From the Research Universities edition of October 12, 2026
ETH Zurich and Roche: active learning plus geometric deep learning builds a 6,865-reaction C–H borylation dataset and predicts reaction sites on ten new substrates
ETH Zurich Department of Biosystems Science and Engineering (Schneider lab) with F. Hoffmann-La Roche; published in Nature Computational Science · AI-guided experimental chemistry and drug discovery research (academic–industry) · Switzerland (Basel/Zurich); Roche compound collection
- Publisher
- Nature Computational Science
- Original publication
- Nature Computational Science, open access, published 9 October 2026; ETH D-BSSE news release 12 October 2026
- Source retrieved
- 2026-10-13
What happened
A closed-loop workflow used an XGBoost-ensemble active-learning oracle (BALD acquisition) to choose which Roche compounds to test for iridium-catalysed C–H borylation. Three prospective rounds (30, 10 and 10 new substrates; approximately 96 reaction configurations tested on the 50 substrates) generated 4,821 reactions, bringing the public dataset to 6,865 reactions on 568 substrates, with regioselectivity annotations for 812 substrates (106 negatives) and 920 products. Scaffold diversity rose from 105 to 147 unique scaffolds. Test MCC improved with most active-learning rounds; on the hardest scaffold split, geometric GNNs reached mean MCC 0.43–0.50 versus 0.35 ± 0.12 for the best XGBoost model, and GNNs beat the T5Chem language model on scaffold and Butina splits. In prospective validation on ten previously unreported substrates, the top-ranked predicted borylation site matched experiment in all ten.
Why it matters
A concrete, peer-reviewed example of AI choosing laboratory experiments in a closed loop, with open data, code and weights. Relevant to university chemistry and drug-discovery groups, research computing (GPU training of GNNs) and core facilities considering automated high-throughput experimentation with industry partners.
Evidence and measured results
Opened the Nature Computational Science article via HTTPS (curl; full open-access HTML text extracted and read, including Results, Discussion, data/code availability and competing interests). Also opened the ETH D-BSSE news release (12 Oct 2026) that announced it.
Limitations and uncertainty
One reaction class (C–H borylation; authors say the architecture comparisons also held for Minisci C–H alkylation) and one company's compound pool; prospective regioselectivity test covers only ten substrates; MCC on scaffold split remains moderate (0.43–0.50); the oracle targets feasibility, not regioselectivity, directly; reduced accuracy on the most structurally intricate substrates; 11 of 13 authors are Roche employees and the senior author declares industry ties; no cost or time-saving figures reported in the paper.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-12; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Chemistry and drug-discovery groups have too little reaction data to model new transformations and limited lab budget to generate it.
- Stakeholders
- Medicinal chemistry PIs
- HTE or chemistry core facility director
- Research computing
- Industry partnership/tech transfer office
- Discovery questions
- Which reaction classes lack enough data for reliable prediction?
- Do you have HTE automation and a structured reaction-data format?
- What data-release terms apply to partner compound libraries?
- Value hypothesis
- Active-learning-guided experiment selection may produce more diverse training data per experiment and better generalisation to new scaffolds; savings were not quantified.
- Potential engagement
- Four-week scoping of a closed-loop active-learning pilot on one reaction class using the open code and data as a baseline.
- Unsupported claims
- Guaranteed reduction in lab cost or time
- Generalisation to other reaction classes without new data
- Accuracy beyond ten prospective substrates
Pre-sales engineering
Role takeaway
- Fit
- Groups with HTE capability and a defined reaction class needing feasibility and site-selectivity prediction.
- Architecture and integration
- Fast XGBoost-ensemble oracle with BALD acquisition feeds HTE batches; geometric GNNs with coordinate-denoising and node-masking auxiliary tasks trained once data accumulate (as described in the paper).
- Prerequisites
- HTE robotics and analytics
- Candidate compound pool
- GPU capacity for GNN training
- Structured reaction-data schema
- Constraints
- Single reaction class validated
- Scaffold-split MCC 0.43–0.50
- Oracle optimises feasibility rather than regioselectivity
- Partner compound libraries may be confidential
- Security
- Protect proprietary structures in shared pipelines; public release only after partner approval.
- Proof of value
- Reproduce reported MCC on the open dataset
- Run one prospective round on an in-house reaction
- Compare active versus random selection on diversity and accuracy
Delivery
Role takeaway
- Work
- Set up data capture, oracle training, acquisition, HTE round scheduling and prospective validation for one reaction class.
- Dependencies
- HTE availability
- Compound supply
- Analytical characterisation capacity
- Partner data agreement
- Ownership
- Lab PI with the HTE/chemistry core facility; research computing for training.
- Skills and adoption
- Train chemists to interpret uncertainty estimates and scaffold-split metrics rather than random-split accuracy.
- Governance checkpoints
- Data and IP agreement before first round
- Review metrics after each round
- Approval before publishing datasets
- Acceptance criteria
- Proposed: per-round MCC and diversity metrics documented
- Proposed: prospective top-site prediction checked on at least ten new substrates
- Proposed: dataset released in a structured format
- Risks
- Narrow chemistry scope
- Over-reliance on moderate scaffold-split accuracy
- Partner restrictions blocking data release
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
The workflow pairs a cheap, fast oracle (scoring 22,253 molecules in about 0.1 s) for experiment selection with heavier GNNs trained once data grow; institutions need HTE automation, a structured reaction-data format and GPU capacity for the second stage.
Governance
Who approves, reviews and stays accountable for outcomes?
Industry co-ownership of compound collections and experiments means data-sharing and IP terms must be settled before closed-loop campaigns; this study released datasets on Zenodo/GitHub and code under GPLv3.
Security and privacy
What data, permissions and controls need testing?
No personal data. Proprietary compound libraries and unreleased structures are the main confidentiality concern in replicating this model with partners.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Chemists need skills in uncertainty-aware model selection and data curation; the open dataset and code lower entry barriers for academic groups without HTE facilities.
Procurement
What should contracts, pricing and exit terms secure?
Replication depends on HTE robotics, compound access and GPU time; joint industry–university agreements should specify data release, as this one did.
Operating model
Which teams own the service once it runs?
Shifts lab planning from chemist intuition alone to model-proposed experiment batches with human scale-up and structural characterisation still required.
Publication history
- 2026-10-12Research Universities · Issue 353 resources
Stable resource ID: eth-roche-active-learning-borylation-prediction-nature-comp-sci-20261009