Significance

Abstract

Results

We compared the two NGS techniques, performing WES with the Agilent Sure Select Human All Exon kit 71 Mb (v4 + UTR) and WGS with the Illumina TruSeq DNA PCR-Free sample preparation kit on blood samples from six unrelated Caucasian patients with isolated congenital asplenia [Online Mendelian Inheritance in Man (OMIM) no. 271400] (6). We used the genome analysis toolkit (GATK) best-practice pipeline for the analysis of our data (17). We used the GATK Unified Genotyper (18) to call variants, and we restricted the calling process to the regions covered by the Sure Select Human All Exon kit 71 Mb plus 50 base pairs (bp) of flanking sequences on either side of each of the captured regions, for both WES and WGS samples. These regions, referred to as the WES71+50 region, included 180,830 full-length and 129,946 partial exons from 20,229 protein-coding genes (Table 1). There were 65 million reads per sample, on average, mapping to this region in WES, corresponding to a mean coverage of 73× (Table S1), consistent with the standards set by recent large-scale genomic projects aiming to decipher disease-causing variants by WES (9, 22, 23). On average, 35 million reads per sample mapped to this region by WGS, corresponding to a mean coverage of 39× (Table S1). We first focused on the analysis of single-nucleotide variants (SNVs). The mean (range) number of SNVs detected was 84,192 (82,940–87,304) by WES and 84,968 (83,340–88,059) by WGS. The mean number of SNVs per sample called by both methods was 81,192 (~96% of all variants) (Fig. S1A). For 99.2% of these SNVs, WES and WGS yielded the same genotype, and 62.4% of these concordant SNVs were identified as heterozygous (Fig. S1B). These results are similar to those obtained in previous WES studies (1, 5, 22). Most of the remaining SNVs (329 of 415) with discordant genotypes for these two techniques were identified as homozygous variants by WES and as heterozygous variants by WGS (Fig. S1B).

Table 1.

Specific regions of the genome covered by WES using the 71-Mb ± 50 bp kit

Exon status	Exons from protein-coding genes		lincRNA	miRNA	snoRNA
	All	CCDS
Fully included	180,830	147,131	554	1,171	252
Partially included	129,946	34,892	855	94	93
Fully excluded	64,921	5,762	25,389	1,782	1,111
Total	375,697	187,785	26,798	3,047	1,456

Open in a separate window

Four types of genomic units were analyzed: exons from protein-coding genes, microRNA (miRNA) exons, small nucleolar RNA (snoRNA) exons, and large intergenic noncoding RNA (lincRNA) exons as defined in Ensembl Biomart (19). We determined the number of these units using the R Biomart package (20) on the GRCh37/hg19 reference. We first considered exons from protein-coding genes (denoted as “All”) obtained from Ensembl. The intronic essential splice sites (i.e., the two intronic bp at the intron/exon junction) were not included in our analysis of exons. Then we focused on protein-coding exons with a known CDNA coding start and CDNA coding end that were present in CCDS transcripts (21). For the counts, we excluded one of the duplicated units of the same type, or units entirely included in other units of the same type (only the longest unit would be counted in this case). We then determined the number of the remaining units that were fully or partly covered when considering the genomic regions defined by the Agilent Sure Select Human All Exon kit 71 Mb (v4 + UTR) with the 50-bp flanking regions.

We then investigated, in WES and WGS data, the distribution of the two main parameters assessing SNV quality generated by the GATK variant-calling process (18): coverage depth (CD), corresponding to the number of aligned reads covering a single position; and genotype quality (GQ), which ranges from 0 to 100 (higher values reflect more accurate genotype calls). We also assessed the minor-read ratio (MRR), which was defined as the ratio of reads for the less covered allele (reference or variant allele) over the total number of reads covering the position at which the variant was called. Overall, we noted reproducible differences in the distribution of these three parameters between WES and WGS. The distribution of CD was skewed to the right in the WES data, with a median at 50× but a mode at 18×, indicating low levels of coverage for a substantial proportion of variants (Fig. 1A). By contrast, the distribution of CD was normal-like for the WGS data, with the mode and median coinciding at 38× (Fig. 1A). We found that 4.3% of the WES variants had a CD of <8×, versus only 0.4% of the WGS variants. The vast majority of variants called by WES or WGS had a GQ close to 100. However, the proportion of variants called by WES with a GQ of <20 (3.1%) was, on average, twice that for WGS (1.3%) (Fig. 1B). MRR followed a similar overall distribution for WES and WGS heterozygous variants, but peaks corresponding to values of MRR of 1/7, 1/6, 1/5, and 1/4 were detected only for the WES variants (Fig. 1C). These peaks probably corresponded mostly to variants called at a position covered by only 7, 6, 5, and 4 reads, respectively. The overall distributions of these parameters indicated that the variants detected by WGS were of higher and more uniform quality than those detected by WES.

Open in a separate window

Fig. 1.

Distribution of the three main quality parameters for the variations detected by WES or WGS. (A) Coverage depth (CD), (B) genotype quality (GQ) score, and (C) minor-read ratio (MRR). For each of the three parameters, we show the average over the six WES (red) and the six WGS (turquoise) samples in SNVs (Left), insertions (Center), and deletions (Right).

Next, we looked specifically at the distribution of these parameters for the variants with genotypes discordant between WES and WGS, denoted as discordant variants. The distribution of CD for WES variants showed that most discordant variants had low coverage, at about 2×, with a CD distribution very different from that of concordant variants (Fig. S2A). Moreover, most discordant variants had a GQ of <20 and an MRR of <0.2 for WES (Fig. S2B). By contrast, the distributions of CD, GQ, and MRR were very similar between WGS variants discordant with WES results and WGS variants concordant with WES results (Fig. S2). All these results indicate that the discordance between the genotypes obtained by WES and WGS was largely due to the low quality of WES calls for the discordant variants. We therefore conducted subsequent analyses by filtering out low-quality variants. We retained SNVs with a CD of ≥8× and a GQ of ≥20, as previously suggested (24), and with an MRR of ≥0.2. Overall, 93.8% of WES variants and 97.8% of WGS variants satisfied the filtering criterion (Fig. S3A). We recommend the use of these filters for projects requiring high-quality variants for analyses of WES data. More than half (57.7%) of the WES variants filtered out were present in the flanking 50-bp regions whereas fewer (37.6%) of the WGS variants filtered out were present in these regions. In addition, 141 filtered WES variants and 70 filtered WGS variants per sample concerned the 2 bp adjacent to the exons, which are key positions for splicing. After filtering, the two platforms called an average of 76,195 total SNVs per sample, and the mean proportion of variants for which the same genotype was obtained with both techniques was 99.92% (range, 99.91–99.93%).

We then studied the high-quality (HQ) variants satisfying the filtering criterion but called by only one platform. On average, 2,734 variants (range, 2,344–2,915) were called by WES but not by WGS (Fig. S3A), and 6,841 variants (5,623–7,231) were called by WGS but not WES (Fig. S3A). We used Annovar software (25) to annotate these HQ variants as coding variants: i.e., variants overlapping a coding exon, this term referring to the coding part of the exon but not the UTR portion. Overall, 651 of the 2,734 WES-exclusive HQ variants and 1,113 of the 6,841 WGS-exclusive HQ variants were coding variants (Fig. S3A). Using the Integrative Genomics Viewer (IGV) tool (26), we noticed that most WES-exclusive HQ variants were also present on the WGS tracks with quality criteria that were above our defined thresholds. We were unable to determine why they were not called by the Unified Genotyper. We therefore used the GATK Haplotype Caller to repeat the calling of SNVs for the WES and WGS experiments. We combined the results obtained with Unified Genotyper and Haplotype Caller and limited subsequent analyses to the variants called by both callers. The mean number (range) of HQ-coding SNVs called exclusively by WES fell to 105 (51–140) per sample whereas the number called exclusively by WGS was 692 (506–802) (Fig. S3B), indicating that calling issues may account for ~80% of initial WES-exclusive coding variants and ~40% of initial WGS-exclusive coding variants. The use of variants identified by both Unified Genotyper and Haplotype Caller (i.e., the intersection) therefore seemed to increase the reliability and accuracy of calls, which served the main purpose of our study. Using variants identified by one or both callers only (i.e., the union), rather than the intersection, would increase the number of false positives but decrease the number of false negatives, and this strategy might be of interest in specific contexts. Using the intersection, we obtained a mean of 74,398 HQ SNVs (range, 72,867–77,373) called by both WES and WGS (Fig. S3B), 19,222 (18,823–20,024) of which were coding variants. The quality and distribution of the CD, GQ, and MRR obtained with this combined calling process were similar to those previously reported for Unified Genotyper (Fig. S4).

We further investigated the HQ-coding variants called exclusively by one method when the intersection of the two callers was used. We were able to separate the variants identified by only one technique into two categories: (i) those called by a single method and not at all by the other, which we refer to as fully exclusive variants, and (ii) those called by both methods but filtered out by one method, which we refer to as partly exclusive variants. Of the HQ-coding variants identified only by WES (105, on average, per sample), 61% were fully exclusive and 39% were partly exclusive. Of those identified only by WGS (692, on average), 21% were fully exclusive and 79% were partly exclusive. We performed Sanger sequencing on a random selection of 170 fully and partly exclusive WES/WGS variants. Of the 44 fully exclusive WES variants successfully Sanger-sequenced, 40 (91%) were absent from the true sequence, indicating that most fully exclusive WES variants were false positives (Table 2 and Dataset S1). By contrast, 39 (75%) of the 52 Sanger-sequenced fully exclusive WGS variants were found in the sequence, with the same genotype as predicted by WGS (including 2 homozygous), and 13 (25%) were false positives (Table 2 and Dataset S1). These results are consistent with the observation that only 27.2% of the fully exclusive WES variants were reported in the 1000 Genomes database (27) whereas most of the fully exclusive WGS variants (84.7%) were present in this database, with a broad distribution of minor allele frequencies (MAFs) (Fig. S5A). Similar results were obtained for the partly exclusive variants. Only 10 (48%) of the 21 partly exclusive WES variants (including 3 homozygous) were real whereas all (100%) of the 24 partly exclusive WGS variants (including 8 homozygous) were real.

Table 2.

Results of Sanger sequencing for 170, 82, and 82 WES and WGS fully and partly exclusive SNVs, insertions, and deletions, respectively

Type of variation	Avg no. per sample	No. successfully sequenced/total no. sequenced	No. (%) of real variants	Estimated no. of real variants*
SNV
WES
FE	64	44/56	4 (9)	6
PE	41	21/27	10 (48)	20
Total	105	65/83	14 (22)	26
WGS
FE	145	52/60	39† (75)	109
PE	547	24/27	24 (100)	547
Total	692	76/87	63† (83)	656
Insertions
WES
FE	5	9/10	3 (33)	2
PE	10	18/21	13 (72)	7
Total	15	27/31	16 (59)	9
WGS
FE	25	17/26	6 (35)	9
PE	36	20/25	11† (55)	20
Total	61	37/51	17† (46)	29
Deletions
WES
FE	7	15/17	6† (40)	3
PE	10	16/23	10 (62.5)	6
Total	17	31/40	16† (52)	9
WGS
FE	15	15/20	6 (40)	6
PE	29	17/22	13 (76)	22
Total	44	32/42	19 (59)	28

Open in a separate window

FE, fully exclusive variations; PE, partly exclusive variations.

^*Estimated numbers of real variants and false positives were computed on the basis of real and false positives proportions applied on the average number of variants per sample.

^†Number includes one variant that was present by Sanger sequencing but the genotype (heterozygous or homozygous) was different between Sanger sequencing and WES or WGS.

Using these findings, we estimated the overall numbers of false-positive and false-negative SNVs detected by these two techniques. WES identified a mean of 26 estimated real-coding variants per sample (including 5 homozygous) that were missed by WGS, and a mean of 79 estimated false-positive variants. WGS identified a mean of 656 estimated real-coding variants per sample (including 104 homozygous) that were missed by WES, and a mean of 36 estimated false-positive variants. We noted that most of the false-positive fully exclusive WGS SNVs were located in the three genes (ZNF717, OR8U1, and SLC25A5) providing the largest number of exclusive variants on WGS. Further investigations of the reads corresponding to these variants on the basis of blast (28) experiments strongly suggested that these reads had not been correctly mapped. Overall, we found that the majority of false-positive WGS fully exclusive variants (11/13) and only a minority of false-positive WES fully exclusive variants (4/40) could be explained by alignment and mapping mismatches. We then determined whether the exclusive WES/WGS SNVs were likely to be deleterious and affect the search for disease-causing lesions. The distribution of combined annotation-dependent depletion (CADD) scores (29) for these variants is shown in Fig. S5B. About 38.6% of the partly exclusive WES variants and 29.9% of the partly and fully exclusive WGS variants, which were mostly true positives, had a Phred CADD score of >10 (i.e., they were among the 10% most deleterious substitutions possible in the human genome) and might include a potential disease-causing lesion. We also found that 54.6% of fully exclusive WES SNVs, most of which were false positives, had a Phred CADD score of >10 and could lead to useless investigations. Overall, these results suggest that WGS is more accurate and efficient than WES for identifying true-positive SNVs in the exome.

We then compared WES and WGS for the detection of indels, following a strategy similar to that used for SNVs, and using Haplotype Caller, which is more appropriate than Unified Genotyper for indel detection (30). The mean number (and range) of insertions detected per sample was 5,795 (5,665–5,984) with WES and 6,443 (6,319–6,763) with WGS (Fig. S3C). On average, 5,313 insertions (85%) were called by both methods. The mean number (range) of deletions detected per sample was 7,530 (7,362–7,735) with WES and 6,259 (6,150–6,548) with WGS (Fig. S3D). On average, 5,383 deletions (74%) were detected by both methods. As for SNVs, the distributions of CD, GQ, and MRR for indels were of higher quality in WGS than in WES (Fig. 1). In particular, the distribution of CD was skewed to the right for both insertions and deletions. After applying the same filters as for SNVs (removing indels with a CD of <8×, a GQ of <20, and an MRR of <0.2), we obtained a mean number of HQ insertions per sample of 4,104 (3,972–4,285) called by both WES and WGS (99.3% with the same genotype), 298 (248–413) called only by WES (5.3%), and 1,197 (974–1,400) called only by WGS (21.4%). We found that 4,121 HQ deletions (3,996–4,308) were called by both methods (99.5% with the same genotype), with the mean number of WES-exclusive deletions (1,189, 1,015–1,419) similar to that of WGS exclusive deletions (1,067, 871–1,215) (Fig. S3). We also investigated the HQ-coding indels, which we defined as indels involving at least 1 bp included in a protein-coding exon. The mean number of HQ-coding insertions per sample called by both WES and WGS was 247 (230–266). On average, 15 HQ-coding insertions per sample were identified only by WES (33% of which were fully exclusive) and 61 were identified only by WGS (41% were fully exclusive). The mean number of HQ-coding deletions per sample called by both WES and WGS was 240 (225–265). On average, 17 HQ coding deletions per sample were identified only by WES (41% of which were fully exclusive) and 44 were identified only by WGS (34% were fully exclusive).

The distribution of HQ indels by size is shown in Fig. S6. Most HQ insertions (57.1%) and deletions (57.3%) involved a single base pair. We hypothesized that indels causing a frameshift in a coding region would be under stronger evolutionary constraints, and we investigated whether such coding regions were enriched in indels of a size corresponding to multiples of 3 bp. Consistent with this hypothesis, we observed that the number of insertions or deletions of a multiple of 3 bp was much higher for coding indels than for noncoding indels: >40% for coding indels and <12% for noncoding indels (Fig. S6 B and C). These percentages were similar for coding indels called exclusively by WES or WGS (Fig. S6 B and C), which suggests that most of these indels called exclusively by one method could be real. We Sanger-sequenced a random selection of 164 coding indels exclusively called by WES or WGS. We found that the Sanger sequences of 32 of the 58 successfully sequenced WES-exclusive indels (55.2%) were consistent with WES findings (Table 2 and Dataset S1). Similarly, 36 (52.2%) of the 69 Sanger sequences obtained for WGS-exclusive indels were consistent with WGS findings (Table 2 and Dataset S1). These Sanger results indicate that, by contrast to what was found for SNVs, the estimated proportion of false positives among exclusive HQ-coding indels was equally high for both WES and WGS, at almost 50%. More indels were detected exclusively by WGS (61 + 44 = 105) than exclusively by WES (15 +17 = 32) so the number of real coding indels per sample detected by WGS and missed by WES was estimated to be higher (57 indels) than that detected by WES and missed by WGS (18 indels).

The results for indels should be interpreted bearing in mind that the Sanger sequencing of indels is more difficult than that of SNVs, for three main reasons. Indels can be complex, with a combination of insertion and deletion. The regions in which indels occur are often the hardest to sequence. For example, it is difficult to identify by Sanger sequencing a heterozygous deletion of 1 adenine (A) in a stretch of 20 consecutive As. And the analysis of the sequencing is harder, especially for heterozygous indels, where it requires reconstructing manually the two alleles from long stretches of overlapping peaks. In this context, calls were difficult to make for a number of indels. We provide more detailed information about the analysis of the Sanger sequencing of indels in Materials and Methods. However, there are three arguments to suggest that our general findings for the WES/WGS comparison are valid. First, it was equally difficult to analyze Sanger sequences for WES-exclusive and WGS-exclusive indels. Second, the proportion of false positives was lower for partly exclusive than for fully exclusive indels for both WES (67.6% vs. 37.5%) and WGS (64.8% vs. 37.5%) (Table 2), as observed for SNVs. Finally, our Sanger results are consistent with the observed similar fractions of WES-exclusive and WGS-exclusive indels reported in the 1000 Genomes database (Fig. S5C). Overall, these results indicate that the proportion of false-positive coding indels is similar for both WES and WGS.

The last step in our study was the comparison of WES and WGS for the detection of CNVs. The methods currently used to identify CNVs from WES data were already known to perform poorly for a number of technical reasons (31), including the fact that CNV breakpoints could often lie outside the regions targeted by the exome kit (32). A recent study comparing four WES-based CNV detection tools showed that none of the tools performed well and that they were less powerful than WGS on the same samples (16). We therefore restricted our analysis to the comparison of two classical WES-based methods, Conifer (33) and XHMM (32), with a well-known WGS-based method, Genome STRiP (34), for the detection of deletions in our six samples. As expected, more deletions were detected by WGS than by WES, with a total of 113 deletions (mean size, 23.7 kb; range 0.8–182.5 kb) detected over the six samples by Genome STRiP, 44 (mean, 45.6 kb ; 0.3–644 kb) detected by Conifer, and 30 (77.6 kb; 0.1–2063 kb) detected by XHMM. Ten of the 113 deletions detected by WGS (9%) were identified by Conifer, and 8 (7%) were detected by XHMM, including 4 detected by both Conifer and XHMM, consistent with the low concordance rates previously reported (16). We hypothesized that most true CNVs are common and should therefore be reported in public databases for CNVs, such as the Database of Genomic Variants (DGV) (35). The vast majority of deletions detected by WGS (105/113, 93%) were present in DGV, suggesting that most CNVs detected by WGS are true CNVs. Ten of the 14 deletions detected by WGS and at least one WES-based method (71.4%) were reported in DGV whereas only 5/34 (14.7%) Conifer-exclusive and 5/22 (22.7%) XHMM-exclusive deletions were present in DGV. For 110 of the 113 deletions detected by WGS, one (24) or both (86) putative breakpoints were located outside the exome-capture regions, providing a plausible explanation. Two of the three deletions with both breakpoints in the exome regions corresponded to the same 1.1-kb deletion identified in two different patients. This deletion was not identified by Conifer and XHMM, probably because only 3.4% of the 1.1 kb was covered by the exome kit. The third deletion was 29.1 kb long and was identified by both Conifer and XHHM. We also observed that most of the deletions (10/14) detected by WGS and at least one WES method belonged to the 20% of regions best covered by the exome kit. These results highlight the importance of both the size and the coverage ratio of the deletion by the exome kit, for optimal detection with the WES analysis methods currently available.

Finally, we investigated in more detail the coding regions and corresponding genes that were either poorly covered or not covered at all by the WES kit we used. We first determined, for each sample, the 1,000 genes with the lowest WES coverage. Up to 75.1% of these genes were common to at least four samples, and 38.4% were present in all six individuals. The percentage of exonic base pairs with more than 8× coverage for these 384 genes was, on average, 73.2% for WES (range, 0–86.6%) and 99.5% for WGS (range, 63.6–100%) (Dataset S2A). These genes with low WES coverage in all patients comprised 47 genes underlying Mendelian diseases, including three genes (IMPDH1, RDH12, NMNAT1) responsible for Leber congenital amaurosis, and two genes (IFNGR2, IL12B) responsible for Mendelian susceptibility to mycobacterial diseases (Dataset S2A). We then focused on the protein-coding exons that were fully outside the WES71+50 region (Table 1). We restricted our analyses to the highest-quality protein-coding exons, those present in a consensus coding sequence (CCDS) transcript (21), with known start and end points of the coding sequence in the cDNA. These CCDS exons comprise a total of 46,227,845 bases that belong to translated regions, including 8,566,582 (18.5%) lying outside the WES71+50 region. The average CADD score for all possible variants was lower among the bases of the CCDS protein-coding exons not targeted by the kit (median, 7.362) compared with the bases targeted by the exome kit (median, 14.87) (Fig. S5D). However, an important proportion (41.5%) of possible variants at bases not targeted by the kit had a CADD score of >10, suggesting that potentially deleterious variations (with high CADD scores) might be missed by WES. We also found that 5,762 CCDS exons (3.1%), from 1,223 genes were located entirely outside the WES71+50 region. Of these genes, 140 were associated with Mendelian diseases (Dataset S2B). We conducted the same analyses with the latest Agilent all-exon kit (v5+utr; 75 Mb), taking 50-bp flanking regions into account. We found that 2,879 CCDS exons (1.5%) were entirely excluded and that these exons belonged to 588 genes, including 50 associated with monogenic disorders, such as BRCA1, the gene most frequently implicated in breast cancer. We also noted that, for these 2,879 CCDS exons, WGS detected a mean (range) of 436 (424–457) HQ SNVs and 16 (12–18) HQ indels.

Discussion

Our findings confirm that WGS provides a much more uniform distribution of sequencing-quality parameters (CD, GQ, MRR) than WES, as recently reported (14). The principal factors underlying the heterogeneous coverage of WES are probably related to the hybridization/capture and PCR-amplification steps required for the preparation of sequencing libraries for WES (36). We also Sanger-sequenced a large number of variations to obtain a high-resolution estimate of the number of false positives and false negatives obtained with WES and WGS (Fig. 2). All these analyses demonstrate that WGS can detect hundreds of potentially damaging coding SNVs per sample (~3% of all HQ-coding variants detected by WGS), about 16% of which are homozygous, including some in genes known to be involved in Mendelian diseases, that would have been missed by WES despite being located in the regions targeted by the exome kit (Fig. 2). The results are less clear-cut for indels and should be interpreted more cautiously because the current methods for identifying indels on the basis of both NGS and Sanger data are less reliable than those for SNVs. Our findings also confirm that WES is not currently a reliable approach for the identification of CNVs, due to the noncontiguous nature of the captured exons, in particular, and the extension of most CNVs beyond the regions covered by the exome kit. In addition to the variants in the targeted regions missed by WES, a large number of exons from protein-coding genes, and noncoding RNA genes were not targeted by WES despite being fully sequenced by WGS (Table 1). It was noticeable that exons pertaining to CCDS transcripts were better covered than those that did not (Table 1). Finally, mutations outside protein-coding exons, or not in exons at all, might also affect the actual exome (the parts covered or not covered by WES) because mutations in the middle of long introns might impair the normal splicing of exons (37). These mutations would be missed by WES but would be picked up by WGS (and selected as candidate mutations if the mRNAs were studied in parallel, for example by RNAseq).

Open in a separate window

Fig. 2.

Diagram of the losses of single nucleotide variants (SNVs) at various levels associated with the use of WES. (A) Exons that were covered by the Agilent Sure Select Human All Exon kit 71 Mb (V4 + UTR) with the 50-bp flanking regions. Exons fully covered are represented by boxes filled entirely in red; exons partly covered, by boxes filled with red stripes; and exons not covered at all, by white boxes. Numbers are shown in Table 1. Exons from protein-coding genes include exons encoding exclusively or partially UTRs, as well as exons mapping entirely to coding regions. (B) Number of high-quality coding SNVs called by WES and WGS (white box), by WES exclusively (red box), or by WGS exclusively (turquoise box). Details for the SNVs called exclusively by one method are provided below the figure. TRUE, estimate based on SNVs detected by Sanger sequencing; FALSE, estimate based on SNVs that were not detected by Sanger sequencing (Table 2). Darker boxes (red, gray, or turquoise) represent homozygous SNVs. Lighter boxes (red, gray, or turquoise) represent heterozygous SNVs.

Overall, our results show that WES and WGS perform very well for the detection of SNVs and indels because more than 96.0% of HQ SNVs and 78.0% of HQ indels in the coding regions covered by the exome kit were called by both methods. The detailed analysis of the variants called exclusively by one approach showed that WGS was slightly but significantly more powerful than WES for detecting variants in the regions covered by the exome kit, particularly for SNVs. In addition, WGS is certainly more appropriate for detecting CNVs because it covers all breakpoints and, of course, could detect variations in RNA- and protein-coding exome regions not covered by the exome kit. WGS currently costs two to three times as much as WES, but most of the cost of WGS (>90%) is directly related to sequencing whereas WES cost is mainly due to the capture kit. Sequencing costs have greatly decreased and are expected to decrease faster than the cost of the capture kit. As an example, if sequencing costs were to decrease by 60% and capture kit costs remained stable, then the cost of WGS would approach that of WES. The cost of data analysis and storage should also be taken into account. In this rapidly changing economic context, specific cost/benefit studies are required and should take into account whether these NGS investigations are conducted for diagnostic or research purposes (15). These global studies should facilitate individual decisions, determining whether the better detection of SNVs, indels, and CNVs merits the additional cost of WGS, if WGS remains more expensive than WES. Finally, we carried out a detailed characterization of the variants and genes for which the two methods yielded the most strongly contrasting results, providing a useful resource for investigators trying to identify the most appropriate sequencing method for their research projects. We provide open access to all of the scripts used to perform this analysis at the software website GITHUB (https://github.com/HGID/WES_vs_WGS). We hope that researchers will find these tools useful for analyses of data obtained by WES and WGS (4, 7, 9, 11, 38), two techniques that will continue to revolutionize human genetics and medicine in the foreseeable future.

Materials and Methods

Supplementary Material

Acknowledgments

Footnotes

References