Efficient Molecular Conformer Generation with SO(3)-Averaged Flow Matching and Reflow

Zhonglin Cao, Mario Geiger, Allan Dos Santos Costa, Danny Reidenbach, Karsten Kreis, Tomas Geffner, Franco Pellegrini, Guoqing Zhou, Emine Kucukbenli

TL;DR: We present a novel flow-matching objective and combine it with reflow to accelerate training and inference of molecular conformer generation

Abstract: Fast and accurate generation of molecular conformers is desired for downstream computational chemistry and drug discovery tasks. Currently, training and sampling state-of-the-art diffusion or flow-based models for conformer generation require significant computational resources. In this work, we build upon flow-matching and propose two mechanisms for accelerating training and inference of generative models for 3D molecular conformer generation. For fast training, we introduce the SO(3)-*Averaged Flow* training objective, which leads to faster convergence to better generation quality compared to conditional optimal transport flow or Kabsch-aligned flow. We demonstrate that models trained using SO(3)-*Averaged Flow* can reach state-of-the-art conformer generation quality. For fast inference, we show that the reflow and distillation methods of flow-based models enable few-steps or even one-step molecular conformer generation with high quality. The training techniques proposed in this work show a path towards highly efficient molecular conformer generation with flow-based models.

Lay Summary: Molecular conformer generation is a critical task in small molecule drug-discovery pipeline. At current stage, generative models for conformer generation are too slow to be applied in industrial-scale use cases. The motivation of our work is to accelerate flow-based generative model for molecular conformer generation, while maintaining high generation quality. We firstly propose a method to training flow-based generative models, called SO(3)-*Averaged Flow*, to improve generation quality and accelerate training. We also proposed to adopt the reflow and distillation algorithms to effectively enable one-step generation of molecular conformers while maintaining high generation quality. Combining these methods, our models show promising performance in efficient conformer generation, making them more practically useful in drug-discovery applications.

Application-Driven Machine Learning: This submission is on Application-Driven Machine Learning.

Primary Area: Deep Learning->Generative Models and Autoencoders

Keywords: Flow-matching, few-shot generation, equivariance, small molecules

Submission Number: 13550