HPODNets

HPODNets predicts associations between human proteins and Human Phenotype Ontology (HPO) terms using deep graph convolutional networks (GCNs) to address incomplete HPO annotations.

Key Features:

• Multiple Network Input: Integrates data from multiple interaction and biomolecular networks to leverage diverse biological relationships.
• Semi-Supervised Learning: Employs semi-supervised learning to enable prediction when only a subset of proteins are labeled with HPO terms.
• Deep Graph Convolutional Network (GCN): Uses an eight-layer deep GCN architecture to capture high-order topological information from input networks.
• Validation and Benchmarking: Evaluated using cross-validation and temporal validation and shown to outperform seven other protein function prediction methods.

Scientific Applications:

• HPO Annotation: Predicts and augments HPO annotations for human proteins to fill gaps in phenotype term assignments.
• Disease Research: Supports studies of disease prevention, diagnosis, and treatment by linking proteins to phenotypic abnormalities.
• Node Label Ranking: Serves as a model for node label ranking problems across multiple biomolecular networks.

Methodology:

Integrates multiple interaction networks, applies semi-supervised learning, and implements an eight-layer deep graph convolutional network (GCN) to capture high-order topology, with performance assessed via cross-validation and temporal validation against seven competing methods.