Reproduce “Flow Matching for Optimal Reaction Coordinates of Biomolecular System” on OpenResearch

About this paper

Mingyuan Zhang, Zhicheng Zhang, Hao Wu, Yong Wang

An unsupervised deep learning algorithm called Flow Matching for Reaction Coordinates (FMRC) identifies optimal reaction coordinates for biomolecular systems by leveraging lumpability and decomposability principles. FMRC effectively compresses high-dimensional dynamics into low-dimensional representations while preserving major slow modes, outperforming state-of-the-art methods in Markov State Model quality and demonstrating potential for enhanced sampling.

Key insights

• The theoretical concepts of lumpability and decomposability for optimal RCs can be effectively translated into a conditional probability framework and learned using deep generative models, specifically Flow Matching.
• A low-dimensional (e.g., 2D) reaction coordinate learned by FMRC can capture a system's major slow dynamics with higher fidelity and less redundancy compared to conventional methods that often require higher dimensions.
• The Flow Matching approach yields a robust and consistent RC learning process, exhibiting significantly lower training variance and requiring minimal hyperparameter tuning compared to other deep learning-based methods.

Summary by alphaXiv.