HiHPO: Multimodal Hierarchical Graph Learning for Predicting Missing Protein-Phenotype Associations

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Protein-phenotype associations are crucial for understanding disease mechanisms and diagnosis.
• The Human Phenotype Ontology (HPO) standardizes phenotype descriptions, but protein-HPO annotations are incomplete and evolving.
• Current prediction methods struggle with hierarchical phenotype semantics and multimodal data, especially with sparse annotations.

Purpose of the Study:

• To develop a robust computational framework for predicting protein-phenotype associations.
• To leverage multimodal biological data and the hierarchical structure of the HPO.
• To improve generalization to fine-grained and sparsely annotated phenotypes.

Main Methods:

• Proposed HiHPO, a multimodal, hierarchy-aware graph contrastive learning framework.
• Integrated protein-protein interaction networks, gene expression profiles, and protein language model embeddings.
• Incorporated HPO hierarchical structure into contrastive representation learning.

Main Results:

• HiHPO consistently outperformed state-of-the-art methods on various validation splits.
• Demonstrated significant advantages for deep HPO terms and newly curated annotations.
• Confirmed the contribution of each data modality and the framework's robustness across different annotation densities.

Conclusions:

• HiHPO advances protein-phenotype association prediction by effectively utilizing multimodal data and HPO semantics.
• The hierarchy-aware approach enhances generalization and performance, particularly for challenging annotations.
• This work holds significant potential for disease mechanism elucidation and phenotype-driven diagnosis.