Circulating Tumor DNA Sequencing Analysis of Gastroesophageal Adenocarcinoma

Gastric cancer and esophageal/esophagogastric junction (EGJ) adenocarcinoma, together gastroesophageal adenocarcinoma (GEA), is a significant global health problem (1). Median overall survival (mOS) of stage IV GEA is 11 to 12 months with optimal palliative chemotherapy (2), and 16 months for erb-b2 receptor tyrosine kinase 2 (HER2 or ERBB2)–amplified tumors treated with trastuzumab plus chemotherapy (3). To date, ramucirumab, an anti-VEGFR2 monoclonal antibody, and pembrolizumab, an anti–PD-1 monoclonal antibody, are the only other approved biologic therapies in subsequent-line therapy (4–9). Development of targeted agents has been limited by low-frequency genomic alterations (GA) and interpatient heterogeneity, exacerbated by immense intrapatient heterogeneity—even at baseline diagnosis (10). Routine tissue-based next-generation sequencing (tissue-NGS) identified that at least 37% of GEA patients harbor gene amplification in receptor tyrosine kinases (RTK), including HER2, MET, EGFR, and FGFR2, and also downstream KRAS (11–14).

These GAs, although each individually relatively infrequent, may have both prognostic and, importantly, predictive significance in GEA patients. This precedent was set by targeting HER2 amplification with trastuzumab. However, only 47% of HER2-positive patients achieved objective response and mOS increased to only 13.8 months, though 16 months in the most strongly HER2-positive patients. Subsequent studies with other anti-HER2 agents were negative for first- and second-line therapy (15–19). These observations likely reflect a combination of factors, including intrapatient heterogeneity in HER2 amplification as well as inherent and/or acquired concurrent molecular resistance mechanisms.

Previously, we identified discordance between coupled synchronous primary and metastatic GEA lesions in 42% of single-nucleotide variants and insertions/deletions, and 63% of gene amplifications (10). However, in a small cohort of patients with “triplet-paired” primary-metastasis-ctDNA, NGS cell-free circulating tumor DNA (ctDNA-NGS) GAs were concordant with metastatic biopsies in 87.5% of cases, as defined by a predefined treatment assignment algorithm, suggesting that this noninvasive approach may be more effective in guiding targeted therapy selection in metastatic disease. The distributions of GAs assessed by tissue-NGS from early (20, 21) and advanced (22, 23) stage GEA patients have been reported. However, these studies relied on single-lesion testing at one time point, and therefore could not account for spatial nor temporal heterogeneity. Thus, we now turned to ctDNA-NGS, in conjunction with tissue-NGS, to obtain a comprehensive and more complete “snapshot” of GAs and their heterogeneity in GEA, in order to understand their implications for targeted therapy.

To accomplish this task, we analyzed the largest landscape cohort of GEA patients who have undergone ctDNA-NGS to date, which included a large clinically annotated subset comprised of patients from the University of Chicago (UC) and Samsung Medical Center (SMC). The goals of this study were several-fold. We first sought to evaluate the detection limit of ctDNA-NGS on clinical samples, and the clinical impact of ctDNA detection on early stage disease recurrence. We next assessed, in advanced disease, whether baseline ctDNA quantity and early serial changes correlated with clinical characteristics and outcomes. We then surveyed the landscape of GEA ctDNA-NGS GAs, including MSI-high, and compared incidences with tissue-NGS cohorts. To corroborate earlier observations, we further characterized heterogeneity between paired tissue-NGS and ctDNA-NGS at baseline and over time. Finally, we assessed the role of ctDNA-NGS in predicting response and resistance of matched inhibitors to various RTK amplifications, including HER2, EGFR, MET, and FGFR2. To our knowledge, this represents the largest and most comprehensive evaluation of the clinical utility of ctDNA-NGS in GEA.

Of 2,326 ctDNA-NGS tests performed between September 30, 2014, and July 11, 2018, on 1,780 gastroesophageal patients, 2,140 tests from 1,630 patients met inclusion criteria for diagnosis with adenocarcinoma of the esophagus, gastroesophageal junction, or stomach (GEA) after filtering out cases with reported non-adenocarcinoma or unknown esophageal carcinoma histologies (Table 1). A large subset of these cases were linked with deidentified patient data from the UC (N = 273 patients, 601 tests) and SMC (N = 96 patients, 97 tests) in Institutional Review Board (IRB)–approved tissue banks. All patient cohorts utilized in this study are described in Supplementary Table S1. This work was conducted in full concordance with the principles of the Declaration of Helsinki. All patients provided written-informed consent, where applicable, or such informed consent was waived by IRB-approved protocols for aggregate deidentified data analysis. Somatic tumor sequencing by Foundation One (Foundation Medicine) was also linked to the UC clinical data using 617 tests from 457 patients, of which 203 patients also had ctDNA-NGS testing performed.

Plasma circulating tumor DNA sequencing (ctDNA-NGS) results were obtained using the Guardant360 test (G360, Guardant Health; ref. 24). The variant allele fraction of somatic alterations in plasma cfDNA is dependent on multiple factors, including mitotic activity/cell turnover rates, vascular access, location and burden of disease, and biological tumor type. These variant allele fractions can also be artificially inflated due to broader genomic context in a sample, including amplification of the mutated gene or LOH at the locus in question. The assay's bioinformatics pipeline attempts to filter out alterations of presumed germline origin using a betabinomial model (25). Absolute plasma copy number was determined utilizing the mode of the normalized number of cell-free DNA (cfDNA) fragments covering each gene to estimate the fragment number corresponding to two copies to derive a baseline diploid value. All values of unique fragments for each gene were then normalized by this baseline value. The baseline derivation was informed by molecule counts data from a large set of normal samples from healthy donors' plasma. Note that the plasma copy number was related to two variables—the copy number in the tissue, and the amount of shedding of tumor DNA into the blood where the tumor DNA, and thus the copy number, was expected to be diluted by abundant leukocyte-derived fragments, the latter having a copy number of 2.0 for each autosomal gene. Centiles of gene copy number reported in the clinical ctDNA-NGS results were denoted by a “+” for absolute plasma copy number greater than 2.1 (<50th percentile), “++” for copy number greater than 2.4 but less than 4 (<90th percentile), or “+++” for copy number greater than 4 (≥90th percentile). In this study, absolute plasma copy number or presence/absence of amplification were used—not these percentiles. Adjusted copy number was calculated from the copy number/[maximum somatic variant allele frequency (maxVAF)+0.01] for each test. Values above the “Global-cohort” median-adjusted copy number for a given gene were considered amplified. Tumor mutational burden (TMB) estimation by ctDNA and tissue-NGS was provided by Guardant Health and Foundation Medicine, respectively, according to previously published methodology (26, 27).

Records from UC and SMC patients who had their initial blood collection for ctDNA prior to stage IV therapy initiation were chart reviewed for disease location at that time and categorized for presence/absence of involvement of: liver, lung, peritoneum, metastatic (M1) lymph node, bone, skin, brain, bone marrow, or other. The relationship between maxVAF and number of GAs with disease sites was evaluated using a Student t test, and across multiple categories using ANOVA. Survival analyses were performed as detailed below.

The percentage of patients with GAs (identified) in 1,627 patients was enumerated among the entire cohort using nonsynonymous GAs from each patient (initial test, if serial tests were available). GA distribution was also assessed within the subset of clinically annotated samples from UC and SMC, representing “Western” and “Eastern” cohorts. All patients with their initial test available (1,627/1,630) were included regardless of the presence of detectable GAs. Frequencies were calculated at the gene level per patient, and GA frequencies of ≥5% were reported. This approach calculated a denominator on a gene-by-gene basis accounting for the genes tested/absent in a given assay version (i.e., if only 900/1,627 assays included gene X, the denominator would be 900). Synonymous mutations were excluded from analysis, and the number of alterations reported was corrected for removal of these synonymous mutations, unless stated otherwise. Differences between proportions of UC versus SMC alterations were performed using a proportion test. Comparison between The Cancer Genome Atlas (TCGA), MSK Impact, and ctDNA-NGS results used TCGA and MSK-Impact data from Cbioportal (accessed on October 14, 2018) in combination with this cfDNA cohort (23, 28). Genes reported were filtered to those available in all three data sets, and comparisons were made using proportion testing (Supplementary Fig. S3; Supplementary Table S4).

The clinically annotated subset of samples was used for most analyses (Supplementary Table S1). Cox proportional hazards models were used for survival analyses and corrected using a likelihood ratio test in the Survival package in R. For gene-by-gene assessment, multiple comparison correction was performed using the Benjamini–Hochberg method. Survival was displayed using Kaplan–Meier curves generated by the SurvMiner R package.

For presurgical and minimal residual disease (MRD) analyses (Supplementary Table S3), patients were classified based upon their diagnosis, perioperative therapy, and surgical dates. A maxVAF detection cutoff of 0.25% was used based upon reported 100% sensitivity for single-nucleotide variants at this level (24), and patients were stratified into ctDNA “detected” or “not detected.” If ctDNA was sampled on multiple dates in a given interval (Supplementary Table S1), the first was used.

To evaluate the utility of serial ctDNA-NGS, patients were included if they had at least 2 serial tests between 20 days prior to and 150 days after stage IV diagnosis. If 2 subsequent tests were available within 150 days, the first was used.

The predictive utility of ctDNA-NGS was evaluated in the untreated “Baseline-cohort” by stratifying patients into “amplified” or “non-amplified” using either unadjusted (reported) or adjusted ctDNA-NGS amplification status. Aggregated adjusted ctDNA and/or tissue oncogene amplification was considered positive if either (i) amplified-adjusted copy number (as above) in the pretreatment ctDNA assay or (ii) tissue-NGS amplification in any patient sample was identified. Of note, tissue-NGS was only available for UC patients.

The majority of immuno-oncologic (IO)-treated GEA patients received IO agents (defined as any anti–PD-1/PD-L1 and/or anti–CTLA-4 antibody) in later lines of therapy. Patients were included in this analysis if ctDNA was collected within 60 days prior to IO initiation in stage IV UC patients.

Intrapatient heterogeneity was determined by identifying untreated stage IV UC patients with tissue-NGS from a primary and metastatic site within 42 days of their initial ctDNA-NGS (n = 34). Common genes to tissue-NGS and ctDNA-NGS panels (n = 72) were then compiled, and GAs were tabulated by gene and patient according to where they were identified (primary, metastasis, blood). GAs identified by tissue-NGS as a “VUS” or “equivocal,” or by ctDNA-NGS as “uncertain significance” were only included if the alternate assay identified the alteration as a non-VUS. Filtered germline GAs not clinically reported by ctDNA-NGS were also included if the GA was also called by tissue-NGS. Analysis was repeated excluding GAs that the ctDNA-NGS assay would be unable to detect due to technical limitations, as manually annotated (Fig. 4B).

All patient cohorts utilized in this study are described in Supplementary Table S1. The “Global-cohort” of ctDNA-NGS included 2,140 tests on 1,630 patients (Table 1). In the Global-cohort, the median age was 63, and 71% of patients were male. The primary anatomical tumor location was 53% gastric cancer versus 47% EGJ. Patient race, tumor grade, clinical HER2 status by conventional tissue testing (29), and tumor stage were unknown for the majority of patients, although disease was indicated as advanced/metastatic at the time of testing per submitted orders. The “Clinically annotated” cohort (N = 369 patients, 698 tests) included 273 patients from the UC and 96 from SMC. Comparing characteristics between the UC and SMC Clinically annotated cohorts, UC patients were older (median, 62 vs. 57.5, P = 0.003), predominantly proximal EGJ tumors (67% vs. 0%, P < 2.2 × 10⁻⁶), and included 5% stage II and 16% stage III patients compared with entirely stage IV patients in the SMC cohort. UC patients were also more frequently HER2-positive by clinical criteria (IHC 3+ or IHC2+/FISH+) with 22% versus 8% of patients positive in at least one tissue sample at any time point in their care (P = 2.3 × 10⁻⁵). These large Global and Clinically annotated cohorts were used for subsequent analyses.

Plasma cfDNA assays depend on shedding of tumor DNA into the circulation (ctDNA), which then mixes with normal plasma cfDNA that is derived from routine nonmalignant cell turnover. The maximal tumor somatic variant allelic frequency (maxVAF) in the plasma reflects the largest mutated ctDNA clone detected among all cfDNA present, and can be used as a proxy to estimate overall ctDNA quantity and to distinguish degree of subclonality of alterations at lower VAFs. However, gene amplifications must also be taken into account (26). In early analyses, we observed that patients who had already initiated therapy within 14 days before plasma collection (n = 12) had a lower mean maxVAF of 5% versus 11.6% in untreated patients (n = 144, P = 0.07), and more of these patients demonstrated undetectable GAs. Though not statistically significant, from this finding as well as observations from serial response assessments discussed below, we concluded that prognostic and predictive ctDNA-biomarker evaluations would be best derived from samples obtained in untreated stage IV patients (n = 144), referred to as the “Baseline-cohort” (Supplementary Tables S1 and S2).

Using the Baseline cohort, we then assessed maxVAF as a surrogate marker for disease volume/burden and confirmed a direct correlation between the number of involved disease sites and maxVAF (Fig. 1A; Supplementary Table S2). Fitting with this, patients with intact primary tumors had a higher mean maxVAF of 10.9% versus 6.5% [P = 0.09; 95% confidence interval (CI), 0.7–9.9; Fig. 1B]. Furthermore, patients with liver involvement (n = 39/144) had a higher mean maxVAF, 19.2% versus 6.2% (P = 0.001; 95% CI, 5.3–20.8), as did those with lung involvement (n = 19/144), 23.3% versus 7.6% (P = 0.01; 95% CI, 3.5–28.0; Fig. 1C). Conversely, those with “peritoneal-only” disease (n = 35/144), an aggressive subset of GEA, demonstrated the lowest mean maxVAF of 2.5% versus 11.9% in “non-peritoneal-only” (P = 5.1e⁻⁶; 95% CI, 5.6–13.6), with many “peritoneal-only” patients having undetectable ctDNA (Fig. 1D). These findings demonstrated that both disease site and burden strongly influence tumor DNA shedding and consequent ctDNA-NGS sensitivity.

Clinical ctDNA-NGS is generally performed in order to identify actionable GAs, but the amount of ctDNA being shed into circulation could itself potentially serve as a prognostic biomarker both in early- and late-stage disease. We tested this hypothesis first in the locally advanced “Pre-Neoadjuvant” cohort of patients at first diagnosis prior to therapy/surgery, and found that those with detectable ctDNA (defined as maxVAF ≥ 0.25%, n = 17/29) had shorter disease-free survival (mDFS) of 15.2 months versus unreached, though this did not reach significance (P = 0.1; HR = 0.2; 95% CI, 0.03–2.1; Fig. 2A; Supplementary Table S3). Importantly, patients with detectable ctDNA (n = 7/22) in samples drawn after curative-intent resection (median, 50 days; range, 20–135 days after surgery) had significantly diminished mDFS of 12.5 months versus unreached (P = 0.03; HR = 0.1; 95% CI, 0.01–1.1; Fig. 2B; Supplementary Tables S3 and S4). Resolution or persistence of detectable ctDNA helped predict nonrecurrent and recurrent disease, respectively, in representative cases (Fig. 2C and D). Sample size was inadequate to formally assess association of ctDNA clearance by neoadjuvant and/or adjuvant therapy. Despite these small numbers, presence and quantity of ctDNA were clearly prognostic in locally advanced disease, and should be validated in future large prospective studies with ctDNA-NGS assays optimized for this purpose.

Following this, because we observed that maxVAF correlated with burden/volume of disease, we hypothesized that higher maxVAF would portend a worse prognosis in the advanced setting. Within the Baseline cohort, those (n = 104) having below-mean (“low”) maxVAF (<9.7%) had an mOS of 14.8 versus 9.4 months for patients (n = 40) with above-mean (“high”) maxVAF (P = 0.1; HR = 0.7; 95% CI, 0.4–1.1; Fig. 2E). We next assessed whether serial ctDNA-NGS analysis could assist with prognostication. In the Baseline cohort, those on first-line therapy who underwent serial ctDNA-NGS (Supplementary Table S1) within their first 150 days from stage IV diagnosis who had a ≥50% decline in maxVAF (n = 23/35) survived a median of 13.7 versus 8.6 months for those that did not (P = 0.02; HR = 0.3; 95% CI, 0.1–0.8; time between serial-collections: median, 68 days; range, 28–108 days; Fig. 2F); individual representative patient cases are shown (Fig. 2G and H). Taken together, the maxVAF dynamics observed suggest that ctDNA-NGS could be used as an early prognostic biomarker, and studies assessing whether altering therapy earlier in “non-responders” may be warranted, akin to PET-directed therapy (30), in an attempt to improve outcomes.

Finally, we assessed whether maxVAF could assist in prognostication of patients treated with immune checkpoint inhibitors (IO) in the IO cohort (Supplementary Table S1). Twenty-seven patients in this IO cohort (any line of therapy: nivolumab, n = 12; pembrolizumab, n = 13; durvalumab + tremelimumab, n = 1; tremelimumab, n = 1) underwent ctDNA-NGS within 60 days prior to IO initiation. Patients with less than the median maxVAF of 3.5% (n = 14/27) had an mOS of 8.8 versus 2.5 months for those higher than the median, from IO initiation to death (P = 0.04; HR = 0.4; 95% CI, 0.1–0.96; Fig. 2I). This suggests that among IO-treated patients, those with higher disease burden have worse outcomes; IO-specific benefit within the low/high disease burden subsets should be confirmed with prospective controlled analyses to account for the generally recognized improved prognosis with low burden disease.

After determining the prognostic insight of maxVAF and its correlations between clinicopathologic features, disease burden/volume, and outcomes, while accounting for these observations, we next assessed the ctDNA-NGS GEA GA landscape at the molecular level. Of the 2,140 assays in the Global cohort, a median of 3 GAs was identified per test (range, 0–80 GAs), and at least 1 nonsynonymous GA was identified in 1,756 (82%) cases (Table 1; Supplementary Table S1). GAs were more commonly identified with proximal primary EGJ versus distal gastric cancer tumors (85% vs. 79%, P = 0.0009). Fifteen patients (0.9%) had ≥20 GAs identified in an individual test, and 10 (0.6%) had ≥20 GAs identified once excluding synonymous mutations. These cases included 4 known MSI-high patients and 1 POLD1 mutation. The mean number of detected GAs between EGJ and gastric cancer primary sites was significantly skewed by the presence of these MSI-high or POLD1-mutated gastric cancer patients. Excluding these few special cases, significantly more GAs were found in EGJ than gastric cancer cases (mean, 3.7 vs. 3.3, P = 0.005). Within the Locally advanced cohort, 81% of tests identified ≥1 GA at diagnosis. Overall, these findings demonstrate that at diagnosis most GEA patients, even in earlier stages, have identifiable ctDNA-GAs.

In addition to providing a survey of GA frequencies per sample, one can also infer TMB from the number of identified GAs, which may have therapeutic implications (31, 32). However, this is challenging using ctDNA due to more limited gene coverage potentially affecting precision, and also ctDNA quantity (directly related to cancer burden and tumor shed at the time of sample collection) influencing the raw number of detected GAs (r² = 0.82; P < 2.2e-16; Supplementary Fig. S1A). Therefore, we corrected this by calculating TMB relative to sequencing coverage and VAF (Supplementary Fig. S1B and S1C) as previously described (26). We then compared paired tissue-NGS and ctDNA-NGS TMB estimates (n = 86), which correlated relatively poorly with one another (r² = 0.15; P < 0.24), though both were now adequately independent of maxVAF after correction (Supplementary Fig. S1D–S1F). Significantly, all 6 patients with known MSI-high tumors demonstrated ctDNA-TMB scores ≥90th percentile of all tested samples, suggesting that for MSI-high tumors, very high ctDNA-TMB is readily detectable. Most importantly, by directly sequencing microsatellite regions, ctDNA-NGS identified 6 of 6 (100%) patients known to be MSI-high (via IHC and tissue-NGS), at a plasma maxVAF range of 0.09% to 47.7%, with obvious clinical implications (7).

We next assessed the detailed genomic landscape of the cohorts, including mutations, amplifications, indels, and splice variants. In the Global cohort, GAs were frequently observed in TP53 (53%), HER2 (17%), EGFR (17%), KRAS (15%), MYC (13%), PIK3CA (13%), and MET (11%; Fig. 3A; Supplementary Table S5A). GAs were further stratified into nonsynonymous mutations (Fig. 3B) and amplifications (Fig. 3C), where events in TP53, ARID1A, APC, and SMAD4 were typically mutations, whereas MYC, HER2, KRAS, EGFR, MET, and FGFR2 events were more often amplifications.

We next compared the UC and SMC cohort GA landscapes, reflecting representative Western and Eastern populations (Fig. 3A–C; Supplementary Table S5B and S5C). More frequent ARID1A mutations and KRAS, EGFR, and PIK3CA amplifications were observed in the UC cohort. Specifically comparing gastric cancer cases (excluding EGJ) among UC and SMC cohorts, a higher incidence of mutations in ARID1A and KRAS was still observed in the UC cohort, whereas mutations in PIK3CA were more common in the SMC cohort.

Finally, we evaluated whether there were significant GA rate differences between early- and late-stage disease, or between tissue-NGS versus ctDNA-NGS testing. Despite having comparatively few early-stage disease samples, within the “Clinically-Annotated” cohort, a direct correlation was observed between disease stage and number of alterations (Supplementary Fig. S2), and likely confounded by disease burden, as elucidated above. For further comparison, we compared tissue-NGS GA incidences from the previously reported TCGA cohorts representing early-stage primary tumors (stages I–III; N = 265; refs. 20, 21), the MSK IMPACT cohort (N = 305) representing predominantly primary tumor biopsies from newly diagnosed stage IV patients (23), and with ctDNA-NGS from the present Global cohort (N = 1,627), reflecting “whole-disease” burden and predominantly pretreated advanced disease (Fig. 3D; Supplementary Table S5D). TP53 mutations were significantly more common in MSK and TCGA patient samples (P = 8.4 × 10⁻¹⁵). Amplifications of MYC (P = 2 × 10⁻⁶), CDK6 (P = 0.003), and CCNE1 (P = 0.0006) were more common in TCGA than in the MSK and Global cohorts (Fig. 1D). HER2 amplification was seen in only 11% of Global cohort patient samples versus 29% in MSK and 25% in TCGA (P = 8.6 × 10⁻¹⁸). Most differences across the three cohorts likely reflected a combination of sample acquisition timing, intrapatient heterogeneity, and/or tumor shed limitations. Specifically, “HER2 conversion” is now well recognized after treatment with anti-HER2 therapy (33–35), and potentially accounts for lower incidence of HER2 amplification in the Global cohort, given that this cohort presumably reflects patients in later lines of therapy after already failing anti-HER2 therapy. This was addressed in more detail in HER2 analyses below. Moreover, acknowledging that some Global cohort cases would have low tumor DNA shed (e.g., Peritoneal-only gastric cancer) and others collected at inopportune time points (e.g., shortly after effective therapy), the analysis was repeated by (i) including only Global cohort cases with GAs detected and (ii) including only patients with a maxVAF ≥ 0.5%, to limit underestimation of ctDNA-NGS GA frequencies relative to tissue-NGS testing (Supplementary Fig. S3A–S3D; Supplementary Table S6A–S6D). Using this approach, TP53 mutation frequency differences lost statistical significance (therefore likely driven by the DNA shedding limitation), though they remained significant for HER2 amplifications (potentially driven by posttreatment HER2 amplification loss in later-line settings). Overall, the GA profiles from these cohorts using tissue-NGS or ctDNA-NGS highlight and contrast the incidences of GAs across tumor stages, treatment time points, tumor sites, and biologic compartments. Notably, there were generally higher incidences of targetable GAs, particularly RTK amplifications (e.g., MET, FGFR2, and EGFR), in the Global cohort than seen with tissue-NGS.

Gene amplification is clinically relevant in GEA due to the predominance of chromosomally unstable disease (CIN; refs. 20, 21). Thus, we specifically assessed the incidence of amplifications across the Global cohort and found 4,136 amplifications in 813 tests from 648 patients (39.8% of Global cohort cases). Focusing on the most immediately therapeutically relevant RTKs, both EGFR and MET demonstrated predominantly low-level ctDNA amplifications, whereas HER2 and FGFR2 included a subset of patients with extremely high-level ctDNA amplifications (Supplementary Fig. S4A). Generally, higher gene copy number in tissue samples has correlated with more clinical benefit from respective targeted therapies (36–38). By ctDNA-NGS, the plasma absolute gene copy-number level could reflect either homogenous amplification throughout all disease sites (in the context of the amount of ctDNA shed or maxVAF), or it could represent heterogeneity with spatially mixed amplified and nonamplified clones, again in the context of ctDNA shedding. In fact, we recently reported the high rate of GA discordance between tissue-NGS on primary and metastatic biopsies, which was most pronounced in RTK amplifications (10). As noted, the absolute level of ctDNA gene amplification is dependent on the plasma maxVAF (point mutations/indels). For instance, we noted that a low-level ctDNA amplification observed in the context of a very low/nondetectable ctDNA maxVAF usually represented very high tissue gene amplification in order for it to be observed in plasma. Reciprocally, low-level gene amplification in the context of very high maxVAF (i.e., high tumor burden) typically did not reflect clinically relevant high level and homogenously distributed gene amplification. Therefore, to address the limitation of tumor shed, plasma gene copy number was normalized by dividing by maxVAF+0.01. This “adjusted” copy-number method increased the ability to discern between high- and low-level tissue-NGS gene amplification in the settings of low or high ctDNA shed (Supplementary Fig. S4B). Overall, ctDNA analysis effectively detected cases with gene amplification, and when accounting for maxVAF, identified patients with RTK amplifications most likely to benefit from matched targeted therapy.

HER2 amplification is the only GA routinely assessed in newly diagnosed advanced GEA patients to date, thus we sought to investigate this GA as it pertained to ctDNA-NGS in more detail. As above, ctDNA-NGS identified 184 HER2-amplified (11.3%) patients within the entire Global cohort (first test result if serially tested). The distribution of amplification level across these ctDNA samples was 33/55/96 patients having “<50th/50th–90th/or >90th” percentile amplification (see Materials and Methods), respectively (gene plasma copy-number range, 2.1–84.1; median, 4.2 copies). To further assess HER2 amplification incidence and concordance with tissue-based analyses, while considering clinical characteristics like treatment timing, we focused on the Clinically annotated cohort. Among the 305 stage IV UC/SMC patients, 18.4% were HER2 amplified by ctDNA-NGS (range, 2.1–68.2 copies; median, 6 copies), and of these 305 patients, 35 of 158 (22.2%) with available tissue-NGS were HER2 amplified. When evaluating only clinically HER2-positive stage IV patients (Supplementary Table S1), only 36 of 58 (62%) of patients had detectable HER2 amplification by ctDNA-NGS (Supplementary Table S7). This was recapitulated in the Baseline cohort where 17 of 28 (61%) of untreated clinically HER2-positive patients also harbored HER2 ctDNA amplification. The discordance between tissue versus ctDNA-NGS HER2 status could be due to tumor shedding limitations but also intrapatient molecular heterogeneity. Thus, we further investigated the degree that each of these factors contributed toward the observed HER2 discordance between tissue-NGS and ctDNA-NGS.

At initial diagnosis, spatial heterogeneity of HER2, along with other GAs, has been recently detailed (10). Here, we sought to further expand on this finding with additional cases, and identified 34 newly diagnosed untreated stage IV GEA patients who had undergone ctDNA-NGS along with tissue-NGS of both baseline primary tumor and a metastatic site (“triplet-pairs”; Supplementary Table S1). When limiting to genes present in both ctDNA and tissue panels (n = 72), any GA was identified in 57%, 58%, and 62% of cases within the primary tumor, metastatic tumor, and ctDNA, respectively (Fig. 4A). However, of the 183 characterized GAs identified, only 48 (26%) GAs were universally concordant within triplet-pairs. Of these, 21 (44%) were mutations in TP53, which represented 81% of the TP53 GAs and were likely “truncal” in the evolutionary phylogenetic tree. Only 2 of 7 triplet-pairs were universally concordant for HER2 amplification. Notably, 14%, 11%, and 22% of GAs were uniquely found in the primary, metastasis, and ctDNA, respectively. Importantly, this analysis did not account for technical limitations of ctDNA-NGS due to the recognized inability to detect large-scale deletion or regions not sequenced. Excluding these tissue-based GAs, 149 GAs were observed across these 34 triplet-paired patients. Now, any GA was identified in 54%, 57%, and 74% in the primary, metastasis, and ctDNA, respectively, with 11%, 8%, and 27% of GAs uniquely detected in the primary, metastasis, and ctDNA, respectively (Fig. 4B). Combining tissue-NGS and ctDNA-NGS increased sensitivity for detection of HER2, EGFR, FGFR2, and MET alterations (Fig. 4C). This highlights the complementary benefit of using ctDNA-NGS together with tissue-NGS to overcome the inherent false-negative rates of either test, either due to spatial heterogeneity (tissue) or technical shedding limitations (ctDNA).

In addition to baseline spatial molecular heterogeneity, ctDNA-NGS may detect acquired resistance over time (temporal heterogeneity). First, we focused on the incidence of persistent HER2 amplification versus conversion to HER2 nonamplified status after failed first-line anti-HER2 therapy using paired pre-/posttherapy tissue and plasma samples. In this “Serial-HER2” cohort, upon disease progression, only 4 of 15 (27%) patients demonstrated persistent HER2 amplification by ctDNA-NGS (Fig. 4D). Two of these ctDNA-amplified patients also demonstrated persistent HER2 IHC 3+ expression. However, another patient retained tissue HER2 amplification, but lacked HER2 ctDNA amplification upon progression—likely a result of low tumor shed in this case. Those with persistent HER2 amplification by either ctDNA-NGS and/or tissue, posttherapy ctDNA-NGS identified additional acquired mutations in KRAS (G12D and T35A), NF1 (N1503S), and PIK3CA (E542K and S1008T), and coamplifications of BRAF, KRAS, PIK3CA, and FGFR1 as likely mechanisms of resistance (Fig. 4E; Supplementary Tables S7 and S8).

Next, we assessed resistance mechanisms to targeted therapy toward other pertinent RTK amplifications, including EGFR, MET, and FGFR2. Resistance mechanisms to anti-EGFR therapy were previously reported, and included loss of EGFR-amplified clones and/or GAs rendering upregulation of various bypass pathways including RTKs and MAPK/PI3K (10, 38). Patients harboring MET- and FGFR2-amplified samples treated with matched TKIs or monoclonal antibodies also revealed upregulation of similar bypass pathways in RTKs and MAPK/PI3K pathway GAs, and redirecting therapy based on observed ctDNA-NGS changes yielded promising results. Based on our findings, exemplified in five cases (Supplementary File S1 and Supplementary Fig. S5), it is apparent that baseline spatial and temporal heterogeneity are interrelated, because preexisting spatially distributed resistant clones were repeatedly selected under targeted therapeutic pressure, yet in some instances these were not identified at baseline and only became apparent over time. ctDNA-NGS identified resistance mechanisms to targeted therapy in evaluated patients upon progression and may direct optimal next-line therapy.

In the context of its role in measuring tumor burden/maxVAF and accounting for interpatient and intrapatient molecular heterogeneity, ctDNA-NGS may identify prognostic and/or predictive GAs. To assess this, the Baseline cohort (n = 144) was again analyzed for key genes (PIK3CA, BRAF, KRAS, HER2, FGFR2, MET, and EGFR) previously reported to have prognostic and/or predictive significance in GEA or other cancers.

Presence of PIK3CA mutation corresponded with shorter survival of 3.8 versus 13.6 months (P = 0.006; HR = 3.4; 95% CI, 1.6–7.2; Fig. 5A). Similarly, BRAF GAs corresponded with an mOS of 5.6 months versus 13.7 months in BRAF–wild-type patients (P = 0.02; HR, 3.0; 95% CI, 1.4–6.7; Fig. 5B). However, none of these or others evaluated remained statistically significant after multiple comparison correction and multivariate analyses (Supplementary Table S9). Within the 144 patient cohort, only 2 of 11 FGFR2-amplified patients and 2 of 11 MET-amplified patients received RTK inhibitors; therefore, survival analysis could not be robustly performed. These data suggest that mutations in PIK3CA and GAs in BRAF portend generally poor prognoses, but should be confirmed in larger clinically annotated homogenously treated studies.

Given that <50% of HER2-positive patients demonstrate response to first-line anti-HER2 therapy, we asked whether incorporating ctDNA-NGS could improve the predictive utility over standard single-lesion tissue-based HER2 testing. Across all “Baseline” stage IV patients having both ctDNA- and tissue-NGS at any time point, 21 of 86 (24.4%) harbored HER2 amplification by at least one approach, but only 13 of 86 (15.1%) were amplified by both (6/8 of discordant patients were identified by tissue-NGS only; Fig. 5C), and an additional 3 of 86 patients were considered clinically HER2 positive, but lacked amplification by tissue- or ctDNA-NGS. Among HER2-targeted patients in the Baseline cohort, 23 patients had received first-line HER2-directed therapy—either lapatinib (n = 7), lapatinib + trastuzumab (n = 1), trastuzumab (n = 14), or trastuzumab and pertuzumab (n = 1). An additional patient was excluded from survival analysis as they had received HER2-directed perioperative therapy before recurrence. Among the HER2-targeted patients, 19 of 23 were clinically HER2-positive by routine tissue analyses, and only 15 were ctDNA-NGS HER2 amplified. Relying solely on reported ctDNA-NGS HER2 amplification, in this small series there was no difference in survival compared with those with or without ctDNA-NGS HER2 amplification—12.7 versus 8.7 months (P = 0.4; HR = 0.6; 95% CI, 0.2–1.7; Fig. 5D). However, this failed to consider the relationship between copy number and maxVAF as noted earlier. After adjustment (copy number/maxVAF+0.01), patients with plasma copy numbers greater than the median (n = 10/23) demonstrated improved mOS of 15.9 versus 9.4 months (P = 0.07; HR = 0.4; 95% CI, 0.1–1.1; Fig. 5E). Further, we identified patients with low tumor DNA shed and therefore HER2 amplification not detected by ctDNA-NGS, as well as molecular heterogeneity missed by single site tumor profiling by comparing those with HER2 amplification present in tissue (n = 11/23) or adjusted ctDNA-NGS (n = 10/23—only 5 both tissue and ctDNA amplified). With this approach, patients with HER2 amplification had a 26.3 versus 7.4 month mOS in this “Complementary-amplified” group (n = 16/23; P = 0.004; HR = 0.2; 95% CI, 0.05–0.6; Fig. 5F). ctDNA identified actionable GAs in cases that would have been missed with tissue-testing alone. These findings focused on HER2 further delineate baseline and temporal molecular heterogeneity of GEA and demonstrate the importance of complementary tissue/plasma-NGS testing to best identify biomarkers of therapeutic relevance.

We recently assessed the prognostic and predictive nature of EGFR amplification in GEA (38). We sought to further evaluate this biomarker and focus on the utility of ctDNA-NGS. There was no difference in survival between the ctDNA-NGS–amplified (n = 12/130) and nonamplified untreated stage IV patients who did not receive EGFR-directed therapy (14.4 vs. 13.3 months; P = 0.6; HR = 1.3; 95% CI, 0.5–3.0; Fig. 5G), suggesting that EGFR amplification does not have specific prognostic implication in this cohort. In the Baseline cohort, 22 of 144 patients had ctDNA-NGS EGFR amplification and an additional 5 patients by tissue-NGS. Of these, 14 received EGFR-directed therapy—ABT806 (n = 12) or cetuximab (n = 2). Among EGFR-amplified patients, those who received EGFR inhibitors (n = 9/27) in any line of therapy showed an mOS of 21.1 versus 14.4 months versus patients who did not (P = 0.01; HR = 0.2; 95% CI, 0.06–0.8; Fig. 5H). This survival benefit was accentuated to 21.1 versus 6.2 months when comparing patients with either an adjusted copy number greater than median or tissue-based amplification (n = 9/14; P = 0.001; HR = 0.05; 95% CI, 0.006–0.4), despite limitations by small sample size (Fig. 5I). These data support that EGFR amplification, again optimally captured in complementary fashion by ctDNA and/or tissue-NGS, is not prognostic but potentially predictive of benefit to anti-EGFR therapy, consistent with previous reports (38, 39).

Herein, we present the largest comprehensive analysis evaluating the utility of ctDNA-NGS from a large commercial database with 2,140 individual tests on 1,630 GEA patients, and a substantial subset (698 tests from 369 patients) having clinical annotation for detailed clinicopathologic and outcomes analyses.

Using these cohorts, we first established an understanding of the detection limit of ctDNA-NGS as it relates to disease burden, disease site, and treatment timing. Though the median maxVAF was quite low, as seen in other studies, there was a long tail of patients with a high maxVAF. Biologically, this may be due to patients with very high maxVAF upon stage IV diagnosis, but technically, can reflect difficulty in filtering germline alterations in patients with high tumor shed and/or genomic instability. However, finding several genes at high level suggests biological origin, rather than technical (26, 40). For patients with low disease burden (few organ sites involved), peritoneal-only disease, and samples obtained shortly after therapy, each demonstrated lower ctDNA yield and in many cases nondetectable ctDNA. The biological reason for lower plasma ctDNA in peritoneal-only disease is uncertain, but may be attributed to less shed into the peripheral vascular system, which was recently noted in patients with peritoneal carcinomatosis in other cancer types (41), and/or different GAs which are not assessed by the ctDNA-NGS panel used. However, patients with peritoneal-only disease often have diffuse or mixed-type histology, and mutations in genes such as CDH1 and RHOA associated with this subtype (the TCGA “genomically stable” molecular subtype) are indeed part of the 73-gene ctDNA-NGS panel (20). It is noteworthy, however, that peritoneal-only disease often has insufficient DNA even for tissue-NGS, likely due to the low viable tumor content within dense desmoplastic tumors from both primary and metastatic biopsies. Therefore, future studies addressing these apparent molecular profiling limitations from both tissue and plasma of this difficult-to-treat subset of GEA patients are needed, as well as peritoneal fluid or lavage as potential sample types.

Regardless, from these limit-of-detection observations, we next determined that residual ctDNA detection after curative-intent resection reliably heralded eventual recurrence and worse prognosis in early-stage disease. This is consistent with reports from other tumor types (42) and suggests that postoperative ctDNA detection in GEA could be an important stratification factor within prospective adjuvant therapy studies. Moreover, via prospective studies, this biomarker may help to select those patients that should and should not receive further adjuvant therapy. However, we must be mindful of false positives in older patients resulting from clonal hematopoiesis. Three patients (all elderly) with detectable mutations after surgery, each at similar low maxVAFs prior to treatment/surgery, have not recurred to date, and none of these mutations were identified by tissue sequencing, which suggests that they may not be tumor-derived at all. Future strategies need to be mindful of both germline and hematopoietic confounding. Similarly, in advanced disease, we observed that baseline ctDNA quantity and early serial changes correlated with clinical characteristics and outcomes. It is possible that ctDNA-NGS may also prove useful here to assess whether patients benefit from changing therapy earlier in these “ctDNA non-responders,” prior to initial restaging CT scans. This hypothesis would be particularly interesting to investigate prospectively—especially when expensive or toxic therapies are employed and could be “fast-failed” early. In addition, this approach depends on having an effective therapeutic option on which to change, which would need to be validated. Our findings are corroborated by others, who also recently noted that changes in maxVAF for GAs reflected response to treatment, with an early spike in the first 1 to 3 days of effective chemo- or targeted therapy followed by order of magnitude drops in maxVAF, reflecting molecular response (43), but differ from that found when trending total cfDNA (44). Finally, as it pertained to levels of ctDNA in the plasma, we developed a framework to optimally identify and understand gene amplifications by adjusting for maxVAF in order to take into account tumor burden, spatial molecular heterogeneity, and DNA shed.

Focusing on the landscape of GAs in GEA as determined by ctDNA-NGS, we demonstrated that at first diagnosis, the vast majority of GEA patients, even in earlier stages, had identifiable ctDNA-GAs, especially after excluding those with peritoneal-only disease and recent therapy. Very importantly, we showed that all known MSI-high cases in our cohort having ctDNA-NGS performed were accurately identified—including a patient with a maxVAF as low as 0.09%. This is the highest sensitivity for plasma-detected MSI-H reported by any method to date and will be a useful tool to identify this relatively infrequent but highly targetable GA where traditionally tissue-based MSI testing is less routinely performed or insufficient tissue is available (7, 32). When comparing the ctDNA-GA landscape between the “Western” UC and “Eastern” SMC cohorts, we noted similar GA incidences, but there were also some interesting differences, even after considering only the gastric cancer UC subset with the gastric cancer SMC cohort. These differences included a higher incidence of KRAS and ARID1A GAs in the UC subset, which was consistent with prior literature (45), whereas the SMC cohort was enriched for PIK3CA mutations. The latter is remarkable because it has been reported that PIK3CA mutation is associated with EBV-positive gastric cancer (20, 46), which may also be more common in Asian countries (47), although the literature is conflicting (45, 48).

Comparisons of the ctDNA-GA landscape with cohorts published during article preparation (49, 50) and previously reported large-scale tissue-based analyses of GEA patients both revealed similar but not identical incidences of various GAs. When dissecting this further, we noted that incidence differences from these three large cohorts were mostly attributable to differences in disease stage, sample acquisition time points, along with differing disease sites and tissue compartments assessed. There were generally higher incidences of targetable GAs, particularly RTK amplifications, in the Global cohort. In this regard, ctDNA accounted for increasing intrapatient molecular heterogeneity, including at baseline and secondary to treatment pressure and evolving resistance. This served to survey the metastatic burden of patients best, in order to determine optimal targeted therapeutic regimens.

Along these lines, molecular heterogeneity both between and within patients has become a formidable hurdle to successful implementation of targeted therapies in GEA (10). Herein, we evaluated the largest “triplet-pairs” cohort reported to date for GEA, and again we uncovered significant discordance, including in routine known and potentially targetable RTKs such as HER2, EGFR, MET, and FGFR2. Together, these 4 RTKs account for approximately 30% to 40% of GEA patients, which make up a large subset of CIN tumors, and therefore a very significant consideration for ensuing targeted therapeutic decisions. Of note, the high frequency of EGFR amplifications in our cohort likely reflects a Western predominance of EGJ CIN tumors at our center. We also demonstrated numerous temporal resistance mechanisms, particularly after specific targeted therapy toward RTK amplifications, which included loss of the RTK amplification itself and/or GAs rendering upregulation of various known pathways to circumvent this inhibition strategy. Serial molecular profiling led to changes in treatment decision for these cases at disease progression points. Although EGFR amplification is not recommended to be routinely assessed by current guidelines, our work builds upon our previous and others' work suggesting benefit for these patients (38, 39, 51). It should be also noted that many patients in our annotated cohorts could not be assessed as “triplet pairs” due to insufficient tissue in either the primary tumor and/or the metastatic site at baseline and also at disease progression. This points toward the practicality of ctDNA-NGS to best assess baseline and temporal heterogeneity due to convenience, expediency, and less-invasive nature of a “liquid-biopsy” in the clinic. Our observations of vast interpatient and intrapatient molecular heterogeneity, spatially at baseline and temporally after therapy, are very much connected. A personalized treatment strategy that incorporates molecular profiling from both the tissue and the plasma at baseline and subsequently over time will likely be necessary in order to successfully improve outcomes of this disease with targeted therapeutics. In fact, we observed that incorporating tissue-NGS and ctDNA-NGS profiling in aggregate identified patients most likely to benefit from anti-HER2 and other targeted therapies. These findings mirror those seen in lung cancer with concurrent tissue- and ctDNA-NGS (52), and a recent report suggested that “first-pass” ctDNA-NGS for lung cancer patients may spare unnecessary redundant testing, with reflex tissue testing only if ctDNA-NGS is unrevealing (53). This may also be applicable for GEA and warrants attention. Ultimately, incorporating ctDNA-NGS may be a strategy to overcome recognized molecular heterogeneity, both at baseline and over time, and prospective innovative trials designs are ongoing to test this hypothesis (54, 55).

This study has some limitations. The Global cohort, albeit large, was relatively limited in clinical utility without the granular clinicopathologic characteristics to contextualize the GA distribution landscape. To address this, we combined two clinically annotated cohorts which provided robust understanding of GA events with clinicopathologic perspective, and subsequent analyses were restricted to samples drawn prior to any therapy to avoid underestimating ctDNA-NGS GAs and to perform tissue-plasma concordance studies more precisely. Another inherent limitation when comparing the 73-gene cfDNA-NGS versus 315-gene tissue-NGS panel is the expected discordance resulting from technical and biological differences between these different tests of distinct biological compartments. Technical limitations leading to discordance between tissue and plasma obviously included nonoverlapping genes, but also some regions of overlapping genes not sequenced on the ctDNA-NGS panel. Another technical limitation is the recognized inability of ctDNA-NGS to discern large-scale deletions among the vast sea of wild-type cfDNA. To account for these limitations and to focus on only those GAs that overlapped, we compared only those regions covered by both panels and excluded large deletions identified by tissue-NGS. This admittedly underestimates the level of “real-life” discordance that the clinical oncologist will observe. However, by doing so, we were able to focus on and identify specific biologic reasons for discordance, including disease burden and tumor site, which was directly related to tumor shed, as well as intrapatient spatial molecular heterogeneity. Finally, despite the relatively large size of the Clinically annotated cohort, inherent to low-frequency GAs, we were unable to definitively evaluate the prognostic importance of individual GAs nor the predictive impact of targeting these infrequent events.

In summary, clinical ctDNA-NGS testing holds promise for GEA—both in the detection of MRD in early-stage disease and as a serial tumor marker. ctDNA-NGS used in conjunction with tissue-NGS may be an approach to best identify actionable GAs and resistance mechanisms in order to overcome intrapatient heterogeneity. However, prospective validation of these findings in future studies is necessary for integration into clinical care.

L.A. Kiedrowski holds ownership interest (including patents) in Guardant Health, Inc. R.B. Lanman is an employee of Biolase, Inc., and holds ownership interest (including patents) in Guardant Health, Inc., Biolase, Inc., and Forward Medical, Inc. D.V.T. Catenacci reports receiving honoraria from Merck, Bristol-Myers Squibb, Astellas, Lilly, Taiho, Five Prime, and Gritstone, and reports receiving speakers bureau honoraria from Guardant Health, Inc., Foundation Medicine, and Tempus. No potential conflicts of interest were disclosed by the other authors.

Conception and design: S.B. Maron, L.A. Kiedrowski, R.B. Lanman, D.V.T. Catenacci

Development of methodology: S.B. Maron, L.A. Kiedrowski, R.B. Lanman, D.V.T. Catenacci

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S.B. Maron, L.M. Chase, S. Lomnicki, S.S. Joshi, J. Johnson, L.A. Kiedrowski, R.B. Lanman, S.T. Kim, J. Lee, D.V.T. Catenacci

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S.B. Maron, S.S. Joshi, J. Johnson, L.A. Kiedrowski, R.J. Nagy, R.B. Lanman, D.V.T. Catenacci

Writing, review, and/or revision of the manuscript: S.B. Maron, L.M. Chase, S.S. Joshi, L.A. Kiedrowski, R.J. Nagy, R.B. Lanman, D.V.T. Catenacci

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S.B. Maron, S. Lomnicki, S. Kochanny, K.L. Moore, S. Landron, L.A. Kiedrowski, S.T. Kim, D.V.T. Catenacci

Study supervision: S.B. Maron, D.V.T. Catenacci

The authors wish to thank all patients for generously participating in all clinical and tissue banking studies. The authors would like to thank Rajesh Acharya for his analytic tool.

This work was supported by Conquer Cancer Foundation Young Investigator Award, AACR Gastric Cancer Fellowship, Paul Calabrese K12 (S.B. Maron, 5K12CA139160), NIH K23 award (CA178203-01A1), UCCCC (University of Chicago Comprehensive Cancer Center) Award in Precision Oncology—CCSG (Cancer Center Support Grant; P30CA014599), Castle Foundation, LLK (Live Like Katie) Foundation Award, Ullman Scholar Award, and the Sal Ferrara II Fund for PANGEA (D.V.T. Catenacci).

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.