This repository contains code to reproduce the results of the RECOMB 2026 submission "Gaining mechanistic insight from geometric deep learning on molecule structures through equivariant convolution".
Installation: git clone and cd into this repository. The package depends on the bioverse for data processing and evaluation pipelines, and on cosmic-torch which contains the Cosmo layers. Install them with pip install -r requirements.txt. Make sure to before install torch and torch-scatter according to their instructions and your system.
Usage: Models can be trained or tested with bioverse [train/test] config.yaml exp=[mnist/beta2d/qm9aph/qm9cv0/qm9gap/qm9hom/qm9lum/qm9mu0]. The figures of the paper can be reproduced with python plot.py exp=[mnist/beta2d].
License: TBD