HiHPO: Multimodal Hierarchical Graph Learning for Predicting Missing Protein-Phenotype Associations.
Understanding protein-phenotype associations is essential for elucidating disease mechanisms and supporting phenotype-driven diagnosis. Although the Human Phenotype Ontology (HPO) provides a standardized framework for phenotypic description, protein-HPO annotations remain incomplete and continuously evolving, posing challenges for robust computational prediction. Existing methods often fail to fully exploit hierarchical phenotype semantics and multimodal biological context, particularly under sparse annotation settings. We propose HiHPO, a multimodal, hierarchy-aware graph contrastive learning framework for predicting protein-phenotype associations. HiHPO integrates complementary biological information from protein-protein interaction networks, gene expression profiles, and protein language model embeddings, while explicitly incorporating HPO hierarchical structure into contrastive representation learning. This design enables the model to preserve semantic relationships among phenotypes and improve generalization to fine-grained and sparsely annotated terms. Extensive evaluations on both random and temporal validation splits demonstrate that HiHPO consistently outperforms state-of-the-art methods, with pronounced advantages on deep HPO terms and newly curated annotations. Additional analyses confirm the contribution of each modality and the robustness of the framework across varying annotation densities. These results highlight the potential of hierarchy-aware multimodal learning for advancing protein-phenotype association prediction and disease-related biomedical research. Code and data are available at https://github.com/ZhuLab-Fudan/HiHPO.