Tumor Genomic Profiling Guides Patients with Metastatic Gastric Cancer to Targeted Treatment: The VIKTORY Umbrella Trial

Recent advances in molecular analysis have revealed that there are patient subsets with differing genomic alterations despite the same histologic diagnosis in gastric cancer (1–3). It has been suggested by previous studies that this interpatient tumor molecular heterogeneity may affect the outcomes of clinical trials, especially with molecularly targeted agents (4, 5). To deliver a more tailored approach for each patient, umbrella or platform clinical trials have been developed (6, 7), which assign treatment arms based on the molecular characteristics of the tumor.

Gastric cancer was the third leading cause of cancer-related mortality in 2018, causing 783,000 deaths worldwide (8). The prognosis of patients with metastatic gastric cancer remains extremely poor, with a median overall survival (OS) of less than 12 months with cytotoxic chemotherapy (9, 10). In addition, gastric cancer is a disease with significant molecular and histologic heterogeneity (1, 3, 11), in which advancements based on “one-size-fits-all” clinical trials have yielded only modest survival benefits. To identify optimal molecular targets and optimal biomarkers, we designed an umbrella trial for second-line (2L) treatment in metastatic gastric cancer based on tumor molecular profiling. We took advantage of an umbrella trial design where patients of a single tumor type are directed toward different arms of the study based on the tumor molecular biomarkers relevant to one or more of the candidate drugs (12). VIKTORY (targeted agent eValuationIn gastric cancer basket KORea, trial NCT#02299648) was designed to classify patients with metastatic gastric cancer based on clinical sequencing and comprised eight different biomarker groups (RAS aberration, TP53 mutation, PIK3CA mutation/amplification, MET amplification, MET protein overexpression, all negative, TSC2 deficient, or RICTOR amplification) to assign patients to one of the 10 associated phase II clinical trials in 2L treatment. The study drugs used were capivasertib (AKT inhibitor), savolitinib (MET inhibitor), selumetinib (MEK inhibitor), adavosertib (WEE1 inhibitor), and vistusertib (TORC inhibitor). The umbrella design was based on the preclinical evidence of known molecular alterations, the prevalence of molecular alterations, and the availability of the targeted agents for clinical trials from AstraZeneca at the time of the study design. The candidate molecular alterations for the umbrella trial at the time of clinical trial design were molecular alterations in the p53, PIK3CA, MET, EGFR, FGFR2, RAS, and DDR pathways (3). Adavosertib is one of the most potent inhibitors targeting WEE1 (13), which is a tyrosine kinase that phosphorylates cyclin-dependent kinase 1 (CDK1, CDC2) to inactivate the CDC2/cyclin B complex (14). Inhibition of WEE1 activity prevents the phosphorylation of CDC2 and impairs the G₂ DNA-damage checkpoint, leading to cancer-cell death. Preclinical studies have demonstrated a very promising antitumor efficacy in vivo, especially in combination with other cytotoxic chemotherapeutic agents (15) including paclitaxel (16). Capivasertib is a selective pan-AKT inhibitor that inhibits the kinase activity of all three AKT isoforms (AKT1–3; ref. 17). Preclinically, sensitivity to capivasertib has been strongly correlated with the presence of PIK3CA mutations in gastric cancer models (18, 19). Savolitinib is a potent small-molecule reversible MET kinase inhibitor that inhibits MET kinase at an IC₅₀ of 4 nmol/L in MET-amplified cancer cells and has been shown to demonstrate promising antitumor activity in patients with gastric cancer (20, 21). Selumetinib (AZD6244, ARRY-142886) is a potent, orally active inhibitor of MEK1/2 that suppresses the pleiotropic output of the RAF/MEK/ERK pathway (22, 23). The tolerability and antitumor efficacy of the combination of selumetinib and docetaxel were demonstrated in KRAS-mutant non–small cell lung cancer (24).

Herein we conducted a prospective clinical sequencing master program that was aligned with 8 prespecified genomic biomarkers and 10 independent biomarker-associated clinical trials in patients with metastatic gastric cancer. We explored whether the biomarker-selected platform trial benefits patients with metastatic gastric cancer in terms of survival. In addition, we investigated PD-L1 score and circulating tumor (ctDNA) change between baseline and post-treatment samples following targeted agents.

Between March 2014 and July 2018, 772 patients with metastatic gastric cancer were enrolled onto the VIKTORY trial. Targeted sequencing was successfully achieved with tissues from 715 patients (92.6%; Fig. 1A and B). Of the 715 tissues, 150 (21.1%) were from fresh tumors, 564 (78.9%) from formalin-fixed paraffin-embedded (FFPE) specimens, and 1 from ctDNA sequencing using Guardant360 (Fig. 2A). Nearly all samples (96.2%) were from the primary gastric tumor specimen; 56.4% of the patients had their tumor sequenced at the time of diagnosis of metastatic gastric cancer and 43.6% of patients were sequenced during first-line (1L) or at the time of progression following 1L chemotherapy. The tissue type, site of biopsy for sequencing, and Epstein–Barr virus (EBV) and mismatch-repair (MMR) status of the 715 patients are summarized in Fig. 2A. A total of 75.9% of patients had poorly differentiated adenocarcinoma. The primary tumor was located in the body (53.2%) or antrum (37.7%) of the stomach in the majority of patients. All patients underwent 1L cytotoxic chemotherapy (>85% with fluoropyrimidine/platinum regimen). In all, 460 of 715 patients (64.3%) were eligible for 2L therapies: 143 of 715 (20.6%) were assigned to one of the umbrella-associated parallel clinical trials in 2L (105 with Biomarker A–E or G; 38 with Biomarker F, unselected), whereas 317 patients received conventional treatment or treatment via other clinical trials (Figs. 1B and 2A).

The tumor profiles of the 715 patients are shown in Supplementary Fig. S1, and the detailed sequencing method is provided in the Supplementary Material. The prevalence of the predefined biomarkers was as follows (Fig. 2B): Biomarker A1: RAS mutation/amplification (81/715, 12.2%; KRAS 62/715, 8.7%; HRAS 6/715, 0.8%; NRAS 19/715, 2.7%); Biomarker A2: high or low MEK signature (49/107, 45.8%); Biomarker B: TP53 mutation (321/715, 44.9%); Biomarker C: PIK3CA mutation/amplification (54/715, 7.6%); Biomarker D: MET amplification (25/715, 3.5%); Biomarker E: MET overexpression by IHC 3+ (42/479, 8.8%); Biomarker F: none of the above (Biomarker A–E); Biomarker G: RICTOR amplification (5/715, 0.7%)/TSC2 deficient (7/715, 0.9%). In addition to the predefined biomarkers, we identified other known molecular targets in gastric cancer (Supplementary Fig. S1): FGFR2 amplification (30/715, 4.2%), EGFR amplification (17/715, 2.4%), MDM2 amplification (8/715, 1.1%), AKT1 amplification (2/715, 0.3%), FGFR1 amplification (10/715, 1.4%), and CCNE1 amplification (14/715, 2.0%). In all, 3.5% were MMR-deficient gastric cancer (18/523) and 4% (20/501) were EBV-positive. Concurrent MMR and EBV status are provided in 105 patients treated according to biomarker status (Fig. 2B, left). In addition, concurrent molecular profiling of each patient according to biomarker (e.g., KRAS mutation and TP53 mutation) and the assigned umbrella arm is summarized in Fig. 2B (right) according to the biomarker priority. The incidence of MET overexpression by IHC (defined by 3+) was 8.8% (42/479) in this cohort: 17 (40.5%) of 42 MET-overexpressed tumors had MET-amplified tumors by next-generation sequencing (NGS) or FISH, and 25 (59.5%) patients had no MET amplification, which concurred with our previous finding on coactivation of MET protein without amplification (25, 26).

The cutoff date for treatment outcome analysis was October 1, 2018. At the time of analysis, enrollment had been completed in all arms or stopped due to early termination of drug development (arms 6, 9, 10) or lack of efficacy at first stage of phase II (arm 7; Supplementary Table S1). Currently, enrollment is completed in phase I of arm 8, and phase II is being considered. Further patient enrollment was halted in arm 5 (savolitinib/docetaxel combination) due to the high efficacy observed with the savolitinib monotherapy arm. The primary endpoint was overall response rate (ORR); assuming ORR of 20% for 2L paclitaxel, experimental arms were considered effective if the combination yielded ≥50% ORR for arms 1–10 except for arm 4 (savolitinib monotherapy arm). The ORR for each umbrella arm was as follows: arm 1 (selumetinib/docetaxel): 28% (7/25, 95% CI: 10.4–45.6), arm 2 (adavosertib/paclitaxel): 24% (6/25, 95% CI:7.3–40.7), arm 3 (capivasertib/paclitaxel): 33.3% (8/24; 95% CI: 14.4–52.2), and arm 4 (savolitinib): 50% (10/20, 95% CI: 28.0–71.9; Supplementary Table S1 for detailed primary endpoints for each arm). The waterfall plots and swimmer plots are provided in Fig. 3. Seven of 25 patients who had a partial response (PR) in the selumetinib/docetaxel arm (arm 1) had KRAS amplification/high MEK (MEK-H), KRAS wild-type (WT)/low MEK (MEK-L), KRAS G12R/MEK-H, KRAS G12D/MEK-L, KRAS G12D/MEK-H, KRAS G13D/MEK-I, KRAS WT MEK-H, and KRAS Q61R/MEK-I, respectively. The longest responder carried a KRAS amp (KRAS WT) with high MEK signature (arm1-005; Fig. 3A, top right). In terms of KRAS mutational status, there was no significant difference in ORR between KRAS-mutant (4 of 11, 36.4%) and KRAS WT (3/14, 21.4%; P = 0.538, χ² test). For the Biomarker B–arm 2 (adavosertib/paclitaxel) umbrella, there were six PRs (6/25) and 3 of these patients responded longer than 6 months (Fig. 3B). For Biomarker C–arm 3 (capivasertib/paclitaxel), there were 8 responders (8/24) with 4 patients responding for more than 6 months (Fig. 3C). For Biomarker D–arm 4 (savolitinib monotherapy), there were 10 PRs (10 of 20) one of whom (arm4-010) had the tumor resected after achieving complete response (CR; Fig. 3D). This patient was a 65-year-old female who was laparoscopically diagnosed with peritoneal seeding at diagnosis. After failing the 1L capecitabine/oxaliplatin treatment and developing rapidly deteriorating malignant ascites, the patient was assigned to savolitinib due to high copy-number MET amplification. After significant tumor reduction following savolitinib, the patient underwent curative resection and achieved pathologic downstaging from M1 disease to T3N2M0 disease. The patient remains in CR, now more than 1 year at the time of manuscript preparation.

Genomic variations are increasingly being utilized as reliable biomarkers for predicting clinical response to therapy for gastric cancer (27–29). To identify genomic variants that significantly correlate with clinical response, we compared the maximal tumor burden change per RECIST 1.1 against single genomic alterations (Fig. 4A). Patients with MET amplifications demonstrated the largest absolute decrease in tumor burden per RECIST 1.1. In addition, PIK3CA helical domain E542K patients had a more profound (≥50%) reduction in tumor burden when compared with patients with other point mutations in PIK3CA—E545G, E545K, E545K, H1047R, C420R, or E453K. Among patients with TP53 mutations, R273C, R175H, R342X, and Y220C demonstrated the most tumor reduction upon adavosertib/paclitaxel therapy. Finally, patients with KRAS G13E and KRAS G12D mutations, KRAS amplification, or MEK-H without KRAS mutation demonstrated the highest tumor burden reduction by selumetinib/docetaxel. Further focused genomic analysis of the Biomarker D (MET amplification) group and treatment response to savolitinib demonstrated that patients with gastric cancer with high MET copy number (>10 MET gene copies by tissue NGS) had high response rates to savolitinib (Fig. 4B). Patient arm4-010 who initially had gastric cancer with peritoneal seeding had a MET tissue NGS copy number of 25.9 and achieved PR following savolitinib, which eventually led to curative surgery, as mentioned previously. Although limited by the small number of patients, 5 responders to savolitinib had PD-L1–positive tumors [range, 3 to 80 for combined positive score (CPS)], including patient arm4-010 (Fig. 4B). Another focused genomic analysis of the Biomarker C (PIK3CA mutation) group and treatment response to capivasertib/paclitaxel showed that 57.1% (4 of 7 PRs) had E542K mutations. Moreover, patients with PIK3CA E542K mutations demonstrated an ORR of 50% (4/8), which was higher than the non-E542K cohort (3/16, 18.8%; P = 0.063 by χ²; Fig. 4C). Toxicity profiles for the four arms are shown in Supplementary Table S2.

We conducted an OS analysis on the biomarker-driven treatment group using the Kaplan–Meier plot in all patients. In all, patients with gastric cancer who had biomarkers identified and were treated accordingly (N = 105) demonstrated better OS (median OS, 9.8 months) when compared with patients who received conventional 2L therapy (N = 266; taxol/ramucirumab, N = 99; taxane-based, N = 105; irinotecan-based, N = 62) treatment (median OS, 6.9 months) with statistical significance (P < 0.001; Fig. 5A). The results from the biomarker-driven treatment cohort retained statistical significance in a multivariate analysis, and this treatment continued to predict better survival (P < 0.0001, HR = 0.58; 95% CI: 0.45–0.76) after correcting for potential prognostic factors such as age, gender, number of involved organs, EBV status, MMR status, and performance status (Fig. 5B). Concordantly, the VIKTORY biomarker-assigned cohort (N = 105) had significantly prolonged progression-free survival (PFS) when compared with the conventional 2L cohort (N = 266; median PFS, 5.7 months vs. 3.8 months, respectively, P < 0.0001; Fig. 5C). The multivariate Cox regression analysis for PFS revealed that being biomarker positive was an independent prognostic factor after adjustment for the several clinically important factors (Supplementary Fig. S2). Hence, when the biomarker was identified and the patient received a matched treatment with targeted agents at an appropriate time, patients had prolonged PFS and OS compared with conventional chemotherapy.

On the basis of the tumor heterogeneity and genomic changes we observed in our previous studies (11, 27, 30), we collected plasma for ctDNA analysis at baseline and every CT evaluation until progression to address tumor evolution. The concordance rate between tumor and ctDNA (tested by Guardant360; Supplementary Table S3) for MET amplification was 89.5%, with 100% specificity and 83.3% sensitivity relative to tissue testing, which increased to 100% if patients without detectable ctDNA were excluded (Fig. 6A). The maximal tumor burden decrease was observed in patients with high adjusted MET copy number by ctDNA, although statistical significance was not reached (Fig. 6B). More importantly, however, increased adjusted plasma copy number for MET amplification was significantly associated with prolonged PFS on savolitinib (Fig. 6C; P = 0.0216) to a significantly greater degree than tissue NGS MET copy number, which may reflect plasma's ability to synthesize the entire tumor cell population. Savolitinib therapy markedly decreased total ctDNA levels in all patients for which baseline and 4-week plasma results were available (Fig. 6D), demonstrating clear biological activity before most radiographic evidence of response. Congruently, adjusted plasma MET copy number was markedly suppressed at 4 weeks in all patients for whom results were available, although 2 of the 6 patients tested retained detectable MET amplification on progression, suggesting additional off-target mechanisms of acquired resistance (Fig. 6E).

We additionally sequenced 55 (from 29 patients) ctDNA samples from arm 1 (13 patients) and arm 2 (16 patients) using a 300-gene AstraZeneca (AZ) panel (Supplementary Table S4 and S5; Fig. 6F and G). Concordance between tumor DNA and ctDNA was observed in 10 of 13 (76.9%) patients for KRAS aberration status (arm 1) and 75% (12 of 16) for TP53 mutation status (arm 2; Fig. 6F and G). Of the 8 baseline/progressive disease (PD) paired ctDNA samples in arm 1, only 2 (25%) had retained baseline genomic alterations at disease progression. Of the 11 baseline/PD paired ctDNA samples in arm 2, 5 (45.5%) patients showed no major alterations at disease progression in the 300-gene panel following adavosertib/paclitaxel treatment. Dynamic changes from baseline to disease progression in ctDNA mutational count using the AZ 300-gene panel are shown in Supplementary Fig. S3.

Finally, we analyzed PD-L1 score in 230 patients, which revealed that 30.4% (70 of 230) had PD-L1 CPS ≥1. In this subset, we had 25 paired biopsy specimens (baseline and at PD to one of the VIKTORY regimens) available for PD-L1 analysis (Supplementary Table S6). All baseline and post-treatment biopsies were obtained from the same primary stomach lesion. Of the 25 paired samples analyzed, there were 2 patients (both treated with selumetinib/docetaxel) who showed a significant increase in PD-L1 (CPS ≥ 10) at progression after 5 to 8 months of selumetinib/docetaxel treatment (Fig. 7A). Arm1-019 patient developed multiple somatic mutations at the time of progression on selumetinib/docetaxel treatment by ctDNA analysis (Fig. 7B).

To our knowledge, this is the first and largest study to use an umbrella platform trial design with preplanned genomic biomarker analyses to assign patients with advanced gastric cancer to molecularly matched therapies. Using a centrally standardized molecular screening protocol, we enrolled 772 patients with gastric cancer and successfully performed tissue analysis for more than 90% (92.6%) of the patients as reported in our previous studies (28, 31). In this study, we demonstrated that when comprehensive molecular screening is linked to seamless immediate access to parallel matched trials, nearly 1 in 7 (14.7%) patients with advanced gastric cancer can receive biomarker-assigned drug treatment. The proportion of biomarker-driven treatment (14.7%) can be increased if the availability of seamless parallel trials is increased (i.e., FGFR2 amplification, EGFR amplification). Importantly, we showed that the biomarker-assigned cohort had encouraging response rates, underscoring the importance of genomically characterizing every patient's tumor for precision therapy.

Of the multiple arms, the highest response rate was observed in arm 4 (MET amplification–savolitinib monotherapy). Savolitinib is a potent small-molecule reversible MET kinase inhibitor that inhibits MET kinase with an IC₅₀ of 4 nmol/L in MET-amplified cancer cell lines. A phase II trial of savolitinib monotherapy in 44 patients with MET-altered papillary renal cell carcinoma showed very promising results, including 8 PRs (32). Our savolitinib monotherapy arm met the prespecified 6-week PFS rate, indicating that this treatment is worthy of phase III exploration in the MET-amplified subset of patients with gastric cancer (3%–5%; refs. 33, 34). Responders were enriched for higher MET copy number (7/10 with MET >10 copies), a biological phenomenon seen in HER2- and EGFR-amplified gastric cancer (35, 36), and adjusted plasma MET copy number was strongly correlated with duration of PFS. Highlighting the importance of genomic biomarker context, concurrent receptor tyrosine kinase (RTK) amplifications in addition to MET amplification resulted in short duration of response or no response to savolitinib. The importance of understanding the concurrent alteration landscape is highlighted by mixed results with prior MET-directed therapies in gastric cancer, likely owing to incomplete biomarker selection (37–39). Although this analysis is lacking functional validation, we speculate tumors with higher MET copy number without other RTK coamplifications are more dependent on MET signaling and may represent the optimal candidates for MET-directed therapies. Of note, patients with gastric cancer with a high level of ctDNA MET amplification (by Guardant360 assay in our study) may benefit more substantially from MET-targeted therapy.

In arm 3 (PIK3CA mutation–capivasertib), we observed moderate antitumor activity with an ORR of 33.3% (95% CI: 14.4–52.2) in 2L gastric cancer, especially when compared with the low response rate (<15%) observed in arm 7 (PIK3CA WT–capivasertib). Capivasertib is a selective pan-AKT inhibitor that inhibits the kinase activity of all three AKT isoforms (AKT1–3; ref. 17). We and others have previously observed differential distribution of PIK3CA hotspot mutations (E542K, E545K, H1047R) according to molecular subtypes—PIK3CA kinase domain H1047R mutations were enriched in microsatellite instability–high (MSI-H) gastric cancer (>80%), whereas helical domain E542K and E545K mutations were enriched in microsatellite-stable (MSS) tumors (1, 40). Given that each molecular subtype (MSI-H, MSS, genomically stable, or mesenchymal subtype) has substantially different survival outcomes (1), we have hypothesized that specific point mutations may show different drug sensitivity to capivasertib. Among arm 3 patients, we observed strikingly different efficacy based on PIK3CA genotype (Fig. 4C). In fact, none of the 4 patients with H1047R PIK3CA mutations responded to capivasertib. In contrast, 4 of the 8 with E542K mutations had durable responses to the capivasertib/paclitaxel combination, and 3 of the 4 patients were EBV-positive (Fig. 3C, green circles). Taken together, capivasertib/paclitaxel demonstrated the highest antitumor activity in MSS gastric cancer with PIK3CA E542K mutations. While this represents the first trial of a pan-AKT inhibitor in PIK3CA-mutated gastric cancer, randomized data will be important to validate our putative composite biomarker (PIK3CA helical domain +/MSS) population.

MAPK pathway alterations are frequent in advanced gastric cancer. We attempted to explore two biomarker selection strategies using selumetinib (AZD6244, ARRY-142886), which is a potent, orally active inhibitor of MEK1/2 that suppresses the pleiotropic output of the RAF/MEK/ERK pathway (22, 23). First, we confirmed that KRAS mutational status did not predict response to selumetinib in patients with gastric cancer, supporting the preclinical data with MEK inhibitors (23). On the basis of a study showing that the RAS pathway can be activated in the absence of KRAS mutation and that the RAS-pathway signature was superior to KRAS mutation status for the prediction of response to RAS-pathway inhibitors (41), a 6-gene MEK signature (DUSP4, DUSP6, ETV4, ETV5, PHLDA1, and SPRY2) was developed and validated in the gastric cancer cohort (42). Given that the prevalence of high MEK signature was only 6.9%, the predictive power of high MEK signature should be tested in a subsequent enriched clinical trial with high MEK signature as a selection biomarker in gastric cancer. Interestingly, we observed the most durable response in a KRAS amplification/MEK-H patient without concurrent KRAS mutation, consistent with recent reports of MEK inhibition in this genomically defined subset (43).

Recent trials have underscored the importance of anti–PD-1 or anti–PD-L1 therapy in gastric cancer treatment especially in patients with metastatic gastric cancer with EBV-positive or high mutational load or MSI-H or PD-L1 combined CPS ≥ 1 by IHC (27, 44). We observed substantial induction (increase of ≥10) of PD-L1 in 8% (2/25) of paired biopsies from primary tumors in the selumetinib/docetaxel arm (Supplementary Table S6). MAPK inhibition by cobimetinib in preclinical tumor models has been shown to promote tumor-infiltrating CD8⁺ T-cell activity (45). In addition, atezolizumab and cobimetinib combination treatment has been shown to increase intratumoral CD8⁺ T-cell infiltration and MHC I expression in patients with MSS colorectal cancer (46). Concordantly, we also observed PD-L1 change with recruitment of intratumoral CD8⁺ lymphocytes following selumetinib/docetaxel treatment. Although a recent cobimetinib/atezolizumab trial has failed to show survival benefit in patients with MSS colorectal cancer (47), selumetinib and anti–PD-1 treatment may be explored in patients with MSS gastric cancer. Congruently, this highlights the nonstatic nature of PD-L1 as a selection biomarker and suggests combination and/or sequential strategies are worth exploration.

Although we are a long way from claiming “VIKTORY” in gastric cancer, we have successfully shown that tumor genomic profiling with matched therapies improves outcomes in 2L treatment, and that platform clinical trials can efficiently identify the optimal biomarker treatment match (e.g., savolitinib to patients with MET-amplified gastric cancer). Nevertheless, this signal needs to be confirmed in an expansion or randomized trial. Exploratory analyses demonstrated that biomarkers such as genomic alterations and/or PD-L1 may not be static, especially during or after treatment. The proportion (14.7%) of patients receiving biomarker-driven treatment in the VIKTORY trial may be improved with more available targeted agents based on genomic alterations (e.g., FGFR2 or EGFR2 amplification) and inclusion of PD-L1 positivity (especially PD-L1 CPS ≥10) may enable interrogation of the potential benefit of anti–PD-L1 treatment with or without targeted agents in future umbrella trials. Finally, although limited by a very small number of patients, we have demonstrated that PD-L1 status changes over time in gastric cancer following selumetinib/docetaxel treatment.

Patients with histologically confirmed metastatic and/or recurrent gastric adenocarcinoma, an Eastern Cooperative Oncology Group (ECOG) performance status of 0 or 1, and at least one measurable lesion according to the RECIST 1.1 were eligible for enrollment in the VIKTORY trial, the molecular screening program, and one of the associated umbrella trial protocols in gastric cancer. Adequate hematologic function, hepatic function, and renal function were required. Patients with other concurrent uncontrolled medical diseases and/or other tumors were also excluded. The trial was conducted in accordance with the Declaration of Helsinki and the Guidelines for Good Clinical Practice (ClinicalTrials.gov Identifier: NCT#02299648). The trial protocol was approved by the institutional review board of Samsung Medical Center (Seoul, Korea) and all participating centers, and all patients provided written informed consent before enrollment.

The main goal of the VIKTORY trial as a molecular screening program was to identify novel molecular subsets for assigning patients into one of the associated biomarker-directed arms (Fig. 1A). There were 10 associated independently operated phase II arms (arms 4 and 8 included a dose-finding phase I trial) with eight biomarkers. Each experimental drug protocol was designed independently from the screening protocol. The eight biomarkers were: Biomarker A1: RAS mutation or RAS amplification; Biomarker A2: high MEK (MEK-H) or low MEK (MEK-L) signature; Biomarker B: TP53 mutation; Biomarker C: PIK3CA mutation or amplification; Biomarker D: MET amplification; Biomarker E: MET overexpression (3+) without MET amplification; Biomarker F: all negative (TP53 WT/PIK3CA WT/RAS WT); and Biomarker G: TSC2 null or RICTOR amplification. There were 10 phase II trials that were associated with the VIKTORY screening protocol: arm 1: selumetinib + docetaxel (Biomarker A1/A2, NCT#02448290); arm 2: adavosertib + paclitaxel (Biomarker B, NCT#02448329); arm 3: capivasertib + paclitaxel (Biomarker C, NCT#02451956); arm 4-1: savolitinib monotherapy (Biomarker D, #02449551); arm 4-2: savolitinib + docetaxel (Biomarker D, NCT#02447406), arm 5: savolitinib + docetaxel (Biomarker E, NCT#02447380); arm 6/7/8: vistusertib + paclitaxel or capivasertib + paclitaxel (Biomarker F, NCT#02449655) or AZD6738 + paclitaxel (NCT#02630199); and arm 9–10: vistusertib + paclitaxel (Biomarker G, NCT#03082833, NCT#02449655), vistusertib + paclitaxel (Biomarker G, NCT#03061708). If patients initially enrolled in the VIKTORY trial were not eligible or refused to participate in one of the associated trials, they were allowed to be treated with conventional chemotherapy, or in non-VIKTORY clinical trials.

FFPE or fresh samples of gastric cancer containing >40% tumor cellularity were used for targeted sequencing. Genomic DNA was extracted using the Qiagen DNA kit for FFPE tissue or the QIAamp DNA Mini Kit for fresh tumor tissues (Qiagen) according to the manufacturer's instructions. The IHC protocol for MET and HER2 used for this trial has been reported previously (48). The remaining tissue samples were reused in case of insufficient DNA amount/quality for molecular analysis, or otherwise stored for further study.

The targeted sequencing method for tissue specimens is provided in the Supplementary Material.

Tissue sections were freshly cut to 4 μm–thick sections and mounted on Fisherbrand Superfrost Plus Microscope Slides (Thermo Fisher Scientific) and then dried at 60°C for 1 hour. IHC staining was carried out on Dako Autostainer Link 48 system (Agilent Technologies) using Dako PD-L1 IHC 22C3 pharmDx kit (Agilent Technologies) with EnVision FLEX visualization system and counterstained with hematoxylin according to the manufacturer's instructions. PD-L1 protein expression was determined using CPS, which was the number of PD-L1–staining cells (tumor cells, lymphocytes, macrophages) divided by the total number of viable tumor cells, multiplied by 100. The specimen was considered to have PD-L1 expression if CPS was ≥1. For CD3 and CD8, IHC staining was performed on tissue sections from FFPE-embedded specimens with VENTANA BenchMark automated staining instrument (Ventana Medical Systems, Inc.). Specimens were incubated with CONFIRM anti-CD3 (2GV6) and CONFIRM anti-CD8 (SP57) rabbit mAbs for 20 minutes and CD3⁺ and CD8⁺ immune cells were visualized using the OptiView DAB IHC Detection Kit.

Antibodies used in this study were specific for MLH-1 (M1, Ventana, ready to use) using Ventana BenchMark XT autostainer (Ventana); MSH2 (G219-1129, 1:500, Cell Marque), PMS2 (MRQ-28, 1:20, Cell Marque), and MSH6 (44/MSH6, 1:500, BD Biosciences) using BOND-MAX autoimmunostainer (Leica Biosystems). In interpretation, loss of nuclear staining in the tumor cells with positively stained internal control was counted as an abnormal result. In cases with loss or suspected as loss of MMR protein, IHC was initially selected and further IHC with the entire block was performed to screen for MMR deficiency. Cases with negative or equivocal nuclear staining were subsequently tested for microsatellite instability using PCR. EBV status was determined by EBER in situ hybridization using standard protocols (27).

ctDNA testing using Guardant360 (Guardant Health) was performed as described previously (49). Briefly, up to 30 ng of cell-free DNA extracted from banked plasma was used for library preparation and enrichment by hybridization capture. Enriched libraries were then sequenced on a NextSeq 550 (Illumina), and the resulting sequence data was analyzed using a locked, previously validated custom bioinformatics pipeline. Plasma copy number was reported as directly observed and adjusted as described previously (50). Change in total ctDNA levels was calculated as described previously (51) and reported as proportional fold change truncated at 10% for graphical purposes.

The molecular tumor board (MTB) was composed of medical oncologists, pathologists, bioinformaticians, and the small-molecule experts from AstraZeneca. The MTB had the responsibilities of scientific validation, prioritization of identified molecular aberrations, and providing guidance on the suitable biomarker-driven experimental arm under the umbrella trial. The process time between biopsy and molecular results was set as 21 to 30 days from our previous study (28, 31). If multiple targets were simultaneously detected in a single patient, the following prioritization was used for patient assignment based on known drivers: (i) PIK3CA mutation/amplification; (ii) RAS mutation/amplification or MEK signature; (iii) MET amplification; (iv) TP53 mutation; (v) RICTOR amplification; (vi) TSC2 null; (vii) MET overexpression by IHC 3+; and (viii) if none of the above biomarkers were present, patients were allocated to the biomarker-negative arms AZD6738/paclitaxel, capivasertib/paclitaxel, phase I portion of docetaxel/savolitinib, other clinical trials, or conventional treatment. The status of enrollment for 10 associated clinical trials (10 phase II studies) is shown in Supplementary Table S1. Currently, patient enrollment has been completed in arms 1, 2, 3, 4, 6, and 7. Further patient enrollment was stopped in arms 4-1 and 5, and arms 9/10 have been closed early due to early termination of the drug for further clinical development.

This trial was designed as two parts: (i) VIKTORY screening protocol for molecular profiling; (ii) parallel phase I/II study with independent statistical assumptions for each arm. For each arm, the primary endpoint was ORR. We adopted Simon's optimal design assuming ORR of 20% for 2L weekly paclitaxel regimen based on robust data from previous studies; the experimental arm (paclitaxel + targeted agents) was considered effective for further development if the combination rendered an ORR of ≥50%. Each arm was designed as a two-stage design, allowing ineffective drugs to be terminated early at stage I. Secondary endpoints were PFS, OS, and correlative biomarker analysis using ctDNA, PD-L1 score, and genomic aberration.

Statistical analyses were performed using the software environment R v3.4.0. The clinical information distribution plots were created using Circos (52). Survival analyses were performed to explore the influences of age, gender, pathology, disease status, the number of metastatic organs, EBV status, MMR status, PD-L1 status, and VIKTORY biomarker status. Survival function curves were visualized using the library and the differences between the levels of each factor were assessed using a log-rank test. Likewise, to model hazard functions and determine the effects of these factors on a patient's survival, Cox proportional hazard models were used. The proportional hazard assumption of Cox models was tested using the R library survival (53). The significance of multiple predictors of survival was assessed by the Cox regression analysis. P < 0.01 was considered to indicate a statistically significant difference. The forest plot of the HRs according to the OS was generated using in-house code. We used the lollipop chart to visualize the maximum change in tumor size per RECIST 1.1.

I. Kozarewa has ownership interest (including stock, patents, etc.) in AstraZeneca. P.G.S. Mortimer is a senior principal medical scientist at AstraZeneca plc and has ownership interest (including stock, patents, etc.) in the same. J.I. Odegaard is a VP, clinical development at Guardant Health. E.A. Harrington has ownership interest (including stock, patents, etc.) in AstraZeneca. S. Luke has ownership interest (including stock, patents, etc.) in AstraZeneca and Illumina. Y.H. Kim has ownership interest in AstraZeneca stock and is a consultant/advisory board member at Orum Therapeutics. A.A. Talasaz is President and COO at Guardant Health, Inc. and has ownership interest (including stock, patents, etc.) in the same. S.J. Hollingsworth has ownership interest in AstraZeneca stock. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J. Lee, S.T. Kim, E.A. Harrington, Y.H. Kim, J.O. Park, H.Y. Lim, A.A. Talasaz, S.J. Hollingsworth, K.-M. Kim, W.K. Kang

Development of methodology: J. Lee, S.T. Kim, J.I. Odegaard, E.A. Harrington, H.Y. Lim, K.-M. Kim, W.K. Kang

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Lee, S.T. Kim, H. Lee, I. Kozarewa, J.I. Odegaard, J. Lee, T. Lee, S.Y. Oh, J.-H. Kang, J.H. Kim, Y. Kim, J.H. Ji, Y.S. Kim, K.E. Lee, T.S. Sohn, J.Y. An, M.-G. Choi, J.H. Lee, J.M. Bae, S. Kim, J.J. Kim, Y.W. Min, B.-H. Min, J.Y. Hong, S.H. Park, J.O. Park, Y.S. Park, H.Y. Lim, K.-M. Kim, W.K. Kang

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J. Lee, S.T. Kim, K. Kim, I. Kozarewa, J.I. Odegaard, E.A. Harrington, J.H. Ji, J. Kim, T.S. Sohn, N.K.D. Kim, S. Luke, Y.S. Park, H.Y. Lim, A.A. Talasaz, K.-M. Kim, W.K. Kang

Writing, review, and/or revision of the manuscript: J. Lee, S.T. Kim, K. Kim, I. Kozarewa, P.G.S. Mortimer, J.I. Odegaard, E.A. Harrington, J.-H. Kang, J.H. Ji, J.Y. An, J.M. Bae, S. Kim, S. Luke, Y.H. Kim, H.Y. Lim, S.J. Hollingsworth, K.-M. Kim, W.K. Kang

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.-H. Kang, Y. Kim, B.-H. Min

Study supervision: J. Lee, E.A. Harrington, J.-H. Kang, J.H. Ji, A.A. Talasaz, K.-M. Kim, W.K. Kang

We would like to thank Drs. Adam J. Bass, Joseph Chao, and Samuel J. Klempner for scientific discussion and critical review of our manuscript. On behalf of the VIKTORY team, we would like thank our patients and their families for their participation. This work was supported by funding from the Korean Health Technology R&D Project, Ministry of Health & Welfare, Republic of Korea (HI14C3418). Support was also provided by a grant from the 20 by 20 Project of Samsung Medical Center (GF01140111). This investigator-initiated trial was also funded by a study-drug donation and partial fund from AstraZeneca.