AI Predicts Borylation Sites on Drug Molecules: ETH & Roche's Active Learning Loop

Background: Predicting Borylation Sites on Drug Molecules

Designing a molecule with potential pharmacological activity is only the starting point of drug development. Whether a candidate molecule can be synthesized and where chemical reactions occur on the molecule determines its progress. In lead optimization, researchers often modify existing molecular scaffolds to rapidly generate series of structurally similar candidates with different properties. C–H borylation converts carbon–hydrogen bonds into boron-containing functional groups, providing handles for subsequent cross‑coupling reactions. However, for complex molecules with multiple similar C–H bonds, predicting whether borylation will occur and at which position remains challenging.

Closed‑Loop Active Learning Framework

The study, published in Nature Computational Science (October 9, 2024) under the title "Advancing chemical reaction prediction in data-scarce drug discovery with active and geometric deep learning," combines active learning with geometric graph neural networks (GNNs) to build a closed loop where the model selects experiments and experimental results update the model. The work comes from ETH Zurich and the Roche Innovation Center.

Initial Data and Binary Reaction Classification

The team curated an initial dataset covering 518 substrates with 2,044 high‑throughput experimentation (HTE) data points. They first trained an XGBoost ensemble model on ECFP4 fingerprints to predict only whether a substrate undergoes borylation under the tested conditions — a binary success/failure task — rather than directly predicting yield or regioselectivity. XGBoost was chosen after baseline tests against Random Forest and CatBoost showed superior classification performance and uncertainty calibration.

Active Learning for Experimental Selection

From Roche's library of 22,253 drug‑like aromatic compounds, the model used the BALD acquisition function (which accounts for predictive uncertainty) to prioritize substrates that would expand the model's knowledge into under‑explored chemical space. Three rounds of experimentation selected 50 new substrates, generating 4,821 additional reaction results. Combined with the initial data, the final dataset comprised 6,865 reaction outcomes across 568 unique substrates.

Geometric Graph Neural Network for Regioselectivity Prediction

With the expanded dataset, the team moved to a geometric GNN that represents molecules as graphs of atoms and bonds, incorporating 3D atomic coordinates to capture spatial structure. The architecture is designed to be equivariant to rotations and translations, preserving reaction‑relevant geometric information.

Self‑Supervised Auxiliary Tasks

Two self‑supervised tasks were added to provide extra training signals without requiring more labeled reaction data:

Node masking: The model predicts the atom type of masked nodes from surrounding context.

Coordinate denoising: The model recovers the original 3D coordinates from perturbed positions.

Performance on Scaffold Splits

As active learning enlarged the dataset, the geometric GNN's performance improved. On the more challenging scaffold split, the final model achieved a Matthews correlation coefficient (MCC) averaging 0.43–0.50, while the best XGBoost model reached 0.35 ± 0.12.

Practical Impact and Limitations

The workflow strings together active‑learning‑driven experiment selection, data accumulation, and geometric deep learning for site prediction. The lighter XGBoost model identifies informative experiments; the richer dataset then trains a more expressive geometric GNN for regioselectivity. The authors emphasize that the current work focuses on C–H borylation and does not claim to predict yields for arbitrary reactions or replace full synthetic planning. Instead, it demonstrates a practical paradigm: models guide experimental resource allocation, and real experiments continuously refine the model's understanding of chemical reactivity.

Paper link: https://www.nature.com/articles/s43588-026-01063-0

Code example

来源：ScienceAI
本文
约1800字
，建议阅读
5
分钟
几何深度学习帮助模型识别复杂分子的反应选择性。