Review of Clinical Next-Generation Sequencing

Next-generation sequencing is an established test method for germline (inherited) and somatic (acquired mutations) genetic mutations in many clinical laboratories. For inherited diseases, testing for germline mutations may include targeted panel, whole exome, whole genome, or mitochondrial DNA sequencing.^8,9  Targeted panel testing, which varies between laboratories, is possible for a wide variety of inherited disorders such as immune deficiencies, bone marrow failure syndromes, blindness, deafness, mitochondrial disorders, renal disorders, neurologic disorders, connective tissue disorders, cardiomyopathies, and cancer predisposition syndromes, among others.^10–17  Targeted panels for genes associated with a clinical phenotype are usually the first line of testing for inherited disorders, while whole exome sequencing is reserved for cases in which targeted testing has been uninformative.^18,19  Whole exome testing often involves testing the child and both parents (trio testing) to assist in the interpretation of variants.¹⁹  In addition, NGS technologies are used in analyzing cell-free DNA in the prenatal setting.^20,21 

Targeted panels for cancer testing also vary between laboratories.⁶  Targeted panels may be broad, including genes for both solid and hematologic malignancies, or may be more focused for a particular type of malignancy (such as myeloid neoplasms).⁶  Any given gene within a panel may be completely sequenced or only partially sequenced (eg, hotspot regions). For both germline and somatic testing, it is important to know the content of the targeted panels when deciding on using a test. Whole exome and whole genome sequencing are not currently used clinically for oncology testing.

Several new applications for NGS have more recently moved into the clinical arena or are being actively researched for clinical use, including circulating tumor DNA testing, human leukocyte antigen (HLA) typing, microbial analysis, RNA sequencing and expression, and methylation. Some of these new uses of NGS may be helped by the unique advantages of new instruments that are now available (see section on New Instrumentation). The use of NGS for HLA typing had some challenges to overcome: differentiating low-frequency alleles from high-frequency artifacts and distinguishing 2 similar alleles as 2 distinct alleles.^22,23  However, newer data analysis techniques such as stepwise threshold clustering have allowed NGS to be explored as a clinical option for HLA typing.^22,24  The use of NGS for identity testing using short tandem repeats (STRs) runs into the same problems as other repeat regions (see Difficult-to-Sequence Areas below); however, again newer data analysis techniques are making headway against this problem and may be applicable to other repeat regions.²⁵  Other uses for clinical NGS include pharmacogenetics, microbial sequencing, and advanced blood group typing (eg, type A1 versus type A2). Further discussion of these topics is beyond the scope of this article.

Cell-free DNA has been used for some time for prenatal testing; however, NGS of circulating tumor DNA (ctDNA), tumor-derived cell-free DNA, is a more recent development that is now clinically available.^20,26,27  This testing is often referred to as a liquid biopsy. The potential applications of sequencing ctDNA include screening or diagnosis of cancer, monitoring for progression or relapse, and guiding therapy for a patient with a known cancer diagnosis. Most research studies have evaluated the ability of ctDNA sequencing to detect somatic mutations in patients with known cancer and the ability to monitor disease.^26–29  Monitoring a known mutation by sequencing of ctDNA has been shown in several studies to correlate with relapse/progression of disease.^29,30  Also, using detection of mutations in ctDNA to help guide therapy for a patient with a known tumor has shown utility, for example, tyrosine kinase inhibitor response with epidermal growth factor receptor (EGFR)–activating mutations in lung cancer.^26,27,31,32  Although ctDNA may have lower sensitivity for detecting mutations than testing tumor tissue, the most common clinical applications of ctDNA appear to be for patients with metastatic cancers when there is insufficient tissue for testing and a repeated biopsy would be associated with significant morbidity and mortality, and when testing of ctDNA is a reasonable alternative.^26,27,31  Using ctDNA to screen for or diagnose early-stage cancers is more problematic.³³  Most studies of patients with known cancer have not contained normal controls, but a limited number of targeted sequencing studies have shown some degree of mutation detection (false positives), albeit usually at a low level, in normal controls.^28,34,35  Low sensitivity to detect early-stage cancer (false negatives) is another limitation. Studies have shown sensitivities in the 30% to 60% range for early-stage tumors, and some tumor types may have a higher false-negative rate, as ctDNA appears to be released owing to apoptosis and necrosis.^36–41  These false-positive and false-negative issues limit the practical use of ctDNA for early cancer diagnosis or screening.

Samples undergo DNA extraction, library preparation, target enrichment, and sequencing⁴²  (Figure 1, A and B).

Almost all DNA extraction methods are acceptable. Extraction methods for formalin-fixed, paraffin-embedded (FFPE) tissue may need special care, and macrodissection or microdissection to enrich for tumor may be required for some cases.⁴³  DNA quantitation is performed by Qubit or Picogreen (Thermo Fisher Scientific, Waltham, Massachusetts) rather than by standard spectrophotometry.⁴⁴ 

Library preparation refers to the process of preparing DNA for use on a sequencer. Although many methods are available, they all result in breaking DNA into fragments and adding adaptors to the ends.^45–48  Adaptors may include molecular bar codes (to allow pooling of patient samples), universal polymerase chain reaction (PCR) primers, hybridization sequences to bind the DNA fragment to a surface, and recognition sites to initiate sequencing. The term library refers to these fragments of DNA with flanking adaptors that are ready for sequencing. The size of the DNA fragment between the adaptors is referred to as the insert size. The insert size may vary and there are different advantages to short insert sizes and long insert sizes. Shorter fragments are more likely to have both ends fall within an exon, which is often the area of interest, while longer fragments are more likely to have 1 end fall in an intron, which may increase detection of structural rearrangements if only exonic regions are being selected (Figure 2). For more details on structural rearrangements, please see Structural Variation and Copy Number Variation.

The resulting library undergoes enrichment for both whole exome analysis and targeted testing or is sequenced directly for whole genome analysis. Enrichment may be performed by hybridization to complementary sequences (sequence capture) or by PCR.⁴⁹  Enrichment by PCR is generally combined with the library preparation step, as the primers that select the regions of interest may also contain the adaptor sequences. The choice of enrichment strategy is often dictated by the clinical use: sequence capture is preferred for large genomic regions, and PCR for smaller regions where greater enrichment is desired.⁴⁹ 

Most clinical sequencing is performed on 1 of 2 main types of instruments: Illumina sequencers (San Diego, California) including the HiSeq, MiSeq, and NexSeq; or the Ion Torrent series of machines including the IonPGM, IonProton, and IonS5 (Thermo Fisher Scientific). These 2 types of machines differ in their chemistry, detection methods, advantages, and disadvantages^42,50–54  (Table 1).

The first sequencing step for both the Illumina and Ion Torrent platforms is to immobilize each DNA fragment and clonally amplify it. Clonal amplification is needed to generate a large enough signal for detection. The Ion Torrent uses a bead emulsion for immobilization and clonal amplification, whereas the Illumina sequencers use a flow cell.^51,53,54  The flow cell or bead contains sequences that hybridize to part of the adaptor on the DNA fragments. The input DNA concentration is critical to ensure only 1 DNA fragment binds per bead and to ensure that the DNA fragments are well spaced out on the flow cell. The clonal amplification step creates a bead or cluster with approximately 1000 identical copies of a unique parent DNA molecule that are physically isolated from other molecules.⁵⁴  For the Ion Torrent the beads are then placed in a well (1 bead per well).⁵³ 

Illumina sequencers use sequencing by synthesis with fluorescent detection^53,54  (Figure 3, A through D). All 4 fluorescently tagged nucleotides are added and compete for the next space. The complementary tagged nucleotide will bind but a blocker prevents addition of more than 1 nucleotide per round (reversible terminator chemistry). The remaining nonbound nucleotides are washed away. Laser excitation leads to a fluorescent emission that is recorded (simultaneously for each DNA fragment cluster). The fluorescent tag and blocker are cleaved, then the next round begins. In each round, 1 base pair is read from each DNA cluster. This process can be repeated on the opposite end of the DNA fragment, which is referred to as paired end reads (Table 2).

Ion Torrent sequencing is different, as only a single base is added in each round (eg, A in round 1, T in round 2).⁵³  When an added base is incorporated, a hydrogen ion is released, accompanied by a pH change that is detected for each bead within a well; if a base is not incorporated there is no voltage generated.^53,54  Incorporation of more than 1 of the same base leads to a proportionately higher voltage signal up to about 6 to 8 bases^51,55  (Figure 3, E). If more than 6 to 8 bases are incorporated, the signal is no longer proportional and the exact amount cannot be determined.^51,55 

The raw data reads from either type of instrument undergo a series of bioinformatics processes (also referred to as a pipeline) to ultimately deliver a variant call file (VCF)^56,57  (Table 3). These processes include demultiplexing (Table 2), quality analysis, mapping of the reads to a reference genome (resequencing), and variant identification/annotation.⁴³  Because of these specialized processes, dedicated bioinformatics personnel may be needed to set up and maintain a clinical NGS service.

The use of bar codes to tag a specimen's DNA fragments allows multiple samples to be pooled and sequenced together, thus decreasing the sequencing cost. However, this process requires a demultiplexing step in which all reads are sorted by bar code/sample before further analysis.⁵⁸  The demultiplexed file with raw reads is referred to as a FASTQ file (Table 3).

Following demultiplexing, the individual reads for a sample are mapped (Table 2) to a reference genome (BAM file Table 3) and any difference between the reference and the sequencing read is noted.⁵⁹  Identical (duplicate) reads are discarded for whole genome sequencing or sequence capture but not for amplicon-based sequencing. If multiple reads show the same difference, a variant is called (the threshold for the number or percentage of reads required is determined by the laboratory and should be validated). For example, a heterozygous single nucleotide variant (SNV) should be present in 50% of reads; however, in actual practice the range has been shown to vary as much as 23% to 74%.⁶⁰  The quality of signal for an individual base read and the mapping quality are also factors considered when calling a variant.⁵⁹  The output file that defines all the variants for a sample and their allelic fractions is referred to as a variant call file⁵⁹  (Table 3). This list of variants undergoes interpretation. The variant call file will contain all variants including common variants, although additional bioinformatics tools can be used to filter out variants meeting certain criteria (minor allele frequency above a threshold or variants previously identified as benign by a laboratory for example).

Before implementation, clinical NGS requires end-to-end validation from DNA extraction through the bioinformatics pipeline, and changes to both the wet bench or informatics portions of the test require revalidation (see Validation and Proficiency Testing sections).⁶¹ 

Variant interpretation is complex when applied to whole genes (as opposed to well-defined hotspots) and a large number of genes. The larger the area of the genome that is sequenced, the greater the probability of encountering rare or novel variants that will require interpretation. This has mainly been a problem in the inherited disease realm, but as oncology testing has moved away from hotspot testing to larger panels, the same issues have besieged it. Several laboratories sign out all or a subset of NGS cases in a consensus conference and share molecular data in molecular tumor boards.⁶² 

There are guidelines for the interpretation of germline variants put forth jointly by the American College of Medical Genetics (ACMG) (now the American College of Medical Genetics and Genomics), the Association of Molecular Pathologists, and the College of American Pathologists (CAP).⁶³  These guidelines assign strength of evidence for various criteria regarding a particular variant and rules for combining all the criteria to classify that variant as pathogenic, likely pathogenic, uncertain significance, likely benign, or benign.⁶³  The criteria include minor allele frequency from population databases and prevalence of a variant in affected individuals, segregation data, functional studies, the type of mutation and its predicted effect, similarity of the mutation to known mutations, computational models of effect, and inheritance factors.^63,64 

These guidelines have limitations and there is subjectivity to interpretation. For example, application of these guidelines to the same set of mutations between several different laboratories achieved a consensus classification 71% of the time after review and training on how to use the guidelines.⁶⁵  Another issue is that the population frequency criteria (absent or rare in population databases) can be problematic for recessive disorders, for variants with decreased penetrance or milder presentations, or in underrepresented ethnicities.⁶⁵  Population databases (Table 4) now contain information on more than 120 000 individuals, so rare pathogenic mutations may be present in a carrier state within the database.⁶⁶  These databases typically excluded patients with severe disease, but milder phenotypes or diseases with an older age of onset were not excluded.⁶⁶  Despite the limitations, these standards are a start and will allow comparisons between laboratories and for research studies. Similar criteria for somatic testing have only been available recently, and utility of these guidelines for standardizing somatic variant interpretation and reporting across laboratories remains to be evaluated.^67,68  Though there are some tools available to assist the implementation of these guidelines for variant classification, use of these guidelines is labor-intensive and we lack automated tools that can evaluate several of these criteria and support the process.⁶⁹ 

As diagnostic panels increase in size, the likelihood of detecting incidental findings also increases, in particular with whole genome and whole exome testing. To fully realize the promise of precision medicine, these incidental findings would need to be incorporated into the clinical care of patients. For example, if a patient is found to have a pharmacogenetics variant leading to decreased metabolism of morphine, during whole exome testing, ideally that information would be available in the future if the patient ever required prescription pain medication. However, there are several issues surrounding reporting of incidental findings, not the least of which is ensuring that patient consent allows for the selection for return of all, some, or no incidental findings.^70,71  A patient may want some incidental results (eg, those that could affect response to a drug); however, he or she may not want other incidental results (eg, carrier status for a disease or a mutation for an adult-onset genetic disease that lacks effective treatment).⁷¹  Having appropriate consents, ensuring that patients understand these consents, and then having the infrastructure in place to mask specific results for individual patients are all challenges to successful implementation.^72,73  Furthermore, there is the issue of which incidental findings are worth reporting from a medical standpoint (eg, should the variant that causes sensitivity/flushing to alcohol be reported).⁷³  In 2013 the ACMG recommended that a minimum set of 52 genes with high penetrance and available intervention be reported if those genes are analyzed; this list was updated to 59 genes in 2016.^74,75  These recommendations created substantial controversy surrounding the issue of informed consent and the patient's right to decline receiving incidental results and testing in minors, which have been incorporated into the updated ACMG recommendations.^76,77  Nevertheless, laboratory policy about how to handle incidental results usually takes these recommendations into consideration.

Another challenging area is determining what genes to test in a given clinical scenario. Although there are guidelines that define the common mutations or genes of interest (testing that is usually reimbursed), the literature and/or clinician interest may suggest other genes (testing that is usually not reimbursed) that may be medically useful.⁷⁸  Commercially and locally available panels often differ to some degree in the genes that are tested or the portions of the genes that are tested, and knowing the pros and cons associated with the different panels is challenging.⁶  Databases or tools to assist with this selection process do not exist. Furthermore, multiple gene mutations in the same tumor that indicate differing prognosis or therapeutic response may be difficult to resolve. Lastly, oncology testing may identify possible germline mutations.^68,79  Although simultaneous testing of matched patient tumor and normal samples is performed in large research studies, in clinical laboratories this practice is difficult because of practical difficulties in obtaining a blood sample for germline testing from patients, and because it doubles the cost of testing, and is not reimbursed.⁶⁸  This is usually addressed by a disclaimer or sometimes by follow-up testing of a germline sample in select cases.⁶⁸ 

It is critical to validate the entire NGS process from end to end (DNA extraction through the bioinformatics pipeline).⁶¹  The validation process should demonstrate the ability to detect different genetic changes such as single nucleotide changes, insertions or deletions of varying sizes, and copy number variations or translocations if applicable. Validation should include patient samples with genetic variants detected by another methodology and may include commercial samples (HapMap or commercial controls); sample types (eg, FFPE, fine-needle aspiration, amniocytes) that will be run in clinical practice should be included as part of the validation. Similar to standard laboratory validations, sensitivity (false negatives), specificity (false positives), and reproducibility (including within run, between run, and with different operators) should be established for all assays. It is not feasible to evaluate these parameters for every possible mutation during validation, but common pathogenic variants should be included in the validation.⁶¹  Limit of detection must also be assessed to establish the minimum amount of DNA needed for the assay and establish the minimum mutant allele frequency. This is especially important for any oncology assay in which tumor percentage and heterogeneity affect the allele frequency, but it is also relevant for the ability to reliably detect mosaicism in an assay to detect inherited disorders.^31,80,81 

During the validation, metrics should be defined to assess the quality of a test run and criteria for repeated testing established.^61,82–84  These metrics may include cutoffs for the insert sizes after library preparation; criteria for assessing adequate target enrichment; library concentration parameters for various steps; expected performance of controls; and metrics for sequencing performance such as clustering, base and mapping quality scores, error rates, GC bias, transition/transversion ratios, total number of sequencing reads, and coverage.^61,82–84  Identifying the need for repeat enrichment before sequencing may be time and cost-effective for a laboratory by avoiding wasted sequencing time and cost. For example, in our laboratory we run quantitative PCR against 3 targeted regions and 3 nontargeted regions for our inherited disease capture to ensure adequate enrichment. If this quality control fails, the sample(s) undergo recapture and reassessment before being sequenced.

Additionally, during validation criteria for supplementary testing should be established.⁶¹  Supplementary testing may include genomic areas that are not reliably sequenced and confirmatory testing for certain variants that do not meet certain quality requirements.^4,85  There should be documentation of areas that cannot be reliably sequenced and a policy for addressing these areas (either as supplemental testing or a disclaimer in the report).⁸⁴  Like any test, NGS is subject to false positives, and the validation process should identify the metrics that require a confirmatory test to verify the presence of a variant identified by NGS.^4,61,85 

After the initial validation, any procedural changes require revalidation. Careful thought should be given to the initial design of an assay, as redesign requires revalidation. Changes that only involve the bioinformatics pipeline may be revalidated by using previous data sets and comparing the output of the old and new bioinformatics process. Changes in any wet bench process require end-to-end revalidation but may use fewer samples than the original validation.⁶¹  The degree of change dictates how many samples should be assessed for a revalidation; a major change should evaluate more samples than a minor change.^61,84 

Clinical Laboratory Improvements Amendments of 1988 require twice yearly proficiency testing (PT) for all clinical assays.⁸⁶  For tests that lack approved PT, the laboratory must verify the accuracy of the test twice yearly. These alternative assessments may include comparison to a national reference, interlaboratory exchange, or in some cases, intralaboratory verification.⁸⁶ 

Ideally, PT materials would cover the assay from start (wet bench aspects) to finish (bioinformatics and interpretation). In addition, data files for testing only the bioinformatics to interpretation portion of the assay would be useful. The advantage of testing the bioinformatics portion is the ability to evaluate the bioinformatics process for multiple variants including variants of various sizes. One challenge to developing this type of PT is making a universal data file that can be recognized and tested through all the different platforms.^82,87  Proficiency testing materials may be analyte specific, which is not sufficient for NGS, or methods based on genomic DNA from individuals, genomic DNA from cell lines, or synthetic DNA.^82,87  Well-characterized materials are currently available from the Genetic Testing Reference Materials Program (GeT-RM) of the Centers for Disease Control and Prevention, from the Genome in a Bottle Consortium of the National Institute of Standards and Technology, and through the CAP Proficiency Testing Program. Estimation of tumor percentage is a necessary part of oncology NGS testing to determine if sufficient tumor is present for testing, and CAP also offers proficiency testing for this step.

The cost for the wet bench portion of NGS is primarily based on (1) library preparation (reagents, labor, necessary equipment), (2) selection strategy (PCR or capture), and (3) sequencer used. Library preparation costs vary widely depending on the method. Reagent costs are determined mainly by the commercial entities supplying the reagents and are often inversely proportional to the labor requirements. Cost of selection will vary depending on the selection strategy used (PCR versus capture), the amount of the genome being targeted (with custom capture–based products often being available in tiers), and labor and equipment required to perform the selection. Library preparation is combined with selection for PCR-based methods, which reduces the combined cost of these 2 steps. Cost of library preparation and selection may also be dependent on batching a fixed number of samples, which may be problematic for a clinical laboratory trying to maintain turnaround times. Lastly, sequencing costs are directly proportional to how much of the sequencer capacity is used for a sample and whether the sequencer is used to full capacity on a given run.

Although costs vary widely depending on the NGS design (size of sequenced area, depth of sequencing, size of sample batches, and scale of the sequencing operation), in general, for all designs, the higher the number of samples included in an analytic run, the lower the per-sample cost. Laboratories can potentially reduce costs by streamlining workflow, choosing the most cost-effective library preparation, increasing sample volume, and if sample volume permits, automating library preparation. Instrument depreciation costs per sample are heavily dependent on instrument usage, and laboratories need to carefully evaluate sample volumes and instrument usage before deciding to purchase capital-intensive sequencing equipment. To minimize capital depreciation costs, we have adopted a model of sharing our high-throughput sequencer used for inherited disease cases with the University of Minnesota Genomics Center that uses the same instrument for research purposes. This increases the total number of samples analyzed on the instrument and dramatically reduces the capital depreciation costs for the clinical samples.

Because of the many variables that affect cost, it is hard to generalize, so we offer our experience of approximate costs for a large capture-based inherited disease panel and a small PCR-based oncology panel. For inherited disease panels, we typically sequence 9 samples and 1 control for 4800 genes (10.5 MB) on 2 lanes of a HiSeq2500 (2×100-bp run). Wet bench costs of sequencing these 9 samples to an average depth of 400× coverage is $12 145 ($1349 per sample). Library preparation accounts for 18% of the cost ($241 per sample), capture-based selection accounts for 18% of the cost ($244 per sample), and sequencing accounts for 64% of cost ($864 per sample). In addition, bioinformatics processing and use of commercial annotation and database software costs $200 per sample when averaged over our sample volume of 800 cases a year. Lastly, Sanger confirmation of NGS variants adds $50 to the total cost of the NGS assay. Thus, our total cost to run a large germline panel of 4800 genes is $1599 per sample if Sanger confirmation is required. By contrast, the wet bench cost for our small PCR-based oncology capture (13.8 kB) is lower, averaging $417 per sample. The breakdown of cost is as follows: 16% ($67) depreciation costs of our lower-throughput sequencer, 21% ($88) labor, and 63% ($263) reagents. However, the number of samples for a given run affects the per-sample cost, as the depreciation, labor, and a subset of reagent costs is divided over the number of samples.

Cost of validation also needs to be considered in developing an NGS assay and this can be a significant upfront cost. Our laboratory had access to sequencing instrumentation and some bioinformatics support, but our initial validation costs for our inherited disease assay were approximately $250 000 to $300 000 in 2012. A significant portion of this initial cost included development of infrastructure including bioinformatics infrastructure. With our infrastructure already in place and with the advancements in the field, subsequent validations of newer versions of the NGS assay have typically cost $50 000 to $70 000.

Although there is a desire to use NGS as a single method to detect all clinically relevant genetic changes, significant limitations currently exist. These limitations include analytic sensitivity of mutation detection, areas of the genome that are difficult to sequence or analyze, limitations in the knowledge of how to interpret novel or rare mutations, limited ability to detect structural gene variation and copy number variation, and integration of genomic information into the medical care of patients. These limitations are discussed in more detail below.

The sensitivity of NGS for SNV detection is approximately 5% to 10%.^43,88,89  Although this sensitivity is acceptable for most inherited disease testing (it may not detect low levels of mosaicism), it limits testing in oncology for minimal residual disease, when a low tumor percentage is present, or to detect low-level mutation due to tumor heterogeneity. Possible reasons for this limited sensitivity include PCR noise compounded by C to T transversions with FFPE tissue, sequencing errors, and systematic errors.^90–92  The general pathologist should be aware that FFPE samples will have higher artifacts than fresh tissue samples; furthermore, small specimens (including cytology specimens) may have limited DNA that can affect testing by NGS methodologies.^44,92  Studies have shown that systematic errors lead to a 4% to 6% error rate; counterintuitively, the rate is higher with increasing coverage.⁹³  Systematic errors may be sequence-specific errors, errors at a particular location of the read (eg, the ends for Illumina sequencers), or related to the base pair content (GC rich for Illumina).^93–95  As neither PCR nor fixation causes insertions/deletions (indels), outside of repeat regions there is better sensitivity for detecting small indels than SNVs.⁸¹ 

There are 2 main methods for improving sensitivity; however, both of these methods decrease the number of useable reads and therefore will increase the sequencing cost to obtain a comparable coverage.⁹⁶  These methods are not in widespread clinical use at this time. The first method is the use of overlapping paired end reads. This method works only for areas where the paired ends overlap, therefore the DNA insert size must be the same size or smaller than the number of reads.⁹⁶  This technique is well suited to amplicon-based sequencing in which the DNA insert size/amplicon size can be strictly controlled.⁹⁶  In this situation, the DNA insert would be completely sequenced by both paired end reads (ie, the forward read and the reverse read). The sequence of these 2 reads should match and any base pair that does not match in both reads is discarded.

The second technique is to use random nucleotide tags, referred to as unique identifiers (UIDs) or primer IDs, as they are often incorporated into the PCR primer.^96,97  This method works with both sequence capture and amplicon-based DNA selection techniques. In this method, random nucleotide tags are added to DNA fragments, assigned if DNA was sheared, or incorporated during the first round or two of PCR for amplicon-based methods. Importantly, these steps occur before amplification and result in a DNA fragment with a random and unique nucleotide sequence at one or both ends. After amplification, multiple identical template molecules will be present and will be sequenced (Figure 4, A through D); therefore, duplicate reads must be retained during analysis.^96,97  All reads that map to the same location and have an identical UID are considered part of a UID family and will be analyzed as a group. Targeted areas should be covered by many different UID families. If a mutation is present in a majority (eg, >95%) of that UID family, the mutation is considered present and is considered 1 read.^96,97  This process is repeated for all other UID families.

Homologous regions, repetitive regions, and GC-rich regions are not reliably interpretable by the current NGS platforms and standard bioinformatics algorithms. Homologous regions, including pseudogenes, are areas of the genome with high sequence similarity that may differ from the gene of interest by only a few base pairs. Fragments of DNA that are sequenced from the target gene and homologous regions may be so similar in sequence as to be indistinguishable; and the shorter the length of sequence, the more likely this is to occur. This is not a problem unique to NGS, as Sanger sequencing is also susceptible to inadvertent sequencing of homologous regions, and test design is important to mitigate the problem. In NGS analysis, fragments of DNA from a target gene and homologous regions will have poor mapping quality, and reads belonging to the homolog can be mismapped to the real gene and vice versa⁹⁸  (Figure 5). Mismapping may lead to both false-positive and false-negative calls (eg, mutations being missed and mutations being erroneously called). Many clinically relevant genes (such as PMS2, STRC) have pseudogenes, are challenging to interpret by NGS, and require specialized methods for target enrichment such as long-range PCR.^98,99  This problem may be solved by newer instruments that have much longer sequencing reads (see New Instrumentation section); however, in current practice, assessment of these areas requires traditional alternative methods.¹⁰⁰ 

For repeat areas, unique sequence flanking the repeat is required to reliably map a sequencing read and determine the size of the repeat. Repetitive regions larger than the size of the DNA insert will not have a flanking sequence and therefore will not be accurately mapped. Smaller repeat sizes will have a unique flanking sequence on at least a proportion of the DNA fragments and therefore will map, although at lower coverage since some reads will not be informative. Even so, enumeration of the repeat size requires specialized bioinformatics algorithms, and errors still occur, which require interpretation. Sources of error include stutter (polymerase slippage leading to small shifts in repeat size) and PCR sequencing mistakes.²⁵  Homopolymers (ie, poly A or poly T) are difficult for the Ion Torrent sequencers, as the degree of change in voltage loses resolution above 6 to 8 base pairs. However, most testing of repeat areas (eg, trinucleotide disorders such as fragile X) continues to use traditional, established methods rather than NGS.

GC-rich regions appear to have higher background noise and lower quality of sequencing. In particular, Illumina sequencers give substitution errors in areas of high GC content and long G/C homopolymers.⁹⁴  GC-rich regions are known to form secondary structures, which may represent part, but perhaps not all, of the problem. There may also be accumulation of G or C fluorophores after washing or out-of-phase sequencing.⁹⁴ 

Validation of an NGS assay should include assessment for areas that cannot reliably be genotyped by NGS methods, and at a minimum these areas should be documented.¹⁰¹  Alternative testing strategies such as Sanger sequencing or long-range PCR may be possible for some regions.⁹⁸ 

Although the technical ability exists to perform whole genome analysis at a reasonable cost, especially for inherited disease, the ability to interpret all of those data lags behind. Sources to help with interpretation include databases (both publicly available, private, or laboratory-specific databases), genetic and medical knowledge, medical literature, patient information, clinical experience, and team discussion. There are different types of databases with different amounts of data within them. A layer-1 database or Clinical Genomic Variant Repository contains only sequence/variant information, a layer-2 database or Genomic Medical Data Repository contains sequence/variant information with clinical/phenotype data, and a layer-3 database or Genomic Medical Evidence Database contains the medical evidence of classification or association with the sequence/variant information.⁷  Most databases contain data for either inherited disease or somatic mutations but not both with the exception of ClinVar and dbSNP (Table 4).

Although databases are extremely useful in the interpretation of variants, there are limitations of the current databases and no database is comprehensive or error free. Many databases lack assurance of the quality of sequence data or other data within the database.⁷  Databases may not be up-to-date or may include conflicting data. Both the medical literature and databases must be used with caution, as some variants have been described as pathogenic by outdated criteria (ie, absent in 100 controls). Furthermore, there is limited knowledge about digenic or multigenic effects.

The significance of variants in introns or untranslated regions is often unknown, and rare or novel exon variants may also be difficult to interpret.¹⁰²  Novel or rare mutations that cause a frameshift or change an amino acid to a stop codon (stop loss or nonsense mutations) will generally be pathogenic if that mechanism has been described for the gene in question, but even then there may be exceptions.⁶³  Missense mutations are more difficult to interpret. Many factors are taken into consideration when interpreting these cases, including details about the specific mutation, details about mutations known to cause disease, similarity to known mutations, whether the mutation is in cis/trans with another known mutation or is de novo, presence/absence in other individuals (such as populations, normal controls, or affected and unaffected family members), and predicted protein effect (using in silico models).^63,103 

Next-generation sequencing performs reasonably well at detection of SNVs and small insertion/deletions (indels), but does not perform as well at detecting structural rearrangements or copy number variations (CNVs) especially when using enrichment to perform targeted NGS.^104,105  Additionally, detecting structural variation and CNVs requires different bioinformatics algorithms from SNV detection. Several clinical laboratories are currently using NGS data to detect CNVs and generally use a combination of 2 or more techniques.^5,104,105 

There are several techniques that have been used to detect CNVs, including depth of coverage (read depth), read pair, split pair, assembly based, or a combination of these techniques.¹⁰⁶  Clinical CNV analysis usually uses some combination of 2 or more of these techniques.^5,104,105  All the methods detect deletions better than duplications, cannot detect CNVs in repeat regions or difficult-to-map areas, and are limited by coverage (although the depth-of-coverage technique is more affected by coverage that the other techniques).¹⁰⁶  False positives are a problem, especially when applying CNV analysis over large areas of the exome, and have been reported to occur 10% to 89% of the time.¹⁰⁷  Recent advances that have incorporated machine-learning techniques hold the promise to reduce false-positive calls.²  However, some areas of the genome are more prone to false positives than others.¹⁰¹ 

Using depth of coverage or read depth to detect CNVs works well with uniform sequencing, which is assumed for standard bioinformatics tools. These tools analyze for increased or decreased coverage to detect duplication/amplification or deletion, respectively. However, coverage varies between runs, within run, and between patients especially when sequence enrichment is used, and spurious calls will be detected when sequencing is not uniform. With sequence enrichment, the pattern of coverage tends to be similar but the absolute coverage varies, requiring some sort of normalization. This may require comparison to a control as well as to control genes within the sample to normalize for the performance of the individual sample.²  The advantage of the read depth technique is the ability to detect large CNVs and to predict the actual copy number; however, this method cannot detect the breakpoint or detect rearrangements.¹⁰⁶ 

Read pair (or mate pair) analysis compares the distance of 2 ends of a read pair to the average insert size. Read pair analysis requires paired end reads, is limited by the insert size, and will only detect smaller CNVs.¹⁰⁶  One advantage of read pair analysis is that it detects both CNVs and rearrangements (translocations and inversions).¹⁰⁶  However, it will only detect duplications/amplifications smaller than the average insert size and deletions smaller than 1 kb, and it cannot accurately estimate the number of copies.¹⁰⁶ 

Split pair (or split read) analysis looks specifically at paired reads where one of the paired reads fails to map or only maps partially. Split pair analysis also requires paired end reads, will only detect smaller CNVs, and does not perform well in regions of low complexity.¹⁰⁶  However, it can pinpoint the breakpoint and detect rearrangements.¹⁰⁶ 

Finally, assembly-based analysis uses de novo alignment of the reads.⁵  De novo alignment (Table 2) matches the individual reads to each other instead of to a reference genome. Because it is computationally intensive, this technique works better for small genomes, such as bacteria, but can be used clinically.⁵ 

There has been intense focus on the meaningful integration of genomics into patient care.¹⁰⁸  Many practical issues need to be solved for this to occur in a widespread manner. Issues include making reports understandable, interfacing genomic results with the electronic medical record (EMR), bioinformatics tools to help categorize variants, handling of incidental findings, and whether and how to offer genetic reevaluation.^109–111  Other issues include data storage, including what data to store (FASTQ, BAM, variant call file), how long to store data, and how to store large data sets securely. The CAP checklist for NGS provides guidance stating that some files must be stored for at least 2 years; these files should allow re-review of the case in the same manner that allowed generation of the original data.⁸⁴  Data storage and processing may be performed locally on a server or through a third party. Cloud companies now offer secure Cloud-based services and storage; however, it is incumbent on the health care facility to ensure services meet all HIPAA (Health Insurance Portability and Accountability Act) requirements for data transfer and storage.

Widespread meaningful integration of large-scale genomic data into the medical record, especially for smaller institutions, is still a challenge. Current laboratory information systems and EMRs can handle discrete data points with an associated interpretive or normal range and can handle interpretive text reports, but they are ill equipped to handle complex genomic data generated by whole genome, whole exome, and large targeted panel NGS.¹¹²  Although laboratory information and EMR systems may evolve, at the current time and for the foreseeable future, ancillary systems are necessary to integrate large amounts of genomic data into the medical record.^112,113  However, implementation of these ancillary systems requires significant time and resources for information technology personnel, in addition to clinicians, laboratorians, pharmacists, and/or pathologists depending on the application.¹¹³  Many of the places that have successfully implemented EMR integration with an ancillary genomic system are academic centers with genomic or information technology expertise and have implemented systems for a specific subset of genomic information (such as pharmacogenomics variants), with fewer institutions offering whole genome or exome testing.¹⁰⁸  Several new companies offer these ancillary systems. Some systems organize, annotate, track variants, and generate a report. These reports are usually pdf or text reports and do not have discrete fields that transfer to the EMR. Even more recently, a few companies have offered clinical decision support tools.

Two new sequencing instruments (sometimes referred to as third-generation sequencers) are currently available for research use that provide longer sequencing reads and are able to read the sequence of a single molecule: the PacBio SMRT (single molecule real time) (Menlo Park, California) and the Oxford Nanopore (Oxford, United Kingdom). These instruments use different underlying chemistry.^114,115  The PacBio SMRT uses multiple wells, each of which has a DNA polymerase affixed to the bottom with 1 long DNA fragment.¹¹⁵  Each fluorescently tagged nucleotide (A, C, G, T) gives a different fluorescent signal when incorporated.¹¹⁵  Illumination and detection occur from the bottom of the well and detection is sensitive enough to detect the single fluorescent signal that is released when a base pair is added to the DNA strand.¹¹⁵  The Oxford Nanopore uses a protein pore inserted into a membrane.¹¹⁴  A current is applied and flows through the pore between the 2 sides of the membrane.¹¹⁴  As the structure (DNA or RNA strand) passes through the pore, the current changes and the degree of change correlates with the individual base (A, C, G, or T) and also correlates with the methylation status of C; therefore, methylation and hydroxymethylation can be detected.^114,116  The PacBio SMRT can also infer methylation status by analyzing the change in DNA polymerase kinetics (the time to incorporate a base and the time between incorporation of 2 bases).

Neither instrument requires amplification steps and thus should reduce the background noise. Both instruments can perform long reads (14 000–40 000 for the PacBio SMRT and 8000–100 000 for the Nanopore), which can overcome issues with pseudogenes and repeat regions and may help with identifying RNA isoforms; however, both have high error rates.^115,117  The errors on the PacBio are random and therefore can be overcome by replicate sequencing of the same molecule and by using a consensus result.^118,119  The errors on the Nanopore are biased (meaning they occur in the same areas) and therefore cannot be overcome by replicate sequencing.¹²⁰  These instruments show promise and may address problems in many clinically relevant regions such as trinucleotide repeat regions, HLA, and homologous regions.¹¹⁵  However, these sequencers have limited adoption in the clinical realm, possibly owing to their higher price and lower throughput and possibly owing to the challenges of clinically validating an instrument with high intrinsic error rates.

Next-generation sequencing is being implemented in clinical laboratories and the use will only increase as the technology, bioinformatics, and resources evolve to address the limitations, improve quality of results, and increase the number of clinically useful applications. Clinical NGS has expanded to detect SNVs as well as structural rearrangements and CNVs, to monitor circulating tumor DNA, and to analyze areas of the genome that previously were challenging for standard bioinformatics algorithms to manage. Further improvements will continue to occur; however, the challenge for clinical laboratories is to ensure testing is clinically relevant, cost-effective, and can be integrated into clinical care.