Machine Learning-Guided Strategy Selection and Condition Prediction for Photoredox Catalysis.
7 affiliations
- 1The Institute for Advanced Studies (IAS), Wuhan University, Wuhan 430072, China.
- 2HOURS Inc., Suzhou 215000, China.
- 3Shandong Laboratory of Yantai Drug Discovery, Bohai Rim Advanced Research Institute for Drug Discovery, Yantai 264117, China.
- 4Beijing National Laboratory for Condensed Matter Physics and Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China.
- 5Department of Biomolecular Systems, Max Planck Institute of Colloids and Interfaces, Potsdam 14476, Germany.
- 6Howard University College of Pharmacy, 2300 Fourth Street NW, Washington, District of Columbia 20059, United States.
- 7State Key Laboratory of Drug Research & Center of Pharmaceutics, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Shanghai 201203, China.
Abstract
Photoredox catalysis offers powerful opportunities for organic synthesis, yet its broader application remains limited by the difficulty of identifying viable synthetic routines and conditions. Here, we present an experimentally validated, machine-learning-assisted workflow that combines literature-precedent retrieval with reaction-condition recommendation to support chemist-guided design of photoredox transformations. With 11,069 curated literature-derived reaction records, a variational autoencoder organizes photoredox reactions in a learned latent space to enable reaction-space navigation and retrieval of traceable synthetic precedents, while random-forest classifiers recommend feasible catalyst and reagent classes for proposed reactant-product pairs. External validation on unseen literature examples and experimental validation of selected model-guided proposals demonstrates the practical utility of the workflow. It further enabled three unreported photoredox transformations, highlighting its possibility for generating experimentally testable synthetic hypotheses.
Related
- Interplay of singlet and triplet excited states in photoredox catalysis.
- Chromoselective Control for Aminoalkylation of CF3‑Alkenes through Low-Energy Photoredox Catalysis
- Designing organic photoredox catalysts as a fragment-based puzzle solved by reinforcement learning
- Yield Prediction and Condition Recommendation on High-Throughput Experimentation Datasets: Encoding Effects and Noise Ceilings
- Folding and Function as Two Sides of One Coin: Statistical Mechanics, Information Theory, and Deep Learning of Binding, Catalysis, and Allostery in an Exactly Enumerable Lattice Protein
BibTeX
@article{ye2026machine,
title = {{Machine Learning-Guided Strategy Selection and Condition Prediction for Photoredox Catalysis.}},
author = {Xiaoyu Ye and Jiaxin Xu and Xiaohui Tian and Dongyuan Fan and Jingdi Ran and Xiaodong Zeng and Yunqi Cai and Peter H Seeberger and Zuo Zeng and Xiang Simon Wang and Yuling Xiao and Yi-Hung Chen},
journal = {JACS Au},
year = {2026},
month = sep,
doi = {10.1021/jacsau.6c00471},
url = {https://europepmc.org/article/MED/42953598},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - Machine Learning-Guided Strategy Selection and Condition Prediction for Photoredox Catalysis.
AU - Xiaoyu Ye
AU - Jiaxin Xu
AU - Xiaohui Tian
AU - Dongyuan Fan
AU - Jingdi Ran
AU - Xiaodong Zeng
AU - Yunqi Cai
AU - Peter H Seeberger
AU - Zuo Zeng
AU - Xiang Simon Wang
AU - Yuling Xiao
AU - Yi-Hung Chen
PY - 2026
DA - 2026/09/24
JO - JACS Au
DO - 10.1021/jacsau.6c00471
UR - https://europepmc.org/article/MED/42953598
AB - Photoredox catalysis offers powerful opportunities for organic synthesis, yet its broader application remains limited by the difficulty of identifying viable synthetic routines and conditions. Here, we present an experimentally validated, machine-learning-assisted workflow that combines literature-precedent retrieval with reaction-condition recommendation to support chemist-guided design of photoredox transformations. With 11,069 curated literature-derived reaction records, a variational autoencoder organizes photoredox reactions in a learned latent space to enable reaction-space navigation and retrieval of traceable synthetic precedents, while random-forest classifiers recommend feasible catalyst and reagent classes for proposed reactant-product pairs. External validation on unseen literature examples and experimental validation of selected model-guided proposals demonstrates the practical utility of the workflow. It further enabled three unreported photoredox transformations, highlighting its possibility for generating experimentally testable synthetic hypotheses.
SN - 2691-3704
ER -