[Paper Review] Symphony: Symmetry-Equivariant Point-Centered Spherical Harmonics for 3D Molecule Generation

Ameya Daigavane, Song Eun Kim+2 more|arXiv (Cornell University)|Nov 27, 2023

Machine Learning in Materials Science4 citations

TL;DR

Symphony introduces an $E(3)$-equivariant autoregressive generative model for 3D molecular geometry that uses point-centered spherical harmonic projections to model radial and angular distributions of atoms simultaneously. It outperforms prior autoregressive models on QM9 and achieves competitive performance with diffusion models, while maintaining faster inference speeds and strong generalization on unseen fragments.

ABSTRACT

We present Symphony, an $E(3)$-equivariant autoregressive generative model for 3D molecular geometries that iteratively builds a molecule from molecular fragments. Existing autoregressive models such as G-SchNet and G-SphereNet for molecules utilize rotationally invariant features to respect the 3D symmetries of molecules. In contrast, Symphony uses message-passing with higher-degree $E(3)$-equivariant features. This allows a novel representation of probability distributions via spherical harmonic signals to efficiently model the 3D geometry of molecules. We show that Symphony is able to accurately generate small molecules from the QM9 dataset, outperforming existing autoregressive models and approaching the performance of diffusion models.

Motivation & Objective

• To develop a generative model that respects the $E(3)$ symmetries of molecules, including global and local symmetries inherent in atomic systems.
• To improve the representation of 3D molecular geometry in autoregressive generation by using higher-degree $E(3)$-equivariant features via spherical harmonics.
• To enable accurate and efficient modeling of both radial and angular distributions of new atoms in a single focus-atom framework.
• To evaluate generalization performance on unseen molecular fragments, moving beyond memorization of training data.
• To achieve faster inference than diffusion models while maintaining high validity and geometric accuracy.

Proposed method

• Symphony uses a point-centered autoregressive generation process where each new atom is predicted relative to a single focus atom, which acts as the local origin.
• It employs higher-degree $E(3)$-equivariant features derived from spherical harmonics $Y_{l,m}$ to represent angular and radial distributions of candidate atoms.
• The model parameterizes the conditional probability distribution of new atom positions using spherical harmonic signals, enabling joint prediction of radial and angular components.
• Message passing with $E(3)$-equivariant features ensures invariance under rotation and translation, preserving molecular symmetry.
• The model uses tensor products to combine features and generate higher-degree irreducible representations of $SO(3)$, enabling richer geometric modeling.
• A novel angular accuracy metric based on the bispectrum is introduced to evaluate local environment matching in generated molecules.

Experimental results

Research questions

• RQ1 Can spherical harmonic projections enable more accurate and efficient modeling of 3D molecular geometry in autoregressive generation?
• RQ2 How does $E(3)$-equivariance with higher-degree features improve generation quality compared to rotationally invariant models?
• RQ3 To what extent can Symphony generalize to completing unseen molecular fragments, indicating true generalization beyond memorization?
• RQ4 How does Symphony's inference speed compare to diffusion models and prior autoregressive models, despite higher computational cost?
• RQ5 Can the use of spherical harmonics reduce discretization errors in bond length distributions compared to lookup-table-based methods?

Key findings

• Symphony outperforms G-SchNet and G-SphereNet on all QM9 metrics, achieving a Jensen-Shannon Divergence (JSD) of 0.0003 for atom type counts and 0.0039 for local environment counts.
• The model achieves a bispectrum MMD of 0.0039 for the 10 most frequent local environments, significantly lower than G-SchNet (0.0150) and G-SphereNet (0.1016).
• Symphony achieves a valid completion rate (VCR) of 98.66% on unseen test fragments, outperforming G-SchNet (98.15%) and showing strong generalization capability.
• Inference speed is 0.293 seconds per molecule on a single A5000 GPU, approximately 3× faster than EDM (0.930 sec/mol), despite higher computational cost from tensor products.
• The model is trained in only ~80 epochs, significantly fewer than G-SchNet (330) and EDM (1100), suggesting improved data efficiency.
• Despite strong performance, Symphony’s bond length distributions are less accurate due to coarse radial discretization, affecting bond order assignment.