The development of next-generation sequencing (NGS) has brought remarkable growth in our understanding of human genome variants through comprehensive characterization. On top of that, various efforts have been made to find associations of these understandings with diseases, including cancers [1].

Particularly, detecting various mutations, including somatic mutations, is essential to comprehend the cancer genome [2]. However, this is one of the most challenging parts in studying cancer, because somatic mutations are sporadic among healthy cells [3], and therefore, variants with low allelic fractions are hard to track down [4]. Fortunately, targeted sequencing has been a great support to overcome these difficulties. Compared to whole-genome sequencing, targeted sequencing has several advantages in many aspects: high coverage data could be generated at a more reasonable price [5]; low-frequency variants could be detected with this ultra-deep sequencing [6]; and different types of mutations, such as single-nucleotide variants, short indels (insertion and deletions), structure variations, and copy number alterations, can be examined via targeted sequencing [7]. Therefore, many laboratories in both academia and the medical industry are making efforts to develop their own gene panels with various sizes.

However, the next predicament is the analysis of the targeted sequencing data, since the variants revealed in the data are hard to discriminate from false-positive errors. In detail, innate sequencing errors and early cycle polymerase chain reaction (PCR) biases during library amplification or target enrichment could be considered super or rare mutations during the variant calling process. To overcome this erroneous algorithm, unique molecular index (UMI) technology was recently developed [8]. UMIs, as 8–9-bp random oligonucleotides, barcode a single DNA molecule and index replicates generated from the same DNA templates during PCR amplification. Several proof-of-concept studies have proven the error correction mechanism of the UMI [6, 9–13], and they revealed that PCR biases and sequencing errors were compensated during consensus sequence extraction from reads with the same UMI tags [6]. Nevertheless, analyzing data with UMI is still a tedious task [8]. In other words, some UMI-analyzing and variant calling tools have limitations in their broad application to the data generated from different hybrid capture platforms or other types of sequencing data, such as single-cell sequencing. In addition, most files from each tool are not compatible with the tools in different pipelines. For example, a recently developed analysis tool, named smCounter2, is a variant caller and an analysis pipeline tool package for targeted sequencing data with UMIs. Purportedly, this tool is specialized to call low-frequency variants. However, the usage of smCounter2 is restricted only to the data generated by the QIAGEN QIAseq DNA target enrichment kit [6].

In this study, to build a versatile analysis pipeline to apply to various types of hybrid capture-based targeted sequencing, our laboratory evaluated various types of analysis tools, like Fulcrum genomics (fgbio, https://github.com/fulcrum-genomics/fgbio), Picard (http://broadinstitute.github.io/picard), and the Genome Analysis Tool Kit (GATK) [14]. We put our effort into optimizing the pipeline so as to not make discrepancies between the files and enable easier interpretation of the results. Furthermore, we show an example of how targeted sequencing experiments can be investigated for their quality from the analysis and how we could take points from the results analysis as feedback for the experiment.

First, total read counts were counted by either Picard tools or fgbio tools. An 8-plexed hybrid capture library generated 100 Gbp data, and approximately 15 Gbp data were generated from each sample. Considering the fact that the target genome size is 0.3 Mbp, the read coverages were calculated as 50,000× to 70,000×, as shown in Table 1. As on-target rate (%) represents how much the probes attach to the target genomic regions and as calculating on-target rates (%) refers to the efficiency of experiments, we examined the on-target rate (%). We were able to get 37% on-target rate, average (Table 1, Fig. 2). On-target rates were slightly increased when target regions were expanded to flanking regions (+100 bp). In addition, each sample showed a similar on-target %, and this may suggest that the hybridization step of targeted sequencing was processed with an even amount of each library.

Before removing sequence duplicates, we estimated duplicate read counts via UmiAwareMarkDuplicateWithMateCigar (Picard). We could measure duplicate rates of >70% of the reads (Fig. 3). Considering the fact that the percent duplicate increases as the sequencing coverage rises [16], an average 73% of duplicates indicates that the targeting efficiency of the probes was high in high-coverage sequencing. In addition, as we were aware that an excess number of PCR cycles during target enrichment brings severe biases [17], we assumed that the number of PCR cycles during enrichment was moderate. Although duplicate reads were not discarded in further steps, by checking duplicate rates, we confirmed that the number of PCR cycles during library preparation and target enrichment was adequate. In addition, we were able to verify the ability to examine the efficiency of our experiments through data analysis.

We then estimated the number of consensus reads to be ready for more accurate variant calling. Through this filtering, we were able to get the consensus coverage down to one-sixth of the raw coverage (Fig. 4). According to the filtering methods, the consensus sequences of each read from different UMI families were called for scanning variants. In addition, by calling and filtering consensus reads, the probability of errors was considered, according to the algorithm of the tool. Furthermore, the reads containing Ns or the reads with low confidence were filtered out for highly confident variant calling.

In this study, we built an analysis pipeline for targeted sequencing data generated based on the hybrid-capture method. As UMI-adopted targeted sequencing data are notorious for their novelty and complexity, we mainly focused on finding tools and optimizing methods for the analysis.

Our data show an on-target rate of 37%, and there is little variation among samples. This may suggest that almost equal amounts of libraries of each sample were used for hybridization. However, because 37% is relevantly low for hybridization efficiency, the estimation of the on-target rates shows that still there is room for improvement.

The reason why on-target rate is important is that by estimating on-target rates and further upgrades of protocols, the quality of targeted hybrid-capture sequencing could be improved. In detail, protocol modifications during hybridization steps could possibly bring about an increase of on-target rates. For example, a slightly excessive amount of input DNA or target probes can increase the off-target effects during hybridization. Furthermore, inconsistent temperature or slightly higher/lower temperature than the proper temperature could bring about larger off-target effects than expected.

With respect to the pipeline, when using CallMolecular ConsensusReads, insertion and deletion errors are not considered in the consensus model. Therefore, realignment steps using other methods, such as IndelRealigner (GATK, v4.0.2.1), should be integrated for better and more precise analysis for identification of short indels. Furthermore, even though we optimized the tools and customized python codes for analyzing UMI data, there are still many tools that could be used in one analysis pipeline. We are still trying to minimize irrelevant steps to simplify the process.

In summary, we have built an analysis pipeline specialized for UMI-adopted hybrid-capture-based data. Given the fact that the precision medicine era has been coming lately and that targeted sequencing and UMI technology help to comprehensively understand the genomewide status of cancer patients, this report suggests that the quality of the experiment can be examined precisely and efficiently by this pipeline, and our laboratory sees its positive potential in being widely used for studies in various clinical fields.