Ontology pre-training improves prediction of molecular properties

Presentation given at the Ontologies4Chem Workshop 2025. Abstract: We presented how we use ontologies to improve Machine Learning models using a method we call “Ontology Pre-Training” (https://www.biorxiv.org/content/10.1101/2025.09.30.679573v1). We transform the ChEBI hierarchy into a classification problem and add it as an additional training step of a chemical foundation model. We show that this method improves performance for many downstream applications of small molecules, such as toxicity, aqueous solubility, and lipophilicity prediction.