Copy number variations (CNVs) are the main genetic structural variations in human cancer genome [1–4]. Accurate inference of CNVs is necessary for identifying cancer-causing genes, and has been of long-standing interest in cancer-focused studies for investigating rules of tumor progression [5–7]. Meanwhile, the advent of next-generation sequencing (NGS) has dramatically furthered our understanding of human diseases with an unprecedented depth, as it allows high-throughput profiling of human genome in nucleotide resolution. Compared to whole-genome sequencing (WGS), whole-exome sequencing (WES) only captures and sequences exonic regions (referred as targets) and allows relatively higher coverage given at the same cost. As always, high efficiency comes with limitations. CNV detection in WES data is likely to has a high false negative rate as a consequence of the uneven distribution of exons across the cancer genome [8].

According to the recent reviews [8, 9], the existed tools showed their specialties in detecting CNVs. However, when analyzing clinical sequencing data, the performances of current CNV detecting algorithms are far from satisfactory. In clinical settings, integrative power in CNV detection is likely to achieve the most stable performance [10]. It should contain following features: 1) Adopted different strategies, current tools show significant divergence in performance. For instance, ADTEx is most likely to detect medium-size CNVs [11], while EXCAVATOR tends to identify CNVs between 1 Mb and 100 Mb [9]. Thus, a new tool that incorporates different methods can deliver a more comprehensive detection. 2) Parameters of the integrative approach should be extensive and easy to modify. CNV detection results are greatly related to parameter settings [8], thus optimal performance of each included method requires the easy modification of parameters. 3) As high precision for CNV detection could not be easily achieved by simply adopting the multiple algorithms, broad annotations should be conducted as a guidance for users in the analysis of datasets.

To these ends, we developed Anaconda (AN Automated pipeline for somatic COpy Number variation Detection and Annotation from tumor exome sequencing data), which successfully satisfied the requirements: 1) Anaconda is designed to be compatible with ease of use and rich features. Running Anaconda only needs one single command “./bin/ANACONDA /path/to/configfile”. Users could easily modify the parameters in config file. Detailed explanation of each parameter could be found at Anaconda website. 2) While utilizing different strategies, users need to locally install and configure the respective running environment for different tools, which sometime is highly challenging for general users. After downloading Anaconda package, by single command “./install”, Anaconda would automatically install and configure the running environment. When running, Anaconda will extract the detected CNV results of the user-selected methods. Consensus results are also generated if CNVs called by multiple methods. 3) To further explore the biological functions beneath shared CNVs, Anaconda can also conduct annotation analysis for the genes that are involved in all CNV regions called by selected tools. Thus, we believe that Anaconda could assist users in a comprehensive and effective manner with their CNV-related projects.

To evaluate the performance gain of Anaconda, we used thirteen simulated samples to evaluate the performance of Anaconda and the individual tool. Each simulated sample contains ten CNVs regions range from one to twenty copies (the size ranges from 500 kb to 4.5 Mb). The definition of true positives (TP), true negatives (TN), false positives (FP) and false negatives (FN) were described in our previous work [18]. The statistical measures of true positive rate (TPR), false discovery rate (FDR) and precision were used to evaluate the performance of individual or combined algorithms. Compared with results from individual software, the approach of integration of different algorithms has more stable performance. The false discovery rate was reduced from 0.0417%–17.7877% to 0.0011%–0.4854%, and the precision was increased from 82.21%–99.96% to 99.51%–100.00% (Additional file 4: Table S1).

To demonstrate the high practicability of Anaconda in detecting and annotating somatic CNVs, and to evaluate the function it presents, we applied Anaconda to analyze a tumor WES dataset downloaded from European Genome Phenome Archive (EGA) with accession number EGAS00001000132. We randomly picked 9 samples, SA018, SA029, SA030, SA031, SA051, SA052, SA065, SA069 and SA071 from this dataset. During the analysis, all samples are conducted with the default parameters. For each sample, all the four calling methods, Control-FREEC, ADTEx, EXCAVATOR and ExomeCNV were applied to call CNVs from WES data. Venn diagrams were plotted (Fig. 2) to compare the overlapping results of called CNVs and genes in called CNV regions.

Distribution of called CNVs and genes are shown in Fig. 3a. Shared CNV regions by 4 callers (tetrad-caller reads) are significantly decreased, ranging from 0.2% in SA029 to 16.8% in SA071. Gene distribution in tetrad-caller read regions is relatively higher than triple, double or single caller reads, as the percentages of gene quantity in tetrad-caller region, over the quantity of all genes is two times higher than the percentage of tetrad-caller CNVs quantity over all CNVs. CNVs called by each tool (Fig. 3b) and gene quantity in accordance with the CNV regions (Fig. 3c) demonstrated great divergence of the performance of each tool. For example, ExomeCNV is likely to call more CNVs than others. CNV regions called by Control-FREEC tend to cover more genes. ADTEx shows a moderate performance in calling CNVs as well as the distribution of genes in its called CNV regions. EXCAVATOR called the least in quantity of CNV regions. These regions share the relatively higher overlapping rate with other tools. For example, in SA018, 82.5% of CNVs called by EXCAVATOR are also the callers by other three tools.

Anaconda is an integrative tool in the detection and annotation of CNVs from whole-exome sequencing data. Utilizing four published tools, Anaconda is able to detect CNVs in a comprehensive manner. Ease in installation and application, Anaconda could satisfy the biologist’s demands in data process. Additionally, pervasive annotation of genes in called CNV regions could serve as a second opinion during the analysis of datasets, compensating the low preciseness caused by the unevenly distributed sequence data. In all, we believe Anaconda could be of great help for users with their CNV-related cancer research.

The package and manual for Anacond a can be freely accessed at http://mcg.ustc.edu.cn/bsc/ANACONDA/. Tools integrated in Anaconda could be find in the referenced articles. WES test dataset is downloaded from European Genome Phenome Archive (EGA) with accession number EGAS00001000132.