Abstract
Along with the promotion and application of genetic techniques, desiring to improve the efficiency of calling copy number variation (CNV) in next-generation sequencing analysis technics (NGS) data turns into a requirement. A single existing method is insufficient for calling all potential CNVs. In this paper, we present cnnCNV, a convolutional neural network (CNV) based framework for calling CNV. Firstly, cnnCNV merges the output of existing CNV calling tools as candidates; secondly, generates images of each candidate region from aligned reads based on multiple detection theories; finally, a trained model can be used to classify candidates into true and false. Our approach was tested on simulated data comparing with existing tools, including Breakdancer, ControlFREEC, CNVnator, Delly and readDepth. Results show that cnnCNV improve precision and sensitivity of CNV calling remarkably. Pseudo-code of image generation can be download at: The pseudo-code of image creation can be downloaded at: https://github.com/StudMs/cnnCNV.