In lab tests, a model predicts which site of a molecule reacts

Chemists who want to know where a molecule will react have long relied on product yield, the share of starting material that ends up as the intended product after several synthesis steps. The measure is blurred by purification and by side reactions, and it changes from one synthesis scenario to another. It also says nothing about which of several reactive groups actually reacted.

At Yokohama National University, a team led by associate professor Hiroaki Gotoh and doctoral student Daimon Sakaguchi chose a simpler route. They mixed complex molecules containing several ketone groups, which are oxygen and carbon atoms joined by a double bond, with different amounts of sodium borohydride in methanol. By measuring which ketone group reacted first, they built a kinetic scale, or reaction-rate map, for each molecule. The team then combined this map with three-dimensional electronic information taken from around each reacting site, and fed both into a machine-learning model.

The model was trained only on simple molecules. Tested on six more complex molecules, each with two ketone groups, it correctly predicted both the main reaction site and the side from which the reaction occurred. The study, published in JACS Au, shows that competition experiments can link laboratory measurements, quantum-chemical calculations and machine learning in a single workflow. The authors stress that these are laboratory results on a defined reaction, with sodium borohydride in methanol.

For a chemist designing a drug candidate, the practical gain would be knowing, before synthesis, which part of a molecule is likely to react and which product will form. Gotoh names drug candidates, natural products and functional molecules as the target, and frames the aim as fewer trial-and-error cycles and faster discovery of new molecules and synthetic routes. The team does not say when such predictions could be used in practice, nor whether the model holds for reactions other than this one. What it does offer is a curated kinetic database and a descriptor workflow that other researchers can reuse to build new models and test their hypotheses on more complex reductions.