ETH and Roche Train AI to Predict Where C–H Borylation Will Occur on a Molecule

Designing a molecule with the right pharmacological activity is only the start. What follows are questions that look simpler but cost far more in practice: can the molecule be synthesized, and exactly where will the reaction take place? Researchers at ETH Zurich and Roche have published a paper in a Nature-family journal in which AI predicts C–H borylation sites, that is, the position where a boron-containing group will attach. The available portion of the material contains no quantitative metrics for the model, so what follows covers the problem setup and its practical significance.

Why regioselectivity can make or break a candidate

At the lead optimization stage, chemists rarely build a molecule from scratch. More often they take an existing scaffold and modify it locally: add a substituent, shift a functional group, then check how activity, solubility and metabolic stability change. This produces a series of close analogs with different properties in a short time.

In this scheme, C–H borylation works as a convenient "handle." A carbon–hydrogen bond becomes a carbon–boron bond, and the resulting organoboron derivative is then easy to build on: it can be used to form new C–C, C–N or C–O bonds. The problem is that a complex molecule has dozens of C–H bonds, and the reaction can occur at any of them. An experienced chemist can often guess the main position, but on large heterocyclic structures with several competing sites, intuition fails. The team then spends weeks on experiments and on separating mixtures of isomers.

What ETH and Roche are doing

The pairing is a logical one: ETH Zurich is strong in computational chemistry and machine learning for molecules, while Roche has real medicinal chemistry problems and experimental data. The model takes a molecule's structure and outputs a prediction of which positions are most likely to undergo borylation. The chemist sees this before starting the synthesis.

The value of such a tool depends on two things: the quality of the training data and the model's ability to handle molecules unlike those it was trained on. For pharma, the second matters more. Real candidates rarely resemble typical test compounds.

What is known about the work

• Organizations: ETH Zurich and Roche.
• Publication: a Nature-family journal.
• Reaction: C–H borylation.
• Model task: predicting the reaction site on a finished molecule.
• Application stage: lead optimization and late-stage modification.
• Accuracy, dataset size and comparison with baseline methods: not stated in the available portion.

Practical significance

For medicinal chemistry, such a prediction reduces the number of "empty" experiments. If the model confidently names one position, the team can plan the synthesis right away. If it shows two or three equally likely positions, the chemist knows in advance that they will need to tune conditions or use protecting groups.

For concrete numbers, check the original paper. The first things to look at are the share of correctly predicted positions on a held-out set and the model's behavior on structures outside the training distribution. If results there prove stable, such models will become a routine part of the pipeline alongside docking and ADMET prediction. For now, this is one more example of AI moving from generating molecules to a more down-to-earth task that is more valuable to the industry: predicting how molecules will behave in the flask.