Computational assessment of feature combinations for pathogenic variant prediction

Abstract

Background Although several methods have been proposed for predicting the effects of genetic variants and their role in disease, it is still a challenge to identify and prioritize pathogenic variants within sequencing studies. Methods Here, we compare different variant and gene-specific features as well as existing methods and investigate their best combination to explore potential performance gains. Results We found that combining the number of “biological process” Gene Ontology annotations of a gene with the methods PON-P2, and PROVEAN significantly improves prediction of pathogenic variants, outperforming all individual methods. A comprehensive analysis of the Gene Ontology feature suggests that it is not a variant-dependent annotation bias but reflects the multifunctional nature of disease genes. Furthermore, we identified a set of difficult variants where different prediction methods fail. Conclusion Existing pathogenicity prediction methods can be further improved.