Property-Guided Molecular Generation and Optimization via Latent Flows

arXiv cs.LG / 3/31/2026

💬 OpinionSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper frames molecular discovery as an inverse-design problem and notes that latent-space optimization can harm chemical validity, structural fidelity, and stability in existing generative approaches.
It introduces MoltenFlow, a modular framework that merges property-organized latent representations with flow-matching generative priors plus gradient-based guidance.
MoltenFlow supports both property-conditioned generation and local latent-space optimization within a unified framework.
The authors report efficient multi-objective molecular optimization with controllable trade-offs under fixed “oracle” query budgets, and they show that a learned flow prior boosts unconditional generation quality.

Abstract

Molecular discovery is increasingly framed as an inverse design problem: identifying molecular structures that satisfy desired property profiles under feasibility constraints. While recent generative models provide continuous latent representations of chemical space, targeted optimization within these representations often leads to degraded validity, loss of structural fidelity, or unstable behavior. We introduce MoltenFlow, a modular framework that combines property-organized latent representations with flow-matching generative priors and gradient-based guidance. This formulation supports both conditioned generation and local optimization within a single latent-space framework. We show that guided latent flows enable efficient multi-objective molecular optimization under fixed oracle budgets with controllable trade-offs, while a learned flow prior improves unconditional generation quality.