A deep-learning project extending Geometric Vector Perceptrons (GVP) — a rotation-equivariant GNN for protein structure — with a TransformerConv attention layer, evaluated on protein design and the ATOM3D benchmark.
Protein structure (DHRS7B homology model) — the kind of 3D macromolecule GVP-GNNs learn from.
Overview
This repo builds on the original Geometric Vector Perceptrons GVP, a rotation-equivariant GNN, and combines a TransformerConv layer attempting to further improve the baseline model. Scripts for training / testing / sampling on protein design and training / testing on all ATOM3D tasks are provided.
Background
GVP-GNNs (Jing et al., 2021) extend standard dense layers to operate on collections of Euclidean vectors, enabling simultaneous geometric and relational reasoning over 3D macromolecular structure. ATOM3D (Townshend et al., 2020) provides a standardized suite of tasks on 3D molecular structures. This project investigates whether adding graph attention (TransformerConv) on top of the equivariant GVP message passing improves the baseline across these tasks.
Links
- Blogpost (full write-up): erencantatar.github.io/DL2-GVP/Blogpost.html · Blogpost.md
- Source code: github.com/erencantatar/DL2-GVP