GC-bias correction.

A common observation across all NGS platforms is the under-representation of DNA fragments from low GC or high GC regions of the genome resulting in a bias in the estimation of RD signals in these regions. In iCopyDAV, two approaches are provided for GC-bias correction: (i) Local Polynomial Regression fitting (Loess) algorithm [44] and (ii) Median approach by Yoon et al [17]. The user also has an option to not opt for GC bias correction. In Loess regression approach, reads in bins with similar GC content within an interval of 0.001 are clubbed together. Loess correction is applied to each bin, computed based on the difference in the average read depth value of the bin and median read depth value for the whole genome (span = 0.75 (default), controls degree of smoothness). The median read depth is computed in such a way that GC adjustments do not have any effect on the overall median of the data. In the Median approach for GC bias correction, the corrected read depth of each bin i is given by the expression, (1) where r_i is the read count of the i^th bin, m_GC is the median read count of all bins having the same G+C content as the i^th bin, and m is the overall median read count of all the bins. The R functions loessGCCorrection [13] and gcCorrectionYoon [17] perform GC bias correction by Loess-regression and Median-based approaches, respectively. GC profile plot generated helps the user to confirm whether the read depth in low and high GC content regions are properly correctly (shown in S1 Fig). Both these approaches require a file containing bin-wise GC content scores which can be obtained from UCSC genome browser [45] for human reference genome (available with 5 bp as bin size). Using this one can compute the GC scores file (gcfile) for any specific bin size and genome assembly using the function prepareData in iCopyDAV (the GC score files for bin size = 100 bp are available for hg18 and hg 19 assemblies at the website).

Handling mappability bias.

Mappability of a region in the reference genome is defined as the probability that a read originating from it is unambiguously mapped back to it. It depends on the length of the reads, the number of mismatches allowed during the mapping step and the reference assembly used. Reads that cannot be uniquely mapped to the genome are considered unmappable, and the genomic positions from where they originate are called as unmappable regions. Bin-wise mappability scores (0–1) can be computed along the chromosome/genome. Regions of low mappability score (< 0.5) correspond to repetitive or low-complexity regions in the genome. For handling mappability bias in iCopyDAV, the user may define the mappability cut-off score, Mth (default = 0.5), with ‘0’ corresponding to no mappability bias correction. Then bins with mappability scores > Mth are only considered for subsequent steps in the pipeline. Mappability score tracks can be obtained from UCSC genome browser and computed for any bin size, similar to the construction of GC score file discussed above or by using gem library [46]. Using function prepareData one can generate mappability file (mapfile) for any desired bin size by using the mappability score file for 100 bp bin size (available at the website).

In iCopyDAV, the module pretreatment calculates raw RD values for the bin coordinate file, corrects the raw RD values of each bin appropriately using one of the above GC bias correction methods and filters low mappable reads. For this bin coordinate file (z), mappability scores (mapfile) and GC content scores (gcfile) are required.

Total variation minimization (TVM) approach.

We have implemented the TVM approach proposed by Duan et al [27] in C++. In this approach, the read depth profile is assumed to be a piecewise constant function of plateaus and basins modelling duplication and deletion events respectively along the length of the chromosome. At every step, two adjacent bins with minimum deviation in their RD signals are merged into a single segment with average RD value of the two merged segments and the process is continued iteratively. At every merging step, a total variation penalized least squares (v) is used to fit the RD profile using the equation: (2) where the first term on the right is the fitting error between original signal y_i and the recovered smooth signal x_i in the i^th bin; the second term is the total variation penalty imposed when a change-point is detected between x_i and x_i+1 and λ is the penalty parameter which controls the trade-off between fitting error and penalty caused by the change-points [27]. Schwarz information criterion (SIC) has been implemented to compute the optimal value of λ [47], given by (3) where mean of y_i in a segment is used as the amplitude x_i i.e., for a given segment spanning i to i+l region, . Here m is the number of segments, σ² variance of noise, and optimal λ is achieved at: (4)

Once is known, the optimal smoothed signal of y_i is and a CNV is identified as a segment with change-points at i and i+l and having amplitude below or above the predefined cut-off values.

Circular binary segmentation (CBS) approach.

CBS is a recursive algorithm that partitions the bins x₁, x₂…x_n into a set of segments such that the read depth values in each segment are similar. The basic approach is as follows. In each step, it identifies an interval i…j such that the read depth values x_i,…,x_j are all similar and significantly different from the rest of the regions, x₁,…,x_i-1 and x_j+1,..,x_n. If no such interval is identified, then the whole region x₁,…,x_n is identified as a segment with no change point detected, else three intervals x_i,…,x_j, x₁,…,x_i-1 and x_j+1,..,x_n are obtained (considering the genome with periodic boundary conditions (circular) effectively results in two segments, x_i,…,x_j and x_j+1,..,x_i-1, with the larger fragment having average copy number 2). To determine if the segment x_i,…,x_j is significantly different from the other regions, the null hypothesis of no difference is assumed and t-statistic is computed for each interval i…j, defined as (5) where n_i..j (= j-i+1) is number of bins in the interval i…j, n-n_i..j is the total number of bins in x₁,…,x_i-1 and x_j+1,…x_n, is the mean read depth in the interval i…j and is the mean read depth for the remaining regions x₁,…,x_i-1 and x_j+1,..,x_n. This procedure is applied recursively to every interval i…j and the interval that maximizes the t-statistic, i.e., T_max = max_1≤i<j≤n |T_i,j| is greater than a threshold (determined using a permutation test), the null hypothesis of no difference is rejected and the interval x_i,…x_j is identified as a copy variant region. The Bioconductor package, DNAcopy [48], is incorporated for circular binary segmentation in the pipeline.

The segmentation step being computationally intensive, TVM is parallelized using OpenMPI [49], while for CBS, parallel implementation of the ‘segment’ function, ParDNAcopy [50] is considered. This is achieved by dividing the reference sequence into ‘p’ equal parts (default p = 32) with an overlap of 20 bins/windows between adjacent parts (to handle copy number variant regions at the boundaries). The parameter p is an optional parameter and can be set according to available CPU cores. The segmentation is applied independently and in parallel on each of the p parts and, after the segmentation, the parts are merged back in proper order to identify segments having similar RD signal. Thus, by utilizing multi-core architecture, iCopyDAV is much faster (~ 8-fold) than most other CNV detection programs. It has been shown that CBS approach is suited for identifying large CNVs [13], while the efficacy of TVM approach is shown in the detection of small CNVs [27]. Thus, by combining the results of the two approaches, one would be able to detect wider range of CNVs in iCopyDAV.

Function runSegmentation with parameter -t for TVM and -d for CBS generates a coordinate file containing both normal and variant regions along with their corresponding RD values.

Analysis of simulated data.

For the simulated data, GC bias correction is carried out using Median-based approach and regions of low complexity regions are removed using mappability threshold, Mth = 0.5 (default). A small bin size of 50 bp is considered. Performance of the two segmentation approaches is evaluated using F-score, defined as the harmonic mean of precision and recall (sensitivity): (8) where, recall is defined as the proportion of true CNV predictions (TP) to the total number of CNVs inserted (P): (9) and precision is defined as the ratio of correctly predicted CNVs (TP) to the total number of CNV predictions (TP + FP): (10)

The criterion for ‘true positive’ considered here is reciprocal overlap of 50% or more between the predicted and actual CNVs. Performance of iCopyDAV is assessed at four levels in the simulated experiment: (i) size of CNVs, (ii) breakpoint accuracy, and (iii) event type (duplication/deletion) as a function of sequencing depth, and (iv) for comparison with three other DoC based approaches.

Performance of CNVs on simulated data.

Size of CNVs: To analyze the efficacy of iCopyDAV in the detection of ‘small’ and ‘large’ CNVs, the inserted CNVs are split into two groups: small (0.1–1 Kb) and large (> 1–5 Kb). CNVs larger than 5Kb are not considered here as most algorithms perform well in the detection of large CNVs. Homozygous deletions and copy gain > 4 are also not considered as these are easy to detect. Figs 2(A) and 2(B) respectively depict the F-scores and precision-recall curves as a function of sequencing depth for the two CNV sets, the ‘dashed’ curves corresponding to the performance of ≤ 1 Kb CNV set, and the solid curves for > 1 Kb CNVs. It is observed from Fig 2(A) that the prediction accuracy for CNVs > 1Kb is quite good, greater than 90% for sequencing coverage ≥10×, while that for CNVs ≤ 1 Kb it is ~50% for sequencing depth (≥ 30×) with both the segmentation approaches. This clearly indicates that larger CNVs can be easily detected even in low coverage data by any segmentation approach, but higher sequencing depth is required for the detection of smaller CNVs (especially for single copy gain/loss events). From Fig 2(A) it is observed that TVM approach performs marginally better than CBS approach in the detection of small CNVs (≤ 1 Kb) in low coverage data (< 30×).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

(a) F-score plots, and (b) Recall & Precision plots as a function of sequencing depth. In (b) Recall and precision values are depicted by ‘closed’ and ‘open’ symbols respectively. Performance of TVM (‘circle’) and CBS (‘triangle’) is shown for the two CNV sets: small (≤ 1Kb) shown as dashed lines and large (> 1Kb) as solid lines. Error bars represent standard deviation in each group. Bin size = 50 bp.

https://doi.org/10.1371/journal.pone.0195334.g002

High recall values indicate that most of the true CNVs are correctly detected, while high precision values indicate fewer false positives. In general, a balance between the two is desired for any algorithm. It is observed from Fig 2(B) that both recall and precision are high and comparable for TVM and CBS approaches in the detection of CNVs of size > 1Kb and approach ‘1’ with increase in sequencing coverage (≥ 20×). However, for detecting CNVs of size ≤1 Kb, we observe that CBS has a better recall while TVM gives better precision (~ 1 even at 5×), with the overall performance being marginally better for TVM approach (Fig 2(A)). This suggests that for the detection of small CNVs (≤1Kb), a combination of the two approaches may be useful.

Breakpoint accuracy: The error in accurate detection of the start/end boundaries of predicted CNVs is defined as the breakpoint error. The breakpoint error in the prediction of 24 CNVs in our simulation experiment as a function of sequencing coverage is summarized as boxplots in Fig 3. As expected, the average breakpoint error and the variation in the error decrease with increase in the sequencing depth. This may probably be due to reliability in the alignment of reads and better resolution of the segmentation approaches at higher sequencing coverage. For 30× and higher sequencing coverage, the median breakpoint error is about ~ 2 bin sizes for both the approaches, but with higher variation observed in the case of CBS approach ~ 109 (62–490) compared to the case of TVM approach, ~105 (54–195) at 30× sequence coverage. Very large variation in the breakpoint error is observed in the case of CBS approach, though the Median value is comparable in the two cases, especially for low coverage data (≤ 30×). This probably could be due to the requirement of larger bin sizes for optimal performance of CBS segmentation approach in low sequence coverage data. Since TVM approach performs well with smaller bin sizes, a better resolution of the breakpoints is observed in this case (≤ 30×).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3.

Box plots representing breakpoint error in CNV detection as a function of sequencing coverage using (a) TVM and (b) CBS segmentation approaches in simulated data. Bin size = 50 bp.

https://doi.org/10.1371/journal.pone.0195334.g003

Type of CNVs: To see if there is any bias in the detection of copy gain or copy loss events, the F-scores are computed for the two segmentation algorithms and depicted in Fig 4 for each case. No difference is observed in the detection of large copy gain and copy loss events (> 1 Kb), and F-scores are observed to approach ~ 1 for sequencing coverage ≥ 20× for both the segmentation approaches. For sequencing coverage < 20×, a slight bias in the detection of copy gain events is observed suggesting that it is easier to detect duplications than deletions in low sequence coverage data. But this bias is likely to be due to copy number 4 events which are easier to detect compared to single copy gain and copy loss events (3 and 1). In the detection of small CNVs (≤ 1 Kb), TVM approach is much better in detecting copy gain events while the performance of both the approaches is very poor in detecting copy loss events (when sequencing depth is ≤ 20×).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Performance of TVM (‘circle’) and CBS (‘triangle’) approaches in predicting copy gain (‘closed’ symbol) and copy loss (‘open’ symbol) events in simulated data is shown.

Dashed line corresponds to detection of small CNVs (≤ 1 Kb) and solid line for large CNVs (> 1 Kb). Bin size = 50 bp.

https://doi.org/10.1371/journal.pone.0195334.g004

Effect of bin size: The bin size is an important parameter and affects CNV detection by DoC-based approaches. We observe that optimal performance for TVM approach is observed on using small bin size (~50 bases), while for CBS approach larger bin sizes are required (see S2 Fig), when sequence coverage is low (≤ 20×). For high sequence coverage data, no such dependence on bin size is observed and a small bin size is desirable for breakpoint accuracy. Since small bin size result in smaller breakpoint error, TVM approach is best suitable when sequence coverage is low as observed above in computing breakpoint accuracy.

Comparison with other DoC-based Methods: The performance of iCopyDAV on simulated data is compared with three DoC-based tools, viz. CNVnator [18], ReadDepth [13] and Control-FREEC [7], using their default parameters for data pre-treatment and segmentation approaches. Bin size is set to 50 bp in Control-FREEC, CNVnator and TVM segmentation algorithm in iCopyDAV; while for ReadDepth and CBS approach in iCopyDAV, optimal bin size is computed using negative binomial distribution for each sequencing depth: 1800 bp (5×), 900 bp (10×), 500 bp (20×), 400 bp (30×), 300 bp (40×) and 200 bp (50×). We expect the difference, if any, in the performance of these tools to arise as a consequence of different pre-treatment and segmentation approaches. For instance, CNVnator uses Median approach for GC bias correction and Mean-shift algorithm for segmentation; Control-FREEC uses Polynomial fitting for GC bias correction and Lasso-regression for segmentation, while ReadDepth uses Loess regression for GC bias correction and circular binary segmentation (CBS). For this comparative analysis, in iCopyDAV, Median approach is used for GC bias correction for both TVM and CBS approaches. Different approaches are used for handling mappability bias in these tools; low mappable regions are filtered by setting a mappability threshold (Mth). In Control-FREEC (Mth = 0.85), iCopyDAV (Mth = 0.5), and ReadDepth (Mth = 0.75 and additionally normalized by Mth), while multi-reads spanning low mappable regions are randomly assigned in CNVnator. All the tools consider raw read depth values for segmentation, except ReadDepth which uses log-transformed values. From Fig 5 we observe that both TVM (cyan circles) and CBS (red triangles) approaches in the iCopyDAV pipeline perform better than the other three DoC-based approaches and the combined predictions from TVM+CBS (black pluses) show further improvement in prediction. For the combined predictions of TVM and CBS, we consider the Union of the set of CNVs detected by the two segmentation approaches, that is, CNVs detected by either of the two approaches. For the CNVs detected by both the approaches, the farthest start/stop boundaries are considered. It may be noted that though both ReadDepth and CBS in iCopyDAV uses the same segmentation approach, the performance of CBS is better than ReadDepth for sequencing coverage < 40× due to the difference in data pre-treatment approaches in two cases.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Performance (F-score) of iCopyDAV with three other DoC-based tools (using default parameters for data pre-treatment and segmentation approaches) shown as a function of sequencing coverage.

Error bars represent standard deviation in each group.

https://doi.org/10.1371/journal.pone.0195334.g005

From the comparison of recall and precision values of CBS and TVM in iCopyDAV with the three DoC-based tools, we observe higher recall with both TVM and CBS approaches for the detection of CNVs ≤ 1 Kb, while for the detection of CNVs > 1 Kb all the DoC-based methods showed recall values approaching ‘1’ with increase in sequencing coverage (S3(A) Fig). The precision ~ 1 with both CBS and TVM in the detection of small and large CNVs (even at 5× sequencing coverage for TVM), while precision values (combined of small and large CNVs) of ReadDepth, CNVnator, and Control-FREEC are ~ 0.71, ~ 0.51 and ~ 0.26, respectively, at 30× coverage (S4(B) Fig). Also, it is observed that in case of ReadDepth and Control-FREEC, a reduction in precision is observed with increase in sequencing depth due to increase in the number of false positives. This clearly indicates the importance of appropriate normalization of read depth signals. We observe that the performance of iCopyDAV is further improved by combining the output of the two segmentation approaches (S5 Fig), especially for small CNVs (≤ 1 Kb).

a. Low sequence coverage data.

For various combinations of data pre-treatment and segmentation approaches in iCopyDAV, the number, type and size of CNVs predicted in low sequence coverage data are summarized in Fig 6. A large variation is detected in the type and size of CNVs. The number of CNVs predicted is very large using CBS approach in the case of no GC bias correction (for both high and low mappability cut-offs), clearly indicating the need for removing GC bias in this case. However, on using TVM segmentation approach, the number and type of CNVs predicted are comparable for Loess GC bias corrected and uncorrected data. The other major observation is that Median GC bias corrected data (with either TVM or CBS segmentation approach) resulted in very small number of CNVs, especially deletions, compared to other combinations. It is observed that for Loess GC bias corrected data, the number of deletion events are reduced on increasing mappability cut-off from Mth = 0.5 to 0.8. On the other hand, the duplications events are seen to increase on increasing mappability cut-off from Mth = 0.5 to 0.8 for all the cases. To test the reliability of predictions for various combinations of data pre-treatment and segmentation approaches, below we discuss recall and precision of the CNV calls by comparing with annotations in Database of Genomic Variations (DGV).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Size distribution of copy gain (solid) and copy loss (striped) events shown for various combinations of data pre-treatment and segmentation approaches in low sequence coverage data (6×) for Chr 1 of NA12878 sample.

Bin size = 300 bp.

https://doi.org/10.1371/journal.pone.0195334.g006

To access the reliability of various data pre-treatment and segmentation approaches in low sequence coverage data, variant calls are benchmarked against six studies reported in DGV for the sample NA12878 (summarized in S2 Table). Of these, four studies are based on SNP-array technique (Altshuler et al [69], Cooper et al [70], McCarroll et al [71] and Redon et al [72]), while the fifth study is a collection of CNVs predicted from four different techniques, viz., Fosmid mapping (Kidd et al [73]), array-CGH (Conrad et al [74]), SNP-array (McCarroll et al [71]) and whole shotgun sequencing (Mills et al [75]) and reported as ‘Gold standard’ in Mills et al [36]. The sixth study is based on sequence-based techniques (part of the 1000 Genome Project [76]) and comprises experimentally validated set of CNVs. For evaluating the performance, a predicted CNV is considered to be a true call (true positive) if ≥ 50% overlap is observed between prediction and annotation or vice versa. Due to large variation in the size of predicted and annotated CNVs, many-to-one (number of small adjacent CNV predictions mapping to a single large annotated CNV) and one-to-many mappings (a single large predicted CNV mapping to a number of small annotated CNVs) is typically observed. In such situations, to avoid over-estimation of recall, all adjacent predictions mapping to a single annotation are merged and counted as a ‘single’ CNV call while computing recall, and a single CNV prediction mapping to m annotated CNVs is counted as m CVN calls to avoid under-estimation while computing precision values. That is, recall is computed as the fraction of number of annotations (l) overlapping with the CNV predictions to the total number of annotations (m), i.e., recall = l/m, and precision is computed as the fraction of the number of true predictions (x) overlapping with the annotations to the total number of CNVs predicted (y), i.e., precision = x/y. Also, due to redundancy among the annotated CNV sets across the six studies, we assess our predictions independently on those studies.

Recall and precision values of CNV predictions for various GC bias correction and segmentation approaches are depicted in Fig 7(A) and 7(B) for two mappability cut-off values, Mth = 0.5 and 0.8, respectively. The recall values are plotted as bar graphs while the precision values are averaged across the six studies and depicted as a ‘black square’ (connected by a dashed line) for each combination of GC bias correction and segmentation approaches.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7.

Performance of iCopyDAV is shown for low sequence coverage data (6×) of Chr 1 of NA12878 for mappability threshold values (a) 0.5 and (b) 0.8. Recall and precision values for various combinations of GC bias correction and segmentation algorithms are computed with respect to the six studies reported in DGV. Bin size = 300 bp.

https://doi.org/10.1371/journal.pone.0195334.g007

From Figs 6 and 7 we observe that the small number of CNV predictions observed in the case of Median + (TVM/CBS) combination resulted in low values of recall and precision for Mth = 0.5. However, on increasing Mth to 0.8, recall and precision improved in case of TVM segmentation approach. For Loess + CBS combination, though the number of CNV predictions is quite high, low recall and precision values observed suggest that majority of these are likely to be false positives. In the case of TVM approach, the recall and precision values for no GC bias correction case are comparable to GC bias corrected (Loess) data, indicating its robustness in CNV detection in raw, untreated data. Thus, Loess + TVM combination in iCopyDAV with Mth = 0.5 results in the best performance for low coverage data. We also observe that duplications are detected more accurately (46%) than deletions (12%) in this case, indicating that detection of deletion events is difficult in low coverage data. Similar observations were made in chromosomes 2 and 21 of NA12878 (results not shown).

b. High sequence coverage data.

A similar analysis is carried out on high sequencing coverage data (35×) of Chr 1, NA12878 sample and the results are summarized in Figs 8 and 9. We again observe a large variation in the type and size of CNVs predicted for different combinations of data pre-treatment and segmentation approaches (Fig 8). This clearly suggests that one should judiciously choose the pre-treatment options along with the segmentation approaches. As in the case of low coverage data, large number of CNVs are predicted (mainly deletions) using CBS approach in the case of no GC bias correction, which are significantly reduced on removing GC bias (much more on using Median approach compared to Loess approach). This clearly indicates the need for GC bias correction when using CBS segmentation approach. However, using TVM segmentation approach, the number of CNV predictions are comparable for GC bias corrected (Loess/Median) and uncorrected data when mappability cut-off, Mth = 0.8. This suggests that TVM segmentation approach is able to handle GC bias and this correction step may be skipped when using higher mappability cut-off values. As in the case of low coverage data, in this case also the Median approach for GC bias correction resulted in fewer CNVs (mainly deletions) with both TVM and CBS approaches compared to other combinations. A clear reduction in the number of CNVs (mainly deletions) is observed on increasing the mappability cut-off from 0.5 to 0.8, except for Median of GC bias corrected data. Thus, a clear dependence of the segmentation results on data pre-treatment choices is observed. As before, we assess the reliability of the CNV calls for various combinations of data pre-treatment and segmentation approaches by comparing with six studies annotated in DGV for the NA12878 sample.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Size distribution of copy gain (solid) and copy loss (striped) events shown for various combinations of data pre-treatments and segmentation approaches in high sequence coverage data (35×) for Chr 1 of NA12878 sample.

Bin size = 300 bp.

https://doi.org/10.1371/journal.pone.0195334.g008

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9.

Performance of iCopyDAV is shown on high sequence coverage data (35×) of Chr 1 of NA12878 sample for mappability threshold values (a) Mth = 0.5 and (b) Mth = 0.8. Recall and precision values for various combinations of GC bias correction and segmentation algorithms are computed independently for the six studies reported in DGV. Bin size = 300 bp.

https://doi.org/10.1371/journal.pone.0195334.g009

In Fig 9(A) and 9(B) recall and precision values are depicted for two mappability thresholds, Mth = 0.5 and 0.8, respectively, in the case of high sequence coverage data. As before, recall values are plotted as bar graphs for the six studies annotated in DGV and an average precision value across all the six studies is depicted as a black ‘square’ for various combinations of GC bias correction and segmentation approaches. As expected, higher recall and precision values are observed in this case (Fig 9) compared to low coverage data (Fig 7). From Fig 9(A) we observe that though high recall values are obtained for all combinations of data pre-treatment and segmentation approaches (except Median+CBS), when mappability cut-off, Mth = 0.5, the precision values are low (except Median+TVM), indicating a large number of false positives. The precision values are observed to improve for GC bias corrected data (Loess/Median) on increasing the mappability threshold to 0.8 as seen in Fig 9(B), and optimal performance is observed for Loess+TVM and Median+CBS combinations with a trade-off in recall and precision values. Thus, it is clear from this analysis that to reduce false positives, appropriate filtering of reads is required depending on the sequence coverage in DoC-based methods. Further, it may be noted that in the case of high sequence coverage data, Median approach for GC bias correction resulted in higher recall and precision values compared to Loess approach or uncorrected data, while it is other way round in the case of low sequence coverage data.

Here, we would like to mention that the six DGV studies considered for evaluating the performance of iCopyDAV are not truly gold standards and each of these methods have their own advantages and limitations in CNV detection. For example, SNP array-based techniques show high sensitivity (up to 1 bp resolution) in identifying CNVs compared to other array-based techniques such as array-Comparative Genomic Hybridization (arrayCGH), BAC and FOSMID. However, this method suffers from limited coverage of the genome with only 3 reported CNVs in Cooper et al to 15 in McCarroll et al (S3 Table), and thus misses out on many true CNVs. Further, the overlap among these four studies itself is quite low, just 2 CNVs. On individually analyzing the performance of iCopyDAV on the four SNP array-based studies reported in DGV (Altshuler et al [69], Cooper et al [70], McCarroll et al [71] and Redon et al [72]), we observe recall values ranging from 67% - 78% and precision from 13% - 70% with TVM approach for GC bias corrected data (Loess), and much higher recall values 75–100% but with very lower precision 11% - 30% on using Median+CBS approach (Mth = 0.8). Low precision values obtained for some of the SNP array-based studies is due to very small number of CNVs reported in these studies (3–15).

In Mills et al [36] study, 43 deletions are referred as ‘gold standard’ and have been collated from four studies using diverse CNV detection methods (Fosmid mapping, array-CGH, SNP-array and whole genome shotgun sequencing). High recall (~ 60%) and precision (~ 52%) values are observed with Median+CBS approach (Mth = 0.8) in detecting copy loss events. The precision improved to ~ 83% on using Loess+ TVM approach but with lower recall (~ 33%), suggesting that Median+CBS approach is better at detecting copy loss events. CNVs reported in 1000 Genome project (1000 GP) are collated from various NGS-based approaches [17,18,77–80] and experimentally verified by PCR (small CNVs) and arrayCGH methods (large CNVs). In iCopyDAV, Median+CBS combination (Mth = 0.8) resulted in best recall (37%) and precision (41%) values for 1000 GP CNVs, while Median+TVM combination (Mth = 0.8) resulted in higher precision (56%) but at the cost of low recall value (18%). A cross comparison of deletions reported in 1000 GP (validated set) with those reported in Mills et al (‘gold standard set’) and iCopyDAV predictions (Median+CBS, Mth = 0.8) using 50% overlap criteria was carried out. We observe that 32 of the CNV-deletions reported in Mills et al dataset had an overlap with 1000 GP CNVs, while 30 CNV-deletions in iCopyDAV overlapped with 1000 GP CNVs, indicating the reliability of the iCopyDAV predictions. We also observed that 33 (size: 700 bp– 431 Kb) of 1000 GP CNV-deletions are not identified by any of the six studies, including iCopyDAV. This disagreement of CNV predictions between different approaches clearly indicates the dependence of the approach used for CNV detection.

Size and Type of CNVs: The size of the CNVs reported in the six DGV studies range from 617 bp to 542 Kb, with median size 5600 bp. Hence, to see if there is any bias in the size of CNVs detected in iCopyDAV, we split the predictions into two sets: ‘small’ (≤ 5 Kb) and ‘large’ (> 5 Kb), and evaluate the performance of these two sets independently. Bias, if any, in the detection of the type of CNVs (copy gain or copy loss events) is also carried out. The F-score values for various size and type of CNVs detected in iCopyDAV are summarized in Table 4 for various combinations of data pre-treatment and segmentation approaches in the case of high coverage data. From the table it may be noted that Median+CBS approach shows optimal performance in the detection of both ‘small’ and ‘large’ CNVs, and also in the detection of ‘copy gain’ and ‘copy loss’ events. The Loess+TVM approach, on the other hand, performs equally well in the detection of ‘large’ CNVs and ‘copy gain’ events, while its performance is poor in the detection of ‘small’ CNVs. This again indicates that appropriate combination of pre-treatment and segmentation approaches need to be considered for optimal performance. On combining the results of the two segmentation approaches (TVM+CBS), we observe that the overall recall is higher with Median approach (0.51) while precision is higher on using Loess approach (0.58) for GC bias correction. No dependence on GC bias correction method (Loess/Median) is observed on the overall performance, which is marginally improved compared to the individual segmentation approaches. Thus, in iCopyDAV the user can merge results from various workflows to improve the predictive power and remove the dependence of pre-treatment methods. On considering the intersection (consensus) of CNVs obtained from Median + (TVM + CBS) and Mth = 0.8, the precision is significantly improved (~ 63%), but at the cost of low recall values (~26%). Similar observations were made on other chromosomes of NA12878 and in high coverage genome sequence data of NA12891 and NA12892 Caucasian samples (results not shown).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Performance of various combinations of GC bias correction and segmentation approaches in the detection of CNVs of different size (small/large) and type (copy gain/loss) are summarized for high sequence coverage data (Mth = 0.8).

https://doi.org/10.1371/journal.pone.0195334.t004

Comparison with other Depth-of-Coverage based Methods: To evaluate the performance of iCopyDAV with other CNV detection tools, the combined set of CNVs obtained using Median + (CBS+TVM) on Chr1 of NA12878 (35×) is compared with three popular DoC-based methods, viz., ReadDepth (RD) [13], Control-FREEC (CF) [7] and CNVnator [18]. The window size is fixed at 300 bp for all the tools (including iCopyDAV) and mappability threshold is set at 0.8 in RD, CF and iCopyDAV (there is no provision to filter reads based on mappability in CNVnator) and the results are summarized in Table 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Comparison of CNVs detected in iCopyDAV (combined approach, median + (CBS +TVM)) with ReadDepth, Control-FREEC and CNVnator (Mth = 0.8, window size 300 bp).

https://doi.org/10.1371/journal.pone.0195334.t005

In Table 5, a large variation is observed in the number (63–265), size (0.6–7.6Mb) and type of CNVs predicted by the four DoC-based approaches. Additionally, CF and CNVnator reported a large CNV event of 18 Mb at the locus chr1:123Mb-141Mb that overlaps with the centromere region, CF reported it as a ‘gain’ while CNVnator reported it as a ‘loss’. In ReadDepth and iCopyDAV this region is filtered out in the pre-treatment step as ‘unmappable’ region, and we do not consider this 18 Mb region for the comparative analysis. We also observe that large number (108) and size (6 Mb) of copy gain events are detected by ReadDepth, while CNVnator results in a large number of copy loss events (189) with a total span of 4.4 Mb. The number of CNVs detected by CF is low (especially copy loss, ~ 12) compared to the other DoC approaches, though the span is much larger (1.9 Mb) compared to that in RD and iCopyDAV. The large number of CNVs observed in CNVnator is due to no option of filtering low mappable regions (~ 204 of the predicted CNVs have an average mappability value < 0.8). On the other hand, the low number of CNVs predicted using CF is likely to be due to narrow range of GC-rich regions considered (> 0.35 and < 0.55). It is observed that ~ 18–23 CNVs predicted by other DoC-based methods fall in regions with GC content < 0.35 and > 0.55.

In Fig 10 is depicted the performance of four DoC based approaches with the six studies reported in DGV. We observe that CNVnator exhibits high recall values for all the six studies, however the precision is the least in this case (~ 0.25), indicating that majority of its predictions are false positives. It depicts high recall with 1000 GP compared to the other methods, possibly because CNVnator was one of the methods considered for annotation in this study. The high precision value of iCopyDAV and recall values comparable with ReadDepth (and only slightly lower compared to CNVnator), clearly indicates the reliability of CNV predictions by iCopyDAV. Since the six studies reported in DGV are not truly gold standard, it is possible that the CNVs predicted by iCopyDAV but not reported in any of the six studies may be true novel CNVs. An analysis revealed 28 CNVs detected by iCopyDAV that are not reported in the six DGV studies. When compared with the other three DoC methods, all of them were detected by at least one of the other three DoC methods and four CNVs (copy gain) are detected by all the four methods (S5 Table). A preliminary analysis of these four CNVs reveals that two of these CNVs at locus Chr1:14323800–143258700 and one CNV at Chr1:146173200–146197800 span the NBPF genes cluster, which is known to have evolutionary significance (discussed below in functional annotation analysis), but is not reported in DGV for NA12878 sample. The fourth CNV at Chr1:85752900–85778400 span the gene DDAH1, a key regulator in the breakdown of asymmetric dimethylarginine (ADMA). It has been shown that raised levels of ADMA are strongly correlated with increased risk of cardiovascular diseases [81].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10. Performance of iCopyDAV with three DoC-based methods, ReadDepth, Control-FREEC, and CNVnator (using default parameters) is shown for high sequence coverage data (35×) of Chr 1 of NA12878 sample.

https://doi.org/10.1371/journal.pone.0195334.g010

Computation time: The iCopyDAV platform is computationally very efficient and can even be run on a desktop. The entire pipeline (data pre-treatment, variant calling and annotation steps) on Chromosome 1 of high coverage data (35×) took ~ 14 minutes CPU time on a 32 AMD Opteron with 2.4 GHz machine. Data pre-treatment step (computing read depth, filtering low mappable regions and correcting for GC bias) is computationally the most demanding step, requiring about ~ 10 minutes, while segmentation and variant calling steps took about ~ 4 minutes and the other steps just a few seconds. On the same machine the processing time required by ReadDepth and CNVnator is ~ 110 minutes and 130 minutes, respectively (~ 8–9×), while the time taken by Control-FREEC is comparable, ~ 15 minutes. We recommend a minimum requirement of 4 CPU machine running at 1.6 GHz for efficient performance of iCopyDAV.

Visualization and functional interpretation of predicted CNVs: The distribution of CNVs along the chromosome can give insights about CNV-enriched regions, and association with other functional elements such as protein-coding genes, etc. which are likely to have phenotypic characteristics. In iCopyDAV, using the function plot, the distribution of predicted CNVs can be visualized along the length of the chromosome or any region of interest (user-defined coordinates). The CNVs obtained using (TVM+CBS) approach for chromosome 1 of NA12878 sample on Median GC bias corrected data and mappability threshold, Mth = 0.8 are shown in Fig 11. A high resolution graph displaying copy gain events in ‘red’ and copy loss events in ‘blue’ on an ideogram (upper) and a 2d-plot (lower) is generated as shown in Fig 11(A). In Fig 11(B) the distribution of the CNVs at the locus 1q21.1 (Chr1:142,400,001–148,000,000) clearly indicates a cluster of CNVs encompassing NBPF-gene family, enriched with segmental duplications.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 11.

The location of CNVs, gain in ‘red’ and loss in ‘blue’ is shown (a) along Chr 1 of NA12878 sample, and (b) at the locus 1q21.1 spanning NBPF gene-family, which is rich in segmental duplications.

https://doi.org/10.1371/journal.pone.0195334.g011

To enable the users to analyze the biological significance of predicted CNVs, functional and structural annotations are provided by the function annotate in iCopyDAV. Four categories of annotations are reported, viz., functional elements (genes, exons, promoters, lncRNAs, miRNAs, etc.), clinical relevance, structural features (segmental duplications, tandem repeats, etc.) and mapping to known CNVs reported in Database of Genetic Variants (DGV). Below we briefly discuss the functional analysis of CNVs obtained by Median+ (TVM+CBS) approach (Mth = 0.8) on using this module. The complete list of CNVs detected on Chr 1 is given in S6 Table along with annotations.

a. Functional elements annotation: Using the function annotate, we identified 46 protein-coding genes (26 gain and 20 loss) mapping to 54 CNVs (35 copy gain and 19 copy loss events). To understand the role of these CNV-genes and identify any genotypic-phenotypic characteristics associated with Caucasian population, we carried out a systematic analysis of these genes. Copy loss of the Late Cornified Envelop-gene cluster (LCE1D, LCE1E, LCE3B, LCE3C and LCE3D) at locus 1q21.3 that is known to be part of epidermal differentiation complex is observed. These are involved in keratinization which in turn effects the melanosome distribution. In a study by de Cid et al [82], copy deletion of LCE3B and LCE3C is shown to be a risk factor for psoriasis in Western European patients. Copy loss of gene BMP8A at 1p34.2 is associated with reduced risk of Ankylosing Spondilytis in Korean populations [83]. Other functional elements encompassing the CNVs include eight long non-coding RNA (lncRNAs), TCONS_l2_00001057, TCONS_l2_00001060, TCONS_l2_00001061, TCONS_00001651, TCONS_00001652, TCONS_00002569, TCONS_00001785 and TCONS_00000355, spanning 5 CNVs (all gain) at loci 1p36.13 and 1q32.1. The functional significance of these lncRNAs is unknown, however, lncRNAs and their interaction with miRNA obtained from lncBase [84] reveal that 3 lncRNAs, TCONS_l2_00001060, TCONS_l2_00001061 and TCONS_00002569 are associated with risk of Breast neoplasm. One enhancer (element_809) spanning copy loss at 1p22.3 is observed in the vicinity of CNV-gene LMO4, which acts as a cell cycle regulator. Copy number variation of miRNA is known to affect the binding and regulation of miRNA target genes involved in various biological processes and may contribute to lethal diseases including cancer. We identified one miRNA-gene, MIR4654, spanning copy loss at 1q23.3 and three miRNA target sites (mir-368, mir-23 and mir-1/206) spanning 3′ UTR of CNV-gene SEC22B at 1q21.1 (gain). Gene SEC22B, regulated by mir-206, has been identified to play a role in the progression and development of Alzheimer’s disease [85]. Thus, we see that a systematic analysis of the functional regions overlapping with CNV predictions can help in identifying predisposition to diseases and in carrying out genotype-phenotype association studies.

b. Clinical relevance: For identifying the clinical relevance of CNV-genes, information from four resources, viz., ClinVar, OMIM, ExAC and DECIPHER, is collated and provided in iCopyDAV. The annotation results from these four resources for Chr 1 of NA12878 sample is summarized in the S6 Table. ClinVar reported three CNVs at loci 1q21.1 (2 copy gain events) and one at locus 1q41 (copy loss) to be ‘Pathogenic’. From literature review it is found that autism spectrum disorders are implicated with gain at 1q21.1 (5100 bp), and Loeys-Dietz syndrome associated with copy loss at 1q41 (600 bp). We also observe 26 CNVs identified at locus 1p36 characterized as ‘uncertain significance’ and association to Ductal Breast Carcinoma (DBC) phenotype from ClinVar. These CNVs (copy gain) span six genes CDK11A and CDK11B, NBPF1, MST1P2, CROCCP2 and ESPNP genes, of which the last three are pseudogenes. Two CDK family genes, CDK11A and CDK11B, are known to play multiple roles in cell cycle progression, cytokinesis and apoptosis, while the gene NBPF1 acts as a tumor suppressor by arresting cell cycle at G1 stage in Neuroblastoma [86]. Six copy loss genes (ST3GAL3, VANG, SLC1A7, LMO4, ERI3, XPR1) exhibit genic intolerance score > 0 in ExAC database, of which three genes (LMO4, ERI3 and XPR1) are also characterized as intolerant with haploinsufficient score ranging from 0–25% in DECIPHER. This implicates possible deleterious effect of copy variation of these genes. For instance, gene LMO4 (1p22.3) is a cell cycle transcriptor regulator and plays an important role in controlling cell cycle and cell proliferation [87]. The under-expression of this gene is linked with reduced breast cancer cell proliferation and cell cycle arrest at G2/M stage [87]. A CNV is found to be spanning intron-2 of the gene XPR1 (1q25.3), which is known to play a role in phosphate homeostasis, mediating phosphate export from the cell. Another four copy loss genes (VAV3, WDR47, NOS1AP and MSH4) are reported to have a haploinsufficient score < 25%, indicating these copy deletions may be pathogenic. Copy deletions are observed in the intronic regions for 9 of these haploinsufficient genes (except LMO4), however, their effect on the phenotype is not yet known. In recent studies, it has been shown that CNVs in introns regulate the expression of the genes directly through change in the intronic length, or variations in the regulatory regions (enchancers or CTCF binding sites) and non-coding functional RNA genes spanning them.

c. Structural annotation: About ~ 36% of predicted CNVs (42 copy gain and 2 copy loss) are observed to be overlapping with segmental duplications (SD) on Chr 1 of NA12878 Caucasian sample. The CNVs in the vicinity of SDs indicate their mechanism of formation via Non-allelic Homologous Recombination (NAHR). NAHR occurs by alignment followed by crossing over between two non-allelic, highly similar DNA sequences (repeats). Repeats on the same chromosome and in the same orientation mediate a duplication and/or deletion. In accordance with previous studies [88,89], NBPF gene-cluster spanning SD segments at 1q21.1 locus is identified: NBPF8 (gain: 3.7), NBPF10 (gain: 3.1), NBPF12 (gain: 5.2), NBPF20 (gain: 5.8) and NBPF25P (gain: 5.3), shown in Fig 11(B). These CNV-genes are reported to be formed as a consequence of human-lineage specific copy amplification which occurred about ~100 million years ago. The CNV-gene, NBPF1 (gain: 5.9) spanning 1p36.2 is a known primate-specific amplification, a common copy gain event observed in human and apes. These recently evolved NBPF gene clusters are characterized by highly conserved domains of DUF1220. O’Blenesis et al [89] showed a high copy number of DUF1220 protein domain in humans (272 copies) compared to chimpanzee (90–125 copies) and similar numbers are observed in our analysis. These amplifications are fixed in the population as a result of positive selection and have been identified to contribute in neuron number, brain size and cognitive ability that occurred specifically in humans (and other primates) [90]. Low copy number of DUF1220 is shown to result in brain growth abnormalities, leading to microencephaly, while high copy numbers of the domains are associated with macroencephaly [91]. A number of CNVs are also observed to be overlapping with tandem and interspersed repeats: 9 copy gain and 5 copy loss events spanning tandem repeats, with an average of 4 tandem repeats per CNV (range: 1–32), and 5 copy gain and 4 copy loss events have their breakpoints overlapping the interspersed repeats, with 1 CNV overlapping Long terminal repeat retrotransposons (LTR) and the rest with LINEs (listed in S6 Table). Six out of nine copy gain events at Chr1:121053000–121055700, Chr1:121056900–121057500, Chr1:121185900–121186800, Chr1:147299700–147302400, Chr1:152953200–152953800 and Chr1:222266700–222268800 have a > 90% overlap with tandem repeats and two copy gain events at Chr1:5657700–5658600 and Chr1:146816400–146819400 have an overlap > 90% with LINES. Thus, the extent of overlap with other known structural variants can help in understanding the mechanism of the predicted CNV (partial overlap at breakpoints), or a common CNV fixed in population (> 90% overlap).

d. Known CNVs: After discarding CNVs having > 90% overlap with other structural variants based on our structural annotation analysis, we have a total of 112 CNVs predicted on Chr 1 of NA12878 Caucasian sample. Of these, 106 (~ 95%) are reported in Database of Genomic Variants (DGV) on considering ≥ 50% overlap criteria. The six CNVs which are not reported in DGV include two copy gain (Chr1:45248700–45249300 and Chr1:81627600–81630300) and four copy loss (Chr1:44519700–44520300, Chr1:87567000–87567900 (spanning LMO4 gene), Chr1:211583100–211583700 and Chr1:219713700–219714300), spanning a total of 6600 bp. Since the majority of CNVs reported in DGV are obtained from array-based techniques, these smaller CNVs, ranging from 600 bp to 2700 bp are likely to be missed by these approaches. Among these six CNVs, copy gain at Chr1:81627600–81630300 (span intron-1 of gene LPHN2) and copy loss at Chr1:211583100–211583700 are also identified as copy variant regions by ReadDepth and CNVnator, respectively.

Thus, using the annotation module in iCopyDAV, we identified 43 CNVs that may be of clinical importance (high priority) and 36 CNVs that may have certain effect on gene expression (medium priority) (see S6 Table). It may be noted that ~ 70% of these CNVs are also predicted by at least one of the three other DoC-based methods (ReadDepth, CF and CNVnator), and ~ 94% also reported in DGV, indicating the reliability of iCopyDAV in identifying and categorizing the variants.