Year: 2026
Type: preprint
Abstract: Although three-dimensional flow-matching models have become powerful tools for efficient sampling of high-quality 3D molecular structures, reward-based reinforcement learning for flow-matching models is complicated by the absence of a tractable trajectory likelihood for hybrid molecular states comprising categorical features and continuous coordinates. Existing policy-optimization strategies address this difficulty by stochasticizing the generative process or constructing likelihood surrogates, introducing additional approximations. Here, we present ReCoFlow, a novel trajectorylikelihood-free reinforcement learning framework for reward-contrastive post-training of pretrained molecular flow-matching models. Within ReCoFlow, reward-favored molecules provide positive endpoint supervision for the flow-matching model, whereas reward-disfavored molecules are used to construct reference-guided negative targets. Additionally, an anchor loss to an exponential moving average (EMA) reference model constrains the drift of the fine-tuned model. In benchmarking studies including physicochemical property optimization, three-dimensional shape optimization, and pocket-conditioned 3D molecular generation, ReCoFlow consistently improved the performance of flow-matching models over reward-weighted and surrogate-ratio optimization schemes while maintaining high molecular validity. These results suggest that ReCoFlow could be an effective reinforcement-learning strategy for goal-directed molecular design.
Source: ChemRxiv
Authors: Bohao Li, Yuedong Yang, Mingyuan Xu, Hongming Chen
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
Award: G431056085
OpenAlex ID: W7212084319