Can active and geometric deep learning improve reaction prediction in data-scarce drug discovery?

Two bottlenecks block reaction prediction

The paper starts from a practical constraint in generative molecular design: a proposed molecule is only useful if it can actually be synthesized. Computational reaction models could help, but the authors identify two linked limits. First, there is too little reference data on the reactions that matter for drug-like molecules. Second, predicting regioselectivity — which specific atom in a molecule reacts — is difficult when a scaffold carries several similarly reactive C–H bonds.

The chosen reaction is C–H borylation, a late-stage functionalization step that converts a C–H bond into a boronate ester. That ester then serves as a handle for downstream cross-coupling, so a single intermediate can be diversified into many analogs. The value of predicting the site correctly is therefore direct: it determines which analog a chemist can build. The paper treats data scarcity and site-level accuracy as one problem, not two.

Active learning chooses the next experiments

Instead of screening compounds at random, the workflow runs three rounds of a closed loop: train a model, let it select candidate substrates, run the laboratory experiments, then update the training data. The selection model is a tree-based ensemble rather than a deep neural network. The authors justify this on computational grounds: the surrogate is retrained after every round and scored against millions of candidate molecules, and tree ensembles remain competitive at a fraction of the compute cost of deep ensembles.

The acquisition function matters as much as the model class. The paper reports retrospective benchmarking of ten acquisition functions using an XGBoost ensemble for condition-agnostic prediction, comparing predictive performance (MCC) against the number of molecules sampled and calibration by expected calibration error. In the prospective campaign, the explorative BALD acquisition function is credited with selecting structurally diverse substrates. The evidence for that diversity is concrete: round 1 sampled all six k-means clusters of a t-SNE space built from GIRAFFE embeddings, the mean Tanimoto distance to the nearest training-set neighbor across rounds 1–3 was 0.69, and 42 previously unseen Bemis–Murcko scaffolds were added, raising unique scaffolds from 105 to 147, a 40% increase.

What the campaign added to the data

The starting point was 626 regioselectivity data points and 2,044 high-throughput experimentation reactions across 518 substrates for binary classification and yield. The three active-learning rounds added 30 new substrates and 186 new molecules, and the final set contained 812 starting materials and 920 borylated products. Of the previously unreported substrates, 80 were positives and 106 were negatives, arising from 1,030 reactions, 348 of which had not been reported before.

The composition changed in ways that matter for modeling. Relative to the initial literature dataset, the expanded set has more negatives (106 versus 0) and more molecules with multiple potential borylation sites (172 versus 147 for two sites, 9 versus 4 for three sites, and 3 versus 1 for four sites). A dataset with no negatives cannot teach a model what failure looks like, and a dataset dominated by single-site substrates cannot teach site discrimination. The final dataset used for analysis contained 6,865 reactions, 4,821 of them from active-learning rounds, with positives defined as yield ≥5% and negatives as yield <5%.

Geometric GNNs win on hard splits

With a larger dataset, the authors moved to geometric graph neural networks, which operate directly on three-dimensional molecular graphs and encode translational, rotational and permutational symmetries through architectural constraints. They benchmarked ten geometric GNNs against several XGBoost encodings across three splits of increasing difficulty: Random, Butina clustering and Bemis–Murcko scaffold. The splits answer different questions. A Random split tests interpolation within familiar chemistry; a Scaffold split tests generalization to unseen molecular scaffolds.

The result depends on which split is used, and this is the comparison readers should hold onto. On the Random split, the best GNN and an XGBoost model using categorical and numerical reaction conditions performed comparably. On the harder Butina and Scaffold splits, GNNs outperformed XGBoost. In the Scaffold split for the final active-learning round, GNNs showed a mean MCC between 0.43 and 0.50, while the best XGBoost model reached an MCC of 0.35 ± 0.12. An XGBoost model trained only on substrate fingerprints performed poorly across all splits, which the authors read as evidence that reaction conditions are essential for accurate prediction.

Two negative findings are equally part of the result. The authors observed no systematic performance difference between GNNs with different internal coordinate systems or symmetry constraints, and no clear advantage for equivariant over invariant models. What did help consistently was adding optional self-supervised auxiliary tasks during training — coordinate denoising and a node-masking strategy — which provided a general performance benefit across models.

Regioselectivity is an atom-level task

Predicting whether a reaction succeeds and predicting where it happens are different problems, and the paper keeps them separate. Regioselectivity was framed as node-level binary classification of whether each sp2 and sp3 carbon is borylated, in a reaction condition-agnostic manner. This is a stricter test than reaction outcome: the model must rank competing sites within the same molecule rather than assign one label to the whole substrate.

The authors are explicit that the active-learning oracle was optimized for binary reaction feasibility, not for regioselectivity, so the two are coupled only indirectly. Their justification is procedural: regioselectivity annotations require a successful borylation followed by scale-up synthesis and structural characterization, none of which exists before substrates are selected. A binary feasibility filter is therefore a prerequisite rather than a substitute. They note that a multitask oracle jointly targeting feasibility and regioselectivity uncertainty would address the downstream objective more directly but at substantially higher computational cost, and leave it to future work.

Prospective tests and a second reaction

The most decision-relevant evidence is prospective rather than retrospective. Tests on unseen substrates featuring challenging N-heteroaryl motifs pinpointed the correct borylation positions in all cases. The authors also ran an in silico late-stage functionalization exercise, taking the top-predicted borylation site of ten initial substrates and diversifying the virtual products through Suzuki–Miyaura coupling against a curated library of aryl-halide building blocks; t-SNE visualizations of the GIRAFFE-embedded products indicated access to a wide range of chemical diversity.

To test whether the benchmarking pattern generalizes beyond borylation, the authors evaluated the same GNN and XGBoost panel on Minisci C–H alkylation, a mechanistically distinct radical-mediated reaction, using 13,490 reactions on 80 heterocyclic substrates with 59 alkyl radical sources. The split-dependent pattern repeated: GNNs and XGBoost with reaction conditions were comparable on the Random split, GNNs outperformed XGBoost on the Scaffold split, and the same held for continuous yield prediction by Spearman correlation. Self-supervised auxiliary tasks again helped across architectures, with the largest improvements on the Scaffold split, and again no single architecture proved systematically superior.

Where the advantage stops

The paper's own limitations sharpen the interpretation. The authors state that at the scale of their experimental dataset, Gaussian processes and deep neural ensembles would also have been computationally tractable and could offer complementary advantages such as principled Bayesian posteriors; choosing XGBoost reflects a trade-off favoring simplicity and competitive uncertainty calibration, not a claim of universal optimality. The models would benefit from larger and more diverse public reaction datasets, particularly to address class imbalance in current binary datasets and to include a higher proportion of drug-like molecules, since the EquiformerV2 model showed reduced accuracy on the most structurally intricate substrates.

Scaffold coverage also remained partial: training-set coverage of the candidate pool's 6,547 unique scaffolds rose from 0.87% to 1.34% over the three rounds. The authors acknowledge that complete coverage is impractical for such a large chemical space and instead point to the prospective validation on unseen scaffolds as evidence that the models generalize within the explored domain. That is a narrower claim than general chemical-space coverage, and it is the claim the data support.

The priority the paper establishes

The paper's most useful conclusion is a ranking of what to invest in. Among the levers tested, architecture choice was the weakest discriminator: invariant, equivariant and unconstrained geometric GNNs showed no systematic differences, and the same held for the mechanistically distinct Minisci reaction. The authors read this as suggesting that, for reaction prediction on drug-like molecules, data quality and training strategy may matter more than architectural expressiveness beyond a baseline level of geometric awareness.

The two levers that did move results were the composition of the training data and the training objective. Active learning improved test MCC with each round in the majority of architecture and split combinations, with the largest gain in the first round, and online self-supervised auxiliary tasks improved performance across most architectures without external unlabeled corpora or a separate pretraining stage. For a team deciding where to spend effort on a data-scarce reaction, the paper's evidence points first to how experiments are selected and how the model is regularized, and only after that to which symmetry-aware architecture is chosen.

Sources used in this answer