Purpose: The oxaliplatin plus S-1 and cisplatin plus S-1 regimens are interchangeably used in the management of advanced gastric cancer. The previously reported G-intestinal (G1) and G-diffuse (G2) intrinsic gene expression signatures showed promise for stratifying patients according to their tumor sensitivity to oxaliplatin or cisplatin.

Experimental Design: The proof-of-concept, multicenter, open-label phase II “3G” trial was done to prospectively evaluate the feasibility and efficacy of using genomic classifiers to tailor treatment in gastric cancer. Patients’ tumors were classified as “G1” or “G2” using a nearest-prediction template method, or “G3” (unclear assignment) when FDR ≥ 0.05. The first 30 patients in the “G1” cohort were assigned oxaliplatin plus S-1 (SOX) chemotherapy; thereafter, subsequently recruited “G1” patients were treated with cisplatin plus S-1 (SP) chemotherapy. “G2” patients and “G3” patients were treated with SP and SOX chemotherapy, respectively.

Results: A total of 48, 21, and 12 patients, respectively, were given “G1,” “G2,” and “G3” genomic assignments. Median turnaround time was 7 days (IQR, 5–9). Response rates were 44.8%, 8.3%, 26.7%, and 55.6% for the “G1-SOX,” “G1-SP,” “G2,” “G3” cohorts, respectively; and was higher in G1 patients treated with SOX compared with SP (P = 0.033). Exploratory analyses using the genomic classifier of Lei and colleagues validated the utility of the metabolic signature as a biomarker for predicting benefit from chemotherapy (log-rank P = 0.004 for PFS), whereas the Asian Cancer Research Group classifier did not demonstrate any predictive value.

Conclusions: This bench-to-bedside effort establishes a reasonable turnaround time for gene expression profiling and possible utility of genomic classifiers in gastric cancer treatment stratification. Clin Cancer Res; 24(21); 5272–81. ©2018 AACR.

Translational Relevance

Precision oncology increasingly hinges on molecular information to facilitate therapeutic decision making. This is the first study, to our knowledge, to prospectively assess a personalized approach in which patients with advanced gastric cancer are assigned to receive one of two standard-of-care first-line chemotherapy regimens on the basis of intrinsic gene expression profiling. Although we found no demonstrable clinical utility of the “intrinsic” classifier in question, exploratory analyses of two additional classifiers based on our prospective data suggested that the metabolic genomic signature developed by Lei and colleagues may predict benefit from chemotherapy. Importantly, this proof-of-concept study demonstrated that prospective gene expression profiling to guide treatment selection is feasible and can yield potentially actionable results with a reasonable turnaround timeframe. Furthermore, the 3G study provides the first prospectively curated gastric cancer dataset comprising high-throughput gene expression data and two defined chemotherapy regimens, which may be valuable for future biomarker studies.

Stomach cancer, the third leading cause of worldwide cancer-related mortality in both sexes, accounts for more than 819,000 deaths each year (1, 2). Approximately half of the global incidence of new cases occurs in Eastern Asia, where the mortality rates are also the highest (1, 2). The internationally accepted first-line therapeutic approach for HER2-negative advanced disease comprises a fluoropyrimidine plus platinum compound, with oral S-1 being a popular substitute for infusional fluorouracil in Asia owing to its favorable toxicity profile and convenient outpatient dosing schedule. Furthermore, on the basis of a phase III randomized study conducted in Japan (3) and the empirical noninferiority of oxaliplatin compared with cisplatin (3–7), the oxaliplatin plus S-1 (SOX) and cisplatin plus S-1 (SP) regimens have generally been considered interchangeable in the first-line setting.

One crucial area of enquiry however is whether it is possible to improve the selection of patient subsets who might derive greater benefit from either regimen. In our previous studies, we demonstrated that gene expression profiling could reliably classify gastric cancer into two intrinsic subtypes: “G-intestinal” (G1) or “G-diffuse” (G2), per a 171-intrinsic gene list (8). Intriguingly, there was evidence that the intrinsic subtypes could predict the likelihood of response to different chemotherapy regimens. Specifically, G-intestinal cell lines were more sensitive to oxaliplatin and 5-fluorouracil, whereas G-diffuse cell lines were more sensitive to cisplatin and 5-fluorouracil (8). We postulated that these findings may inform the choice of platinum agents in the first-line management of advanced gastric cancer.

We thus initiated the translational “3G” trial to test the benefit of using a genomic classifier to tailor the choice of cisplatin plus S-1 versus oxaliplatin plus S-1, and to determine the practicality of routine gene expression profiling to personalize chemotherapy regimens in the advanced gastric cancer setting.

Study design and participants

The “3G” study was a proof-of-concept, multicenter, open-label phase II trial done at three academic hospitals in Singapore and South Korea. Patients were eligible for enrollment if they were 21 years and older, had a histologically or cytologically confirmed diagnosis of locally advanced, metastatic, or recurrent gastric cancer and were indicated to receive first-line platinum-fluoropyrimidine doublet chemotherapy regimen. Participants had to have at least one measurable lesion according to the revised RECIST version 1.1 (9), Eastern Cooperative Oncology Group performance status 0 to 2, a life expectancy of at least 3 months, and adequate bone marrow, renal, and hepatic function determined by tests done within 14 days of treatment initiation. Patients were ineligible if they had received adjuvant chemotherapy within 6 months of the study, prior radiotherapy to target lesions, had symptomatic or progressive CNS metastases or other uncontrolled medical disorders. Full inclusion and exclusion criteria are detailed in the appendix.

All patients provided written informed consent before enrollment, and the study was done in accordance with the Declaration of Helsinki and International Conference on Harmonisation and Good Clinical Practice guidelines. The protocol was approved by the Institutional Review Board at each study site and complied with local laws and regulations.

Interventions

We obtained fresh endoscopic biopsies of the primary tumor in situ within 3 weeks prior to treatment initiation to establish tumor molecular profiles in eligible patients. Patients without adequate tumor tissue were not entered into the screening phase. Pathologic assessments of tumor histology and grade were not mandated by protocol as this is not the routine practice for endoscopic biopsies, but data were collected if available.

We allocated patients with the intestinal (“G1”) and unknown (“G3”) genomic subtypes to standard SOX chemotherapy (intravenous oxaliplatin 100 mg/m2 on day 1 plus S-1 given orally at 40 mg/m2 twice daily on days 1–14), and patients with the diffuse (“G2”) genomic subtype to standard SP regimen (intravenous cisplatin 60 mg/m2 on day 1 plus oral S-1 at an identical dose schedule) every 21 days. Treatment was continued until disease progression, unacceptable toxicity, or withdrawal of consent by the patient. Protocol-specified dose adjustments were based on treatment-related toxicities, graded according to National Cancer Institute Common Terminology Criteria for Adverse Events (version 3.0), and creatinine clearance, calculated with the Cockcroft–Gault formula. The full advice regarding dose modifications are detailed in the protocol (Supplementary Methods and Appendix).

Gene expression profiling and treatment assignment

Biopsy samples were immediately placed in RNA solution and shipped over for gene expression profiling and bioinformatics analysis at Duke-NUS Medical School, Singapore. The median tumor content of endoscopic samples was 60% [interquartile range (IQR), 40%–80%]. Total RNA was extracted using Qiagen RNeasy Mini Kit (Qiagen) according to the instructions of the manufacturer. Biotinylated cRNA were prepared according to the Affymetrix protocol from 250 ng of total RNA (GeneChip 3′ IVT PLUS Reagent Kit User Manual; Affymetrix Inc.). Following fragmentation, samples were hybridized on the Affymetrix Human Genome U133 plus Genechips (HG-U133 Plus 2.0; Affymetrix Inc.) for 16 hours at 45°C. GeneChips were washed and stained using the GeneChip Fluidics Station 450 (Affymetrix Inc.). Scanning was done using the GeneChip Scanner 3000 7G (Affymetrix Inc.). Raw Genechip data were processed with the robust multiarray average (RMA) algorithm (10). We did molecular subtype assignment of individual tumor samples using the nearest-template prediction method. Patients’ tumors were classified as “G1” (oxaliplatin-sensitive, intestinal subtype) or “G2” (cisplatin-sensitive, diffuse subtype) based on their resemblance to a previously published set of gene expression signatures (10, 11). Participants who met the inclusion criteria but whose tumor specimens had unclear (false discovery rate ≥0.05) or unavailable genomic assignment were allocated to the “G3” genomic arm. We also performed a post hoc analysis to assess the accuracy of initial classification, by normalizing individual patient tumors against previously published gene expression signatures (8) and reclassifying them using a dataset comprising tumors collected during this study. Unsupervised hierarchical clustering of genes was performed on the basis of the Euclidean distance with complete linkage. A separate analysis was done to assess additional predictive biomarkers using other published genomic classifications of gastric cancer developed in Asian cohorts (11, 12). Microarray data used for this study have been deposited in the Gene Expression Omnibus under accession number GSE100935.

Dose adjustments

For first-time occurrences of grade 3 neutropenia or thrombocytopenia, or any intolerable grade 2 nonhematologic toxicity, all study agents were withheld for up to 21 days until recovery to grade 1 or less; if the same adverse effect recurred, the doses of all drugs were reduced by 25%. And if grade 4 neutropenia or thrombocytopenia, or any grade 3 nonhematologic toxicity or febrile neutropenia was recorded, treatment was interrupted until toxic effects resolved and resumed at 25% reduced dosage at the next cycle. Chemotherapy was permanently discontinued if grade 4 nonhematologic toxicities were encountered or if adverse effects had not resolved after interrupting treatment for 21 days. Furthermore, for second and subsequent cycles, the doses of cisplatin and S-1 were algorithmically adjusted according to creatinine clearance: Cisplatin dosage was lowered by 25% and 50% if creatinine clearance was between 51–59 and 41–50 mL/minute, respectively; S-1 dosage was lowered by 25% if it was between 41 and 59 mL/minute; and both cisplatin and S-1 were discontinued permanently if it fell below ≤40 mL/minute. Up to two dose reductions of any chemotherapeutic agent were allowed, and dose escalation was not permitted after dose reduction. The full advice regarding dose modifications are detailed in the protocol (Appendix).

Outcomes

The primary endpoint was best overall response as per RECIST (version 1.1) criteria (9) and was assessed by CT scans or MRI done at 6-week intervals. Tumor assessments were performed by radiologists who were masked to treatment assignment, and outcomes were locally reviewed. Planned secondary endpoints were PFS, defined as the time from enrollment until documented disease progression or death from any cause, and feasibility, as assessed by the number of working days from sample receipt to return of genomic results. At the data cutoff of May 3, 2016, the median follow-up for overall survival (OS) had been 41.9 months and OS data were considered relatively mature. Hence, a decision was made to undertake an unplanned analysis of OS, defined as the time from enrollment to death from any cause, as this would afford more clinical evidence to support or refute the utility of the genomic classifier being investigated.

Statistical analysis

On the basis of A’Hern's single-stage phase II design (13), we calculated that a sample size of 30 patients per study arm would provide approximately 90% power to demonstrate a response rate of at least 70%, under the null hypothesis that the proportion of patients with an objective response is 40% or less (14, 15), with a one-sided type I error rate of 2.5%. We initially estimated that approximately one third of eligible patients would be classified as G3. Hence, to enroll 60 subjects in arms G1 and G2, an estimated 30 subjects would be accrued into the G3 arm, which would provide a 95% confidence interval (CI) of 23% to 59% assuming a historical response rate of 40%. When the planned target accrual of 30 patients was reached in the G1 arm in January, 2013, one study site amended their local protocol such that further patients with the G1 subtype could be treated with cisplatin plus S-1 per the institutional standard of care.

Analyses were done on a modified intention-to-treat basis, which took into account patients’ initial genomic assignment and the chemotherapy regimen they were ultimately given. Hence, the modified intention-to-treat population comprised four arms: “G1 subtype-SOX regimen,” “G1 subtype-SP regimen,” “G2 subtype-SP regimen,” and “G3 subtype-SOX regimen.” It is important to note that this study was not designed to compare between groups, and the following between-group comparisons were undertaken on a post hoc basis. Differences in the proportions of patients with an objective response were analyzed using Fisher exact test, and exact binomial distributions were used to provide 95% CIs and P values. We estimated PFS and OS with the Kaplan–Meier method. Patients who did not experience progression or death by the time of analysis were censored at their last follow-up. The log-rank test was used as the primary test of an overall difference in Kaplan–Meier curves. We used Cox regression analyses to generate HRs and 95% CIs and estimated the median follow-up duration of OS using the reverse Kaplan–Meier (16). P values of less than 0.05 (two-sided) were considered to indicate nominal statistical significance. STATA (version 13.0; StataCorp) and R version 3.1.2 were used for all analyses. This trial is registered with ClinicalTrials.gov, number NCT01100801.

Enrollment started on July 2, 2010, and stopped on April 2, 2015, after the recruitment of 81 patients in the 3G trial because the G1 arm had completed accrual and further patients classified as G-intestinal had to be treated with institutional standard of care (cisplatin plus S-1) instead of the protocol-specified therapy (oxaliplatin plus S-1), whereas patients being accrued into the G2 and G3 arms were fewer than anticipated.

Genomic profiling and stratification was successfully performed in 69 (86%) patients, of whom 48 (70%) and 21 (30%) patients were determined to have the G-intestinal and G-diffuse signatures, respectively. Twelve (14%) of the 81 patients had ambiguous or unavailable results and were assigned to the G3 arm (Fig. 1). Of 48 patients with the G-intestinal subtype, 5 patients did not receive treatment as planned due to closure of the G1 arm (n = 4) and tumor bleeding (n = 1). Two of 21 patients with the G-diffuse subtype carcinoma did not receive treatment as planned due to assignment to trastuzumab-containing regimen for HER2-positive disease (n = 1) and receiving treatment in another country (n = 1). The median time from sample receipt to return of molecular results was 7 working days (IQR, 5–9). Baseline characteristics of the four study groups are shown in Table 1. Median age in the overall evaluable cohort was 59 years (range, 32–76), and most participants had baseline ECOG performance statuses of 0 or 1 (41.1% and 54.8%).

Figure 1.

Flow of participants through the “3G” study.

Figure 1.

Flow of participants through the “3G” study.

Close modal
Table 1.

Baseline characteristics

CharacteristicsG1, SOX (n = 30)G1, SP (n = 13)G2, SP (n = 19)G3, SOX (n = 12)
Age, years 
 Median 57 50 62 61 
 Range 32–75 39–74 33–74 35–76 
Sex 
 Male 15 13 
Country of enrollment 
 Singapore 19 
 South Korea 11 13 14 
ECOG PS at screening 
 0 15 
 1 13 14 
 2 
Histology type 
 Adenocarcinoma 10 10 10 
 Diffuse 
 Signet ring cell 
 Intestinal 
 Mixed 
 Not reporteda 
Histology grade 
 Poorly differentiated 19 
 Moderately differentiated 
 Well differentiated 
 Not reporteda 13 
CharacteristicsG1, SOX (n = 30)G1, SP (n = 13)G2, SP (n = 19)G3, SOX (n = 12)
Age, years 
 Median 57 50 62 61 
 Range 32–75 39–74 33–74 35–76 
Sex 
 Male 15 13 
Country of enrollment 
 Singapore 19 
 South Korea 11 13 14 
ECOG PS at screening 
 0 15 
 1 13 14 
 2 
Histology type 
 Adenocarcinoma 10 10 10 
 Diffuse 
 Signet ring cell 
 Intestinal 
 Mixed 
 Not reporteda 
Histology grade 
 Poorly differentiated 19 
 Moderately differentiated 
 Well differentiated 
 Not reporteda 13 

Abbreviation: ECOG PS, Eastern Cooperative Oncology Group performance status.

aPathology assessments were not routinely done or mandated for endoscopic biopsies.

We ascertained the initial classification accuracy post hoc and identified 2 patients each from the G1 arm and G2 arm who should have been assigned to the opposite arm, giving a misclassification rate of 4 in 62 (6.5%; Fig. 2). However, the misclassified patients were included in the evaluable set on an intention-to-treat basis as they all went on to receive treatment. Five and 2 patients from the G1 and G2 arms, respectively, did not receive any study treatment and were not included in the evaluable set. We next extended the analysis to investigate the concordance between histologic and genomic subtypes. We observed a nominal agreement rate (intestinal and diffuse vs. G1 and G2, respectively) of 0.582 (95% CI, 0.429–0.735; P < 0.001), which was statistically significant after correcting for chance agreement (Gwet AC1 statistic, 0.290; 95% CI, 0.106–0.570; P = 0.042).

Figure 2.

Heatmap with dendrograms of gene expression based on initial classification (A) and post hoc reclassification (B).

Figure 2.

Heatmap with dendrograms of gene expression based on initial classification (A) and post hoc reclassification (B).

Close modal

At the cut-off date for PFS of May 3, 2016, a total of 407 treatment cycles had been administered and the total person-time for PFS was 498.0 months. All four study groups had each received a median of five treatment cycles. The response rates in the G1 SOX, G1 SP, G2 SP, and G3 SOX groups were 44.8% (95% CI, 26.5%–64.3%), 8.3% (0.00%–38.5%), 26.7% (7.8%–55.1%), and 55.6% (21.2%–86.3%), respectively, by blinded radiological assessment (Table 2); hence, no study group met the primary endpoint of demonstrating a 70% or higher response rate. The proportion of patients in the G1-SOX group who had an objective response was significantly higher than in the G1-SP group (P = 0.033), thus appearing to support the clinical utility of our genomic classifier. On the other hand, analyses of the secondary endpoints provided refuting evidence (Fig. 3A). However, it should be noted that the original observation of utility in our previous studies was based on in vitro cell proliferation assays (6), which may not translate to survival. In this regard, the appropriate clinical phenotype to cell proliferation may be objective response as opposed to survival. On the basis of 72 events, PFS was found to significantly differ between study groups by the log-rank test (P = 0.027), and patients in the G1 arm who received cisplatin/S-1 treatment had superior PFS (median, 9.8 months; IQR, 3.2–21.1) compared with the G1 SOX (median, 4.1 months; IQR, 1.9–7.0; HR, 2.51; 95% CI, 1.21–5.22; P = 0.013), G2 SP (4.2 months; IQR, 2.6–5.9; HR, 3.10; 95% CI, 1.37–7.05; P = 0.007) and G3 SOX (3.1 months; IQR, 1.4–6.3; HR, 3.08; 95% CI, 1.28–7.40; P = 0.012) arms. No other between-group comparisons were statistically remarkable.

Table 2.

Treatment response to platinum-based chemotherapy according to genomic subtype

G1, SOX (n = 30)G1, SP (n = 13)G2, SP (n = 19)G3, SOX (n = 12)
Partial response (%) 13 (44.8) 1 (8.3) 4 (26.7) 5 (55.6) 
Stable disease (%) 10 (34.5) 10 (83.4) 9 (60.0) 4 (44.4) 
Progressive disease (%) 6 (20.7) 1 (8.3) 2 (13.3) 0 (0.0) 
Not evaluated 
G1, SOX (n = 30)G1, SP (n = 13)G2, SP (n = 19)G3, SOX (n = 12)
Partial response (%) 13 (44.8) 1 (8.3) 4 (26.7) 5 (55.6) 
Stable disease (%) 10 (34.5) 10 (83.4) 9 (60.0) 4 (44.4) 
Progressive disease (%) 6 (20.7) 1 (8.3) 2 (13.3) 0 (0.0) 
Not evaluated 
Figure 3.

Kaplan–Meier curves of PFS (A) and OS (B).

Figure 3.

Kaplan–Meier curves of PFS (A) and OS (B).

Close modal

At the time of analysis, we recorded 61 deaths and the median duration of follow-up for OS was 41.9 months (IQR, 25.8–46.7). OS was numerically longer in the G1 SP group than in the G1 SOX group, suggesting that contrary to our initial hypotheses, patients with the G-intestinal subtype appeared to derive greater benefit from cisplatin plus S-1 (median OS 21.1 months; IQR 20.6–27.0) compared with oxaliplatin plus S-1 (7.6 months; 3.6–20.8), although the difference between Kaplan–Meier curves was not statistically significant (P = 0.308). Median OS duration was 7.5 (IQR, 3.2–13.3) and 9.4 (IQR, 3.2–12.6) months in the G2 SP and G3 SOX arms, respectively (Fig. 3B).

In an exploratory biomarker substudy, we also retrospectively-prospectively evaluated two published genomic signatures that were also developed in Asian populations with gastric cancer (11, 12) to serve as decision-support aids (Fig. 4A). The 3G trial data corroborated the survival advantage of patients whose tumors harbored the “metabolic” genomic subtype, as reported by Lei and colleagues (11). The response rates per the “mesenchymal,” “metabolic,” and “proliferative” subtypes (11) were 27.3% (95% CI, 10.7%–50.2%), 43.5% (23.2%–65.5%), and 40.0% (16.3%–67.7%). Median PFS in patients differed by the log-rank test (P = 3.9 × 10−3) and was significantly longer in tumors exhibiting the metabolic signature (median, 7.0 months; IQR, 2.2–5.9) compared with the mesenchymal (3.3 months; IQR, 2.2–5.9; HR, 0.327; 95% CI, 0.166–0.646; P = 0.001) and proliferative (4.6 months; IQR, 2.6–9.8; HR, 0.505; 95% CI, 0.249–1.02; P = 0.058) subtypes, respectively (Fig. 4B). Median OS were 7.0 (IQR, 3.2–12.6), 20.6 (6.2–27.0), 7.2 (9.8–14.2) months, respectively (Fig. 4D), and was longer in patients with the metabolic subtype than the mesenchymal subtype (HR, 0.515; 95% CI, 0.269–0.987; P = 0.046).

Figure 4.

Exploratory biomarker analyses using other published genomic classifiers. Clustering of patients by genomic and histological subtypes, platinum agent, and tumor response (A), PFS per Lei et al. classifier (B), PFS per ACRG classifier (C), OS per Lei et al. classifier (D), OS per ACRG classifier (E).

Figure 4.

Exploratory biomarker analyses using other published genomic classifiers. Clustering of patients by genomic and histological subtypes, platinum agent, and tumor response (A), PFS per Lei et al. classifier (B), PFS per ACRG classifier (C), OS per Lei et al. classifier (D), OS per ACRG classifier (E).

Close modal

There were no notable findings of exploratory biomarker analyses based on the Asian Cancer Research Group's (ACRG) genomic classifier (12). Response rates for the “MSI,” “EMT,” “P53-null,” and “P53-positive” subgroups were 0%, 16.7% (95% CI, 0.04%–64.1%), 42.9% (24.5%–62.8%), and 36.0% (18.0%–57.5%). Median PFS were 1.4 (IQR, 1.4–10.2), 5.5 (4.6–9.6), 4.5 (2.6–7.0), and 4.2 (2.4–10.4) months, respectively (Fig. 4C), and median OS were 1.4 (IQR, 1.4–10.2), 12.6 (10.5–20.0), 9.6 (4.5–20.8), and 11.7 (4.6–24.4) months, respectively (Fig. 4E). No molecular signature of the ACRG classifier showed any predictive ability for assessing outcomes from chemotherapy treatment.

In another exploratory analysis, we evaluated whether dimensionality reduction of the intrinsic gene expression signatures from the 171-gene set (8) would affect the performance characteristics of the classifier within our patient cohort. We evaluated three different thresholds (Q = 0.0015, 0.001, and 0.0005) corresponding to 136, 113, and 77 genes, respectively (data not shown). Concordance between the reduced gene set classifier and our original classifier was 89% to 92%. This suggests that an abridged gene set, which may be more cost-efficient and amenable to be utilized in the clinical setting, could potentially be used to reproduce the original intrinsic subtypes classifier without substantially compromising its performance characteristics.

To our knowledge, this is the first clinical trial to prospectively stratify patients with advanced gastric cancer based on their tumor gene expression profiles, and to use the information in therapeutic decision making. The clinical utility of the genomic classifier in question was not proven, as neither the G1 nor G2 arms achieved a response rate of at least 70%, and did not predict differential benefit from oxaliplatin or cisplatin. The results of this prospective clinical trial refute the findings of our preclinical article, which suggested that G-intestinal cell lines were more sensitive to oxaliplatin, whereas G-diffuse cell lines were more sensitive to cisplatin (8). However, there were other important findings from this study. The success rate of genomic profiling [86% (69/81 patients, with a misclassification rate of 6.5%)] and acceptable turnaround time (median of 7 working days; IQR, 5–9) in the 3G trial demonstrates that this approach of personalized medicine is clinically implementable in the advanced gastric cancer setting. For the translation of molecular information into predictive clinical applications, a reasonable turnaround time for actionable results is fundamental; hence, this study significantly adds to the present evidence base (17).

In this study, we did analyses on a modified intention-to-treat basis with four study groups: G1 SOX, G1 SP, G2 SP, and G3 SOX, because after the target accrual was reached in the G1 arm, subsequent patients who met the inclusion criteria and harbored the G-intestinal subtype were allowed to receive the institutional standard of care (cisplatin plus S-1) instead of the study regimen (oxaliplatin plus S-1). The considerably slower accrual of patients into the G2 arm could represent a lower incidence of G-diffuse gastric cancers in the population and/or the correlation of the G-diffuse genomic profile with diffuse-type Lauren classification, which may be more likely to have nonmeasurable disease. However, this allowed us to examine whether oxaliplatin plus S-1 was superior to cisplatin plus S-1 in patients with G1 subtype tumors, an additional hypothesis that was not prespecified, yet afforded more evidence to evaluate the efficacy of our genomic classifier for treatment stratification. Although a significantly greater proportion of tumor responses occurred in the G1 SOX group than the G1 SP group (P = 0.033), PFS and OS were significantly or numerically longer in the G1 SP group (9.8 months vs. 4.1 months, P = 0.013; and 21.1 vs. 7.6 months, P = 0.308, respectively). Whether these conflicting endpoints represent inherent biological and clinical manifestations of the intestinal genomic subtype (for instance, a propensity to respond well initially to oxaliplatin plus S-1 but develop resistance soon after) or more likely, chance occurrence or a large proportion of G1 SP patients achieving long-term stable disease, remains to be investigated. Indeed, a response rate of 8.3% in the G1 SP group is certainly peculiar and far lower than would be expected in molecularly unselected patients treated with the SP regimen, for which historical response rates of 35% to 45% have been obtained (3–6). It should also be noted that the intrinsic subtypes of gastric cancer convey prognostic connotations (8), which could have an impact on analyses of PFS and OS. Furthermore, the original observation of utility that led to this trial was based on in vitro cell proliferation assays (8), which may not translate to survival. Rather, the appropriate clinical analogue to cell proliferation may be objective response, which is the primary endpoint of this trial.

Nonetheless, the results are provocative. As with genomic profiling studies, one major limitation is that a single endoscopic biopsy may not be adequately representative given that tumors are known to exhibit significant spatiotemporal heterogeneity in their molecular profiles (18–23). Whether such intra- and intertumoral heterogeneity underpins the subset of tumors with “unclear” (G3) genomic assignment, and whether multiple tumor samples from the same patient could be both G1 and G2 simultaneously is beyond the scope of this study, but represents an interesting area for further investigation. There are other considerations to take into account: First, we report the results of several analyses, such as OS and the comparison of the G1 SOX and G2 SP groups, which had not been preplanned into the design of the study protocol. Together with the fact that accrual into the G2 and G3 groups were lower than anticipated or required may limit the ability of our study to reach statistically powered conclusions. Furthermore, the comparison of G1 and G2 groups was not randomized and is, thus, subject to various sources of bias and error. However, these exploratory analyses were still undertaken because the additional evidence they afforded to support or refute the usefulness of the genomic classifier in question was deemed to outweigh the risks of insufficient statistical power. Second, P values reported here are nominal without adjusting for multiple testing, which again underscore the need to regard positive findings in our trial as exploratory. Third, baseline imbalances between the four study groups (Table 1), although not statistically evaluated, could confound our analyses.

Although the primary analysis casted uncertainty on the clinical usefulness of a genomic classifier based on intrinsic subtypes (8), we also did post hoc genomic reassignment to evaluate the two other published genomic classifiers developed from Asian cohorts of gastric cancer patients with purported prognostic or predictive utility (11, 12). The Asian Cancer Research Group's molecular taxonomy, which assigns gastric tumors as “MSI,” “EMT,” “P53-null,” and “P53-positive,” showed no ability to predict benefit from chemotherapy. However, the genomic classifier proposed by Lei and colleagues, which stratifies tumors as “mesenchymal,” “metabolic,” and “proliferative” depending on which biological pathways and gene sets are upregulated (11), showed some evidence of prognostic prowess as we were able to recapitulate the PFS advantage in patients with the metabolic subtype as reported previously (11). Nonetheless, owing to the limited sample size and aforementioned limitations of our data, these findings should be interpreted with caution and require replication in a large cohort before they can be incorporated in clinical practice.

The clinical usefulness of any genomic classifier is determined in part by its internal and external validity, reproducibility, and stability, and these are parameters that need to be optimized prior to phase III testing and clinical approval. In this study, we identified two patients each in the G1 and G2 arms who should have been assigned to the opposite arm, based on a post hoc analysis wherein we normalized individual patient tumors against previously published gene expression signatures (8) and reclassified them using a dataset comprising tumors collected during this study. Considering that “misclassification rate” in this study represents the percentage of samples that were classified differently when a modified bioinformatics pipeline (i.e., a different sample class prediction algorithm) was used, a “misclassification” rate of 6.5% may be interpreted to indicate good stability of the G1 and G2 gene expression signatures (8). Nevertheless, the potential for bona fide misclassification in prospective studies of gene expression profiling is a valid concern due to the need for rapid turnaround time and other issues including nonoptimal sample acquisition and handling. Hence, future prospective trials should put in place robust measures for auditing and scrutiny of sources of error.

The overlap or lack thereof between the Lauren subtypes and genomic classifier warrants discussion. It could be argued that basic histopathology could render genomic testing obsolete if a high level of concordance between histopathology and molecular testing is observed. In this study, we observed an agreement rate of 58.2%, which is statistically significant (P = 0.042 after correcting for chance agreement), but nonetheless connotes a clinically unacceptable level of discordance between the two platforms. In addition, it is worth noting that previous reports have indicated that intra- and interpathologist agreement based on traditional histopathologic assessment (Lauren classification) can differ up to 5% to 23% of the time, whereas histologic diagnosis based on biopsies and actual specimens has been reported to differ in 16% to 26% of cases (24–26), thus underscoring a continued need to develop novel approaches (beyond basic histopathology) to classify gastric cancers reliably and reproducibly.

Predictive genomic classifiers, such as those evaluated in the 3G trial, hold the potential to improve the ratio of the number needed to harm (NNH) to the number needed to treat (NNT). However, to justify their place in clinical practice, the benefits have to be weighed against the additional costs of adopting a new technology. In this study, a rough estimate of the total cost per sample is approximately SGD $500 (∼USD $400). On the basis of data generated by the 3G trial, follow-up studies have been planned to clarify the health economic value of implementing genomic classifiers in the clinical management of advanced gastric cancer.

Conclusions

In conclusion, we have shown that genomic profiling to guide chemotherapy selection in the advanced gastric cancer setting is feasible, even though our data do not support our genomic classifier in question because it did not predict differential benefit from oxaliplatin and cisplatin or meet the primary objective of generating an objective response rate of 70% in any arm. However, the metabolic genomic signature developed by Lei and colleagues warrants further investigation as a promising biomarker of clinical benefit. Integration of gene expression profiling with techniques such as next-generation sequencing might yield greater chances of identifying stable and reproducible genomic signatures for diagnostic, prognostic, and predictive applications in the clinic. This trial, which is the first to use gene expression profiling for disease stratification and treatment assignment in advanced gastric cancer, therefore significantly adds to the growing literature in personalized medicine strategies.

S. Choo reports receiving speakers bureau honoraria from Bristol-Myers Squibb and Lilly Oncology, and is a consultant/advisory board member for Bayer, Bristol-Myers Squibb, Celgene, Eisai, Shire, and Sirtex. C. Tan reports receiving speakers bureau honoraria from and is a consultant/advisory board member for Astra Zeneca, Boehringer Ingelheim, Bristol-Myers Squibb, Eisai, Eli-Lilly, Merck, MSD, Novartis, and Takeda. No potential conflicts of interest were disclosed by the other authors.

Neither the funders nor sponsors of the trial participated in study design, in patient accrual or data analysis, or in the preparation of this manuscript. The corresponding author had full access to all of the data and the final responsibility to submit the manuscript for publication.

Conception and design: W.P. Yong, S.Y. Rha, I.B.-H. Tan, J.B.-Y. So, A. Shabbir, H.-S. Kim, H.C. Chung, K.G. Yeoh, P. Tan

Development of methodology: W.P. Yong, S.Y. Rha, I.B.-H. Tan, S.-P. Choo, N.L. Syn, J.B.-Y. So, A. Shabbir, H.C. Chung, P. Tan

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): W.P. Yong, S.Y. Rha, I.B.-H. Tan, S.-P. Choo, B.R. Asuncion, J.B.-Y. So, A. Shabbir, C.-S. Tan, H.-S. Kim, M. Jung, H.C. Chung, M.C.H. Ng, D.W.-M. Tai, M.-H. Lee, J. Wu, P. Tan

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): W.P. Yong, S.Y. Rha, I.B.-H. Tan, S.-P. Choo, N.L. Syn, V. Koh, S.-H. Tan, B.R. Asuncion, R. Sundar, H.-S. Kim, H.C. Chung

Writing, review, and/or revision of the manuscript: W.P. Yong, I.B.-H. Tan, S.-P. Choo, N.L. Syn, V. Koh, R. Sundar, J.B.-Y. So, A. Shabbir, C.-S. Tan, H.C. Chung, K.G. Yeoh, P. Tan

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): W.P. Yong, N.L. Syn, V. Koh, R. Sundar, A. Shabbir, H.C. Chung

Study supervision: W.P. Yong, S.Y. Rha, A. Shabbir, H.C. Chung

Other (pathology analysis): B.R. Asuncion

This work was supported by the National Research Foundation Singapore under its Translational and Clinical Research (TCR) Flagship Programme grant, administered by the Singapore Ministry of Health's National Medical Research Council (TCR/009-NUHS/2013) and awarded to the Singapore Gastric Cancer Consortium (SGCC). This research is also supported by the Cancer Science Institute of Singapore, NUS, under the National Research Foundation Singapore's and the Singapore Ministry of Education's Research Centres of Excellence initiative, as well as a grant from the National R&D Programme for Cancer Control, Ministry of Health and Welfare, Republic of Korea (1520190). The S-1 drug used in this study was supplied by Taiho Pharmaceutical.

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.

1.
Ferlay
J
,
Soerjomataram
I
,
Ervik
M
,
Dikshit
R
,
Eser
S
,
Mathers
C
, et al
GLOBOCAN 2012 v1.0, Cancer Incidence and Mortality Worldwide: IARC CancerBase No. 11
.
Lyon
,
France
:
IARC
; 
2013
.
2.
Fitzmaurice
C
,
Allen
C
,
Barber
RM
,
Barregard
L
,
Bhutta
ZA
,
Brenner
H
, et al
Global, regional, and national cancer incidence, mortality, years of life lost, years lived with disability, and disability-adjusted life-years for 32 cancer groups, 1990 to 2015
.
JAMA Oncol
2017
;
3
:
524
.
3.
Yamada
Y
,
Higuchi
K
,
Nishikawa
K
,
Gotoh
M
,
Fuse
N
,
Sugimoto
N
, et al
Phase III study comparing oxaliplatin plus S-1 with cisplatin plus S-1 in chemotherapy-naïve patients with advanced gastric cancer
.
Ann Oncol
2015
;
26
:
141
8
.
4.
Hamada
C
,
Yamada
Y
,
Azuma
M
,
Nishikawa
K
,
Gotoh
M
,
Bando
H
, et al
Meta-analysis supporting noninferiority of oxaliplatin plus S-1 to cisplatin plus S-1 in first-line treatment of advanced gastric cancer (G-SOX study): indirect comparison with S-1 alone
.
Int J Clin Oncol
2016
;
21
:
668
75
.
5.
Montagnani
F
,
Turrisi
G
,
Marinozzi
C
,
Aliberti
C
,
Fiorentini
G
. 
Effectiveness and safety of oxaliplatin compared to cisplatin for advanced, unresectable gastric cancer: a systematic review and meta-analysis
.
Gastric Cancer
2011
;
14
:
50
5
.
6.
Wagner
AD
,
Syn
NL
,
Moehler
M
,
Grothe
W
,
Yong
WP
,
Tai
B-C
, et al
Chemotherapy for advanced gastric cancer
.
Cochrane Database Syst Rev
2017
;
8
:
CD004064
.
7.
Cunningham
D
,
Starling
N
,
Rao
S
,
Iveson
T
,
Nicolson
M
,
Coxon
F
, et al
Capecitabine and oxaliplatin for advanced esophagogastric cancer
.
N Engl J Med
2008
;
358
:
36
46
.
8.
Tan
IB
,
Ivanova
T
,
Lim
KH
,
Ong
CW
,
Deng
N
,
Lee
J
, et al
Intrinsic subtypes of gastric cancer, based on gene expression pattern, predict survival and respond differently to chemotherapy
.
Gastroenterology
2011
;
141
:
476
85
.
9.
Eisenhauer
EA
,
Therasse
P
,
Bogaerts
J
,
Schwartz
LH
,
Sargent
D
,
Ford
R
, et al
New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1)
.
Eur J Cancer
2009
;
45
:
228
47
.
10.
Pepper
SD
,
Saunders
EK
,
Edwards
LE
,
Wilson
CL
,
Miller
CJ
. 
The utility of MAS5 expression summary and detection call algorithms
.
BMC Bioinform
2007
;
8
:
273
.
11.
Lei
Z
,
Tan
IB
,
Das
K
,
Deng
N
,
Zouridis
H
,
Pattison
S
, et al
Identification of molecular subtypes of gastric cancer with different responses to PI3-Kinase inhibitors and 5-Fluorouracil
.
Gastroenterology
2013
;
145
:
554
65
.
12.
Cristescu
R
,
Lee
J
,
Nebozhyn
M
,
Kim
K-M
,
Ting
JC
,
Wong
SS
, et al
Molecular analysis of gastric cancer identifies subtypes associated with distinct clinical outcomes
.
Nat Med
2015
;
21
:
449
56
.
13.
A’Hern
RP
. 
Sample size tables for exact single-stage phase II designs
.
Stat Med
2001
;
20
:
859
66
.
14.
Ajani
JA
,
Rodriguez
W
,
Bodoky
G
,
Moiseyenko
V
,
Lichinitser
M
,
Gorbunova
V
, et al
Multicenter phase III comparison of cisplatin/S-1 with cisplatin/infusional fluorouracil in advanced gastric or gastroesophageal adenocarcinoma study: the FLAGS trial
.
J Clin Oncol
2010
;
28
:
1547
53
.
15.
Koizumi
W
,
Narahara
H
,
Hara
T
,
Takagane
A
,
Akiya
T
,
Takagi
M
, et al
S-1 plus cisplatin versus S-1 alone for first-line treatment of advanced gastric cancer (SPIRITS trial): a phase III trial
.
Lancet Oncol
2008
;
9
:
215
21
.
16.
Schemper
M
,
Smith
TL
. 
A note on quantifying follow-up in studies of failure time
.
Control Clin Trials
1996
;
17
:
343
6
.
17.
Sotiriou
C
,
Piccart
MJ
. 
Taking gene-expression profiling to the clinic: when will molecular signatures become relevant to patient care?
Nat Rev Cancer
2007
;
7
:
545
53
.
18.
Bass
AJ
,
Thorsson
V
,
Shmulevich
I
,
Reynolds
SM
,
Miller
M
,
Bernard
B
, et al
Comprehensive molecular characterization of gastric adenocarcinoma
.
Nature
2014
;
513
:
202
9
.
19.
Syn
NL
,
Teng
MWL
,
Mok
TSK
,
Soo
RA
. 
De-novo and acquired resistance to immune checkpoint targeting
.
Lancet Oncol
2017
;
18
:
e731
41
.
20.
Teo
Y-Y
,
Sim
X
,
Ong
RTH
,
Tan
AKS
,
Chen
J
,
Tantoso
E
, et al
Singapore genome variation project: a haplotype map of three southeast asian populations
.
Genome Res
2009
;
19
:
2154
62
.
21.
Chia
N-Y
,
Tan
P
. 
Molecular classification of gastric cancer
.
Ann Oncol
2016
;
27
:
763
9
.
22.
Syn
NL-X
,
Yong
W-P
,
Goh
B-C
,
Lee
S-C
. 
Evolving landscape of tumor molecular profiling for personalized cancer therapy: a comprehensive review
.
Expert Opin Drug Metab Toxicol
2016
;
12
:
911
22
.
23.
Pectasides
E
,
Stachler
MD
,
Derks
S
,
Liu
Y
,
Maron
S
,
Islam
M
, et al
Genomic heterogeneity as a barrier to precision medicine in gastroesophageal adenocarcinoma
.
Cancer Discov
2018
;
8
:
37
48
.
24.
Berlth
F
,
Bollschweiler
E
,
Drebber
U
,
Hoelscher
AH
,
Moenig
S
. 
Pathohistological classification systems in gastric cancer: diagnostic relevance and prognostic value
.
World J Gastroenterol
2014
;
20
:
5679
84
.
25.
Flucke
U
,
Mönig
SP
,
Baldus
SE
,
Zirbes
TK
,
Bollschweiler
E
,
Thiele
J
, et al
Differences between biopsy- or specimen-related Laurén andWorld Health Organization Classification in Gastric Cancer
.
World J Surg
2002
;
26
:
137
40
.
26.
Shibata
A
,
Longacre
TA
,
Puligandla
B
,
Parsonnet
J
,
Habel
LA
. 
Histological classification of gastric adenocarcinoma for epidemiological research: concordance between pathologists
.
Cancer Epidemiol Biomarkers Prev
2001
;
10
:
75
8
.