TL;DR: On Oct. 9, 2026, ETH Zurich and Roche reported a closed-loop workflow that used active learning to choose experiments and geometric deep learning to predict C–H borylation outcomes and sites.
A loop that picks experiments, then predicts reactions
A team from ETH Zurich and Roche Pharma Research and Early Development published an integrated closed-loop workflow for data-scarce reaction prediction on Oct. 9, 2026, according to Nature Computational Science. The paper appeared as an open-access article in Nature Computational Science with DOI 10.1038/s43588-026-01063-0. The authors were Mason Minot, Yannick Stenzhorn, Jens Wolfard, Sebastian Strobel, Philippe Jablonski, Daniel Zimmerli, Martin Binder, Uwe Grether, Rainer E. Martin, Alex T. Müller, David F. Nippa, Kenneth Atz and Gisbert Schneider. Affiliations listed ETH Zurich Department of Biosystems Science and Engineering in Basel and Roche Pharma Research and Early Development at Roche Innovation Center Basel. The manuscript history listed receipt on Sept. 23, 2025, acceptance on Sept. 1, 2026, and first online publication on Oct. 9, 2026.
The study addressed generative molecular design and synthetic accessibility, according to Nature Computational Science. The abstract stated that computational models for reaction prediction are limited by lack of reference data and difficulty predicting regioselectivity. The team presented a workflow combining efficient data acquisition with high-accuracy deep learning. A tree-based ensemble guided acquisition of laboratory experiments. The campaign produced a benchmark for C–H borylation. The dataset supported training and evaluation of geometric graph neural networks. Augmentation with self-supervised auxiliary tasks improved outcome and regioselectivity prediction. Prospective tests on unseen substrates with N-heteroaryl motifs pinpointed correct borylation positions in all cases.
According to Bioengineer.org, the work was framed as a response to synthesis as a bottleneck in drug discovery. The outlet summarized the result as correctly identifying sites on all ten prospectively tested substrates. The collaboration was described as ETH Zurich working with Roche Pharma Research and Early Development. The reaction focus was iridium-catalyzed C–H borylation for late-stage functionalization.
Why synthesis, not design, set the limit
The paper argued that rapid identification and optimization of pharmacologically active molecules are increasingly limited by cost and speed of experimental data generation rather than molecular design algorithms, according to Nature Computational Science. Early hit-finding and multiparameter optimization face time and material constraints in choosing reactions, substrates and conditions. High-throughput experimentation partially alleviates the bottleneck by producing findable, accessible, interoperable and reusable reaction data, known as FAIR data. Experimental bandwidth must still be allocated judiciously.
Active learning, in which a model selects the next most informative experiments, was presented as a remedy that improved sample efficiency in hit discovery and multiparameter optimization. Most prior approaches targeted rapid identification of high-scoring molecules from retrospective screening data or computational oracles. Chemical reaction optimization posed a different goal: mapping the reaction landscape, including negative outcomes, to build models that generalize to unseen substrates. It also demanded atom-level annotations such as regioselectivity that require experimental characterization in a closed loop between prediction and laboratory work. The authors described that potential as largely untapped.
Late-stage functionalization generates analogs by installing functional groups on fully elaborated scaffolds without designing new synthetic routes from scratch, according to Bioengineer.org. That approach allows chemists to probe structure-activity relationships and adjust absorption, distribution, metabolism, excretion and toxicity profiles on accelerated timelines. C–H borylation was described as attractive because resulting boronate esters serve as versatile handles for downstream cross-coupling transformations. The same coverage noted that drug-like molecules often carry multiple similarly reactive C–H bonds, making prediction of which bond reacts a formidable challenge.
How the team split fast selection from accurate prediction
The workflow assigned two distinct roles to two model families, according to Nature Computational Science. A tree-based ensemble served as the selector for iterative active learning. Geometric graph neural networks served as production models once a suitable dataset had accumulated. The stated rationale was that data collection rewards rapid iteration, while a finished dataset rewards maximum predictive performance.
Deep neural ensembles estimate epistemic uncertainty by training several independent models and rank among the strongest baselines for active learning, but their accuracy carries steep computational cost that parameter-sharing and low-rank refinements only partially reduce. That overhead was described as untenable where the surrogate is retrained after every iteration and evaluated on millions of candidates to select points for labeling. Tree-based methods remain competitive with deep learning at a fraction of the compute budget, making them suited as an oracle for iterative active learning in data-scarce medicinal chemistry.
Geometric graph neural networks operate directly on three-dimensional molecular graphs, encoding translational, rotational and permutational symmetries through architectural constraints. Such symmetry-aware networks have delivered state-of-the-art performance across quantum property prediction, protein-ligand binding and increasingly reaction outcome and selectivity tasks. Self-supervised learning was introduced to extract transferable chemical priors that reduce labeled-data burden. Typical schemes mask or perturb atomic identities, bond types or 3D coordinates and task the model with reconstructing input or predicting added noise, applied either as online auxiliary objectives or offline pretraining.
Choosing the selector: XGBoost and BALD
The team benchmarked ensembles of Random Forest, CatBoost and XGBoost classifiers as the decision-making oracle, according to Bioengineer.org. Deep neural ensembles were judged computationally prohibitive for retraining after every iteration and scoring millions of candidates. Tree-based methods proved competitive at lower cost. The XGBoost ensemble delivered the best combination of classification performance and uncertainty calibration. It scored the full pool of 22,253 drug-like aromatic compounds in roughly 0.1 seconds.
The team then benchmarked ten acquisition functions, the mathematical rules that determine which molecules an active learner requests next. Bayesian Active Learning by Disagreement, known as BALD, consistently emerged as the top performer. Its explorative character prioritized structurally diverse substrates rather than repeated sampling of molecules similar to those already tested. That behavior broadened scaffold coverage in training data.
The candidate pool contained 22,253 drug-like aromatic compounds with 6,547 unique scaffolds, according to detailed coverage of the study. Across three prospective rounds, the system selected 50 previously unreported substrates. Laboratory teams tested approximately 96 reaction configurations on almost every selected substrate. Round one included 30 substrates, followed by ten substrates in each of rounds two and three. The effort generated 4,821 new reactions and expanded the dataset to 6,865 reactions spanning 568 unique substrates for yield and binary success prediction.
What 4,821 new reactions added
Quantitative analysis confirmed explorative behavior, according to Bioengineer.org. Mean Tanimoto distance between selected substrates and nearest training-set neighbors was 0.69. The campaign added 42 previously unseen Bemis-Murcko scaffolds. Scaffold diversity expanded from 105 to 147 unique frameworks, a 40 percent increase. Coverage of the candidate chemical space increased from 0.87 percent to 1.34 percent of scaffolds, leaving most structural families untested. In the first active-learning round, selections covered all six clusters used to divide candidate chemical space.
The final public resource contained 6,865 high-throughput reaction outcomes across 568 substrates, according to detailed reporting on the paper. It also contained atom-level regioselectivity information for 812 substrates, including 106 negative examples, and 920 borylated species. Only 24 percent of new reactions met the study's positive threshold of at least 5 percent yield. Negative results were retained because mapping failure teaches a classifier where reactions do not work. Datasets were released through Zenodo under DOI 10.5281/zenodo.20773622, while code and model weights were released through Zenodo under DOI 10.5281/zenodo.20783136.
Human chemists ran chosen experiments, characterized products and returned results to the dataset. Trained personnel handled air- and moisture-sensitive reagents, heated reactions and performed scale-up and structural characterization before labels could be trusted. The loop was repeated three times. That structure made the work model-guided experimental design rather than a laboratory operating without people.
From rapid iteration to 3D models
Once the active-learning campaign enriched the dataset, the focus shifted from rapid iteration to predictive performance, according to Nature Computational Science. Researchers transitioned to geometric graph neural networks operating on three-dimensional molecular graphs. The team benchmarked ten architectures, including invariant networks such as SchNet, equivariant networks such as PaiNN, Ponita and EquiformerV2, and unconstrained designs such as FAENet, against XGBoost baselines.
Evaluation used three increasingly challenging splits: random, Butina clustering and Bemis-Murcko scaffold. Random splits allow training and test molecules to be structurally similar. Cluster and scaffold splits hold out broader molecular families and test generalization to unseen series. On random splits, the best graph network and a condition-aware XGBoost model performed comparably. On harder scaffold splits, graph networks pulled ahead.
On the final scaffold split for binary reaction success, graph models produced mean Matthews correlation coefficients between 0.43 and 0.50, while the best condition-aware XGBoost model reached 0.35 with a standard deviation of 0.12. For reaction-site prediction, the best condition-aware graph model also outperformed the T5Chem language-model comparator on harder splits. One Ponita configuration reached an MCC of 0.59 with a 0.05 standard deviation on the scaffold split, compared with 0.38 with a 0.08 deviation for T5Chem. Adding self-supervised tasks — masking atom identities and reconstructing perturbed coordinates — usually helped, especially on unfamiliar scaffolds.
Ten unseen molecules, ten correct sites
The prospective test used unseen substrates featuring challenging N-heteroaryl motifs, according to Nature Computational Science. The abstract reported that tests pinpointed correct borylation positions in all cases. Detailed coverage specified ten prospectively tested substrates. The highest-ranked site from the graph model matched the experimentally observed site in all ten.
The substrates were distinct from training chemistry rather than close copies. The result illustrated practical utility for drug optimization within the tested workflow. It did not establish reliability across other reactions, laboratories or drug programs because ten examples are too few for broad generalization. The study also demonstrated performance on a single reaction class, C–H borylation, and did not show how the loop behaves where selectivity follows other rules.
Open release of data, code and weights gives independent groups a route to check reported splits and build comparisons rather than accepting a company-only benchmark. The paper and repository release the data, while a separate Zenodo record provides code and model weights. That openness matters because reaction models have been held back by scarce data and by difficulty in predicting regioselectivity. A generated molecule is only worth proposing if it can be made, so synthesizability screening remains tied to experimental validation.
Who did the work and where it stands
The collaboration linked Gisbert Schneider's group at ETH Zurich with Roche laboratories, according to author statements indexed by LinkedIn on Oct. 9, 2026. The ETH side built models while Roche laboratories generated experimental data and jointly shaped data structure. The program included ETH postdocs embedded in Roche laboratories. Yannick Stenzhorn was identified as a first author, with Mason Minot as co-first author. David F. Nippa, a Roche author, described the collaboration as active since 2022.
Competing interests disclosed that Gisbert Schneider is a co-founder of inSili.com in Zurich and Xanadys in Zurich and a scientific consultant to the pharmaceutical industry. Yannick Stenzhorn, Jens Wolfard, Sebastian Strobel, Philippe Jablonski, Daniel Zimmerli, Martin Binder, Uwe Grether, Rainer E. Martin, Alex T. Müller, David F. Nippa and Kenneth Atz were listed as full employees of F. Hoffmann-La Roche. Other authors declared no competing interests. Text and data mining and version of record were valid from Oct. 9, 2026.
The stated goal is to map the whole reaction landscape, negative results included, so models generalize to substrates not yet run. Whether the loop holds on reaction classes beyond C–H borylation remains to be tested. Whether chemists wire the regioselectivity predictor into generative design pipelines to screen proposed molecules for synthesizability is a next integration question. A published dataset size and numeric baselines now allow alternative methods to be sized against the reported scaffold-split scores.