In a 2026 study in Nature Computational Science, researchers paired model-guided experiments with geometric deep learning to predict whether drug-like molecules undergo C–H borylation and, for selected molecules, which carbon atoms react. Their results show how a closed-loop workflow can make better use of scarce reaction data—not that AI can replace laboratory validation.
Why is reaction prediction difficult for drug-like molecules?
Drug discovery increasingly uses computational design to propose compounds, but a promising structure is useful only if chemists can make it. Predicting synthesis is especially challenging when a molecule has several similar carbon–hydrogen (C–H) bonds: a reaction may occur at one position, another position, or not at all under the tested conditions.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Drugs: From Discovery to Approval | $59.12 | Buy on Amazon |
| 2 |
|
Basic Principles of Drug Discovery and Development | $268.00 | Buy on Amazon |
| 3 |
|
Textbook of Drug Design and Discovery | $55.19 | Buy on Amazon |
| 4 |
|
Computational Drug Discovery and Design (Methods in Molecular Biology, 2714) | $139.46 | Buy on Amazon |
| 5 |
|
Drugs: From Discovery to Approval | $135.33 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
The study focuses on C–H borylation, a late-stage functionalization that adds a boron-containing group to a molecule. The resulting boronate esters can provide handles for later cross-coupling and compound diversification. The central problem is therefore twofold: estimate reaction feasibility and identify the reacting atom, known as regioselectivity.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallHow does the active-learning workflow acquire data?
Rather than treating data collection and model training as separate stages, the authors used a loop: select experiments with a model, run them in the laboratory, add the results to the dataset, and train outcome-prediction models on the expanded data.
#1 Best Overall
- Start with existing measurements. The active-learning selection process began with data on 518 substrates. The initial binary dataset was imbalanced: 7% of reactions were negative, which the authors identify as a source of performance variability.
- Rank candidate experiments. The team benchmarked Random Forest, CatBoost, and XGBoost ensembles, then used XGBoost as the selection oracle. Its experimental-set classification performance and uncertainty calibration made it a practical choice; the authors do not claim it is universally optimal.
- Test selected molecules. The candidate pool contained 22,253 drug-like aromatic compounds available at Roche. Selection filters included more than 14 heavy atoms and a free aromatic C–H bond. The researchers screened 30 substrates in the first prospective round and 10 in each of the next two.
- Update the training data. Across three rounds, approximately 96 reaction configurations were tested on 50 previously unreported substrates. Those experiments generated 4,821 reactions, which were added to earlier data.
In the authors’ reported yield and binary-outcome dataset, the combined total was 6,865 reaction records across 568 unique substrates. A reaction counted as positive under the binary rule if its yield was at least 5%; 24% of records met that threshold. That percentage describes this dataset and labeling rule, not a general success rate for C–H borylation.
The selection oracle’s immediate objective was binary feasibility, not choosing a specific reacting atom. The authors describe the connection between that acquisition objective and the later atom-level regioselectivity task as indirect.
What did geometric deep learning add?
With the expanded data, the researchers trained geometric graph neural networks (GNNs) to predict reaction feasibility and atom-level regioselectivity. Unlike a fingerprint-only representation, a geometric model can use a molecule’s three-dimensional structure. The comparison included ten geometric GNNs and XGBoost models using molecular fingerprints, with a condition-aware XGBoost model providing the stronger tree-based comparison.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesThe GNNs also used online self-supervised tasks during training: node masking, which asks a model to recover masked molecular information, and coordinate denoising, which asks it to learn from perturbed atomic coordinates. These auxiliary tasks draw learning signals from the labeled reaction data itself; the method does not require a separate unlabeled molecular dataset or an offline pretraining stage.
Rank #3
How well did the models generalize?
The paper evaluated models using random, Butina-clustered, and Bemis–Murcko scaffold splits. A random split can place related compounds in both training and test sets. Cluster and scaffold splits hold out more distinct molecular groups, offering a harder test of whether a model transfers to series or scaffolds it has not seen.
| Evaluation or model comparison | Reported result | What it indicates |
|---|---|---|
| Final active-learning round, Bemis–Murcko scaffold split | Geometric GNN mean MCC: 0.43–0.50; strongest condition-aware XGBoost: 0.35 ± 0.12 | On this paper-specific held-out scaffold test, the GNNs performed better on average. |
| Butina-cluster and scaffold splits | GNNs outperformed XGBoost, according to the authors | The advantage was clearest on these more challenging splits. |
| Random split | Results were more comparable | The apparent gap depends on how compounds are divided between training and test data. |
| Fingerprint-only XGBoost | Performed poorly across the tested splits | Adding reaction conditions strengthened the XGBoost comparison; fingerprint-only results should not be treated as representative of the best tree-based model. |
MCC, or Matthews correlation coefficient, summarizes binary classification performance while accounting for both classes; higher values indicate stronger classification. The reported 0.43–0.50 and 0.35 ± 0.12 values apply to the final active-learning round’s scaffold split, not to every dataset or deployment. The paper reports no single GNN architecture as consistently best, and no systematic advantage based on the tested internal coordinate systems or symmetry constraints. Self-supervised tasks generally improved performance across most tested models, rather than guaranteeing gains in every setting.
For the experiment-selection stage, the authors also report that their XGBoost ensemble scored 22,253 candidate molecules in approximately 0.1 seconds. That is the reported time for this pool and workflow, not a general benchmark across hardware or candidate sets.
Free tools Windows power users keep installed
One-click scans. No signup required.
Can the models identify where borylation occurs?
The regioselectivity dataset included 812 starting materials and 920 borylated products, expanded with products from the active-learning rounds and other Roche borylation experiments. The authors report that prospective tests on unseen substrates with challenging N-heteroaryl motifs identified the correct borylation positions in all cases tested. This is an encouraging result for that reported test set, not evidence of perfect accuracy across drug-like chemical space.
Best Value
The authors also report reduced accuracy for EquiformerV2 on the most structurally intricate substrates. That finding reinforces the need to examine performance by molecular structure and test split rather than relying on one overall score.
What are the study’s limits and practical implications?
- Data balance and diversity matter. The initial binary data had few negative examples, and the authors call for public reaction datasets with greater size and diversity.
- Acquisition and regioselectivity are not the same task. Choosing experiments for feasibility does not directly optimize atom-level site prediction.
- Results depend on the evaluation design. GNN advantages were clearest on cluster and scaffold splits, while random-split results were more comparable.
- No architecture wins everywhere. Model choice remains dependent on the data, molecular structures, and evaluation objective.
- Predictions still need experiments. The work helps prioritize experiments and estimate outcomes; it does not establish that a predicted reaction or site will succeed in a new laboratory setting.
The authors identify the SURF-formatted yield and regioselectivity datasets as Zenodo record 10.5281/zenodo.20773622. Their reference implementation is in the GitHub repository minotm/active-drug-discovery; code and model weights are also identified as Zenodo record 10.5281/zenodo.20783136. The paper states that the reference implementation and weights are released under GPLv3.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
