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.

What problem does this paper attempt to address?

The paper aims to address the following issues: 1. **Generating 3D Molecular Structures**: In drug discovery, material design, and fundamental scientific research, generating atomic systems with diverse geometries and desirable properties is crucial. Directly enumerating and validating all possible 3D structures is computationally infeasible and cannot effectively characterize atomic systems. Therefore, machine learning methods are needed to explore this domain by learning to generate 3D molecular structures. 2. **Overcoming Existing Challenges**: Effective 3D molecular generation models must learn to represent and produce highly relevant geometries to represent chemically valid and energetically favorable configurations. These models need to overcome several challenges: - Chemical Rules: The generation models are trained on 3D structures obtained from computationally intensive quantum mechanical calculations and must learn to follow chemical rules, generating stable molecular structures based solely on examples. - Atomic Position Precision: The stability of atomic systems depends on the precise placement of individual atoms, and any omission or misplacement of atoms can lead to significant property changes and instability. - Intrinsic Symmetry: Atoms of the same element are indistinguishable, so there is no consistent method to order atoms within an atomic system; additionally, atomic systems lack a global coordinate system (global symmetry), and local symmetry causes geometric patterns to repeat in different positions and orientations. 3. **Introducing the New Model Symphony**: This paper introduces a new autoregressive generative model called Symphony, which utilizes high-order E(3)-equivariant features and spherical harmonic projections to construct molecules while respecting the E(3) symmetry of molecular fragments. Compared to existing autoregressive models, Symphony uses spherical harmonic projections to parameterize the distribution of new atomic positions, allowing it to use a single "focus" atom as the selected origin during the generation process while predicting the radial and angular distribution of possible atomic positions. 4. **Experimental Results and Comparison**: Experiments on the QM9 dataset show that Symphony outperforms existing autoregressive models on various metrics and approaches the performance of diffusion models. Additionally, a bispectral-based metric is introduced to evaluate the angular accuracy of the generated local environments, demonstrating that Symphony can generate valid molecules with a high success rate, even under unseen molecular fragment conditions.

Symphony: Symmetry-Equivariant Point-Centered Spherical Harmonics for 3D Molecule Generation

Lift Your Molecules: Molecular Graph Generation in Latent Euclidean Space

3DMolNet: A Generative Network for Molecular Structures

Equivariant score-based generative diffusion framework for 3D molecules

Symmetry-Informed Geometric Representation for Molecules, Proteins, and Crystalline Materials

Nonadiabatic Dynamics Via the Symmetrical Quasi-Classical Method in the Presence of Anharmonicity.

Advances in Deep Learning-Based 3D Molecular Generative Models

Symmetry-adapted modeling for molecules and crystals

Learning Harmonic Molecular Representations on Riemannian Manifold

Equivariant representations for molecular Hamiltonians and N-center atomic-scale properties

Equivariant Shape-Conditioned Generation of 3D Molecules for Ligand-Based Drug Design

Equivariant Flow Matching with Hybrid Probability Transport for 3D Molecule Generation

Symmetry-Constrained Generation of Diverse Low-Bandgap Molecules with Monte Carlo Tree Search

Efficient 3D Molecular Generation with Flow Matching and Scale Optimal Transport

Tokenizing 3D Molecule Structure with Quantized Spherical Coordinates

3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction

Learning Neural Generative Dynamics for Molecular Conformation Generation

MolSym: A Python package for handling symmetry in molecular quantum chemistry

Symmetrical Graph Neural Network for Quantum Chemistry, with Dual R/K Space

Molecular Conformation Generation via Shifting Scores

Learning Joint 2-D and 3-D Graph Diffusion Models for Complete Molecule Generation