Dice Question Streamline Icon: https://streamlinehq.com

Performance of equivariant architectures on MD-based training

Evaluate the predictive performance and sample efficiency of E(n)-equivariant graph neural network architectures, such as EGNN and EGMN, when trained on molecular dynamics simulation frames for protein-ligand binding affinity prediction, in comparison to convolutional and recurrent baselines.

Information Square Streamline Icon: https://streamlinehq.com

Background

Prior work (Libouban et al.) trained convolutional and recurrent architectures on MD frames to augment PDBbind and reported improvements for binding affinity prediction, but did not consider equivariant architectures.

The authors explicitly note that the impact and performance of equivariant architectures in this MD context remains unclear, motivating their investigation of EGNN and EGMN modules in SurGBSA.

References

This study did not consider equivariant architectures so it is not clear how these architectures would perform in this context.

— SurGBSA: Learning Representations From Molecular Dynamics Simulations (2509.03084 - Jones et al., 3 Sep 2025) in Section 1 (Introduction)