Ontology pre-training improves machine learning-based predictions for metabolites
Deep learning has recently led to multiple breakthroughs in longstanding biological problems. Recent advances have shown that integration of expert knowledge can improve deep learning model predictive performance in complex domains such as biology. Bio-ontologies offer a rich source of curated biological knowledge that can be harnessed to this end. Here, we describe an intuitive and generalisable approach to embed the knowledge contained in a classification hierarchy derived from a bio-ontology into a deep learning model as an intermediate training step between general-purpose pre-training and task-specific fine-tuning: a process that we call'ontology pre-training'. We show that this approach leads to an improvement in predictive performance and a reduction in training time for a broad range of prediction tasks related to metabolite functions, activities and properties in living systems, using a range of datasets derived from MoleculeNet. We see the biggest improvement for regression tasks, e.g., prediction of lipophilicity and aqueous solubility of molecules, and a robust improvement for most classification tasks. Our approach can be adapted for a wide range of knowledge sources, models and prediction tasks.
Authors
- Charlotte Tumescheit (ORCID: https://orcid.org/0000-0002-7563-5575)
- Till Mossakowski (ORCID: https://orcid.org/0000-0002-8938-5204)
- Simon Flügel (ORCID: https://orcid.org/0000-0003-3754-9016)
- Janna Hastings (ORCID: https://orcid.org/0000-0002-3469-4923)
- Martin Larralde (ORCID: https://orcid.org/0000-0002-3947-4444)
- Fabian Neuhaus (ORCID: https://orcid.org/0000-0002-1058-3102)
- Martin Glauer (ORCID: https://orcid.org/0000-0001-6772-1943)
Institutions
- Osnabrück University (DE)
- University of Zurich (CH)
- Leiden University Medical Center (NL)
- Otto-von-Guericke-Universität Magdeburg (DE)
Publication Details
- Journal
- PLoS Computational Biology
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1371/journal.pcbi.1014925
- Citations
- 1
- Primary Topic
- Computational Drug Discovery Methods
- Type
- article
- Field-Weighted Citation Impact
- 1.31
Funders
- National Science Foundation
- Deutsche Forschungsgemeinschaft
- Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung