Validation of a Plasma-Based Comprehensive Cancer Genotyping Assay Utilizing Orthogonal Tissue- and Plasma-Based Methodologies

This article is featured in Highlights of This Issue, p. 3475

Modern systemic cancer therapy is evolving away from cytotoxic polychemotherapy to embrace novel therapeutics targeting molecular characteristics specific to individual patients' tumors that deliver superior clinical outcomes at reduced toxicity and overall cost (1, 2). Indeed, precision oncology is the preferred standard of care in National Comprehensive Cancer Network (NCCN) treatment guidelines (3) for advanced non–small cell lung cancer (NSCLC), colorectal cancer, melanoma, breast cancer, gastric and gastroesophageal adenocarcinoma, and gastrointestinal stromal tumors and is rapidly maturing in multiple other indications. Because of the increasing diversity of targeted therapies and associated molecular biomarkers, serial testing via traditional singleplex methods, such as PCR and FISH, has become impractical in some diseases (4). Congruently, the 2017 NCCN NSCLC practice guidelines explicitly recommend testing via “broad molecular profiling” by next-generation sequencing (NGS) methods validated to detect multiple variants simultaneously across all four genomic variant classes [base substitutions (single-nucleotide variants (SNV)), insertions/deletions (indel), gene amplifications (copy number alterations (CNA)), and gene rearrangements (fusions); ref. 4].

Even when broad molecular profiling is adopted, however, precision oncology has historically remained dependent on tissue-based biomarker assessment, which limits patient benefit in four primary ways. First, tissue sampling, whether by core needle biopsy, surgical biopsy, or fine-needle approaches, is necessarily invasive and burdened with the morbidity, mortality, costs, and delays associated with these procedures (5–7). Second, tissue sampling has variable but substantial failure rates due to patient ineligibility, procedural/sampling failure, or material insufficiency, which range from 20% to more than 50% in second-line advanced lung cancer patients (8–10). Third, tissue sampling is unavailable for a substantial minority of patients due to medical contraindication, patient unwillingness, and/or logistical concerns; congruently, extant data on genotyping rates demonstrate only moderate adherence to practice guidelines in the first line (65%–75%) and poor adherence in the second and later (<25%; refs. 11, 12). Fourth, tissue sampling is necessarily limited to a single, small tumor focus, which may not accurately represent all clinically relevant biomarkers due to spatial and temporal heterogeneity (13–16). Taken together, these data describe critical and avoidable missed treatment opportunities for patients with advanced solid tumors (8, 14, 17).

The advent of tumor genotyping from peripheral blood (“liquid biopsy”) via circulating cell-free tumor DNA (ctDNA) obviates these limitations and expands access to standard-of-care precision oncology to patients previously ineligible due to barriers associated with tissue sampling (4, 8, 9, 18–22). Despite this, comprehensive, structured validation studies of relevant patient populations of meaningful size using validated orthogonal comparator methods are lacking, with existing studies failing to adhere to established comparison study design principles (23). Here, we describe the analytic (technical characterization using controlled and often contrived samples) and clinical (characterization using real-world patient samples, contexts, and comparators) validation of a comprehensive ctDNA sequencing assay for clinical genotyping of advanced solid tumors, including large-scale comparisons with orthogonal clinical plasma- and tissue-genotyping methods. We further report our experience applying this technology in the clinical care of 10,593 consecutive patients in our Clinical Laboratory Improvement Amendment (CLIA)–certified, College of American Pathologists (CAP)–accredited, New York State Department of Health–approved laboratory, which represents the largest published dataset of clinical liquid biopsy experience to date.

All patient samples were collected and processed in accordance with the Guardant360 Clinical Blood Collection Kit instructions (Guardant Health, Inc.). Briefly, 2 × 10 mL of peripheral venous blood was collected in Streck Cell-Free DNA BCT (Streck, Inc.), packaged into a provided padded thermal insulation block within a secondary biohazard containment barrier, and shipped overnight to Guardant Health at ambient temperature. Although samples were accepted up to 7 days after blood collection, the median draw-to-receipt interval was 1.43 days. Upon receipt, samples were accessioned and proceeded immediately to plasma isolation (median accessioning-to-isolation interval 17 minutes). To isolate plasma, whole blood was centrifuged (1,600 × g for 10 minutes at 10°C), and the resulting supernatant was clarified by additional centrifugation (3,220 × g for 10 minutes at 10°C). Clarified plasma was transferred to a fresh tube and stored at 2°C for immediate extraction or stored at −80°C.

All cell-free DNA (cfDNA) extraction, processing, and sequencing was performed in a CLIA-certified, CAP-accredited laboratory (Guardant Health, Inc.) as described previously (24). Briefly, cfDNA was extracted from the entire plasma aliquot prepared from a single 10-mL tube described above (QIAamp Circulating Nucleic Acid Kit, Qiagen, Inc.), which yielded a median of 48.5 ng of cfDNA (range, 2–1,050 ng). Five to 30 ng of extracted cfDNA (66% of samples processed with 30 ng input) was labeled with nonrandom oligonucleotide barcodes (IDT, Inc.) and used to prepare sequencing libraries, which were then enriched by hybrid capture (Agilent Technologies, Inc.), pooled, and sequenced by paired-end synthesis (NextSeq 500 and/or HiSeq 2500, Illumina, Inc.). Contrived analytic samples were generated using similarly prepared cfDNA from healthy donors (AllCells, Inc.) and cfDNA isolated as above from the culture supernatant of model cell lines (ATCC, Inc.; Sigma, Inc.) and serially size-selected using Agencourt AMPure XP beads (Beckman Coulter, Inc.) until no detectable gDNA remained. Separate sequencing controls are utilized for SNVs and CNAs/fusions/indels (CFI). The SNV control comprises a mixture of healthy donor cfDNA pooled to target germline SNVs to 0.5%, 2.5%, and 6% allelic fraction. The CFI control comprises cell lines with known CNAs, fusions, and indels diluted into healthy donor cfDNA.

All variant detection analyses were performed using the locked clinical Guardant360 bioinformatics pipeline and reported unaltered by post hoc analyses. All decision thresholds were determined using independent training cohorts, locked, and applied prospectively to all validation and clinical samples. As described previously (24), base call files generated by Illumina's RTA software (v2.12) were demultiplexed using bcl2fastq (v2.19) and processed with a custom pipeline for molecule barcode detection, sequencing adapter trimming, and base quality trimming (discarding bases below Q20 at the ends of the reads). Processed reads were then aligned to hg19 using BWA-MEM (arXiv:1303.3997v2) and used to build double-stranded consensus representations of original unique cfDNA molecules using both inferred molecular barcodes and read start/stop positions. SNVs were detected by comparing read and consensus molecule characteristics to sequencing platform- and position-specific reference error noise profiles determined independently for each position in the panel by sequencing a training set of 62 healthy donors on both the NextSeq 500 and HiSeq 2500. Observed positional SNV error profiles were used to define calling cutoffs for SNV detection with respect to the number and characteristics of variant molecules, which differed by position but were most commonly ≥2 unique molecules, which in an average sample (∼5,000 unique molecule coverage) corresponds to a detection limit of approximately 0.04% variant allele fraction (VAF). Indel detection used two methods. For short (<50–70 bp) indels, a generative background noise model was constructed to account for PCR artifacts arising frequently in homopolymeric or repetitive contexts, allowing for strand-specific and late PCR errors. Detection was then determined by the likelihood ratio score for observed feature-weighted variant molecule support versus background noise distribution. Detection of indels ≫50 bp relies on secondary analysis of soft-clipped reads using methods described in the fusion section below and is only performed to detect specific genomic events (e.g., MET exon 14–skipping deletions). Reporting thresholds were event-specific as determined by performance in training samples but were most commonly at least one unique molecule for clinically actionable indels, which in an average sample corresponds to a detection limit of approximately 0.02% VAF. Fusion events were detected by merging overlapping paired-end reads to form a representation of the sequenced cfDNA molecule, which was then, aligned, mapped to initial unique cfDNA molecules based on molecular barcoding and alignment information, including soft clipping. Soft-clipped reads were analyzed using directionality and breakpoint proximity to identify clusters of molecules representing candidate fusion events, which were then used to construct fused references against which reads soft-clipped by the aligner on the first pass were realigned. Specific reporting thresholds were determined by retrospective and training set analyses but were generally ≥2 unique postrealignment molecules meeting quality requirements, which in an average sample corresponds to a detection limit of approximately 0.04% VAF. To detect CNAs, probe-level unique molecule coverage was normalized for overall unique molecule throughput, probe efficiency, GC content, and signal saturation and robustly summarized at the gene level. CNA determinations were based on training set–established decision thresholds for both absolute copy number deviation from per-sample diploid baseline and deviation from the baseline variation of probe-level normalized signal in the context of background variation within each sample's own diploid baseline. Per-sample relative tumor burden was determined by normalization to the mutational burden expected for tumor type and ctDNA fraction and reported as a Z-score.

In the analytic validation of the assay, we adhered to established CLIA, Nex-StoCT Working Group, and Association of Molecular Pathologists/CAP guidance regarding performance characteristics and validation principles. To define clinically meaningful performance, the entire assay workflow was mapped against possible clinical sample contexts, points and mechanisms of possible failure or inaccuracy identified, and validation studies designed to systematically examine each of these critical areas as per Clinical and Laboratory Standards (CLSI) guidance EP23-A. As such, the validation cohort comprises broad variant diversity and is heavily enriched for alterations with low VAFs and/or near decision boundaries. All studies were conducted using contrived materials and subsequently verified using patient samples, both of which were themselves validated using orthogonal methods and reference materials wherever possible. In addition, all studies were conducted at standard input amounts using the locked clinical Guardant360 bioinformatics pipeline and results reported unaltered by post hoc analyses. All decision thresholds were predetermined using independent training cohorts, locked, and applied prospectively to all validation and clinical samples.

To assess sensitivity for sequence variants (SNVs, indels, and fusions), cfDNA samples with orthogonally confirmed (e.g., ddPCR, tissue genotyping, exome sequencing, published characterization studies) variants were titrated into donor cfDNA targeting ≥5 prespecified VAFs at both standard (30 ng) and minimum (5 ng) cfDNA inputs. Multiple replicates of each titration point were then processed and used to calculate the detection probability for both clinically actionable and background variants and define the 95% limit of detection (LOD) both empirically and by probit regression. CNAs, in contrast, are not detectable at the level of individual molecules and are instead defined by statistical overrepresentation of reference sequences in the cfDNA. As such, we utilized the CLSI “Classical Approach” (CLSI EP17-A2) to statistically determine copy number LOD based on variation observed in titration series replicates and normal samples by calculating the limit where estimated copy number and Z-score values jointly have a 95% probability of exceeding the decision threshold.

Accuracy studies comprised orthogonally verified contrived and clinical samples representing the entire reportable range; however, cohorts were heavily enriched for variants at near and below the LOD and in unusual sequence contexts. Prevalence-adjusted PPV as a function of allelic frequency was defined as follows: |PP{V_i} = {\frac{{( {F_i^c - F_i^N} )}}{{F_i^c}}}$|⁠, where |$F_i^c$| is per-sample frequency of somatic calls in clinical samples within VAF range |i$|⁠, and |$F_i^N$| is per-sample frequency of false positive calls in known negative samples in allelic range |i$|⁠. A total of 2,585 previously analyzed consecutive clinical samples were used to define somatic alteration frequency by VAF. Frequency of false positive calls was defined using a set of 408 validation runs (324 LOD pools and 84 healthy cfDNA donor runs across both NextSeq and HiSeq instruments) for which truth was orthogonally defined. Analytic specificity was determined by analysis of pooled samples at high dilution that had been orthogonally defined at high VAF. Precision was determined utilizing in-process positive controls from the clinical sequencing workflow and pooled clinical samples.

For in-house assay ddPCR, sequence variant analyses were performed using the Bio-Rad droplet digital PCR (ddPCR) platform with QX200 droplet reader (Bio-Rad, Inc.). SNVs and indels were quantified using variant-specific PCR primer pairs and wild-type or mutant-specific fluorescent hydrolysis probes (Bio-Rad, Inc., IDT, Inc.). Copy number determination was assessed using a compendium of multiplexed PCR assays against reference genes in combination with target-specific PCR assays. Copy number was determined from the average of the reference gene signal relative to the total target signal. For external clinical ddPCR studies, deidentified clinical samples were submitted to the Dana-Farber Cancer Institute Translational Research Laboratory for ddPCR testing as described previously (25). The variant test was determined from available assays (EGFR L858R, exon 19 deletions, and T790M; KRAS G12X/G13D; BRAF V600X; ESR1 D538G and Y537C; and PIK3CA H1047R) by the Laboratory's choice based on a provided clinical diagnosis; no variant or result information was provided. For external tissue concordance, tissue genotyping was conducted as part of routine clinical care and comprised a variety of methods based on PCR, sequencing, ISH, or IHC.

All studies were conducted in accordance with recognized ethical guidelines (e.g., Declaration of Helsinki, CIOMS, Belmont Report, U.S. Common Rule) and with a waiver of patient consent by an Institutional Review Board–approved protocol.

The assay, Guardant360, utilizes NGS-based digital sequencing (DS; ref. 24) to comprehensively profile 73 cancer-related genes (Supplementary Fig. S1) comprising both therapeutically relevant biomarkers, including those for all approved and many investigational drugs and clinical trials, and cancer-specific biomarkers that may be used to establish both ctDNA presence and quantity. Briefly, cfDNA is extracted from stabilized whole blood, labeled at high efficiency with nonrandom oligonucleotide adapters (“molecular barcodes”), and used to prepare sequencing libraries, which are then enriched using hybrid capture and sequenced (Fig. 1). Sequencing reads are then used to reconstruct each individual cfDNA molecule present in the original patient sample with high-fidelity using proprietary double-stranded consensus sequence representation.

The combination of high-efficiency (80%–90%) input molecule labeling, optimized assay conditions, and hybridization probe design allows DS to recover 60%–85% of all input molecules postalignment (Fig. 2A), depending on cfDNA input, across the entire range of G:C content (Fig. 2B) with highly uniform coverage (13% quartile-based coefficient of variation with >96% of exonic panel positions covered above 0.2× median panel coverage; Fig. 2C). In addition, DS's in silico molecule reconstruction suppresses substitution and indel errors intrinsic to NGS workflows by more than three orders of magnitude relative to standard sequencing approaches, from approximately 10^−3.5 to 10^−6.5 per base (Fig. 2D and E), which allows accurate variant calling down to VAFs of 0.01%/1 molecule for indels and 0.04%/2 molecules for SNVs and fusions (summarized in Supplementary Table S1). Critically, >50% of molecules are reconstructed from both strands of the original cfDNA molecule, greatly increasing consensus sequence fidelity and specificity over other previously published approaches (26–32), which integrate both strands in only 12% or fewer of recovered unique molecules.

To validate the sensitivity and accuracy of SNV detection, donor cfDNA comprising 43 nonredundant germline variants defined by orthogonal exome analysis of leukocyte genomic DNA (gDNA) and patient sample cfDNA comprising 12 actionable somatic SNVs confirmed by ddPCR was titrated to 5 prespecified VAFs in triplicate bracketing the estimated 95% limit of detection (LOD₉₅) in 753 total observations. An additional patient cfDNA pool comprising 67 diluted germline variants and 22 actionable somatic SNVs was also processed at VAFs representative of clinical somatic variant prevalence. Within the titration series, the LOD₉₅ was determined empirically to be 0.3% (0.20% by probit) for actionable and slightly above 0.4% (94% detection rate, 0.34% by probit) for variants of uncertain significance (Fig. 3A; Supplementary Fig. S2A). Overall, detection above the LOD₉₅ was high (122/122, 100%), with high prevalence–adjusted PPV (99.2%). Importantly, even at VAFs below the LOD₉₅ (0.05%–0.25%), SNV detection remained moderate (35/55, 63.8%) and accuracy high (PPV 96.3%; Fig. 3C), which is critical for ctDNA analyses where many clinically relevant alterations are present at very low levels. When variants associated with clonal hematopoiesis (identified by reproducible detection on replicate analysis within and between sequencing platforms) were excluded, PPV across the validation dataset rose to 99.96%. Similarly, in 667 serial replicates of a well-defined cfDNA reference material used as an in-process sequencing control, only 20 of 42,496 total variants (VAF, 0.5%–6%) detected were unexpected, corresponding to an analytic PPV of 99.95% and a per-sample specificity of 97%. Importantly, in all studies, PPV of clinically actionable SNVs was 100% both above and below the LOD₉₅. SNV detection also maintained quantitative accuracy at VAFs near the LOD₉₅; across 871 observations between 0.05% and 1.0%, the correlation between observed and expected VAFs was high (r² = 0.84, y = 1.06; Supplementary Fig. S3A).

To validate indel detection performance, samples comprising 3 ddPCR-confirmed driver and 13 tumor suppressor indels across 11 genes were diluted to create an 8-point titration series comprising 200 near-LOD observations. The LOD₉₅ was established empirically at 0.2% and 0.7% for driver and tumor suppressor indels, respectively (0.11% and 0.43% by probit; Fig. 3A; Supplementary Fig. S2B). These data were supplemented with samples comprising 25 additional indels across 10 genes verified by either literature (cell lines), tissue genotyping (patient samples), or ddPCR (both) to determine an indel identification accuracy of 100% (34/34). Prevalence-adjusted PPV for all indel calls was estimated at 98.2% both above and below the LOD₉₅. PPV for actionable indels remained 100% across the entire reportable range of VAFs. Per-sample analytic specificity using donor cfDNA was 100% (42/42) and 98% (49/50) for undiluted and pooled donors and 99.7% (665/667) in sequencing control material described above.

To assess fusion detection performance, samples comprising 4 cell line- and 8 patient-derived cfDNA fusions were titrated to generate 93 near-LOD observations. The LOD₉₅ was determined empirically to be 0.2% (0.22% by probit; Fig. 3A; Supplementary Fig. S2C). Including additional cell lines and patient samples, analytic accuracy was determined to be 95% (20/21 unique fusions), and analytic PPV was 100% both above and below the LOD₉₅. Per-sample analytic specificity using donor cfDNA was 100% for undiluted (42/42) and pooled (50/50) donors and in sequencing control material (667/667).

To validate CNA detection performance, cell line cfDNA harboring known amplifications of ERBB2, MET, EGFR, MYC, CCND1, and CCNE1 was used to bracket the expected LOD₉₅ in 120 near-LOD observations. LOD₉₅ estimates constructed for each gene independently ranged from 2.24 copies (ERBB2) to 2.76 (CCND2, the smallest gene on the panel), median 2.44 (Fig. 3A; Supplementary Fig. S2D). Using multiplex-baselined ddPCR, we validated CNA accuracy in 14 cell lines comprising 43 amplifications (3–74 copies, median 4) in 10 genes. Importantly, DS results correlated closely with copy number results derived from both ddPCR (r² = 0.98, slope = 1.03) and Cancer Cell Line Encyclopedia microarray measurements (r² = 0.84, slope = 1.16; Supplementary Fig. S3B and S3C). As no unexpected positives were observed in the 257 characterized samples within the validation dataset, CNA detection PPV was determined to be 100%. Using well-characterized cfDNA CNA-specific sequencing control material, analytic PPV was measured to be 98.3% (226/230). Analytic specificity using healthy donor cfDNA was 100% for undiluted (42/42) and pooled (50/50) donors and 99.9% (3,712/3,716) in CNA-specific sequencing control material.

To determine run-to-run precision, we analyzed positive sequencing controls included in each batch of clinical samples. Across 375 consecutive runs of one SNV control lot, precision of exonic SNVs was 99.9% for 6% VAF (n = 749/750), 99.5% for 2.5% VAF (n = 8,583/8,625), and 97.6% for 0.5% VAF (n = 6,954/7,125). Across those same 375 runs, which comprised multiple copy number-fusion-indel control lots, indel (unique n = 4, 1.0%–7.5% VAF), fusion (unique n = 6, 1.0%–2.0% VAF), and copy number precision (unique n = 6, 2.4–5.2 copies) was 100%. Quantitative VAF and copy number measurements were consistent over more than 370 runs (Fig. 3B; Supplementary Fig. S4) with 95% of SNVs, indels, fusions, and CNAs within 32.5%, 26.1%, 73.6%, and 5.2% of targeted levels, respectively.

To validate clinical accuracy, we submitted 222 consecutive EGFR, KRAS, BRAF, and/or ESR1-positive NSCLC, colorectal cancer, or breast cancer specimens received for clinical testing to the Dana-Farber Cancer Institute (Boston, MA) for blinded analysis using a highly validated clinical ddPCR panel (Fig. 4A and B; Supplementary Table S2; ref. 25). DS and ddPCR were concordant for 269 positive calls [0.1%–94% VAF, positive percent agreement (PPA) 99.6%] and 308 negative [negative percent agreement (NPA) 97.8%]. Only a single variant, KRAS G12C, was detected by ddPCR but not DS (at 0.18%, below DS's LOD₉₅), while DS detected 3 EGFR exon 19 deletions and 4 T790M SNVs (0.10%–0.39% VAF) at or below ddPCR's reportable range (25), most probably due to DS's greater per reaction cfDNA input (DS interrogates up to 30 ng in a single reaction, whereas ddPCR spreads this across 3 separate reactions). Importantly, all discordant results occurred below either one or both assays' LOD and/or reportable range. Moreover, clinical follow-up of such samples demonstrated support for positive calls in 6 of 8 discordant cases (Supplementary Table S3).

We similarly assessed the accuracy of copy number determination relative to ddPCR in 70 samples with detectable CNAs in 8 genes (copy numbers 2.33–48.0) and 49 samples without. Together, ddPCR and DS calls were concordant for 63 of 70 positive calls and all 49 negative (PPA 90%, NPA 100%; Fig. 4A). Importantly, all 7 discordant calls were near the decision threshold of one or both assay platforms; outside of this region, concordance was 100%.

To validate quantitative accuracy, we compared DS's VAFs and copy numbers to ddPCR measurements (Fig. 4B and C). As expected, SNV and indel VAFs and CNA copy numbers were highly correlated (r² = 0.994, 0.995, and 0.943 respectively; y = 0.944, 0.922, and 0.980, respectively).

To assess ctDNA-tissue genotyping concordance, we conducted a blinded, IRB-approved retrospective review of samples received for clinical testing, distinct from the studies above, in which we extracted external molecular testing results from submitted pathology reports and determined whether variants identified by DS were present by established clinical genotyping methods. From an initial cohort of 6,948 consecutive NSCLC patients, we identified 543 for whom DS calls could be compared with external tissue genotyping results. Across seven genomic biomarkers (EGFR, ALK, ROS1, RET, BRAF, MET, and KRAS) clinically relevant for primary therapy selection, the positive concordance was 92% to 100% (Fig. 5). Importantly, post hoc clinical follow-up of the 3 patients positive for ALK fusions by DS but negative by FISH revealed that these patients had not only been treated with crizotinib irrespective of the diagnostic discordance but had also responded clinically as assessed by clinician judgment (33).

Validation studies, such as those described above, are critical to establish technical performance metrics and verify clinical performance; however, clinical utility is predicated on the ability to apply a test in real-world clinical care and return results in a reliable and timely manner. Moreover, test performance on large numbers of patients can often provide insights into test accuracy than smaller, controlled validation studies. To these ends, we analyzed consecutive clinical samples submitted to our CLIA-certified, CAP-accredited clinical laboratory. As our test is intended only for systemic therapy selection, all samples are derived from patients with advanced tumors. Within the study period, 10,593 samples were submitted with 10,585 (99.9%) comprising sufficient cfDNA for analysis and 10,547 successfully reported (99.6% analytic success rate; Fig. 6A). ctDNA detection rate overall was high (85.9%), predominantly driven by NSCLC (87.7%), colorectal (85.0%), and breast (86.8%); however, certain tumors, for example, primary central nervous system malignancies, were less likely to comprise detectable ctDNA, due to smaller typical tumor volumes and anatomic considerations, such as the blood–brain barrier (Fig. 6B).

Alteration prevalence was markedly enriched at low VAFs (median VAF 0.46%) even when restricted to targetable driver mutations (Fig. 6C), emphasizing the need for highly sensitive diagnostics. At least one variant with therapeutic or clinical trial connotations was identified in 72.0% (7,591/10,547) of samples, and although individual patients typically comprised fewer than two per sample, the superset comprised 3,479 unique alterations, underscoring the importance of broad genomic profiling. Congruent with previous reports of high ctDNA–gDNA concordance (8, 9, 24, 34, 35), per-variant and VAF-normalized relative tumor mutation burden landscape analyses closely match those of previous reports (Fig. 6D; Supplementary Fig. S5; refs. 36–38). At least one of 155 total unique variants associated with FDA-approved on-label therapies was identified in 16.7% (1,766/10,547) of samples, and at least one of 133 total unique variants associated with resistance to these approved drugs was identified in 13.9% (1,467/10,547). In colorectal cancer patients, 66.7% (548/819) comprised alterations associated with resistance to anti-EGFR antibody therapy, only 54.9% (301/548) of which were detectable by common (codons 12/13) KRAS testing (Fig. 6E). In 4,521 NSCLC patients, 34.5% comprised at least one alteration directly targetable with standard-of-care therapeutics (Fig. 6F), compatible with previously reported alteration frequencies (36, 37).

Precision oncology is associated with improved outcomes, reduced adverse effect profiles, and reduced overall cost; however, its application is limited by reliance on tissue-based biomarker assessment, which shackles it to invasive biopsy procedures associated with substantial cost, morbidity, mortality, and failure rates. In contrast, ctDNA-based liquid biopsies decouple biomarker assessment from tissue, enabling tumor-specific genotyping using safe, inexpensive, and near-painless peripheral blood sampling. Although promising, successful implementation of liquid biopsy is fraught with challenges including ctDNA's short fragment lengths (∼165 bp), scant material (median 48 ng/10 mL whole blood vs. micrograms in typical tissue sections), low tumor representation (VAFs typically 10- to 100-fold lower than tissue), and absent reference materials and methods.

We have developed an NGS-based ctDNA diagnostic that overcomes these challenges to provide highly accurate tumor-specific genotyping for all guideline-recommended advanced solid tumor somatic biomarkers from a single peripheral blood draw. This test demonstrates exceptional sensitivity (LOD₉₅ ≤ 0.3% for SNVs, indels, and fusions and near 2.2 copies for CNAs), while maintaining high PPV (>98%) even in sample cohorts enriched for alterations at or below the applicable LOD. Specificity and precision are similarly high, allowing accurate variant identification as low as 0.02% to 0.04%, which is critical for clinical ctDNA analysis as many relevant alterations are present at very low levels. Performance was verified in 349 clinical samples using internal and blinded external comparison to clinical ddPCR, which demonstrated very high qualitative and quantitative concordance. Analysis of 543 matched tissue genotyping results demonstrated high PPVs relative to standard-of-care tissue diagnostics. Of note, clinical response to targeted therapy in the three ALK fusion ctDNA-positive/tissue-negative patients is congruent with previous reports (8, 39, 40), highlighting a weakness of method comparison with imperfect reference methodologies as compared with the gold standard of clinical response. Clinical sensitivity was not assessable in this tissue genotyping comparison study; however, other studies have reported PPAs relative to tissue varying substantially between 52% and 100% (8, 9, 21–24, 34, 41–44), with most between 70% and 90%, depending on the specific biomarker, tissue test, clinical indication, overlap of comparator panels and capabilities, and time elapsed between tissue and plasma sampling (21, 23).

When used clinically for 10,593 consecutive patients, DS demonstrated robust technical success (>99.6%), ctDNA detection (85.9%), and rapid result delivery (median time from sample receipt to report, 7 days). Although at least one variant with therapeutic or clinical trial connotations was identified in 72.0% of samples, these comprised 3,479 unique alterations in aggregate, underscoring the importance of broad genomic profiling. Importantly, 16.7% of samples harbored variants with on-label treatment recommendations, whereas 13.9% harbored variants predictive of resistance to the same. Underscoring the importance of highly sensitive diagnostics, >50% of alterations were detected below 0.5% VAF.

These studies comprise the largest published ddPCR and tissue concordance series and demonstrate near-perfect concordance for driver alterations. Similarly, mutational prevalence and mutational burden analyses mirror findings from previous tissue-based reports (36–38). These high concordances strongly support existing evidence demonstrating the validity and utility of DS-guided treatment, which includes 17 clinical outcome studies for genomically targeted therapies (8, 9, 19–22, 34, 42–48) and TMB-based (49) immunotherapy. In contrast to limited/hotspot ctDNA panels, Guardant360's 73-gene design also allows interpretation of negative findings (required for effective biopsy prevention) and early detection of immunotherapy resistance (50, 51) through inclusion of all major driver oncogenes and tumor suppressors.

These data demonstrate applicability of ctDNA-based genotyping to treatment paradigms based on tissue genotyping; however, ctDNA has the potential to not only match but to surpass tissue genotyping in four important ways: First, ctDNA, as a circulating analyte, is derived from the entire tumor burden and thus has potential to capture tumor heterogeneity that tissue-based methods, which sample only a small focus, cannot achieve. Indeed, plasma-based detection of ALK fusions not present in focal tissue biopsies in the tissue comparison study above and those patients' subsequent therapeutic responses illustrates the potential impact of tumor heterogeneity. Second, the noninvasive nature of liquid biopsy enables longitudinal analyses not feasible with invasive tissue biopsies. Testing for acquired therapy resistance is a current application of longitudinal analysis with proven therapeutic relevance; however, other potential applications include therapy monitoring, early prediction of immunotherapy efficacy, and disease surveillance. Third, as ctDNA production is proportional to tumor growth, the fastest-growing tumor clones shed the most material, which makes ctDNA enriched for the most biologically aggressive, and likely most clinically relevant, tumor cells. Fourth, the quantitative accuracy of ctDNA-based genotyping enables discrimination of dominant versus subclonal alterations over both time and space and may potentially allow more accurate therapeutic targeting (13–16).

In summary, we present the analytic and clinical validation of a highly accurate noninvasive option for comprehensive tumor genomic profiling of advanced solid tumors, including high concordance to both orthogonal clinical plasma- and tissue-based genotyping assays, and present real-life clinical use data in >10,000 advanced cancer patients, demonstrating its feasibility and value in delivering both established and novel benefits of precision oncology to patients for whom tissue is unavailable.

J.I. Odegaard is an employee of Guardant Health. J.J. Vincent is an employee of and holds ownership interest (including patents) in Guardant Health. S.R. Fairclough holds ownership interest (including patents) in Guardant Health. O.A. Zill holds ownership interest (including patents) in Guardant Health. C.E. Lee holds ownership interest (including patents) in Guardant Health. L.A. Kiedrowski holds ownership interest (including patents) in Guardant Health. C.P. Paweletz reports receiving commercial research grants from Guardant Health and is a consultant/advisory board member for AstraZeneca, Bio-Rad, and DropWorks, Inc. R.B. Lanman is an employee of and holds ownership interest (including patents) in Guardant Health. D.I. Chudova is an employee of and holds ownership interest (including patents) in Guardant Health. A. Talasaz is an employee of and holds ownership interest (including patents) in Guardant Health. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J.I. Odegaard, S. Mortimer, J.V. Vowles, K.C. Banks, H. Eltoukhy, R.B. Lanman, D.I. Chudova, A. Talasaz

Development of methodology: J.I. Odegaard, J.J. Vincent, S. Mortimer, J.V. Vowles, S.R. Fairclough, R. Mokhtari, C.P. Paweletz, R.B. Lanman, D.I. Chudova

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.I. Odegaard, J.J. Vincent, S. Mortimer, J.V. Vowles, B.C. Ulrich, S.R. Fairclough, R.J. Nagy, C.P. Paweletz, R.B. Lanman, D.I. Chudova, A. Talasaz

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.I. Odegaard, J.J. Vincent, S. Mortimer, J.V. Vowles, B.C. Ulrich, K.C. Banks, S.R. Fairclough, O.A. Zill, M. Sikora, R. Mokhtari, D. Abdueva, R.J. Nagy, L.A. Kiedrowski, H. Eltoukhy, R.B. Lanman, D.I. Chudova

Writing, review, and/or revision of the manuscript: J.I. Odegaard, S. Mortimer, J.V. Vowles, B.C. Ulrich, K.C. Banks, S.R. Fairclough, O.A. Zill, R.J. Nagy, C.P. Paweletz, H. Eltoukhy, R.B. Lanman, D.I. Chudova, A. Talasaz

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.I. Odegaard, J.V. Vowles, C.E. Lee, L.A. Kiedrowski, D.I. Chudova

Study supervision: J.I. Odegaard, S. Mortimer, H. Eltoukhy, D.I. Chudova

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.