Self-supervised learning of atomic and molecular representations with 3D equivariant graph neural networks

At the American Chemical Society Spring National Meeting & Expo, Atomwise members were selected to present their research and work. Learn what our Atoms have been working on below and visit Atomwise at ACS Spring 2022 National Meeting for other presentation sessions. 

Aryan Pedawi, PhD 
Atomwise

Atomwise Co-Authors: Kate Stafford, Andreana Rosnik, Dr Saulo H de Oliveira, Brandon Anderson, Jon Sorenson, Ph.D.

Title: Self-supervised learning of atomic and molecular representations with 3D equivariant graph neural networks

Division: Computers in Chemistry

Abstract

Equivariant graph neural networks operating on 3D atom point clouds have emerged as a powerful paradigm for a variety of tasks where the 3D molecular structure is of primary importance, such as machine learned force-fields or protein-ligand binding affinity prediction. Equivariant networks can efficiently learn local geometric information, and sometimes require significantly less training data to achieve high quality predictions than approaches that do not encode for physical symmetries by construction. In parallel, computer vision and natural language processing have witnessed significant advances in the applicability of self-supervised pre-training, making use of larger libraries of unlabeled data to learn rich representations that transfer well to downstream tasks. In this work, we propose and investigate a number of pre-text tasks for learning self-supervised representations of molecular or biological data in the context of their 3D scene. We explore tasks defined on 3D conformations of ligands, of proteins, and of protein-ligand interactions. Finally, we study how the learned representations transfer to the more data-scarce supervised tasks of interest compared to baselines like supervision-only.

View Poster

Join our team