Prioritizing Clinically Relevant Copy Number Variation from Genetic Interactions and Gene Function Data

Abstract

It is becoming increasingly necessary to develop computerized methods for identifying the few disease-causing variants from hundreds discovered in each individual patient. This problem is especially relevant for Copy Number Variants (CNVs), which can be cheaply interrogated via low-cost hybridization arrays commonly used in clinical practice. We present a method to predict the disease relevance of CNVs that combines functional context and clinical phenotype to discover clinically harmful CNVs (and likely causative genes) in patients with a variety of phenotypes. We compare several feature and gene weighing systems for classifying both genes and CNVs. We combined the best performing methodologies and parameters on over 2,500 Agilent CGH 180k Microarray CNVs derived from 140 patients. Our method achieved an F-score of 91.59%, with 87.08% precision and 97.00% recall. Our methods are freely available at https://github.com/compbio-UofT/cnv-prioritization. Our dataset is included with the supplementary information.

Fig 1. Importance of Model Features.

(a) Histogram of CNV lengths (on log scale) for harmful and benign CNVs within our dataset shows that harmful CNVs are more likely to be longer, and hence likely affect more genes and gene functions. (b-d) Precision (b), recall (c) and f-measure (d) for predicting harmful versus benign CNVs relative to the number of closest neighbors considered within the gene interaction network. Both precision (b) and f-measure (d) improve as we expand the number of neighbors considered, but stabilize or slightly descend after 10 neighbors. We also see an improvement from utilizing the patient phenotypes uniform model in precision and accuracy as we add the ranking as a source for weighing our features.

Fig 2. Precision, recall and f-measure for CNVs when combining the three following features length, DGV and gene.

Length is the CNV length. DGV is a measure of the CNV’s frequency in the Database of Genomic Variants. Gene is the feature derived from the previous machine learning step in this method.

Fig 3. Databases, ontologies and known associations used to identify CNV-phenotype correlations.

Our approach integrates 3 types of information: 1) CNVs an their non-exhaustive frequency in healthy individuals, 2) genes and gene interactions, with their respective functions (each gene within a CNV is weighted by its likelihood of contributing to the phenotypes; via semantic similarity within the GO ontology), and 3) phenotypic descriptions and relationships between them as specified by HPO, with their non-exhaustive associations to disease genes (via OMIM). For an individuals variants and known HPO phenotypes, genes affected by these variants are highlighted within the gene interaction network, while the phenotypes are emphasized in the phenotype ontology layer.