CNVscore calculates pathogenicity scores for copy number variants together with uncertainty estimates accounting for learning biases in reference Mendelian disorder datasets.

Copy number variants (CNVs) are a major cause of rare pediatric diseases with a broad spectrum of phenotypes. Genetic diagnosis based on comparative genomic hybridization tests typically identifies [~]8-10% of patients as having CNVs of unknown significance, revealing the current limits of clinical interpretation. The adoption of whole-genome sequencing (WGS) as a first-line genetic test has significantly increased the load of CNVs identified in single genomes. Alongside short- and long-read sequencing technologies, a number of pathogenicity scores have been developed for filtering and prioritizing large sets of candidate CNVs in clinical settings. However, current approaches are often based, either explicitly or implicitly, on clinically annotated reference sets, which are likely to bias their predictions. In this study we developed CNVscore, a supervised-learning approach combining tree ensembles and a Bayesian classifier trained on pathogenic and non-pathogenic CNVs from reference databases. Unlike previous approaches, CNVscore couples pathogenicity estimates with uncertainty scores, making it possible to evaluate the suitability of a model for the query CNVs. Comprehensive comparative benchmark tests across independent sets and against alternative methods showed that CNVscore effectively distinguishes between pathogenic and benign CNVs. We also found that CNVs associated with CNVscores of low uncertainty were predicted with significantly higher accuracy than those of high uncertainty. However, the performance of current scoring approaches, including CNVscore, was compromised on CNV sets enriched in highly uncertain variants and presenting unconventional features, such as functionally relevant non-coding elements or the presence of disease genes irrelevant for the clinical phenotypes investigated. Finally, we used the CNVscore framework to guide CNV scoring model selection for the French National Database of Constitutional CNVs (BANCCO), which includes clinical diagnosis annotations. The CNVscore framework provides an objective strategy for leveraging the uncertainty on bioinformatic predictions to enhance the assessment of CNV pathogenicity in rare-disease cohorts. CNVscore is available as open-source software from https://github.com/RausellLab/CNVscore and is integrated into the CNVxplorer webserver http://cnvxplorer.com.