursday, August 24th, at 9am (Hybrid event).

Vous êtes cordialement invité.e.s à la présentation du sujet de recherche de Chence Shi, jeudi 24 août à 9h00. (Présentation hybride).


Title:  Deep Generative Modeling of Atomistic Systems

Date: August 24th, 2023 at 9:00am-11:00am PM EST

Location: Auditorium 1 - 6650 Rue Saint Urbain




PrésidentBengio, Yoshua
Directeur de rechercheTang, Jian
Beaini, Dominique



“Artificial intelligence (AI) has gained growing attention in scientific discoveries, catalyzing the emergence of a new research direction known as AI for Science (AI4Science).

Focusing on a specific facet of AI4Science, this research proposal studies the deep generative modeling of atomistic systems, which encompasses a lot of fundamental challenges in computational biology and drug discovery, spanning from the generation of molecular graphs and conformations to protein sequences and structures.

These tasks share common hurdles, which offer an opportunity for simultaneous investigation. One central challenge involves capturing the symmetries inherent in atomistic systems, including permutation equivariance of graph data and roto-translation equivariance of structural data, etc.

While extensively explored in existing literature, prior approaches to these tasks often fail to adequately capture system symmetries, rely on inefficient traditional methods, or require domain knowledge for intricate feature engineering.

This proposal introduces novel deep generative models expressly tailored for atomistic systems, directly baking symmetry inductive bias into models' design.

The ultimate goal is to elucidate intricate geometric interactions and interplay of different particles within atomistic systems, grounded in physics' first principles, with the hope of facilitating the process of drug discovery.

Notably, our models circumvent the need for domain-specific expertise and frequently outperform previous methods in terms of efficiency.

Extensive experiments prove the effectiveness of the proposed models.”


The keywords are “Generative Models, Graphs Machine Learning, Deep Learning, Drug Discovery”.