Abstract
Generating 3D molecular structures is a fundamental challenge in drug discovery and materials science. While diffusion models have set the standard for quality, their iterative nature leads to high computational costs. We introduce GOAT (Geometric OptimAl Transport), a generative framework designed to accelerate synthesis by unifying atom coordinates and types into a single equivariant latent space. By solving a joint optimal transport problem and implementing a "purification" mechanism for transport paths, GOAT achieves a 10x speedup over traditional diffusion models and a 2x speedup over state-of-the-art flow-matching methods, while simultaneously enhancing the chemical novelty of generated molecules.
Introduction
Current state-of-the-art 3D molecule generation typically relies on diffusion models, which require hundreds to thousands of denoising steps. Newer flow-matching techniques attempt to find "straight-line" trajectories from noise to data but face two primary hurdles: the multi-modality of molecular data (combining continuous 3D coordinates with categorical atom types) and sub-optimal coupling between noise and target distributions. These issues often result in winding, inefficient paths that are difficult for neural networks to learn and slow to simulate during inference.
Methodology
GOAT addresses these challenges through three core innovations:
- Joint Latent Space: Instead of treating coordinates and atom types as separate entities, GOAT utilizes an Equivariant Autoencoder to map both modalities into a unified, continuous latent space. This representation respects molecular symmetries (rotation and translation), ensuring that the underlying identity of a molecule remains invariant to its spatial orientation.
- Geometric Optimal Transport (OT): Within the latent space, the model solves a bilevel optimization problem to identify the shortest possible path—the optimal coupling—between the base noise distribution and the target molecular data. This process accounts for rotations, translations, and atom permutations to minimize the overall transport cost.
- Coupling Purification: To ensure high-quality trajectories, GOAT incorporates a purification mechanism that filters out chemically nonsensical intermediate states during training. This ensures the model learns efficient, "straight-line" paths that lead to stable molecular structures.
Experiments and Results
The model was benchmarked on the QM9 (small molecules) and GEOM-DRUG (large, drug-like compounds) datasets.
Efficiency and Speed
GOAT demonstrates remarkable efficiency, requiring only 90 sampling steps to generate valid molecules, compared to 200 steps for EquiFM and 1,000 steps for standard diffusion models (EDM). In terms of wall-clock time, GOAT achieves a generation speed of 0.12s per significant molecule on QM9, outperforming all baselines.
Quality and Novelty
Beyond speed, GOAT significantly improves the diversity of generated structures. On the QM9 dataset, it achieved a 78.6% novelty score, a substantial increase over previous leaders like EquiFM (57.4%). This indicates that the model is effectively exploring the chemical space rather than merely memorizing training data.
Transport Cost Reduction
By optimizing the coupling between noise and data, GOAT achieves a staggering reduction in transport cost—approximately 89.65% lower than previous hybrid methods. This reduction simplifies the learning task for the neural network, leading to the observed gains in both speed and validity.
Conclusion
GOAT represents a significant step forward in efficient 3D molecule generation. By unifying molecular modalities and optimizing transport paths through geometric optimal transport, it overcomes the speed-quality trade-offs inherent in previous generative models. Future work may explore model distillation to further compress inference times, potentially enabling real-time molecular design and screening.
Summary updated