A unified deep generative framework for surrogate-guided molecular discovery across diverse molecular spaces.
TL;DR
A high-accuracy surrogate predictor based on message-passing neural networks is decoupled from heterogeneous generators, and integrated into a closed-loop strategy of generate, score, select, and regenerate to progressively enrich high-performing candidates without modifying the underlying model architectures.
Abstract
A unified deep generative framework for surrogate-guided molecular discovery across diverse molecular spaces is presented to connect forward property prediction with property-oriented candidate generation. A high-accuracy surrogate predictor based on message-passing neural networks is decoupled from heterogeneous generators, including sequence models and diffusion models, and integrated into a closed-loop strategy of generate, score, select, and regenerate to progressively enrich high-performing candidates without modifying the underlying model architectures. The framework is evaluated on three chemically distinct tasks involving the detonation velocity of energetic materials, the binding affinity of BACE-1 inhibitors, and the effective permeability of macrocyclic compounds. Consistent enrichment of molecules satisfying predefined property thresholds is achieved across all datasets. Diffusion models perform better in small and rigid systems, whereas SELFIES-based sequence models exhibit stronger structural stability in macrocyclic and topologically complex spaces. These results demonstrate the generality and transferability of the unified surrogate-guided molecular discovery framework.