Comparison of methods to detect copy number alterations in cancer using simulated and real genotyping data

BACKGROUND: The detection of genomic copy number alterations (CNA) in cancer based on SNP arrays requires methods that take into account tumour specific factors such as normal cell contamination and tumour heterogeneity. A number of tools have been recently developed but their performance needs yet to be thoroughly assessed. To this aim, a comprehensive model that integrates the factors of normal cell contamination and intra-tumour heterogeneity and that can be translated to synthetic data on which to perform benchmarks is indispensable. RESULTS: We propose such model and implement it in an R package called CnaGen to synthetically generate a wide range of alterations under different normal cell contamination levels. Six recently published methods for CNA and loss of heterozygosity (LOH) detection on tumour samples were assessed on this synthetic data and on a dilution series of a breast cancer cell-line: ASCAT, GAP, GenoCNA, GPHMM, MixHMM and OncoSNP. We report the recall rates in terms of normal cell contamination levels and alteration characteristics: length, copy number and LOH state, as well as the false discovery rate distribution for each copy number under different normal cell contamination levels. Assessed methods are in general better at detecting alterations with low copy number and under a little normal cell contamination levels. All methods except GPHMM, which failed to recognize the alteration pattern in the cell-line samples, provided similar results for the synthetic and cell-line sample sets. MixHMM and GenoCNA are the poorliest performing methods, while GAP generally performed better. This supports the viability of approaches other than the common hidden Markov model (HMM)-based. CONCLUSIONS: We devised and implemented a comprehensive model to generate data that simulate tumoural samples genotyped using SNP arrays. The validity of the model is supported by the similarity of the results obtained with synthetic and real data. Based on these results and on the software implementation of the methods, we recommend GAP for advanced users and GPHMM for a fully driven analysis.