Skip to content
DraconiticResearch
Search Sources API
194,161 studies31 sourcesUpdated 12 min ago
PreprintCC BYChemRxiv

Contrastive Reward Based Reinforcement Learning for Goal Directed 3D Molecular Design

  • Bohao Li1,2,3
  • Yuedong Yang1,2
  • Mingyuan Xu3
  • Hongming Chen3,4,5
5 affiliations
  1. 1School of Computer Science and Engineering
  2. 2Sun Yat-Sen University
  3. 3Guangzhou National Laboratory
  4. 4School of Pharmaceutical Sciences
  5. 5Guangzhou Medical University
Open at sourcePDFDOICite

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.

Related

  1. AET5: A transcriptome-guided molecular generation framework with contrastive self-supervised learning
    Zhikang Yuan, Xin Zhang, Gaoming Lin +6
    PLOS Computational BiologyPeer-reviewedCC BY 4.0Full textMolecular biology & gene regulationBioinformatics & computational biologyPLOS
  2. A Molecular Generation–Validation Workflow Combining Reversible Junction Tree Reinforcement Learning and Semi-empirical Quantum Chemistry: A Case Study on Molecular Solar Thermal Fuels
    Qingyu Zhu, Ryan Lingg, Reuben Szabo +2
    ChemRxivPreprintCC BYFull textMaterials & polymersChemRxiv
  3. Development of a Deep Reinforcement Learning Approach for Transition State Search in Molecular Systems: A Transferable Model Across Structural and Dimensional Variations
    Takechika Kikkawa, Shusuke Yamanaka, Takashi Kawakami +1
    arXivPreprintarXiv licenceFull textPhysical chemistryarXiv
  4. SelfTICA: contrastive learning of dynamical representations for rare-event sampling and characterization
    Kai Zhu, Jintu Zhang, Pietro Novelli +2
    arXivPreprintarXiv licenceFull textPhysical chemistryComputational physicsarXiv
  5. End-to-end multimodal structure elucidation from raw spectra combining contrastive learning and evolutionary algorithms
    Adrian Mirza, Luc Patiny, Kevin Maik Jablonka
    Nature CommunicationsPeer-reviewedCC BYFull text14 citedAnalytical chemistryOpenAlex

Cite

BibTeXRIS
BibTeX
@misc{li2026contrastive,
  title = {{Contrastive Reward Based Reinforcement Learning for Goal Directed 3D Molecular Design}},
  author = {Bohao Li and Yuedong Yang and Mingyuan Xu and Hongming Chen},
  howpublished = {ChemRxiv},
  year = {2026},
  month = sep,
  doi = {10.26434/chemrxiv.15008550/v1},
  url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008550/v1},
  note = {CC BY},
}
RIS
TY  - UNPB
TI  - Contrastive Reward Based Reinforcement Learning for Goal Directed 3D Molecular Design
AU  - Bohao Li
AU  - Yuedong Yang
AU  - Mingyuan Xu
AU  - Hongming Chen
PY  - 2026
DA  - 2026/09/09
PB  - ChemRxiv
DO  - 10.26434/chemrxiv.15008550/v1
UR  - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008550/v1
AB  - 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.
ER  -