Bias-Corrected Targeted Next-Generation Sequencing for Rapid, Multiplexed Detection of Actionable Alterations in Cell-Free DNA from Advanced Lung Cancer Patients Free

Genotype-directed targeted therapies are revolutionizing cancer care. Genomic alterations in genes such as EGFR, ALK, KRAS, and BRAF have been validated as powerful predictive biomarkers in the management of non–small cell lung cancer (NSCLC), colorectal cancer, and melanoma; it is now standard to test for these mutations to personalize treatment decisions (1–7). Development of new genotype-directed therapies is widespread in solid tumor oncology, leading to increasing application of next-generation sequencing (NGS) panels that can test tumor biopsies for a wide range of potentially targetable mutations (8, 9). However, routine use of NGS for tumor genotyping presents practical challenges including the availability of adequate biopsy specimens, slow turnaround time, and the need for repeat biopsies after development of drug resistance (9). Given these challenges, it is clear that there is an unmet need for noninvasive assays that can broadly detect actionable genomic alterations.

Many groups, including our own, have investigated noninvasive tumor genotyping of cell-free plasma DNA (cfDNA) as an alternative to tissue genotyping (10–15). Rather than studying circulating cells, these technologies study the free floating DNA contained in the plasma; in advanced cancer patients, a portion of this cfDNA may be derived from the tumor. Plasma genotyping has the potential to be less invasive and faster than tumor genotyping, while also allowing serial assessment of genotype during development of treatment resistance. We recently reported on a highly specific and rapid droplet digital PCR (ddPCR) assay for quantifying the concentration of EGFR and KRAS mutations in cfDNA of advanced NSCLC patients (16, 17). Such PCR-based plasma assays test for mutations at a single site in a gene, but are limited by their inability to detect more complex genomic alterations such as chromosomal rearrangements and their inability to multiplex across several genes. Others have studied NGS of cfDNA using PCR amplicons or tagged DNA baits to enrich for target DNA sequences; however, many such assays are unable to detect rearrangements, whereas other assays rely on massive sequencing and computational processing resulting in unacceptable costs and slow turnaround time (11, 18). Although detection of mutant cfDNA present at low concentration is possible with these approaches despite the more abundant wild-type (germline) DNA, a universal challenge with these highly sensitive genotyping assays is the risk of false positives due to PCR artifact.

In this study, we piloted a novel-targeted NGS approach for the detection of driver mutations and rearrangements in cfDNA from advanced NSCLC patients. Taking cues from traditional hybrid capture approaches that isolate genomic subsets by pull down with probes to genes of interest, our methodology improves on key steps during library generation to reduce sequencing demands and turnaround times. First, to maximize on-target reads to ∼90%, a two-step pull-down process was used that includes both a thermodynamically controlled hybridization step and a kinetically controlled extension step under conditions that neutralize GC bias. Then, to improve signal-to-noise ratio, tags were connected to each captured DNA fragment, so that every read is anchored to its clonal family and to its pull-down probe of origin, facilitating identification of low-frequency mutant alleles and quantification of subtle changes in gene copy numbers (Fig. 1 and Supplementary Methods). We hypothesized that this approach would allow for accurate detection of a broad range of targetable genotypes, including insertions/deletions and rearrangements, in cfDNA from advanced NSCLC patients. Our goal was to leverage a desktop sequencing platform to enable a rapid turnaround time and facilitate widespread clinical adoption.

Targeted NGS of cell-line DNA and plasma cfDNA was performed at Resolution Bioscience as described in Supplementary Methods. Chimeric gene fusions were detected using tiled probes that allow sequencing-based discovery of de novo rearrangements (Supplementary Fig. S1 and Supplementary Table S1).

For comparison to plasma NGS, plasma ddPCR was performed using an established and validated assay which has been described previously (16). Briefly, this assay emulsifies extracted plasma cfDNA into thousands of droplets which subsequently undergo individual PCR with custom fluorescently labeled probes designed to detect EGFR L858R, EGFR exon 19 deletions, or KRAS G12X (16). Individual droplets are then read in a flow cytometer and the number of positive droplets are quantified (Bio-Rad). Each sample is analyzed in triplicate.

The targeted NGS panel was validated using genomic DNA from 14 independent, genetically-annotated cell lines harboring four gene fusions, 19 point mutations, and two insertions/deletions (Supplemental Table S2). Cell lines were combined into two separate DNA pools, each containing the genomes of seven cell lines, and systematically blended with normal, wild-type DNA to produce admixtures at 2.5%, 1.0%, 0.4%, and 0.1% dilutions. Prior to NGS, DNA pools were acoustically fragmented to an average size distribution centered around 165 bp and purified by two-sided SPRI to give fragment profiles of 150 to 200 bp that closely approximate cfDNA. Cell lines for the analytical validation experiment were obtained from ATCC (A549, H2228, SK-MEL-2, H1666, SW48), RIKEN (Lc-2/ad), the Broad institute (H1781, SW480, HCT116, H2347, HCC78), and the NCI (H3122). PC-9 and H1975 cells were obtained from the laboratory of Dr. Pasi Jänne. All cell lines were validated to be correct by short tandem repeat (STR) analyses.

Patients were identified during their routine lung cancer care at Dana-Farber Cancer Institute. Patients were deemed eligible if they had advanced NSCLC with a known tumor genotype, either untreated or progressive on therapy. Tumor genotyping was performed as part of routine care, either using conventional genotyping assays (PCR, FISH) or a targeted NGS panel when available (5, 19). All patients consented to plasma collection and analysis on an IRB-approved prospective plasma collection protocol or associated correlative science protocols. Following clinical validation, plasma NGS was explored in two patients with high suspicion of a targetable genotype missed on tumor genotyping.

Plasma was collected prior to initiation of therapy for untreated or progressive advanced NSCLC. Whole blood was collected into 10 mL EDTA containing “purple-top” vacutainer tubes, centrifuged for 10 minutes at 1,200 × g and the plasma supernatant was further cleared by centrifugation for 10 minutes at 3,000 × g. Cleared plasma was stored in cryostat tubes at −80°C until use. cfDNA was isolated using the QIAmp Circulating Nucleic Acid Kit (Qiagen) according to the manufacturer's protocol. DNA was eluted in AVE buffer (100 μL) and stored at −80°C until use.

To ensure data integrity in these experiments, the samples were identified only by sample key. Only the clinical team (GRO, AGS, RSA) that identified patients for study had access to the tissue genotype results. The teams involved in ddPCR and NGS analysis were blinded during data acquisition (plasma isolation, cfDNA extraction, library generation, NGS, and ddPCR analysis) and unblinding was done after NGS and ddPCR results had been reported to the clinical team.

A probe set was developed that covers portions of 11 genes known to be targetable oncogenic drivers in NSCLC. Selected coding regions of eight genes were sequenced (KRAS, EGFR, ALK, HER2, BRAF, NRAS, PIK3CA, MET, and MEK1). In addition, intronic probes were designed to detect genome-level rearrangements that create chimeric gene fusions in ALK, ROS1, and RET. The coding regions of the tumor suppressor TP53 were also included as a control because this gene is commonly mutated in NSCLC. Analysis of the performance of this probe set on plasma DNA showed >80% on-target percentage, which compares favorably with the less than 50% on-target percentage seen in previously published data using standard hybridization selection on plasma DNA (Supplementary Fig. S2 and ref. 18).

The targeted NGS panel was initially validated with dilutions of cell line DNA. Four different dilutions of two DNA pools were sequenced, each derived from seven cell lines harboring previously characterized mutations (Supplemental Table S2). Dilutions of 2.5% to 0.1% resulted in calculated allelic frequencies ranging from 4.5% to 0.01% (Supplementary Fig. S3). Variant calling algorithms (Supplementary Methods) were able to identify mutations that were present at 0.1% or greater with sensitivity and specificity of 88% and 100%. Diagnostic performance improved to 100% sensitivity and specificity for mutations present at an allelic frequency 0.4% or greater (Supplementary Fig. S3 and Table S2). This preliminary analysis of the cell lines allowed the setting of thresholds that ensured high specificity in the subsequent analysis of patient samples.

After validating the NGS platform with DNA dilutions, plasma samples from 48 patients with advanced NSCLC were studied, blinded to the tumor genotyping results (Supplemental Table S3). The median age of these patients was 57, 61% were female and 92% had extra-thoracic metastatic disease. Mean reads per sample was 8.9 million, with a mean coverage per base of 983 unique reads.

The sensitivity of the NGS platform was first studied in 29 of the 48 patients known by tumor genotyping to harbor EGFR and KRAS driver mutations readily assayed with ddPCR (Fig. 2A). Using the tumor genotyping as the gold standard, ddPCR of plasma had a sensitivity of 86%, whereas NGS had a sensitivity of 79%, not significantly different (P = 0.43). Both methods were more sensitive when more cfDNA was available. The allelic fraction of the mutant allele in cfDNA, calculated as the number of mutant reads over wild-type reads, was closely correlated for plasma NGS and plasma ddPCR (Pearson correlation = 0.93, P < 0.001; Fig. 2B).

Detection of rare mutations and rearrangements in cfDNA was next studied in a blinded fashion in 20 of the 48 patients with NSCLC known to harbor a rare mutation or rearrangement on tumor genotyping (Fig. 2C); these included 19 new cases plus one previously studied above, harboring both a KRAS mutation and a PIK3CA mutation. Plasma NGS was able to blindly detect six of eight cases with rearrangements and four cases with rare targetable HER2 and EGFR mutations. Sensitivity in this cohort was 75% (15/20), similar to the prior analysis (Fig. 2C). For each of the six rearrangements detected, the exact breakpoints and fusion partners could be mapped to the genome (Fig. 3A and Supplementary Fig. S4).

Specificity was studied in these 48 patients, each with tumor genotyping positive for an oncogenic driver mutation in EGFR, KRAS, ALK, ROS1, RET, BRAF, or HER2. These oncogenes are established as nonoverlapping on tumor genotyping and are therefore ideal gold standards for assessment of false positives (5, 7). Specificity of the plasma NGS platform was 100% for these seven driver genotypes, with a false positive rate of 0% (95% confidence interval, 0%–10%).

Detection of resistance mutations was explored in 15 of the 48 patients who had plasma collected after development of acquired resistance to a tyrosine kinase inhibitor (Table 1). Of 12 patients with acquired resistance to erlotinib or afatinib, plasma NGS detected T790M in eight. One of the 12 patients had been refractory to erlotinib and afatinib despite harboring an EGFR exon 19 deletion, and tumor NGS had identified high MET amplification. Blinded to the tumor genotype, plasma NGS similarly detected MET amplification, as evidenced by a significant increase in MET copies compared with control (Fig. 3B). No resistance mutations were identified in the remaining three EGFR-mutant cases. Studying two patients who had developed acquired resistance to crizotinib, plasma NGS identified two point mutations in the ALK kinase domain in one case with an ALK rearrangement, and no ROS1 resistance mutations in the other case with a ROS1 rearrangement. One patient was studied who had previously developed T790M-positive resistance to erlotinib, and subsequently developed acquired resistance to the investigational EGFR kinase inhibitor AZD9291 (20); plasma NGS identified two different DNA mutations encoding for EGFR C797S (Fig. 3C), a mutation recently described as a common mediator of acquired resistance to AZD9291 (21). Fourteen of 15 cases had resistance biopsies available for genotyping. Tumor and plasma results were concordant in 12 (86%), whereas in two cases plasma NGS did not detect an acquired T790M mutation detected in tumor. Altogether, sensitivity for the 62 known driver and resistance mutations from the 48 cases was 77% (48/62).

Finally, the plasma NGS assay was explored in two advanced NSCLC patients with high clinical suspicion of possessing a targetable genomic alteration which had been missed on prior tissue genotyping. The initial patient was a 66-year-old female never-smoker who had responded durably to empiric treatment with erlotinib and developed resistance; EGFR genotyping of a rebiopsy using a commercial PCR assay had identified T790M without any sensitizing mutation evident. Plasma ddPCR was first performed and detected 1508 copies/mL of an exon 19 deletion. Plasma NGS then confirmed the presence of a novel double-deletion in exon 19 of EGFR which would have been missed by many commercial PCR assays because these often detect only common exon 19 deletion variants (Fig. 3D). The second patient was a 28-year-old female never-smoker who had progressed on multiple lines of therapy, for whom previous tumor genotyping (including NGS) had revealed no targetable alterations despite four biopsies. Plasma NGS revealed high-level MET amplification which was subsequently confirmed by fluorescent in situ immunohistochemistry (MET:CEP7 ratio >5), and led to the initiation of crizotinib. For each case, turnaround time from blood draw to result reporting in this initial feasibility study was six business days.

In this blinded study, a targeted NGS of cfDNA from advanced NSCLC patients was able to accurately detect a broad range of targetable genomic alterations in NSCLC, including point mutations, insertions/deletions, and rearrangements, with no false positives. This report is the first, to our knowledge, to describe the accurate detection of ALK, ROS1, and RET rearrangements using a single plasma assay without prior knowledge of fusion partners. This represents a dramatic advance over PCR-based plasma genotyping assays which are limited to the detection of hotspot mutations in coding regions (10). This assay was also able to detect both canonical and novel resistance mechanisms, including MET amplification and EGFR C797S (22, 23).

Importantly, this approach uses widely available equipment such as standard EDTA-containing vacutainers and a desktop sequencing platform: any accredited molecular pathology lab with a MiSeq could, with the right technical expertise and bioinformatics support, implement this assay to guide the care of advanced NSCLC patients.

Although others have also studied plasma NGS of lung cancer, this is the first to describe comprehensive and blinded detection of a broad range of alterations with one clinic-ready assay. Detection of hotspot mutations using plasma NGS was described by Couraud and using the IonTorrent platform, achieving a sensitivity of 58% and 87% specificity but without the ability to detect rearrangements or amplifications (11). Newman and others recently demonstrated more comprehensive detection of lung cancer mutations in cfDNA and tumor tissue by hybrid capture using biotinylated oligonucleotide probes on a HiSeq (18); however, this study noted an inefficient capture of fusion rearrangements. Finally, targeted sequencing of cfDNA using PCR amplicons has been successful for detection of single nucleotide variant in multiple types of cancer (15); however, PCR amplicons are not trivial to multiplex and are inherently blind to gene rearrangements. The NGS platform described here overcomes the weaknesses of each of these prior approaches, allowing multiplexed detection of a broad range of alterations with no false positives using an efficient platform. Furthermore, turnaround time from time of blood draw can be as short as 6 days.

Intuitively, we found that the sensitivity of plasma NGS is improved in specimens with a higher quantity of cfDNA, although in some instances targetable genotypes could be detected in specimens with a relatively low number of GE sequenced indicating high DNA shed from the tumor. In this pilot study, sensitivity of plasma NGS was 77%, comparable to most plasma genotyping assays which have reported sensitivity in the range of 70% (24). Sensitivity improves with higher DNA concentration, highlighting how important it will be to understand why the amount of total cfDNA in plasma varies across such a wide range, and whether this variance is due to fluctuations in tumor biology or in methods of extraction. For example, we and others have described previously that plasma genotyping assays are more sensitive in lung cancer patients with extra-thoracic metastases (13, 15, 17, 18). Even with a moderately high sensitivity, the lack of false positives with this assay results in a 100% positive predictive value, meaning that this plasma NGS assay could be used as a screening step before a biopsy sample is taken for genomic analysis: if positive, the results are reliable and can potentially obviate the need for a biopsy, and if negative, a biopsy for further testing may still have value.

Our detection of various resistance mechanisms in patients with acquired resistance to TKIs, including the detection of two different EGFR C797S clones in one case, highlights the potential power of plasma NGS for understanding the heterogeneity of the resistant state. In addition, the C797S alleles identified with plasma NGS were clearly in cis with T790M (Fig. 3), a detail that would be difficult to decipher with a PCR assay and may have treatment implications (22). It is increasingly appreciated that resistant tumors are made up of clones with diverse biologies that may respond differently to therapy (25, 26). We recently showed that three molecular subtypes of acquired resistance to the investigational EGFR inhibitor AZD9291 are apparent by use of serial plasma ddPCR: although all patients started with T790M plus a sensitizing mutation, some lose T790M at resistance whereas some gain C797S (21). However, this analysis required five separate PCR assays: T790M, 19 deletion, L858R, and assays for two C797S variants. Alternatively, one plasma NGS assay can detect these five alterations plus detecting any novel resistance mechanisms that emerge. Our data further suggest that quantification of allelic fraction is similar using plasma NGS and ddPCR, suggesting either could be used to serially monitor plasma genotype concentration.

The findings described in this report are achieved by applying a novel bias-corrected capture technology that builds upon standard sequencing approaches to maximize the efficiency of on-target versus off-target and redundant sequencing reads. By reducing PCR artifacts, this assay can accurately detect mutation present in as few as 0.1% of sequencing reads; in contrast, some tumor NGS platforms are unable to accurately call a mutation unless detected in >10% of sequencing reads. Bias-corrected targeted NGS is an approach that can be applied to any sequencing platform to improve on-target coverage and reduce noise. Here, we apply bias-corrected targeted NGS to desktop sequencing on a MiSeq platform in order to develop a plasma assay with the potential to be rapid and clinic-ready, in contrast to more time intensive assays utilizing HiSeq platforms (18). Yet this NGS approach could also be applied to clinical sequencing panels or discovery efforts to improve sequencing coverage and reduce artifact, particularly when studying clinical specimens from small biopsies with scant DNA.

In conclusion, we have developed and successfully piloted a plasma NGS assay that, using a novel capture and analysis technique, can detect targetable mutations and rearrangements in plasma from advanced NSCLC patients. This is the first plasma NGS assay to demonstrate blinded, multiplexed detection of such a broad range of actionable alterations with no false positive results. By using a widely available desktop sequencing platform and standard vacutainers with the potential for a rapid turnaround time, this assay has the potential for broad uptake and application. Through reducing the barriers between NSCLC patients and genotyping, we hope that plasma NGS will be able to facilitate delivery of targeted therapies and improve outcomes for patients with advanced NSCLC.

A.G. Sacher reports receiving travel reimbursement from AstraZeneca and Roche. C.K. Raymond and L.P. Lim have ownership interest in Resolution Bioscience. P.A. Jänne reports receiving a commercial research grant from AstraZeneca and Astellas; has ownership interest in Gatekeeper Pharmaceuticals; and is a consultant/advisory board member for AstraZeneca, Chugai Pharmaceutical, Merrimack Pharmaceuticals, Pfizer, and Roche. G.R. Oxnard is a consultant/advisory board member for Ariad, AstraZeneca, Clovis, and Sysmex. No potential conflicts of interest were disclosed by the other authors.

Conception and design: C.P. Paweletz, C.K. Raymond, G.R. Oxnard

Development of methodology: A.G. Sacher, C.K. Raymond, A. O'Connell, L.P. Lim, G.R. Oxnard

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A.G. Sacher, C.K. Raymond, R.S. Alden, S.L. Mach, Y. Kuang, P.A. Jänne, G.R. Oxnard

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C.P. Paweletz, A.G. Sacher, C.K. Raymond, R.S. Alden, A. O'Connell, Y. Kuang, L. Gandhi, L.P. Lim, G.R. Oxnard

Writing, review, and/or revision of the manuscript: C.P. Paweletz, A.G. Sacher, C.K. Raymond, R.S. Alden, A. O'Connell, S.L. Mach, Y. Kuang, L. Gandhi, P. Kirschmeier, J.M. English, L.P. Lim, P.A. Jänne, G.R. Oxnard

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A. O'Connell, J.M. English

Study supervision: P. Kirschmeier, G.R. Oxnard

This work was supported in part by the Department of Defense, Conquer Cancer Foundation of ASCO, Phi Beta Psi Sorority, Stading-Younger Cancer Research Foundation, Expect Miracles Foundation, Harold and Gail Kirstein Lung Cancer Research Fund, and US National Institutes of Health grant R01CA135257 (P.A. Jänne).

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.