EnsembleCNV: an ensemble machine learning algorithm to identify and genotype copy number variation using SNP array data

• Published in: Nucleic acids research 2019-04, Vol.47 (7), p.e39-e39
• Main Authors: Zhang, Zhongyang, Cheng, Haoxiang, Hong, Xiumei, Di Narzo, Antonio F, Franzen, Oscar, Peng, Shouneng, Ruusalepp, Arno, Kovacic, Jason C, Bjorkegren, Johan L M, Wang, Xiaobin, Hao, Ke
• Format: Article
• Language: English
• Subjects: Bioinformatik (beräkningsbiologi) (tillämpningar under 10610); Data- och informationsvetenskap (Datateknik); Datasets as Topic; DNA Copy Number Variations - genetics; Genome, Human - genetics; Genotyping Techniques - methods; Humans; Machine Learning; Medicin och hälsovetenskap; Medicinsk genetik; Medicinska och farmaceutiska grundvetenskaper; Methods Online; Naturvetenskap; Polymorphism, Single Nucleotide - genetics; Quality Control
• ISSN: 0305-1048; 1362-4962; 1362-4962
• DOI: 10.1093/nar/gkz068

Abstract The associations between diseases/traits and copy number variants (CNVs) have not been systematically investigated in genome-wide association studies (GWASs), primarily due to a lack of robust and accurate tools for CNV genotyping. Herein, we propose a novel ensemble learning framework, ensembleCNV, to detect and genotype CNVs using single nucleotide polymorphism (SNP) array data. EnsembleCNV (a) identifies and eliminates batch effects at raw data level; (b) assembles individual CNV calls into CNV regions (CNVRs) from multiple existing callers with complementary strengths by a heuristic algorithm; (c) re-genotypes each CNVR with local likelihood model adjusted by global information across multiple CNVRs; (d) refines CNVR boundaries by local correlation structure in copy number intensities; (e) provides direct CNV genotyping accompanied with confidence score, directly accessible for downstream quality control and association analysis. Benchmarked on two large datasets, ensembleCNV outperformed competing methods and achieved a high call rate (93.3%) and reproducibility (98.6%), while concurrently achieving high sensitivity by capturing 85% of common CNVs documented in the 1000 Genomes Project. Given this CNV call rate and accuracy, which are comparable to SNP genotyping, we suggest ensembleCNV holds significant promise for performing genome-wide CNV association studies and investigating how CNVs predispose to human diseases.