Nitesh Kumar Sharma

Copy number variations (CNVs) are a significant class of variants having role in the etiology of numerous disease manifestations. It is still challenging to find CNVs in genomic data, and the approaches used currently have an unacceptably high false positive rate. Before moving to downstream analysis or experimental validation, interventions of human specialists to carefully check the original CNV calls for weeding out false positives is required. Here, we present a deep learning-based tool designed uptake this human intervention while validating CNV calls, with emphasis on the calls made by PennCNV tool, which is one of the most reliable CNV callers reported in literature. An ensemble model was developed that outperformed traditional machine learning techniques, with improved accuracy of 0.9807 in CNV calls and an ideal area under the receiver operating characteristic curve of 0.9985. The model's improvement resulted in reducing the false positives and instances when the CNV association results couldn’t be replicated.This study also presents a CNV prediction server (http://backlin.cabgrid.res.in/eqcnvdb/index.php) based on ensemble deep learning technique with minimum false discovery rate (FDR) which can be used in other related species. This work is the first genome-wide, chromosome-wise, breed- wise CNV Atlas of Indian Equine breeds, EqCNVDb, available at http://backlin.cabgrid.res.in/eqcnvdb/index.php. The output of this study can befurther evaluated for horse breed signature, evolutionary studies including adaptive response of equine germplasm against biotic and abiotic stresses.