Do we need equivariant models for molecule generation?

TL;DR

This paper examines whether non-equivariant CNNs with rotation augmentation can match the performance of complex equivariant GNNs in molecule generation, potentially simplifying models without sacrificing quality.

Contribution

It introduces a loss decomposition to analyze learned equivariance and evaluates factors affecting performance, pioneering the study of learned equivariance in generative molecular models.

Findings

Non-equivariant CNNs can learn equivariance with sufficient training.

Model size, dataset size, and training duration influence performance.

Rotation augmentation can enable CNNs to match equivariant GNNs in molecule tasks.

Abstract

Deep generative models are increasingly used for molecular discovery, with most recent approaches relying on equivariant graph neural networks (GNNs) under the assumption that explicit equivariance is essential for generating high-quality 3D molecules. However, these models are complex, difficult to train, and scale poorly. We investigate whether non-equivariant convolutional neural networks (CNNs) trained with rotation augmentations can learn equivariance and match the performance of equivariant models. We derive a loss decomposition that separates prediction error from equivariance error, and evaluate how model size, dataset size, and training duration affect performance across denoising, molecule generation, and property prediction. To our knowledge, this is the first study to analyze learned equivariance in generative tasks.