Predicting instances of pathway ontology classes for pathway integration

Area of Science:

• Bioinformatics
• Computational Biology
• Ontology Engineering

Background:

• Biological pathway analysis requires improved integration of data from disparate sources.
• Ontologies offer a framework for organizing complex biological data.
• The Pathway Ontology (PW) serves as a unifying resource for pathway information.

Purpose of the Study:

• To develop a method for associating pathway instances from different databases with appropriate classes in the Pathway Ontology.
• To leverage machine learning for automating the mapping of biological pathways to ontology classes.

Main Methods:

• A supervised machine learning approach using neural networks (NN) was employed.
• NN models were trained to predict mappings between Reactome pathways and Pathway Ontology classes.
• A baseline bag-of-words (BOW) model was used for comparison.
• Model performance was evaluated by curators on a subset of Reactome pathways.

Main Results:

• The neural network model generated 10,952 class recommendations for 2222 Reactome classes.
• Compared to the BOW model, the NN model achieved higher recall (0.78 vs. 0.42) but lower precision (0.39 vs. 0.49).
• Approximately 78% of Reactome pathways received pertinent recommendations from the NN model.

Conclusions:

• The neural predictive model effectively assists Pathway Ontology curators in mapping Reactome pathways.
• This methodology reduces manual curation effort.
• The approach enhances the organization and integration of data from pathway databases using the Pathway Ontology.