Sparse reaction data gets a closed loop: active learning picks the experiments, and geometric GNNs call every borylation site on unseen substrates
Synopsis
Late-stage functionalization in medicinal chemistry is limited by scarce reaction data, especially when several similar C–H bonds compete for borylation. A Roche and ETH team used an XGBoost-ensemble active learning loop to pick 50 previously unreported substrates from a 22,253-molecule pool across three rounds of high-throughput experiments, expanding C–H borylation data to 6,865 reactions over 568 substrates. Geometric graph neural networks trained on that set beat XGBoost in most cases on random, Butina and Bemis–Murcko scaffold splits, and in a prospective test on ten unseen substrates with N-heteroaryl motifs the top predicted site matched experiment in all cases.
Interpretation
Training data for reaction prediction now comes from experiments the model chooses, not only from curated literature. Earlier active learning mostly chased high-scoring molecules in retrospective screening data, while reaction optimization needs coverage of the reaction landscape with atom-level labels, a potential left largely untapped. On the initial 518-substrate C–H borylation data, the XGBoost ensemble had the highest average MCC, balanced accuracy and F1 across ten random seeds and hyperparameter configurations, plus the best expected calibration error, so it served as the active learning oracle.
Three prospective rounds turned candidate-pool molecules into trainable reaction records. The initial dataset held only 626 regioselectivity molecules and 2,044 high-throughput reaction points across 518 substrates. The three rounds tested roughly 96 reaction configurations on 50 previously unreported substrates, adding 4,821 reactions for a combined 6,865 reactions over 568 unique substrates; scoring the full 22,253-molecule pool takes about 0.1 s.
Geometric graph neural networks hold up better than fingerprint models on out-of-distribution splits. XGBoost trained on ECFP4 fingerprints alone performed poorly across all splits, showing that reaction conditions must enter the model. On the scaffold split of the final active learning round, the GNNs reached a mean MCC between 0.43 and 0.50, while the best XGBoost with categorical and numerical conditions reached 0.35 ± 0.12.
Regioselectivity can be predicted atom by atom and fed directly into virtual library design. Predicting the borylation position on drug-like molecules with several similarly reactive C–H bonds had been a formidable challenge. The selected EquiformerV2 with coordinate denoising matched the most probable predicted site to the observed site on all ten unreported substrates, with confidences of 0.98 and 0.83 for compounds 2 and 3.
Perspective
The demonstrated scope is iridium-catalysed C–H borylation: 6,865 reactions over 568 substrates, plus prospective validation on ten unseen substrates carrying motifs such as N-heteroaryls; the same model panel was also benchmarked on Minisci C–H alkylation with 13,490 reactions over 80 heterocyclic substrates. The intended readers are computational and medicinal chemists working on reaction prediction and late-stage functionalization.
The active learning objective is binary reaction feasibility, only indirectly coupled to regioselectivity, because regioselectivity labels require successful borylation and scale-up characterization first; the authors leave a multitask oracle targeting both feasibility and regioselectivity uncertainty to future work. They also note that at their experimental data scale Gaussian processes and deep ensembles would be computationally tractable, so choosing XGBoost is a trade-off favouring simplicity and calibration rather than a claim of universal optimality; EquiformerV2 showed reduced accuracy on the most structurally intricate substrates, and predicted probabilities are not calibrated to yield.