Ontology pre-training improves machine learning-based predictions for metabolites

Recent advances in the field of machine learning have shown that integration of expert knowledge improves performances, in particular for 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 machine learning model as an intermediate training step between general-purpose pre-training and task-specific fine-tuning in 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 predictive tasks relevant to understanding metabolite functions 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.