InpherNet provides attractive monogenic disease gene hypotheses using patient genes indirect neighbors

Close to 70% of patients suspected to have a Mendelian disease remain undiagnosed after genome sequencing, partly because our current knowledge about disease-causing genes is incomplete. Although hundreds of new diseases-causing genes are discovered every year, the discovery rate has been constant for over a decade. Generating an attractive novel disease gene hypothesis from patient data can be time-consuming as each patients genome can contain dozens to hundreds of rare, possibly pathogenic variants. To generate the most plausible hypothesis, many sources of indirect evidence about each candidate variant may be considered. We introduce InpherNet, a network-based machine learning approach to accelerate this process. InpherNet ranks candidate genes based on gene neighbors from 4 graphs, of orthologs, paralogs, functional pathway members, and co-localized interaction partners. As such InpherNet can be used to both prioritize potentially novel disease genes and also help reveal known disease genes where their direct annotation is missing, or partial. InpherNet is applied to over 100 patient cases for whom the causative gene is incorrectly given low priority by two clinical gene ranking methods that rely exclusively on human patient-derived evidence. It correctly ranks the causative gene among its top 5 candidates in 68% of the cases, compared to 9-44% using comparable tools including Phevor, Phive and hiPhive.