Reaction-Level Machine Learning for Condition Optimization in Suzuki–Miyaura Cross-Coupling
- 1Independent Researcher
Abstract
Selecting the ligand, base and solvent for a Suzuki–Miyaura coupling remains largely empirical, and highthroughput experimentation (HTE) screens are usually mined only for one reaction’s best condition. This study tests whether condition–yield relationships learned from some reactions transfer to reactions a model has never seen. Using a published HTE dataset of 5,760 reactions (15 quinoline couplings × 384 ligandbase–solvent combinations), models were trained in MATLAB on categorical condition descriptors with whole reactions held out: a global strategy trained on 12 reactions and was tested on 3; a focused strategy trained on 3 reactions sharing a boronic acid and was tested on a fourth. Selected models were compared with a majority-class predictor and a no-learning condition-average ranker, with 95% bootstrap confidence intervals. Ranking transferred: high-yield (≥62%) precision–recall AUC was 0.21–0.34 where only 6–8% of conditions were high-yield and 0.92–0.96 where most were, above the base rate on every held-out reaction. However, the condition-average ranker matched the trained models within overlapping confidence intervals on all four held-out reactions, so what transfers is mainly the persistence of good-on-average conditions. At a fixed 62% cut-off, a RUSBoosted ensemble flagged 58–96% of high-yield conditions against 6–27% for the untuned average. Accuracy misled: naive Bayes scored 94.8% where always predicting low yield scores 94.3%. Exact yields did not transfer (external R2 at or below zero on most reactions), and focusing the training set left ranking unchanged (PR-AUC 0.943 vs 0.942). Reaction-level splits, minority-class metrics and a condition-average baseline should be standard when evaluating models on HTE data.
Related
- RxnOptBench: Benchmarking LLMs for Reaction-Condition Optimization in Organic Methodology
- pdb2reaction: End-to-End Reaction-Path Elucidation from PDB Structures Using Machine-Learning Interatomic Potentials
- Construction of a Curated Literature Dataset and Machine Learning-Guided Exploration of Supported Metal Catalysts for Low- to Intermediate-Temperature Reverse Water–Gas Shift Reaction
- Machine Learning-Guided Strategy Selection and Condition Prediction for Photoredox Catalysis.
- Machine Learning and Corona Effect Coupling, a New Tool for Water Quality Control: Highlighting the Impact of Microwaves on Water Samples
BibTeX
@misc{manne2026reactionlevel,
title = {{Reaction-Level Machine Learning for Condition Optimization in Suzuki–Miyaura Cross-Coupling}},
author = {Pranav Manne},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010362/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010362/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Reaction-Level Machine Learning for Condition Optimization in Suzuki–Miyaura Cross-Coupling
AU - Pranav Manne
PY - 2026
DA - 2026/10/13
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010362/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010362/v1
AB - Selecting the ligand, base and solvent for a Suzuki–Miyaura coupling remains largely empirical, and highthroughput experimentation (HTE) screens are usually mined only for one reaction’s best condition. This study tests whether condition–yield relationships learned from some reactions transfer to reactions a model has never seen. Using a published HTE dataset of 5,760 reactions (15 quinoline couplings × 384 ligandbase–solvent combinations), models were trained in MATLAB on categorical condition descriptors with whole reactions held out: a global strategy trained on 12 reactions and was tested on 3; a focused strategy trained on 3 reactions sharing a boronic acid and was tested on a fourth. Selected models were compared with a majority-class predictor and a no-learning condition-average ranker, with 95% bootstrap confidence intervals. Ranking transferred: high-yield (≥62%) precision–recall AUC was 0.21–0.34 where only 6–8% of conditions were high-yield and 0.92–0.96 where most were, above the base rate on every held-out reaction. However, the condition-average ranker matched the trained models within overlapping confidence intervals on all four held-out reactions, so what transfers is mainly the persistence of good-on-average conditions. At a fixed 62% cut-off, a RUSBoosted ensemble flagged 58–96% of high-yield conditions against 6–27% for the untuned average. Accuracy misled: naive Bayes scored 94.8% where always predicting low yield scores 94.3%. Exact yields did not transfer (external R2 at or below zero on most reactions), and focusing the training set left ranking unchanged (PR-AUC 0.943 vs 0.942). Reaction-level splits, minority-class metrics and a condition-average baseline should be standard when evaluating models on HTE data.
ER -