A chemist who wants to improve a complex molecule faces an awkward economy. Rebuilding the entire structure to change one small part can be laborious. Late-stage functionalisation offers another route: modify a molecule that already exists. C–H borylation replaces a hydrogen at a carbon site with a boron-containing group that can serve as a handle for further chemistry. The difficulty is getting that handle into the intended position.[2]
An ETH Zurich–Roche study published on October 9 combines active learning with geometric neural networks. After three experimental rounds, its dataset held 6,865 reaction records on 568 substrates; a separate collection for site prediction covered 812 substrates.[1]
The word “site” carries much of the practical problem. A drug-like molecule can contain several carbon–hydrogen bonds surrounded by different chemical groups. Some positions are physically crowded; others have a different electronic environment. The researchers’ earlier work combined small-volume parallel reactions with molecular graphs, including three-dimensional information, to study how these differences affect reactivity and selectivity.[2]
Think of the intended modification as an address. Attaching a useful group somewhere on a molecule does not necessarily create the particular analogue a chemist wants. A forecast that says a reaction will occur and a forecast that identifies its likely address answer related, but distinct, questions.
The experiment that resolves a disagreement
Active learning adds another decision: which reaction should the laboratory run next? The original Bayesian Active Learning by Disagreement method frames this as an information problem. A useful query is one whose answer would resolve disagreement among plausible models. This differs from selecting the example most likely to succeed, and from selecting an outcome that is simply noisy and difficult to predict.[4]
Imagine two proposed experiments. For one, every plausible model expects a good yield. For the other, half expect a reaction and half expect very little product. If the immediate aim is to obtain material, the first may be the sensible choice. If the aim is to learn which explanation of the chemistry is right, the second may be more revealing. This is an illustration of the selection logic, not an experiment reported by the team.
In the new study, an XGBoost ensemble guided experimental selection; geometric neural networks then learned site predictions from the expanded evidence. The two model families had different jobs.[1]
The released data preserve that experimental history in separate files for the starting collection and three active-learning rounds, alongside a regioselectivity file. Keeping those stages distinguishable makes the growing evidence base available for inspection and reuse.[3]
Which bond, and how much product?
In prospective tests on ten substrates not previously reported for this borylation, the highest-ranked site matched experiment in all ten. Those site probabilities were not calibrated as product yields. Across the larger reaction dataset, only 24% of entries met the study’s positive-outcome threshold of at least 5% yield. These are different measurements, with different denominators.[1]
A recorded negative result therefore means that a particular combination did not clear a defined threshold. It does not establish that the molecule can never be modified. A different catalyst, solvent or other condition creates a different experimental question. Recording the conditions is essential to making a negative result informative.
The prospective sample is small, and the authors report weaker performance on the most complex molecular structures. The useful outcome is a better-supported list of experiments to attempt, with laboratory confirmation still determining the products.[1]
For a chemist with a valuable compound and several possible places to alter it, this is a concrete kind of assistance. The model can help decide where to spend the next vial of material. Just as consequentially, the experiment that follows can improve the next decision—even when it produces less material than hoped.
Sources & further reading
- Advancing chemical reaction prediction in data-scarce drug discovery with active and geometric deep learning Nature Computational Science · 9 Oct 2026 · Accessed 17 Oct 2026
- Enabling late-stage drug diversification by high-throughput experimentation with geometric deep learning Nature Chemistry · 23 Nov 2023 · Accessed 17 Oct 2026
- Advancing chemical reaction prediction in data-scarce drug discovery with active and geometric deep learning: dataset Zenodo · 20 Jun 2026 · Accessed 17 Oct 2026
- Bayesian Active Learning for Classification and Preference Learning arXiv / University of Cambridge · 24 Dec 2011 · Accessed 17 Oct 2026