Improved detection of intra-specific genomic structural variation using paired end high throughput resequencing data and Support Vector Machine

Several bioinformatics methods have been proposed for the detection and characterization of genomic structural variation (SV) from ultra-high throughput genome resequencing data. Recent surveys show that comprehensive detection of SV events of different types between an individual resequenced genome and a reference sequence is best achieved through the combination of methods based on different principles (split mapping, reassembly, read depth, insert size, etc). The improvement of individual predictors is thus an important objective. Here we propose a new a method that combines deviations from expected library insert sizes and additional information from local patterns of read mapping and uses supervised learning to predict the position and nature of structural variants. We show that our approach provides greatly increased sensitivity with respect to other tools based on paired end read mapping at no cost in specificity, and it makes reliable predictions of very short insertions and deletions in repetitive and low complexity genomic contexts that can confound tools based on split-mapping of reads.