SupSiam: Non-contrastive Auxiliary Loss for Learning from Molecular Conformers

TL;DR

This paper introduces SupSiam, a non-contrastive auxiliary loss for learning from molecular conformers, enhancing Euclidean neural networks' smoothness and reliability in drug-activity prediction tasks.

Contribution

It proposes a positive-pair only auxiliary task for Siamese networks that improves the training of E3NNs and extends manifold smoothness to probabilistic and regression settings.

Findings

Auxiliary task improves manifold smoothness in E3NNs.

Method maintains performance metrics in drug-activity prediction.

Analysis reveals effects of task-weighting, latent dimension, and regularization.

Abstract

We investigate Siamese networks for learning related embeddings for augmented samples of molecular conformers. We find that a non-contrastive (positive-pair only) auxiliary task aids in supervised training of Euclidean neural networks (E3NNs) and increases manifold smoothness (MS) around point-cloud geometries. We demonstrate this property for multiple drug-activity prediction tasks while maintaining relevant performance metrics, and propose an extension of MS to probabilistic and regression settings. We provide an analysis of representation collapse, finding substantial effects of task-weighting, latent dimension, and regularization. We expect the presented protocol to aid in the development of reliable E3NNs from molecular conformers, even for small-data drug discovery programs.