Active Learning Enables Extrapolation in Molecular Generative Models

2025-01-03

AI Summary

Existing molecular generation models exhibit poor generalization to unseen chemical spaces, struggle to simultaneously optimize molecular performance and synthetic accessibility, and suffer from inaccurate out-of-distribution (OOD) property prediction. To address these limitations, we propose a closed-loop active learning framework for molecular generation, integrating a graph neural network–based generator, density functional theory (DFT)-derived quantum-chemical feedback, conditional generation, and an active learning strategy to enable co-iterative optimization of the generator and property predictor. Our approach achieves, for the first time, significant extrapolation in molecular property space (+0.44σ beyond the training distribution), improves OOD classification accuracy by 79%, and generates thermodynamically stable molecules at 3.5× the rate of the second-best baseline—substantially outperforming state-of-the-art methods.

Technology Category

• Machine Learning: Active Learning
• Natural Language Processing: Generation
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications

Abstract

Although generative models hold promise for discovering molecules with optimized desired properties, they often fail to suggest synthesizable molecules that improve upon the known molecules seen in training. We find that a key limitation is not in the molecule generation process itself, but in the poor generalization capabilities of molecular property predictors. We tackle this challenge by creating an active-learning, closed-loop molecule generation pipeline, whereby molecular generative models are iteratively refined on feedback from quantum chemical simulations to improve generalization to new chemical space. Compared against other generative model approaches, only our active learning approach generates molecules with properties that extrapolate beyond the training data (reaching up to 0.44 standard deviations beyond the training data range) and out-of-distribution molecule classification accuracy is improved by 79%. By conditioning molecular generation on thermodynamic stability data from the active-learning loop, the proportion of stable molecules generated is 3.5x higher than the next-best model.

Problem

• Molecular Generation
• Predictive Analytics
• Synthetic Feasibility

Innovation

• Active Learning
• Molecular Stability
• Efficiency Improvement