mims-harvard/SHEPHERD

Leverages knowledge-guided graph neural networks trained on simulated rare disease patients to enable label-efficient diagnosis without real annotated data. SHEPHERD integrates phenotype (HPO), gene (Ensembl), and disease (MONDO) ontologies through a rare disease knowledge graph, supporting three interconnected tasks: causal gene discovery, patient matching, and disease characterization. Built on PyTorch Geometric, it accepts patient phenotype profiles and candidate genes/patients/diseases to identify associations through graph-based reasoning.

No commits in the last 6 months.