HPOFiller

HPOFiller predicts missing protein-phenotype associations in the Human Phenotype Ontology (HPO) to improve mapping between proteins and HPO terms for genetic disease research.

Key Features:

• Graph Convolutional Network (GCN) architecture: Uses graph convolutional networks to model relationships in graph-structured biological data.
• S-GCN (Semantic GCN): Processes the protein-protein interaction network and the HPO semantic similarity network to produce semantic embeddings using network weights.
• Bi-GCN (Bipartite GCN): Operates on the protein-phenotype bipartite graph and implements message passing between proteins and phenotypes to refine associations.
• Iterative refinement: Iteratively runs S-GCN and Bi-GCN over three interconnected networks (protein-protein interaction, HPO semantic similarity, protein-phenotype bipartite) to progressively refine embeddings.
• Empirical validation: Evaluated with cross-validation and temporal validation and reported to outperform existing state-of-the-art methods.
• Batch normalization importance: Batch normalization was identified as a critical component contributing to improved predictive performance.
• Ablation study: Ablation analyses were performed to assess the contribution of model components, highlighting the role of batch normalization.
• Literature corroboration: High-ranking predictions are supported by literature evidence linking predicted associations to published findings.

Scientific Applications:

• Disease-Gene Association Discovery: Suggests potential unknown disease-gene links by leveraging known disease-HPO term associations and predicted protein-HPO mappings.
• Enhanced Disease Diagnosis and Treatment: Provides more comprehensive protein-phenotype annotations that can inform research into prevention, diagnosis, and treatment of genetic disorders.

Methodology:

Implements S-GCN and Bi-GCN models with message passing to extract embeddings from the protein-protein interaction network, HPO semantic similarity network, and protein-phenotype bipartite graph through iterative refinement; applies batch normalization and evaluates performance using cross-validation and temporal validation.