Reproduce “Towards Predicting Equilibrium Distributions for Molecular Systems with Deep Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shuxin Zheng, Jiyan He, Chang Liu, Yu Shi, Ziheng Lu, Weitao Feng, Fusong Ju, Jiaxi Wang and 10 more

Microsoft Research AI4Science developed Distributional Graphormer (DiG), a deep learning framework that predicts equilibrium distributions of molecular systems by simulating a diffusion process. It efficiently samples diverse, chemically plausible structures for proteins, ligands, catalysts, and carbon polymorphs, demonstrating computational speedups of orders of magnitude compared to traditional molecular dynamics and achieving competitive accuracy against experimental and DFT data across various tasks.

Key insights

  • Modeling molecular equilibrium distributions as a reverse diffusion process allows for efficient, independent sampling of diverse molecular conformations from a simple prior distribution, bypassing rare event challenges.
  • Integrating physics-informed pre-training (PIDP) enables the model to learn complex energy landscapes and their corresponding score functions without requiring extensive ground-truth equilibrium samples.
  • The framework's adaptability to various molecular representations (coarse-grained for proteins, all-atom for ligands/catalysts) and its handling of non-Euclidean spaces enhance its generalizability across diverse molecular systems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.