New transferable machine learning model predicts chemical reactions before they happen

Chemists at Yokohama National University have developed an experimental modeling approach that could dramatically reduce the trial and error that has long defined synthetic chemistry. The method, described in a study published on September 14, 2026, in JACS Au, combines simple competition experiments with three-dimensional electronic data and machine learning to predict not only where a chemical reaction will occur on a complex molecule, but also from which side the reaction will proceed. The work addresses one of the most persistent challenges in organic chemistry: the fact that a molecule’s behavior in one environment may differ dramatically in another, making predictions costly and time-consuming.

The ability to precisely control reaction rates and selectivity is of fundamental importance in organic chemistry, according to corresponding author Hiroaki Gotoh, associate professor in the Faculty of Engineering at Yokohama National University. His stated ultimate goal is to predict where and from which side drug candidates, natural products and functional molecules will react before synthesis is ever attempted. Such foresight would reduce trial and error and accelerate the discovery of new molecules and synthetic routes, a benefit that extends from preventing unwanted medication reactions to improving the efficacy of industrial processes.

The problem with conventional approaches lies in how reactivity is typically measured. Chemists usually assess reactivity by product yields after multistep synthesis processes, a technique that involves purification steps and competing side reactions that all influence the final amount of product. This makes product yield an imprecise measurement, particularly because the same molecule can produce different yields in different synthesis scenarios. Gotoh explained that yield-based evaluations often require optimized conditions, and the numerous parameters that influence reaction yields are frequently difficult to identify and quantify. Consequently, he noted, there is a pressing requirement to establish methods for measuring reaction rates in complex, multistep reaction systems and for analyzing their correlations with molecular structure.

The research team focused its attention on complex molecules containing multiple ketone groups, which are functional units consisting of double-bonded oxygen and carbon atoms that react to reduce a complex molecule and create new molecules. In molecules with more than one ketone group, traditional yield measurements cannot reveal which ketone group actually reacted, nor from which side of that group the reaction occurred. First author Daimon Sakaguchi, a doctoral student in the Gotoh laboratory, said the team wanted to predict both where a ketone reduction would occur and which product would form, while overcoming the fact that conventional yield data do not cleanly measure a molecule’s inherent activity.

The solution came from competition experiments, simple experimental designs that Sakaguchi described as capable of revealing quite a lot of information. By mixing the complex molecules in varying reagent amounts of sodium borohydride in methanol, the researchers could determine which ketone group reacted first. This allowed them to develop a kinetic scale, essentially a reaction-rate map, of the ketone groups within each molecule. The competition format provides comparable reaction-rate data across different substrates, sidestepping the confounding variables that plague yield-based measurements in multistep systems.

Crucially, the kinetic data were then combined with three-dimensional electronic information measured from around each reacting site. This structural descriptor captures how the molecular architecture surrounding each ketone controls its reactivity, providing the machine learning model with a physically meaningful picture of the chemical environment rather than a mere catalog of outcomes. The researchers developed a model to predict chemical reaction rates from this combined dataset, linking laboratory measurements, quantum-chemical calculations and machine learning into a single workflow.

The results exceeded expectations in a way that speaks to the transferability of the approach. The model was trained only on simple molecules, yet it correctly predicted both the main reaction site and the side from which the reaction occurred in all six more complex molecules containing two ketone groups. This transferability is the study’s headline achievement, because a predictive model that works only on the molecules it was trained on offers little value. A model that extends from simple training molecules to more complex, multifunctional targets suggests the underlying kinetic scale and electronic descriptors capture genuine chemical principles rather than dataset-specific patterns.

Sakaguchi emphasized that the reported curated kinetic database and descriptor workflow provide a foundation on which researchers can perform further analyses, construct new predictive models and test mechanistic hypotheses for complex reduction systems. In practical terms, this means other laboratories can adopt the competition experiment format, generate their own comparable rate data, and build on the published database to extend predictions to new reaction classes. The workflow connects experimental measurement with computational chemistry in a way that each component validates and enriches the others.

The implications for drug discovery and industrial chemistry are considerable. Unwanted reactions in pharmaceutical synthesis can render medication candidates ineffective or even harmful, while inefficient industrial processes waste energy, materials and time. A method that predicts reaction outcomes before synthesis could prevent these failures at the design stage, allowing chemists to select the most promising synthetic routes and conditions in advance. The approach may also help chemists understand how chemicals behave when exposed to one another or in different environments, a question that becomes increasingly important as molecules move from laboratory conditions to biological or industrial settings.

The study, titled Competition-Derived Relative Reactivity and 3D Electronic-State Analysis of Site- and Facial Selectivity in NaBH4/MeOH Ketone Reductions, was published in JACS Au on September 14, 2026, with the DOI 10.1021/jacsau.6c01269. Collaborators on the paper include Taisei Kawasaki, Mayu Itakura and Chihiro Tada, all affiliated with the Department of Chemistry and Life Science at Yokohama National University. The research was funded by the Tokyo Ohka Foundation for the Promotion of Science and Technology, the Next Generation AI Human Resource Development Project, the Japan Science and Technology Agency and the Japan Society for the Promotion of Science. As the curated database grows and the workflow is tested across new reaction types, the promise of predicting chemical behavior before synthesis moves closer to routine practice, potentially reshaping how chemists approach the design of molecules and synthetic routes for years to come.

Subject of Research: Chemistry

Article Title: Transferable chemical prediction approach may help reduce trial and error

Article References: Transferable chemical prediction approach may help reduce trial and error. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: Transferable, chemical, prediction, approach, help, reduce, trial, error, scientific research