Penalized weighted low-rank approximation for robust recovery of recurrent copy number variations

BMC Bioinformatics

Research Article

Xiaoli Gao¹ 

[1] Department of Mathematics and Statistics, University of North Carolina at Greensboro, 1400 Spring Garden St, Greensoboro, NC, USA;

关键词: Copy number variation;    Fused lasso;    Low-rank approximation;    Recurrent copy number variation;    Penalized weighted approximation;   

DOI  :  10.1186/s12859-015-0835-2

 received in 2015-09-22, accepted in 2015-11-23,  发布年份 2015

来源: Springer

【 摘 要 】

BackgroundCopy number variation (CNV) analysis has become one of the most important research areas for understanding complex disease. With increasing resolution of array-based comparative genomic hybridization (aCGH) arrays, more and more raw copy number data are collected for multiple arrays. It is natural to realize the co-existence of both recurrent and individual-specific CNVs, together with the possible data contamination during the data generation process. Therefore, there is a great need for an efficient and robust statistical model for simultaneous recovery of both recurrent and individual-specific CNVs.ResultWe develop a penalized weighted low-rank approximation method (WPLA) for robust recovery of recurrent CNVs. In particular, we formulate multiple aCGH arrays into a realization of a hidden low-rank matrix with some random noises and let an additional weight matrix account for those individual-specific effects. Thus, we do not restrict the random noise to be normally distributed, or even homogeneous. We show its performance through three real datasets and twelve synthetic datasets from different types of recurrent CNV regions associated with either normal random errors or heavily contaminated errors.ConclusionOur numerical experiments have demonstrated that the WPLA can successfully recover the recurrent CNV patterns from raw data under different scenarios. Compared with two other recent methods, it performs the best regarding its ability to simultaneously detect both recurrent and individual-specific CNVs under normal random errors. More importantly, the WPLA is the only method which can effectively recover the recurrent CNVs region when the data is heavily contaminated.

【 授权许可】

CC BY   
© Gao 2015