GenPhenia: using deep neural networks to accelerate rare-disease diagnosis

Traditional molecular diagnosis of rare genetic diseases often struggles with complex genotype–phenotype relationships, such as pleiotropy and locus heterogeneity. Despite advances in next-generation sequencing, variant interpretation remains a major bottleneck, frequently requiring labor-intensive expert review to establish phenotypic specificity. To address this challenge, we developed GenPhenia, a deep learning–based framework designed to assist phenotype-driven gene prioritization using clinical features encoded with the Human Phenotype Ontology (HPO). GenPhenia leverages large-scale synthetic clinical history datasets and a graph neural network architecture with message-passing strategies combined with BERT-based phenotype embeddings to learn phenotype–gene associations. The model was validated using both ClinVar-derived datasets and real-world clinical cohorts. On the challenging MCRD benchmark dataset, GenPhenia achieved 60% Top-1 recall and 80% Top-10 recall, substantially outperforming previously reported methods, which achieved less than 30% Top-10 recall under the same conditions. Clinical case studies across endocrinology, cardiology, and immunology further demonstrate that the model effectively refines gene rankings as phenotypic specificity increases during the diagnostic process. GenPhenia provides a scalable and highly sensitive computational approach for phenotype-driven gene prioritization that aligns with the iterative nature of clinical phenotyping and genomic diagnosis. The framework may facilitate more efficient variant interpretation and support the integration of artificial intelligence methods into clinical genomics workflows for rare disease diagnosis.