Equivariant Graph Neural Networks for 3D Macromolecular Structure

TL;DR

This paper introduces equivariant graph neural networks tailored for 3D macromolecular structures, demonstrating superior performance on biological benchmarks and showcasing the benefits of transfer learning in this domain.

Contribution

The work extends geometric vector perceptrons to develop equivariant GNNs for 3D structures, achieving state-of-the-art results on multiple biological tasks.

Findings

Outperforms all reference architectures on 3 of 8 tasks

Ties for first on 2 tasks in the ATOM3D benchmark

Transfer learning enhances downstream task performance

Abstract

Representing and reasoning about 3D structures of macromolecules is emerging as a distinct challenge in machine learning. Here, we extend recent work on geometric vector perceptrons and apply equivariant graph neural networks to a wide range of tasks from structural biology. Our method outperforms all reference architectures on three out of eight tasks in the ATOM3D benchmark, is tied for first on two others, and is competitive with equivariant networks using higher-order representations and spherical harmonic convolutions. In addition, we demonstrate that transfer learning can further improve performance on certain downstream tasks. Code is available at https://github.com/drorlab/gvp-pytorch.