Ontology Pre-training for Poison Prediction

Chebifier: automating semantic classification in ChEBI to accelerate data-driven discovery

“…It was also shown that this kind of training provides domain-relevant knowledge beyond the scope of the ontology. 32 A similar approach based on large language models has been proposed to classify entities within the DOLCE upper-level ontology, 33,34 and a graph-based representation learning strategy has been used to suggest classifications in the SNOMED clinical vocabulary. 35 …”

“…A more advanced use case in the context of data-driven discovery is that the Chebifier model can be further fine-tuned to enable the domain knowledge which the model has learned from the ontology to be harnessed to improve performance for other, more specific prediction and discovery tasks. We have previously shown that this approach has value for example in toxicity prediction, 32 and a similar approach can also be used to extend the Chebifier model for better performance in a sub-domain of interest. For example, if a task requires an improved and more fine-grained classification of lipids than Chebifier can provide out of the box, then it is possible to further fine-tune Chebifier's model by providing a set of examples for the kinds of lipids that are relevant for the task.…”

“…For example, we have shown that ontology pre-training improves performance of a prediction model in predicting chemical toxicity, even though the pre-training did not include any biological information directly relevant to the toxicity prediction task. 32 This illustrates the broad-ranging importance of incorporating domain knowledge into datadriven discovery methods. A similar approach in the domain of proteins is OntoProtein, 36 which uses the Gene Ontology to augment the information available for a protein language model.…”

“…It was also shown that this kind of training provides domain-relevant knowledge beyond the scope of the ontology. 32 A similar approach based on large language models has been proposed to classify entities within the DOLCE upper-level ontology, 33,34 and a graph-based representation learning strategy has been used to suggest classications in the SNOMED clinical vocabulary. 35 A benet of our approach over and above enabling chemicals to be classied into an ontology, is that it encodes the knowledge from the ontology into a computable format that is then able to be applied through transfer learning to improve other classication tasks.…”