Noninvasive Diagnosis of Actionable Mutations by Deep Sequencing of Circulating Free DNA in Lung Cancer from Never-Smokers: A Proof-of-Concept Study from BioCAST/IFCT-1002

Somatic mutation analysis of known or potential druggable oncogenes has gradually become part of the routine practice in thoracic oncology (1, 2). In non–small cell lung cancer (NSCLC), the type of somatic mutations and the genes affected are related to smoking status. Indeed, lung cancers in never-smokers (LCINS) harbor a specific mutation profile characterized in particular by mutations in the EGFR tyrosine kinase domain (3). Thus, EGFR mutation status is now used to select patients eligible for a treatment with a tyrosine kinase inhibitor (4–7).

In thoracic oncology, tumor samples are often limited as they are mainly obtained through minimally invasive procedures such as bronchial (ultrasono-) endoscopy or CT-guided trans-mural punctures. Testing a wide range of molecular biomarkers in addition to pathologic diagnosis may thus be challenging because of the limited amount of material that can be extracted from these small samples (8, 9).

Circulating free DNA (cfDNA) are small double-stranded fragments of DNA found in plasma. In patients with cancer, the release of cfDNA in the bloodstream is thought to be related to apoptosis, necrosis, as well as active secretion from tumor cells (10). CfDNA recently attracted growing interest in oncology for multipurpose use (11). It has been considered as a prognostic and predictive biomarker (12, 13), and as a “liquid biopsy” to perform noninvasive testing for biomarker detection (14–20). However, the main challenge remains technical because, in many cases, tumor DNA (tDNA) may only represent a very small fraction of cfDNA (21, 22). The level of detection of tumor mutations in cfDNA is thus limited by the sensitivity of the detection method used.

Massive parallel sequencing comprises a new set of promising technologies in the field of cancer research and personalized medicine. Several studies showed that these technologies can be applied to the detection of somatic mutation and suggested that they can be used in clinical settings (23).

To date, no study has reported an inclusive multiplex analysis of relevant biomarkers in NSCLC from plasma cfDNA using massive parallel sequencing. Here, we used the IonTorrent Personal Genome Machine (PGM) for the deep sequencing of the most clinically relevant hotspot somatic mutations in tDNA and plasma cfDNA of never-smoker NSCLC patients.

Patients were consecutive cases recruited in the BioCAST/IFCT-1002 study (ClinicalTrials.gov identifier NCT01465854). BioCAST is a prospective, multicentric cohort sponsored by the French intergroup IFCT, designed to describe clinical, pathologic, and molecular epidemiology of LCINS in a French population. The detailed protocol is reported elsewhere (24). Briefly, newly diagnosed cases of NSCLC from self-declared nonsmokers were included and surveyed. All patients provided written informed consent for participation in the study and consent for blood and tumor samples collection. Overall, from December 2011 to January 2013, 384 patients were included in 75 active centers. IFCT ensured the quality insurance of data management according to international standards.

The present study was planned to be restricted to the 100 first consecutives cases of patients recruited with complete blood sample and tumor molecular analysis (up to May 2012).

The BioCAST study was approved by the Sud-Est IV Ethics Committee of Lyon and by the National Advisory Committee on Information Processing in Health, Paris. Sample collection was declared to French Ministry of Superior Education and Research and to the National Committee for Ethic in Informatics. This ancillary study was also approved by the International Agency for Research on Cancer (IARC) Ethics Committee.

tDNA samples were obtained from a network of genetic laboratories. This network was initiated by the French National Cancer Institute (INCa) to perform routine molecular analysis (2). The performance and concordance of sample processing and mutation screening methods across participating centers were previously investigated in blinded cross-validation studies (25). Standards for sample collection and processing used in each laboratory are described in Supplementary Table S1. Most of the paraffin-embedded samples were first microdissected. DNA was extracted using commercial kits.

All patients included in BioCAST underwent a blood sampling before receiving any cancer treatment. Samples were shipped at room temperature and within 24 hours from each center to the BioCAST central laboratory in Paris (Centre d'étude du polymorphisme humain - Fondation Jean Dausset). Blood samples were immediately processed upon receipt (mean time between delivery and processing: 19 ± 13 minutes) to isolate plasma. Plasma was isolated from EDTA tubes by centrifugation at 1,600 g during 10 minutes at 4°C. Plasmas were then aliquoted and stored at −80°C. Plasma samples were shipped to IARC according to international standards. CfDNA was extracted from aliquots (1 mL) of plasma using the QIAamp circulating nucleic acid kit (Qiagen) with the QIAvac 24 Plus vacuum manifold, following the manufacturer's instructions. CfDNA was quantified by picogreen.

A set of 12 primers pairs was designed using Primer3 software (v4.0) with default parameters (26) to amplify sequences of 98 to 125 bp covering hotspot regions of EGFR, BRAF, ERBB2 (HER2), KRAS, and PIK3CA (see Supplementary Table S2 for primer sequences and targeted regions). For amplicon production, 2 ng of DNA were used in multiplex PCR reactions (Supplementary Table S3) with the GoTaq HotStart DNA polymerase (Promega Corporation) and with the following program: 30 seconds at 94°C, 3 cycles of 30 seconds at 58.5°C, 30 seconds at 72°C, then, the annealing temperature was decreased from 0.5°C every 3 cycles until reaching 55.5°C; then 15 cycles of 30 seconds at 94°C, 45 seconds at 50°C, 30 seconds at 72°C, and a final extension of 7 minutes at 72°C.

One microliter of each PCR multiplex reaction was loaded on a gel to check levels of amplification and adjust for the quantity of each multiplex to be pooled by sample for equalizing multiplex representation. Multiplex pools were purified with Agencourt AMPure beads (Beckman Coulter Incorporated) and quantified by Qbit (Invitrogen Corporation). Library preparation was done using 100 ng of multiplex pools and the NEBNext End Repair Module (New England Biolabs) following the manufacturer's instructions. Individual barcodes (designed in-house and produced by Eurofins MWG Operon) were ligated to each multiplex pool for sequencing.

The libraries were sequenced with the IonTorrent PGM sequencer (Life Technologies) at deep coverage (aiming for 10,000X for plasma and 1,000X for tumors) using the Ion OneTouch 200 Template Kit v2 DL and Ion PGM Sequencing 200 Kit v2 with the 314 or 316 chip kits (all produced by Life Technologies), following the manufacturer's instructions. Six runs were performed to process all samples. For cfDNA, amplicons that failed to reach the targeted depth of coverage and were negative for an expected mutation (according to routine biomarker analysis of tDNA) were repeated in an independent run to obtain a minimum coverage of 10,000X.

The sequencing reads were aligned to the human reference hg19 genome with the IonTorrent Suite V3.4.2. The IonTorrent Variant Caller (ITVC) v3.4 was used for the detection of mutations in tDNA, setting a frequency above 5% for a variant to be called. Variant annotation was done with ANNOVAR (27). For the detection of ultra-low frequency variations in cfDNA, an in-house program was used to extract, from BAM files (Binary sequence Alignment/Map format), various statistics on read counts at the targeted hotspot positions. These statistics were used to filter variants based on three criteria: (i) minimum number of reads carrying the mutation ≥ 5 (a priori choice), (ii) variant allele frequency ≥ 0.2%, and (iii) variant reads strand orientation ratio ≥ 0.33 and ≤ 3 (a priori choice based on the range of strand bias values, 99% of 3,760 observations within this range, observed in plasma samples for reference alleles at positions of mutations reported in the COSMIC database within the analysed regions). To determine the minimum allele frequency threshold, we calculated the average variant allele frequencies (0.13%) for 272 non-hotspot mutations reported in the COSMIC database within the analyzed regions as an estimate of the background sequencing noise. Mutations were called if above this threshold. Mutations identified with these parameters were further checked by manual inspection of BAM files using the Integrative Genomics Viewer (IGV) 2.2 (Broad Institute, Cambridge, MA; ref. 28). Low allele frequency mutations (below 1.5%) in cfDNA were confirmed by independent PCR reactions and resequencing of the amplicons indicative of the presence of mutations.

Categorical variables are expressed in percent and compared with χ² tests if expected count in each category is at least 5. Plausibility of normal distribution assumption for continuous variables was assessed with the one-sample Kolmogorov–Smirnov test. Continuous variables are expressed in mean and SD or median and interquartile range (IQR) if non-normally distributed. Differences in distribution of continuous variables between two independent samples were assessed by Mann–Whitney U test, or Kruskal–Wallis one-way ANOVA when comparing more than two independent samples. Correlations between non-normal distributed variables were assessed with the Spearman correlation coefficient (rho).

Tests using plasma cfDNA were compared with tDNA analysis, which was considered as the reference. True positives were considered when both tDNA and cfDNA samples carried the same mutation and true negatives when both samples had no mutation; false positives were considered when a mutation was found in cfDNA but not in tDNA; and false negatives when a mutation was found in tDNA but not in cfDNA. Sensitivity and specificity were calculated as appropriate. In addition, we computed the mutation detection rate (n mutation detected in cfDNA/n mutation detected in tDNA) and concordance rate [(true positive + true negative)/n]. All analyses were performed with IBM SPSS Statistics version 20 (IBM Corp). All statistical tests were conducted two sided and a P value <0.05 was considered to indicate statistical significance. No correction of P values was performed in the course of multiple testing; however, results of all performed tests were thoroughly reported, allowing for an informal adjustment for multiplicity while reviewing the data (29). To express uncertainty of estimated probabilities, 95% Wilson confidence intervals (CI) were reported.

One hundred and seven patients were included in this study (Supplementary Fig. S1). CfDNA extraction failed in one sample. The main characteristics of the remaining 106 patients are shown in Table 1. Most of patients were women (88%) with adenocarcinoma (86%) and stage IV disease (75%). Characteristics of patients included in the molecular analysis (n = 68), and those with a stage IV disease (n = 50) are also shown in Table 1. These two subpopulations were similar to their respective complement populations except that adenocarcinomas were more frequently represented (P = 0.035 and 0.005, respectively).

Overall, the median cfDNA concentration in plasma samples was 67.1 ng/mL (IQR = 122.11 ng/mL). CfDNA concentration was not associated with the histologic subtype, tumor size (“T”), or node status (“N”). However, cfDNA concentrations were significantly higher in stage IV cases when compared with stages I to III (median concentration 92.3 ng/mL vs.34.7, respectively; P = 0.002), and in patients with higher number of metastatic sites (Table 2 and Supplementary Fig. S2). CfDNA concentrations were also significantly correlated, although modestly, both to stage and number of metastatic sites (Table 2).

Among the 107 patients recruited, 68 had both cfDNA and tDNA available for molecular analyses. The 68 matched cfDNA/tDNA samples were analyzed by deep sequencing (Supplementary Fig. S1). In tDNA, a hotspot mutation was found in 50 samples (74%) while 18 samples (26%) were wild-type at hotspot positions (Supplementary Table S4). Five mutations identified by next-generation sequencing (NGS) were not detected by routine clinical biomarker analyses. Among them, two had not been tested for the corresponding amplicons by routine clinical biomarker analyses, and three had variant allele frequencies ≤25% (which is the usual limit of level of detection by biomarker assays used in clinical settings). In addition, one tumor mutation expected from clinical testing was not confirmed by NGS. As shown in Fig. 1, mutations were mainly located in EGFR exon 19 (44%) and EGFR exon 21 (18%). No mutation was found in PIK3CA exon 20, BRAF exon 11, KRAS exon 3, and ERBB2 exon 19. No sample was found to carry multiple mutations. Most mutations were insertions or deletions (n = 29; 58%). The median mutant allele frequency was 20% (±29%) in tDNA sample.

In cfDNA, seven of the 50 (14%) mutations expected from tDNA results were called by the manufacturer's ITVC. However, using the in-house program and filtering strategy described in the Materials and Methods section, 28 (56%) tumor mutations were found in cfDNA with allele frequencies ranging from 0.2% to 44% (median ± IQR = 2.0% ± 12%; Fig. 1 and Supplementary Table S4). Three cfDNA samples had hotspot mutations (two in EGFR and one in PIK3CA) that were not called in tDNA and were thus considered as false positives. However, examination of the tDNA BAM files with IGV showed that two of these mutations were present at allele frequencies below 5% (allele frequency threshold value used for variant calling in tDNA).

Results of the comparative analyses of cfDNA and tDNA, using tDNA as reference, are presented in Table 3. When evaluating each amplicon independently, sensitivity ranged from 55% for EGFR exon 19 to 100% for EGFR exon 18. Overall, considering all amplicons, the sensitivity was 58% (95% CI, 43%–71%), the mutation detection rate was 52%, and the concordance rate was 68%. We estimated the specificity of the assay by taking into account only cases that reached at least 10,000X of coverage for cfDNA and 50X in tDNA (Table 4). We obtained a specificity value of 86% for PI3KCA exon 9, 88% for EGFR exon19, 100% for other amplicons, and 87% (62%–96%) overall.

Although it did not reach statistical significance, there was a trend for a better detection rate of point mutations compared with insertions/deletions: in the 19 patients with a tDNA mutation but no mutation detected in cfDNA, six tDNA mutations (32%) were point mutations and 13 (68%) were insertions or deletions compared with 15 and 16, respectively, for true positives (P = 0.376). Among cfDNA-positive patients, no correlation was found between mutant allele frequency in plasma and mutant allele frequency in tumor (Spearman rho = −0.227, P = 0.246), nor between mutation concentration in plasma (computed as mutant allele frequency in cfDNA multiplied by cfDNA concentration in ng/mL) and mutant allele frequency in tumor (Spearman rho = −0,051, P = 0.797). In a posthoc analysis restricted to stage IV cases only (n = 50), the sensitivity was higher 65% (Supplementary Table S5) but we found no significant difference or correlation between the detection of mutation in cfDNA and disease stage or any other clinical variable (not shown).

In this proof-of-concept study, we assessed the performance of the NGS IonTorrent PGM platform for the detection of tumor mutations in cfDNA using samples collected in conditions close to routine practice. Assay's sensitivity was 58% (26 mutations on 50 detected). CfDNA concentration was positively associated with tumor stage and with the number of metastatic sites, in agreement with other studies in lung and other cancer types (11–13). NGS for the detection of tumor mutation in cfDNA may thus be clinically meaningful because biomarker analysis is currently only useful for the selection of treatment in patients with lung cancer with metastatic disease.

To our knowledge, this study is the first to report the use of the Ion Torrent PGM for the detection of tumor mutations in cfDNA. The main advantage of this technology is that it allows the simultaneous screening of a broad range of known hotspot mutations in a large number of samples in a cost and time effective way compatible with clinical practice. Deep sequencing (10,000X) allowed the detection of mutations with frequencies as low as 0.2% in cfDNA. However, the manufactured PGM variant caller was not designed to call low abundant mutations diluted in wild-type genomic DNA (30). We thus had to use an in-house variant calling strategy, which improved the overall sensitivity from 16% (ITVC approach for all amplicons together) to 58% (Supplementary Table S6). This limitation of the manufacturer variant caller should be addressed in the future by dedicated software to allow the implementation of the test in clinical settings. Accurate detection of insertions and deletions is another limitation of NGS technologies, including the IonTorrent PGM (31). Given that the expected deletions and insertions were of defined size and location, they could be unequivocally detected with IGV, even those with an allele frequency as low as 0.2%. However, we found a trend for a lower detection rate of insertions/deletions compared with point mutations. This may be due to the principle of the technology that may fail to properly align reads to the reference genome in the presence of insertions or deletions. Using a NGS technology that uses paired-end sequencing may improve the detection of insertions/deletions in this setting. However, this would need to be investigated in a dedicated study directly comparing such technology with IonTorrent.

Several studies have previously investigated the potential use of cfDNA for the noninvasive detection of diagnostic biomarkers in lung cancer using various methods (Table 5). However, most have focused only on one biomarker and included small series of patients. In addition, the majority has only included advanced stages (IIIB/IV). The reported sensitivity of these studies ranged from 36% to 100%, and the specificity from 80% to 100%. Our results are thus consistent with those previously published in the field. However, our method has the advantage to assess several biomarkers simultaneously and to be easily scalable to a higher number of biomarkers.

Only a handful of studies have used NGS technologies for the detection of tumor somatic mutations in body fluids (22, 23, 32–34). Narayan and colleagues, for example, showed that a deep-sequencing approach can be an appropriate strategy for the detection of low abundant point mutations in surrogate tissues (32). Our work, assessing a larger panel of mutations, including both point mutations and insertions/deletions, and performed on a larger series of patients, further supports and extends the applicability of the NGS for the detection of aberrant genetic events in cfDNA.

In the present study, two EGFR and one PIK3CA hotspot mutations were detected in cfDNA but were not called in tDNA. Two of these mutations were found at a 4% allelic fraction in tDNA by manual inspection with IGV. We have set a 5% allelic fraction threshold for calling variant in tDNA to look for mutations representative of the tumor content. Results on these samples suggest that mutations poorly represented in tDNA can still be found in cfDNA. However, it is also possible that the analyzed tumor biopsy was not representative of the tumor content (sampling bias). Indeed, recent findings provide strong evidence that a tumor may host different subclonal populations carrying distinct somatic mutation profiles (35, 36). This point is particularly critical for the molecular analysis of small biopsies (36). Some studies have isolated circulating tumor cells to detect tumor-specific molecular alterations (37, 38). Although this approach would require less sensitive techniques for detecting relevant mutations, it may have the same drawback as biopsies because these cells may also represent a limited set of subclones of the tumor. Therefore, applying highly sensitive NGS-based techniques on cfDNA (that may come from different tumor subpopulations) may provide a better assessment of the spectrum of alterations present in the entire tumor. However, this hypothesis remains to be tested. Indeed intra-tumor heterogeneity about EGFR remains controversial, as a recent study suggested that heterogeneous distribution of EGFR mutations is extremely rare and may simply be artefacts due to differences in EGFR DNA sequences amplification (39).

In some samples, tumor mutations could not be detected in cfDNA even if the coverage obtained for the amplicons carrying the expected mutation was high (13 samples had above 30,000X coverage, including 5 samples above 60,000X). In other cases, some mutations were detected in cfDNA while the obtained coverage was low (between 400X and 800X). Thus, although high sensitivity allows the detection of low abundant mutations, parameters related to the tumor biology may be the most important factors influencing the capacity to detect tumor mutation in cfDNA. Our results show that metastatic stage is one major factor. Stage IV cases have higher concentration of cfDNA. These results suggest that a part of cfDNA may come from metastatic sites and that cfDNA reflects the tumor burden. However, the origin of cfDNA is still poorly understood and further studies would be needed to fully address this question.

A possible limitation of our study resides in suboptimal cfDNA samples. Indeed, the BioCAST study protocol allowed storage of blood samples at room temperature for up to 24 hours before plasma isolation. Thus, plasma DNA may have been contaminated with genomic DNA released from necrotic white blood cells, resulting in the dilution of tDNA in plasma, negatively affecting the detection of tumor mutations in cfDNA (10). This contamination with genomic DNA could also explain the absence of correlation between mutation allelic frequency in the tumor and mutation detection rate in cfDNA or mutation allelic frequency in plasma. Nonetheless, the detection rate obtained in these suboptimal conditions shows the potential of this approach in samples obtained in more controlled clinical settings. Indeed, plasma sample quality could be easily addressed in future cohorts by processing (centrifuging and freezing) samples within 3 hours after blood collection, which is achievable in routine clinical settings (40).

One of the strengths of our study is that the analyzed never-smoker population carries a high frequency of actionable mutations of various types that represent the most relevant driver mutations in the field of thoracic oncology. We were thus able to test our NGS IonTorrent PGM platform approach on a panel of deletions, insertions, and point mutations (12 hotspot regions targeting five genes) that have not been evaluated in other cfDNA-based studies.

In conclusion, we showed that targeted NGS with the IonTorrent platform for the detection of tumor mutations in cfDNA is applicable to clinical samples in lung cancer. The assay could therefore be a good alternative for initial molecular diagnosis when the obtained histologic sample is poor and could be used as a noninvasive test for assessing regularly the efficacy of targeted therapy by monitoring mutations in blood. Further studies on prospective validation cohorts are required to determine whether an optimized version of the assay may be applicable in routine clinical practice.

S. Couraud is a consultant/advisory board member for Astra Zeneca, Boehringer Ingelheim, and Roche. G. Zalcman is a consultant/advisory board member for AstraZeneca, Boehringer Ingelheim, Bristol-Myers Squibb, Eli Lilly, Pfizer, and Roche. No potential conflicts of interest were disclosed by the other authors.

The funding sources had no role in the design, analysis, and interpretation of the results, and thus, the authors were independent from the funding source.

Conception and design: S. Couraud, F. Vaca-Paniagua, H. Blanché, N. Girard, J. Trédaniel, M. Vincent, F. Morin, M. Olivier

Development of methodology: S. Couraud, F. Vaca-Paniagua, S. Villar, J. Oliver, N. Girard, M. Vincent, J. McKay, F. Morin, M. Olivier

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S. Couraud, F. Vaca-Paniagua, S. Villar, H. Blanché, J. Trédaniel, L. Guilleminault, R. Gervais, N. Prim, M. Vincent, S. Larivé, P. Foucher, B. Duvert, F. Le Calvez-Kelm, J. McKay, F. Morin, M. Olivier, P.-J. Souquet

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S. Couraud, F. Vaca-Paniagua, S. Villar, J. Oliver, T. Schuster, N. Girard, M. Vallee, F. Le Calvez-Kelm, J. McKay, G. Zalcman, M. Olivier, P-.J. Souquet

Writing, review, and/or revision of the manuscript: S. Couraud, F. Vaca-Paniagua, S. Villar, J. Oliver, T. Schuster, N. Girard, J. Trédaniel, L. Guilleminault, N. Prim, M. Vincent, J. Margery, S. Larivé, P. Foucher, M. Vallee, F. Le Calvez-Kelm, F. Morin, G. Zalcman, M. Olivier, P.-J. Souquet

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S. Couraud, N. Girard, P. Missy, F. Morin, G. Zalcman

Study supervision: S. Couraud, F. Vaca-Paniagua, N. Girard, P. Missy, F. Morin, G. Zalcman, M. Olivier

Other (organization and management of the collection of biologic samples): H. Blanché

The authors thank Stéphanie Labonne and William Lebossé, who were involved in the BioCAST study as senior and junior project managers, respectively.

The BioCAST/IFCT-1002 study was supported by research grants from AstraZeneca, Boehringer Ingelheim, Eli Lilly, Pierre-Fabre, Pfizer, and Roche. This work was also supported by CONACyT 169082 (to F. Vaca-Paniagua).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.