Towards standardization of next-generation sequencing of FFPE samples for clinical oncology: intrinsic obstacles and possible solutions

The advent of next-generation sequencing and other genomic technologies makes it routinely possible today to investigate DNA specimens extracted from human tumors for gene mutations, chromosomal aberrations, differential expression of mRNA and epigenetic alterations. Such detected molecular changes are extensively used for integrative analyses aimed at the identification of key dysregulated pathways and importantly, the establishment of molecular classifiers indispensable for targeted therapy and personalized management of malignant disorders [1].

Due to relatively high prevalence of the mutations in the epidermal growth factor receptor (EGFR) that has led to the development of EGFR tyrosine kinase inhibitors (EGFR-TKIs), non-small cell lung cancer (NSCLC) has become a proving ground for the development of novel approaches for molecular typing [2]. Unfortunately, it has been established that EGFR-responsive patients with activating EGFR mutations eventually develop resistance to the treatment and progress either due to commonly acquired T790M mutations or to a variety of other molecular changes [3]. On the other hand, initial success in exploiting EGFR-TKI sensitivity paved the way for the development of other small molecular inhibitors as well as widespread introduction of the mutation analyses into the molecular subtyping of other tumors. Examples of actionable mutations include RAS (KRAS and NRAS), BRAF and PI3 K in colorectal cancer [4], BRAF or NRAS as well as KIT in malignant melanoma [5], and others. These and other genetic alterations continue to gain importance in companion diagnostics associated with recently marketed targeted therapies and investigational drugs [6‐8]. As a consequence of efforts invested in this area, the number of clinically actionable variations that contribute to sensitivity or resistance of tumors to each of these drugs is dramatically expanding alongside available therapeutic solutions, especially in cases of metastatic and refractory diseases. Moreover, the numerous case reports demonstrating clinical response to the drugs prescribed off-label based on matching molecular evidence ensure the spread of a molecular-informed therapeutic decision-concept into oncologist office routine [9‐12].

This avalanche of changes was precipitated by the replacement of conventional real-time polymerase chain reaction (PCR) or direct sequencing by Sanger with Next Generation Sequencing (NGS) which has recently revolutionized the field of molecular diagnostics. In addition to higher sensitivity and remarkable throughput, NGS allows gaining non-traditional kinds of information about extracted DNA specimen, including the prevalence of shorter fragments [13] and the degree of its sequence heterogeneity within each locus [14]. Nevertheless, a majority of NGS applications are currently limited to research use only, while its implementation in clinical practice awaits rigorous validation requiring establishing Clinical Laboratory Improvement Amendments (CLIA) and College of American Pathologists (CAP)-compliant performance characteristics. Thus, rapid entering of NGS into the space of clinical diagnostics is complicated by a number of loosely defined unknowns, including a coverage required for confident detection, data quality control, benchmarking bioinformatics pipelines currently producing discordant variant calls and overall validation of the robustness of NGS techniques.

Besides the technical obstacles outlined above, there are basic issues with existing “gold standard”: a number of novel, potentially very important insights into the tumorigenesis were gained by massive parallel sequencing for the first time, as they originated from observations that cannot be performed earlier due to intrinsic shortcomings of Sanger sequencing and conventional PCR. As example we can point to the clinical significance of intratumor heterogeneity with therapy resistance mutations being present with low prevalence already at the pretreatment assessment [15]. Currently, there is no consensus on how to differentiate these mutations from the FFPE artifacts [16]. Another set of important issues are incidental findings that may be revealed by NGS reads even if they were targeted for detecting of the mutation spectrum defined beforehand [17]. Examples of such incidental findings may include novel, not yet characterize mutations within target gene, common variants that may influence risks of other diseases or even germline mutations [18]. Such findings are generally hard to interpret in terms of their clinical relevance and lack a consensus on whether they should be communicated to physician or the patient at all. Finally, NGS allows one to detect mutations, which remain unseen if not being targeted by analysis. For instance, owing to implementation of the NGS into clinical practice, EGFR kinase domain duplication mutation has recently emerged as novel EGFR TKI sensitizing variant [19].

Despite recent publications of a number of clinical NGS guidelines clinical test development by a variety of organizations including American College of Medical Genetics [20] and College of American Pathologist [21], overcoming the obstacles mentioned above remains challenging, especially when the detected mutations are somatic, as the majority of guidelines are established only for the reporting of germline mutations. It is likely that both general recommendations and relevant protocols for sequencing, data analysis and clinical interpretation will continue to evolve in the process of sharing the outcomes of the tailoring of commercially available tests to the needs of particular practice or country, or panel customization. In this paper we describe particular obstacles we encountered while analyzing the clinical NGS dataset obtained using the TruSeq Amplicon—Cancer Panel (TSACP) and possible solutions to the problems presented.

Advances in target therapy development have introduced an opportunity of informed management of the malignant disorders using the guidance of their molecular profiles. Most conveniently, NGS techniques are capable of the detection of multiple genetic alterations at once, while retaining both an accuracy, informativeness and cost-efficiency. Despite the advantages that NGS techniques provides over conventional methods, their implementation often meets difficulties and should be accompanied equivalency of its analytical performance. For numerous targeted panels, a number of published NGS protocols have already demonstrated satisfactory results. Nevertheless, some of the obstacles remain.

In this study, TSACP technology was tested in the retrospective analysis of somatic mutations in 26 FFPE specimens collected from NSCLC patients. All hotspot alterations detected in clinically relevant genes, including EGFR, PIK3CA, BRAF, KRAS, NRAS, ERBB2, ALK and MET, were successfully validated by Sanger sequencing or Real-Time PCR. EGFR mutations were detected in eight patients out of 25, while high-prevalent mutations of KRAS were detected in two, and low prevalent in additional three samples. Some findings were less than typical. For example, one of the specimens harbored both activating EGFR p.L858R mutation and PIK3CA p.Glu545Lys mutation that was present at allele frequency of 15%. Another specimen with the deletion of EGFR exon 19 also harbored KRAS p.Gly12Val mutation that was detected at mutant allele frequency of 6%. Though cases of concomitant presence of EGFR and KRAS or PIK3CA mutations have already been described, they are very rare. In both cases, compound mutations were below the limits for the detection by Sanger sequencing [38]. Importantly, the shortcomings of Sanger sequencing and other conventional methods influence historic accumulation of the data on cancer related variants with lower prevalence and, therefore, provide a ground for underrating of these, potentially important, findings. Widespread use of high-sensitive mutation detection techniques should be accompanied by appropriate data analysis and their deposition into panel-specific databases that may facilitate further insights on co-occurrences of various mutations, their allelic frequencies and clinical significance of each finding. The latter aspect remains the most challenging due to common underreporting of the minor alleles.

Clear clinical need of identification of the low-prevalent mutations remains, at least in part, unmet due to use of FFPE for the tumor specimen long-term preservation. FFPE is convenient, cost-effective, and efficient for immunohistochemical staining and morphology analyses. In the meantime, it is unsuitable for the high resolution mutation analysis due to DNA damage artifacts introduced by fixation and paraffin-embedding procedures. In this study, the tissues were fixed using uniform protocol implemented at two independent laboratories. These FFPE samples are routinely produced by clinical workflow.

With the average coverage of 2084×, at least in theory, we were capable to detect mutations at allele frequencies of 1%. According to TCGA data, both lung adenocarcinoma and squamous cell lung cancer specimens harbor approximately 9 mutations per Mb, or, on average, 0.36 mutations per 40 kb, or the size of the TSACP panel [64, 65]. Considering that the sequencing efforts are concentrated in areas intrinsically important for tumorigenesis and, therefore, undergoing the selective pressure, one may expect the mutation rates a bit higher. In our experiment, we detected a bit over 500 variants per 40 kb of sequenced DNA in each sample, on average. In both histological types of lung cancer samples described in TCGA, the proportion of C > T substitutions was at approximately 20%, while in experimentally assessed DNA specimens extracted from FFPE blocks the percentage of C > T base substitutions was substantially higher (p < 0.01 for both lung adenocarcinoma or squamous cell carcinoma, by t-test), in one sample reaching 74%. Moreover, a majority of detected variants were present at frequencies of 0–10% (Fig. 1). Hence, we have to conclude that these changes are more likely to represent DNA degradation artifacts rather than to be a consequence of intratumor heterogeneity [66].

As it was already shown previously, in a majority of FFPE specimens, DNA degradation artifacts are seen in the allele frequency range of 0–10% and represented predominantly by C:G > T:A nucleotide substitutions. It is tempting to dismiss these artifacts by applying simple allele frequency threshold. Nevertheless, in our study we observed that some of the specimens produced a number of artifactual findings presenting as high frequency variants. We noted that the processing of these specimens resulted in remarkably low pre-normalization concentrations of libraries. One of these specimens harbored hotspot AKT1 p.E17K (or p.Glu17Lys) variant at allele frequency of 4%, which significance was overshadowed by highly prevalent artifacts. Therefore, limiting the analysis only to known hotspots while assuming that probability of artifactual detection of the molecular lesion within known hotspot is low would not resolve the reporting dilemma.

In attempt to determine the applicability of using simple mutation allele frequency cut-offs for sorting out FFPE derived artefacts, the relationship of allele frequencies and mutation counts at these frequencies was studied. Figure 3 illustrates that the DNA specimens that resulted in the production of libraries with high pre-normalization concentrations demonstrate a steady decrease in respective mutation counts along with growing frequencies of individual alleles. In contrast, the sequencing of DNA specimens that resulted in the production of libraries with low pre-normalization concentrations identified subset of called variants, with allele frequencies ranging between 9 and 20% (p < 4e−3). This subset could not be clearly filtered by recommended reportable allelic frequency ranges for an analysis of FFPE derived DNA specimens. Thus, it is more difficult to deal with, as it could be identified only when taking into account the range of pre-normalization concentrations for libraries entered into the sequencing run.

In this situation, an assessment of pre-normalization concentrations of libraries may be part for the sample-specific calculation of the detection cut-offs. Moreover, these cut-offs may be even mutation-specific: if the frequency of allele in question is below 10%, but its nature is different from C > T, the probabilities that the variant originated in damaged DNA base would be relatively low. On the other hand, there is some evidence that the DNA degradation artifacts are not exclusively represented by substitutions C:G > T:A. For example, when we applied Fisher’s exact test to mutation counts for all nucleotide substitutions detected in our samples and their frequency ranges, we observed that A > T, T > C, G > T and G > C substitutions may also be associated with increased probability of being FFPE derived artefacts, while for G > C and T > C substitutions these probabilities were significantly decreased (data not shown).

This means that instead of simple empirical rules based on allele frequency cut-offs or nucleotide substitution types, this task may require supervised machine learning approaches executed on paired tumor samples obtained from the same patient and validated on the similar datasets. In any case, the necessity of further research in this field is clear since formalin fixation remains the most common technique for long-term preservation of cancer specimens. An alternative is in introduction of additional library preparation steps that may reduce the counts of FFPE-related artifacts, for example, by treating samples with uracil-DNA glycosylase [16]. It should be also noted that an introduction of additional sample processing steps may influence sequencing outputs, the interpretation of these outputs, and, therefore, clinical decision. In particular, the treatment of samples with uracil-DNA glycosylase may significantly impact mutation allele frequencies [67], which, in turn, are paramount for evaluating potential benefits from the targeted therapy [68]. Therefore, an introduction of additional enzymatic processing requires a comparative study and a through validation before its adoption into the routine.

Despite relatively small size of the study, a single false positive and two false negative calls of the EGFR mutations were uncovered. Importantly, using of two or more combinations of aligning software and further indel realignment with variant callers may help to solve these discrepancies. It is important to note that, for the specimens of question, using of unified and, therefore, simplified analysis protocols would definitely result in non-optimal management of the disease. Aiming at eventual replacement of conventional molecular diagnostics with high-throughput NGS-based methods of NGS, we should remember that commonly used analytic pipelines may remain inefficient when dealing with insertions and deletions, thus, justifying the need for in-house pipelines and scripts.

In addition, a total of five presumably germline coding frame variants were identified, including KDR p.Q472H (or p.Gln472His), KIT p.M541L (or p.Met541Leu), TP53 p.P72R (or p.Pro72Arg), HNF1A p.G226A (or p.Gly226Ala) and ATM p.F858L (or p.Phe858Leu). For a majority of these variants, the frequencies in human populations were high, thus indicating that these variant were likely inherited form the parent. Nevertheless, it may be possible that the presence of one or more of these variants in patient’ genome may have an impact on tumorigenesis or on the therapy outcomes. For instance, the presence of KDR p.Q472H variant was previously shown to alter tumor angiogenesis and vascularization [69]. Moreover, in vitro studies showed that the sensitivity to variant-harboring cells to VEGFR2 inhibition is higher than that in the cells with wild-type genotype [69, 70]. As compared to the most common TP53-72P variant, the presence of TP53-72R protein augments apoptosis up to 15-folds, thus, to some degree protecting individual from neoplastic development and, possibly, modifying the response to conventional chemotherapy [71]. Though prognostic value of this polymorphism have already been shown in several case series and prospective studies in several tumor types [72, 73], its predictive role remains not well understood [74‐76]. Furthermore, KIT p.M541L variant was reported to confer an enhanced proliferative response to low levels of stem cell factor, though evidences supporting its predictive effect are controversial [77‐83]. Finally, despite relatively low frequency of the ATM p.F858L variant in human populations described within the 1000 Genomes Project (0.5%), according to previous publications, its nature is likely germline. Interestingly, in large population based cohorts, this variant was associated with an increase in risks for prostate [84], breast [85] and colorectal [86] carcinomas as well as chronic lymphocytic leukemia [87]. Additionally, in patients with childhood T-lineage acute lymphoblastic leukemia, carrying ATM p.F858L was associated with a variety of negative predictors, and worsened outcomes [88].

It is clear that accumulation of the knowledge on the prevalence of these variants in tumor specimens may lead to eventual recognition of these incidental findings as relevant to either personalized selection of therapeutic strategy, or to the risks of the cancer development in the relatives of the proband. No database currently accumulates these kinds of findings in depersonalized form. Therefore, in a majority of the cases, these germline variants remain unreported and, therefore, missed for further analysis. In our opinion, creation of panel-specific incidental finding databases is warranted, as it may hasten overall understanding of tumorigenesis.

Reporting of entire mutational spectrum revealed by targeted sequencing is questionable, at least until the clinically-driven guidelines on reporting of somatic mutations are established. The need for the development of panel-specific databases allowing analysis of co-occurrence and relevance of somatic mutations, CNVs and germline variants in de-identified form is evident. Further standardization of sequencing protocols, especially their data analysis components, may require assay-, disease-, and, in many cases, even sample-specific customization that could be performed only in cooperation with clinicians.