Evolutionary training-free guidance in diffusion model for 3D multi-objective molecular generation

2025-05-16

AI Summary

This work addresses the challenge of multi-objective property optimization for 3D molecular generation. To overcome the limitations of conventional methods—namely, expensive retraining upon introducing new property constraints and restricted structural flexibility—we propose a training-free, evolution-guided diffusion generative framework. Our approach directly embeds evolutionary operations (crossover and selection) into the noise space of a pre-trained unconditional 3D diffusion model, enabling joint optimization of user-specified multi-objective properties and structural constraints during denoising. It supports arbitrary 3D fragment insertion and unified trade-off handling for conflicting objectives, enhanced by multi-objective Bayesian optimization for efficient latent-space navigation. Experiments demonstrate superior accuracy over state-of-the-art conditional diffusion models on both single- and multi-objective generation tasks, with a 5× speedup in sampling. The framework is successfully applied to protein–ligand co-design and joint optimization of quantum chemical properties.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Deep Generative Models & Autoencoders

Application Category

• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

Discovering novel 3D molecular structures that simultaneously satisfy multiple property targets remains a central challenge in materials and drug design. Although recent diffusion-based models can generate 3D conformations, they require expensive retraining for each new property or property-combination and lack flexibility in enforcing structural constraints. We introduce EGD (Evolutionary Guidance in Diffusion), a training-free framework that embeds evolutionary operators directly into the diffusion sampling process. By performing crossover on noise-perturbed samples and then denoising them with a pretrained Unconditional diffusion model, EGD seamlessly blends structural fragments and steers generation toward user-specified objectives without any additional model updates. On both single- and multi-target 3D conditional generation tasks-and on multi-objective optimization of quantum properties EGD outperforms state-of-the-art conditional diffusion methods in accuracy and runs up to five times faster per generation. In the single-objective optimization of protein ligands, EGD enables customized ligand generation. Moreover, EGD can embed arbitrary 3D fragments into the generated molecules while optimizing multiple conflicting properties in one unified process. This combination of efficiency, flexibility, and controllable structure makes EGD a powerful tool for rapid, guided exploration of chemical space.

Problem

• Generating 3D molecular structures with multiple property targets
• Avoiding expensive retraining for new property combinations
• Embedding structural constraints while optimizing conflicting properties

Innovation

• Evolutionary guidance in diffusion sampling process
• Training-free framework with pretrained diffusion model
• Optimizes multiple conflicting properties simultaneously