Automated Shared Phenotype Discovery in Undiagnosed Cohorts for Rare Disease Research.

Rare disease diagnosis is challenging in large part due to incomplete knowledge of gene-to-phenotype associations. One way to address this is to adopt a gene-to-patient paradigm wherein one selects an in-silico predicted pathogenic variant, identifies individuals with the variant, and then determines if the individuals have a shared phenotype. Most studies following this paradigm determine presence of a shared phenotype through manual review of ontology terms in the patient record. We propose a novel automated method to identify the shared phenotype via genetic search using a fitness function that compares the similarity of phenotype term embeddings generated by advanced NLP models applied to the term's text descriptions. Leveraging Human Phenotype Ontology resources, we generated a library of simulated patients across 5,076 Mendelian diseases. Applying our approach to these simulated disease cohorts, we found that the solution phenotypes included a closely matching term for the majority of terms in the disease phenotype under variable conditions of annotation imprecision and noise. We anticipate these methods can aid gene-to-phenotype association discovery for rare diseases by enabling a scalable gene-to-patient research paradigm.