Score-based 3D molecule generation with neural fields

2025-01-15

AI Summary

Generating large molecules—such as drug-like compounds and macrocyclic peptides—in continuous 3D space without prior structural constraints remains a fundamental challenge. Method: This work introduces the first unconditional, all-atom generative framework based on atomic density fields and neural fields. It proposes a novel “walk-jump” sampling paradigm: geometrically aware density field evolution via Gaussian-smoothed Langevin MCMC (“walk”) is combined with single-step score-matching denoising (“jump”) for accelerated convergence; conditional neural field encoding and continuous density field decoding eliminate reliance on predefined molecular topology. Contribution/Results: It is the first to integrate neural fields with score matching for 3D molecular generation. The method supports molecules of arbitrary size—including macrocyclic peptides with >100 atoms—with linear computational scalability. On drug-like molecule generation, it achieves state-of-the-art performance, improves sampling speed by over 10×, and strictly satisfies both geometric validity and chemical feasibility.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Deep Generative Models & Autoencoders
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications

Abstract

We introduce a new representation for 3D molecules based on their continuous atomic density fields. Using this representation, we propose a new model based on walk-jump sampling for unconditional 3D molecule generation in the continuous space using neural fields. Our model, FuncMol, encodes molecular fields into latent codes using a conditional neural field, samples noisy codes from a Gaussian-smoothed distribution with Langevin MCMC (walk), denoises these samples in a single step (jump), and finally decodes them into molecular fields. FuncMol performs all-atom generation of 3D molecules without assumptions on the molecular structure and scales well with the size of molecules, unlike most approaches. Our method achieves competitive results on drug-like molecules and easily scales to macro-cyclic peptides, with at least one order of magnitude faster sampling. The code is available at https://github.com/prescient-design/funcmol.

Problem

• 3D molecular structure generation
• large molecules
• continuous space

Innovation

• 3D Molecular Representation
• FuncMol Model
• Open Source Code