Transfer learning across ontologies for phenome-genome association prediction

Area of Science:

• Bioinformatics
• Computational Biology
• Genetics

Background:

• Predicting gene-phenotype associations is crucial for understanding diseases.
• Existing methods struggle to model hierarchical phenotype structures and leverage sparse data effectively.

Purpose of the Study:

• To develop novel computational methods for improved gene-phenotype association prediction.
• To leverage hierarchical phenotype information and functional gene annotations.

Main Methods:

• Introduced Dual Label Propagation (DLP) to model hierarchical phenotype paths in the Human Phenotype Ontology (HPO).
• Developed a transfer learning framework (tlDLP) integrating Gene Ontology (GO) functional annotations with DLP.
• Utilized protein-protein interaction networks for simultaneous reconstruction of GO and HPO associations.

Main Results:

• Both DLP and tlDLP improved cross-validation predictions of gene-phenotype associations in HPO.
• Transfer learning significantly enhanced predictions for phenotypes with sparse known associations.
• Demonstrated improved predictions using phenotype paths and GO transfer learning.

Conclusions:

• The proposed DLP and tlDLP methods effectively model hierarchical phenotype structures and integrate functional annotations for accurate gene-phenotype association prediction.
• Transfer learning provides substantial benefits for predicting associations with limited prior data.