Do we need equivariant models for molecule generation?

2025-07-13

AI Summary

Is explicit equivariant modeling strictly necessary for high-quality 3D molecular generation? This work challenges the prevailing assumption that architectural equivariance is indispensable. Method: We investigate whether non-equivariant convolutional neural networks (CNNs) can implicitly learn rotational equivariance through extensive rotation-based data augmentation. We introduce a novel loss decomposition framework to quantitatively analyze how model size, dataset scale, and training duration affect the emergence of equivariance—first such systematic study in molecular generation. Experiments span denoising, conformational generation, and property prediction. Results: With sufficient rotational augmentation, non-equivariant CNNs match or surpass state-of-the-art equivariant graph neural networks (GNNs) across all tasks, while exhibiting greater training stability, lower computational overhead, and superior scalability. Our findings establish a lightweight, efficient paradigm for 3D molecular generation without architectural equivariance constraints.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Natural Language Processing: Generation

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI

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.

Problem

• Assessing if non-equivariant CNNs match equivariant GNNs in molecule generation
• Analyzing how model and dataset size affect equivariance learning
• Investigating learned equivariance in generative tasks for molecular discovery

Innovation

• Non-equivariant CNNs with rotation augmentations
• Loss decomposition separates prediction and equivariance errors
• Analyzes learned equivariance in generative tasks