Teaching a chemical model where a reaction will start

Two reactions can use the same ingredients and produce different molecules because the reagent acts at a different site or approaches from a different direction. Chemists call the preference between competing pathways selectivity; stereoselectivity concerns the preference between products with different spatial arrangements. The practical question is which product the reaction will favour.[2][3]

Yokohama National University researchers have trained a model to predict the dominant site and approach direction for a particular ketone reduction. Their JACS Au paper appeared on 14 September 2026. A ketone contains a carbon–oxygen double bond; a diketone contains two, giving the reagent competing sites.[1]

Let the molecules compete

The experimental idea has a useful precedent. In 2022, University of Victoria researchers studied a different reaction, nucleophilic aromatic substitution, by making two starting compounds compete for one reaction partner. Measuring how much of each starting compound remained let them infer relative reaction rates without calibrating the analytical response of every possible new product. They used those rates to build a model from calculated electronic properties.[4]

Competition therefore helps turn an experimental bottleneck into a measurement design: compare starting materials that can already be measured, then connect their behaviour to molecular structure. It also requires care when interpreting the numbers. Consumption and reaction speed are different quantities.

Consider our hypothetical example under the competition model’s shared-reagent assumptions. Compound A finishes with half its initial amount; B finishes with a quarter. Using the logarithms of those remaining fractions gives ln(¼) ÷ ln(½) = 2, so B has twice A’s rate constant. Simply comparing the amounts consumed would give 75% ÷ 50% = 1.5. That tempting calculation answers the wrong question: the amount available to react changes as the reaction proceeds.[4]

From relative speed to a preferred site

The Yokohama experiments paired ketones competing for sodium borohydride in methanol. Measurements of the remaining molecules established relative rates. The team combined 83 monoketone reaction faces with three-dimensional electronic descriptors and trained a Lasso regression model. The descriptors distinguish the local electronic environment around each potential approach.[1]

Across six diketones—one new experiment and five literature cases—the model identified the dominant initial site and face each time. Product proportions remained approximate. Other reagents or solvents would require new data and validation. The demonstrated gain is a reaction-specific ranking of the competing routes through a molecule.[1]

Report an error or suggest a correction