Predicting Substrate Reactivity in Oxidative Homocoupling of Phenols Using Positive and Unlabeled Machine Learning

Aug 5, 2026

A positive and unlabeled machine learning (PU learning) model was trained to predict substrate reactivity in oxidative homocoupling of phenols under varied conditions. Validation used two descriptor sets: 28-dimensional descriptors linked to reactivity factors and extended-connectivity fingerprints. Model parameters were tuned using experimental data, achieving strong prediction accuracy on existing data across conditions. For 30 unlabeled datasets, predictions matched experiments for about 83.3–86.7% of substrates, outperforming a model trained with both positive and negative reactivity labels.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

AccessFree to download

Abstract

Introduction

Machine learning for reaction outcome prediction

PU learning as an approach for small labeled datasets

Figures and workflow

Data collection and descriptor calculation

Model training and reactivity prediction

This article is licensed under CC-BY-NC-ND 4.0 [http://pubs.acs.org/journal/acsodf](http://pubs.acs.org/journal/acsodf) Article Predicting Substrate Reactivity in Oxidative Homocoupling of Phenols Using Positive and Unlabeled Machine Learning Takafumi Nishii, Kaname Ichizawa, Haruka Nagano, Hiroya Mukai, Daimon Sakaguchi, and Hiroaki Gotoh * Cite This: ACS Omega 2025, 10, 49805−49815 Read Online Downloaded via YOKOHAMA NATL UNIV on December 16, 2025 at 03:54:03 (UTC) . See [https://pubs.acs.org/sharingguidelines](https://pubs.acs.org/sharingguidelines) for options on how to legitimately share published articles. ACCESS Metrics & More Article Recommendations *sı Supporting Information ABSTRACT: A positive and unlabeled machine learning (PU learning) model was trained to predict substrate reactivity in the oxidative homocoupling of phenols under different conditions. We demonstrated its effectiveness by conducting validation using two descriptor sets: 28-dimensional descriptors considered to influence reactivity and extended-connectivity fingerprints. We performed parameter tuning of the model using our experimental data and determined that the optimized parameters provided excellent prediction accuracy for the existing experimental data, regardless of the reaction conditions. Furthermore, the prediction results obtained using 30 types of unlabeled data matched the experimental results for approximately 83.3−86.7% of substrates, and the prediction accuracy of the PU learning model was shown to be superior to that of a model trained with both positive and negative reactivity data. Because negative data are not required to train a PU learning model, it can be applied to reactions reported in many previous studies, informing the cost-effective synthesis of molecules based on model-predicted results. ■ INTRODUCTION Many researchers have utilized machine learning to predict the outcomes of organic synthesis reactions. These approaches typically collected considerable quantities of experimental data using high-throughput experimentation (HTE) or automated flow synthesis. 1−4 Ahneman et al. 1 focused on the Buchwald − Hartwig cross-coupling of aryl halides and 4-methylaniline to construct a random forest model using 4608 experimental data points obtained from varying combinations of substrates, bases, and additives through ultra-HTE; this model predicted the test yield with an accuracy of R2 = 0.92. Nielsen et al.2 focused on the deoxygenative fluorination of alcohols to construct a random forest model using 640 experimental data points obtained from varying combinations of substrates, bases, and sulfonyl fluorides through HTE; this model predicted the test yield with a root mean squared error (RMSE) of 7.4%. Furthermore, Granda et al.3 constructed a fully connected neural network model using 5,760 experimental data points collected from Suzuki−Miyaura couplings by an automated flow synthesis platform developed by Perera et al.;5 they reported a predicted test yield RMSE of 11%. Yarish et al.4 developed descriptors and constructed a graph neural network that predicted the test yield from Ahneman et al. 1 with an accuracy of R2 = 0.93 and that from Perera et al.5 with an RMSE of 10.35%. However, as HTE and automated flow synthesis have not been implemented widely,6 a method providing individual organic chemists with the ability to predict the feasibility of reactions using untested substrates based on a small quantity of experimental data could be beneficial. This approach would be particularly useful for identifying substrates that lead to highly synthesizable bioactive molecules or reducing costs by predicting the reactivities of expensive substrates before expending them. Positive and unlabeled machine learning (PU learning) offers a promising approach for achieving predictions from small quantities of experimental data.7 Figure 1a illustrates how PU learning identifies positive and negative r

FAQs

What machine learning approach is used to predict phenol reactivity?

The study uses positive and unlabeled (PU) learning, which learns from positive and unlabeled data rather than requiring explicit negative examples.

How was the model validated in the study?

Validation was performed using two descriptor sets—28-dimensional descriptors and extended-connectivity fingerprints—followed by parameter tuning on experimental data.

How accurate were the PU learning predictions on unlabeled data?

Predictions on 30 unlabeled datasets agreed with experimental results for approximately 83.3–86.7% of substrates, and the PU model outperformed a model trained with both positive and negative data.

Aug 5, 2026

Upload by Valentina