IMPROVE-DD: Integrating multiple phenotype resources optimizes variant evaluation in genetically determined developmental disorders

Summary: Diagnosing rare developmental disorders using genome-wide sequencing data commonly necessitates review of multiple plausible candidate variants, often using ontologies of categorical clinical terms. We show that Integrating Multiple Phenotype Resources Optimizes Variant Evaluation in Developmental Disorders (IMPROVE-DD) by incorporating additional classes of data commonly available to clinicians and recorded in health records. In doing so, we quantify the distinct contributions of sex, growth, and development in addition to Human Phenotype Ontology (HPO) terms and demonstrate added value from these readily available information sources. We use likelihood ratios for nominal and quantitative data and propose a classifier for HPO terms in this framework. This Bayesian framework results in more robust diagnoses. Using data systematically collected in the Deciphering Developmental Disorders study, we considered 77 genes with pathogenic/likely pathogenic variants in ≥10 individuals. All genes showed at least a satisfactory prediction by receiver operating characteristic when testing on training data (AUC ≥ 0.6), and HPO terms were the best predictor for the majority of genes, though a minority (13/77) of genes were better predicted by other phenotypic data types. Overall, classifiers based upon multiple integrated phenotypic data sources performed better than those based upon any individual source, and importantly, integrated models produced notably fewer false positives. Finally, we show that IMPROVE-DD models with good predictive performance on cross-validation can be constructed from relatively few individuals. This suggests new strategies for candidate gene prioritization and highlights the value of systematic clinical data collection to support diagnostic programs.