Determinants of Response and Intrinsic Resistance to PD-1 Blockade in Microsatellite Instability–High Gastric Cancer

Proficient cellular mismatch repair (MMR) machinery is required for the detection and replacement of single-nucleotide mismatches and the correction of small insertions and deletions that may occur during DNA replication, particularly at repetitive sequences known as microsatellites. Loss of MMR activity (dMMR) is associated with tumor development and high microsatellite instability (MSI-H). These tumors harbor large numbers of frameshift and single-nucleotide variants and are characterized by a high tumor mutational burden (TMB). Across tumor types, MSI-H tumors are associated with increased tumor-infiltrating lymphocytes and enrichment of PD-L1 expression. These clinicopathologic observations partly underlie immunogenicity and the clinical responsiveness to immune checkpoint inhibitor (ICI) therapy (1–6). In gastric cancer, MSI-H is observed in 9% to 22% of nonmetastatic cases and in 3% to 14% of advanced cases (7–11).

The clinical efficacy of PD-1 antibodies in metastatic gastric cancer has been well validated. Among the total of 67 MSI-H patients enrolled in the Keynote-059, -061, and -062 trials, the overall response rates to pembrolizumab monotherapy were 57%, 47%, and 57%, respectively (12). These outcomes were superior to those of patients treated with chemotherapy in trials with a comparator arm [HR of median overall survival 0.37; 95% confidence interval (CI), 0.14–0.97 in Keynote-062 trial; refs. 13–15)]. A similar responsiveness is seen in first-line colorectal cancer and in later-line therapy for MSI-H tumors (16, 17). The well-documented MSI-H activity has led to tissue-agnostic approval for pembrolizumab (Keytruda; Merck MSD), a selective, humanized, high-affinity IgG4-κ mAb designed to bind to PD-1 and prevent interactions between PD-1 and its ligands (e.g., PD-L1; ref. 18). Notably, regardless of tumor type, the objective response rate to anti–PD-1 therapies in MSI-H tumors appears to peak at slightly above 50%, suggesting that nearly half of the tumors are intrinsically resistant. The genomic and immunologic explanations for the variable response in gastric cancer remain a significant knowledge gap although recent studies have highlighted the importance of DNA sensing and cyclic GMP-AMP synthase-stimulator of IFN genes (cGAS–STING) pathway activation, which could trigger antitumor innate immune response elements (14, 19–22).

Although small heterogenous series have suggested that within MSI-H cancers, TMB heterogeneity, randomly acquired alterations in antigen presentation machinery, and intratumoral heterogeneity may attenuate the antitumor response (23–27), prospective examination is lacking. To generate a more granular understanding of the genomic and immunologic determinants of the PD-1 response in MSI-H gastric cancer, we conducted a phase II trial in which we newly enrolled 19 patients with advanced MSI-H gastric cancer and treated with pembrolizumab monotherapy as an expansion to our previously published heterogenous cohort (28). All patients underwent pretreatment tissue biopsy with whole-exome sequencing (WES) to infer genomic variations and TMB, followed by RNA sequencing (RNA-seq) to categorize patients according to transcriptional signatures and tumor microenvironment (TME) components. Early on-treatment multi-region tissue samples were performed to examine clonal architecture changes as a putative response biomarker. To explore potential noninvasive biomarkers, we performed serial collection of peripheral blood mononuclear cells (PBMC) to evaluate the dynamic changes in immune cell populations.

Nineteen patients with metastatic gastric cancer were enrolled in this study between August 2017 and December 2019 (Table 1). Cohort A was published previously (N = 61; ref. 28), and cohort B, reported here, comprised patients with MSI-H or MMR-deficient gastric cancer (N = 19). The median age of the patients was 69 years (range, 47–90 years). All patients were of Korean ethnicity. Sixteen (84.2%) patients received pembrolizumab as second-line therapy, and three (15.8%) received pembrolizumab as third-line treatment for metastatic disease. There were no new safety signals, and the toxicity profile was aligned with that of previous reports (Supplementary Table S1). All patients had tumors located within the gastric body or antrum. About two thirds of the patients (68.4%) had more than two sites of metastatic involvement. Fourteen patients (87.5%) had confirmed PD-L1 positivity [combined positive score (CPS) ≥ 1] metastatic gastric cancer (Table 2), and there were no Epstein–Barr virus (EBV)-positive patients. All 19 patients underwent pretreatment biopsy (16 stomach/primary tumor, 2 peritoneal, and 1 distant lymph node) before study entry. Moreover, 15 specimens passed quality control for WES, and 16 specimens were of sufficiently high quality for RNA-seq (low DNA/RNA quality, n = 3; low tumor volume, n = 1; Supplementary Table S2).

At the data cutoff (June 3, 2020), 18 patients were evaluable for response with a median follow-up of 19.5 months (range, 6.1–33.6). Complete response (CR) was achieved in 3 patients [16.7%, 95% confidence interval (CI): -0.6–33.9], 7 patients [38.9%; 95% CI, 16.4–61.4] achieved confirmed partial response (PR), and 6 patients [33.3%; 95% CI, 11.6–55.1] had stable disease, resulting in an overall response rate (ORR) of 55.6% (95% CI, 32.6–78.5) and a disease control rate (DCR) of 88.9% (95% CI, 74.4–103.4; Fig. 1A and B; Table 2). Among the complete responders, two patients completed 35 cycles of pembrolizumab and remained in CR off pembrolizumab (EP-62 and EP-67). The depth of response was notable; 5 of the 7 PRs experienced more than a 50% reduction in tumor burden per Response Evaluation Criteria in Solid Tumors (RECIST) 1.1 (Fig. 1A). Among all enrolled patients (n = 19), 7 patients were dead, and 11 patients continued to receive study treatment at data cutoff. The median overall survival and the median progression-free survival was the same at 26.9 months (Supplementary Fig. S1A and S1B). The median duration of response among those with a PR was 21.2 months (range: 6.3–26.9 months; Fig. 1B). Moreover, patients who achieved more than 50% of tumor reduction compared with baseline within 6 months maintained a durable response to pembrolizumab (Fig. 1C). Among seven patients who achieved SD, three patients (42.9%) progressed within 6 months (Fig. 1C). In all evaluable patients (n = 18), 13 patients (72.2%) showed durable clinical benefit (DCB), as defined by more than 6 months of progression-free survival. The patients who showed DCB had superior PFS (26.9 months vs. 4.0 months, P < 0.0001) and OS (26.9 months vs. 4.5 months, P < 0.001; Supplementary Fig. S1C and S1D).

We analyzed the pre- and post-pembrolizumab somatic genomic landscapes of 15 patients with MSI-H metastatic gastric cancer. We first sought to identify pretreatment genomic alterations (mutations and copy-number variations) that were significantly associated with a clinical response to pembrolizumab (Fig. 2A; Supplementary Fig. S2A). Although the majority of patients achieved some level of disease control, we observed CDH1 and JAK2 mutations in patients EP-76 and EP-77, who obtained stable disease as their best response. Similarly, patient EP-63, who demonstrated refractory peritoneal carcinomatosis and harbored CDH1, RHOA, ERBB2, and FGFR2 alterations, did not achieve significant tumor reduction. AXIN1 and PTCH1 mutations, which are related to the Wnt/β-catenin pathway, were identified in nonresponders. Taken together, there was an overrepresentation of genomic alterations that are previously associated with innate PD-1 resistance in patients with a lesser magnitude of benefit although these associations did not reach statistical significance owing to small numbers. Such differences might be due to the high accumulation of passenger events driven by DNA mismatch repair deficiency in MSI-H tumors, and we observed wide-spread pathogenic mutations in MMR pathway members (Fig. 2A; Supplementary Table S3), and a highly enriched dMMR mutational signature in our cohort as predicted (Fig. 2B).

The TMB, measured as the total number of nonsynonymous somatic mutations per megabase, trended toward correlation with the overall frequency of somatic MSI events (P = 0.05238, Pearson χ² test; Supplementary Fig. S2B). Among our entire cohort with evaluable tissue (n = 15), the median TMB among responders (37 mutations per megabase, Mut/Mb, n = 6) was significantly higher than that among nonresponders (24 Mut/Mb, n = 8; P = 0.0018, Mann–Whitney test; Fig. 2C). This was consistent with a recent study in MSI-H colorectal cancer (29, 30). Thus, the degree of TMB elevation appeared to be predictive of the response to pembrolizumab within MSI-H gastric cancer. Similarly, patients with a high number of nonsynonymous mutations showed the greatest depth of response (Fig. 2D). Using ROC curve analysis, the optimal cutoff of TMB to distinguish between responder and nonresponder was determined as 26 Mut/Mb (sensitivity, 100%; specificity, 75%, AUC, 0.917; P = 0.0098). In addition, patients with a TMB greater than 26 Mut/Mb (n = 9) showed significantly longer progression-free survival (PFS, not reached vs. 4.3 months; HR, 0.06; 95% CI, 0.007–0.45; P = 0.0077) compared to those with TMB less than 26 Mut/Mb (n = 6; Fig. 2E). Among nonsynonymous mutations, missense, nonsense, and splice mutations were significantly higher in responders than nonresponders (Fig. 2F; Supplementary Fig. S2C). Other genomic features including ploidy, genome instability, intratumoral heterogeneity, technical variation (tumor purity), and IHC PD-L1 expression (CPS scoring) showed no significant ability to distinguish responders from nonresponders (Supplementary Fig. S2D–S2F). Genomically, within our cohort of patients with MSI-H metastatic gastric cancer, only the pretreatment degree of TMB elevation was a significant predictor of response to pembrolizumab.

The cancer immunoediting hypothesis posits the immune system effectively eliminates highly immunogenic cancer clones that present neoantigens derived from tumor mutations and consequently selects clones that are poorly immunogenic or immune-evasive (31). To investigate tumor evolution under the therapeutic pressure of pembrolizumab in MSI-H metastatic gastric cancer, we collected serial multi-region biopsies from pre- and post-pembrolizumab cases through mapped endoscopic biopsies and analyzed the molecular evolutionary dynamics during pembrolizumab treatment. Intriguingly, the patients who achieved CR or PR showed a significant mutation burden reduction across multiple lesions after pembrolizumab therapy (P = 0.0069, Wilcoxon signed-rank test), while nonresponding patients exhibited a minimal reduction in the number of mutations (Fig. 3A).

To further understand mutational selection in spatiotemporal tumor evolution after pembrolizumab treatment, we estimated the relative rates of nonsynonymous and synonymous mutations (the dN/dS ratio) using trinucleotide mutational signatures, sequence composition, and variable mutation rates across a set of homologous protein-coding genes as described previously (32, 33). This approach describes the survival advantage of mutant clones (dN/dS = 1, neutral selection with clonal drift; dN/dS > 1, positive selection with adaptive advantage; dN/dS < 1, negative selection resulting in clonal loss). Interestingly, in response to pembrolizumab treatment, genome-wide missense dN/dS ratios were significantly lower only in responding patients (P = 0.027, Wilcoxon signed-rank test; Fig. 3B), suggesting reduced clonal expansion with decreased adaptive fitness and a reduction in subclonal diversity.

Next, we sought to explore whether longitudinal changes in the genetic composition of MSI-H gastric cancer tumors in response to pembrolizumab-affected mutant clone size. We used PhylogicNDT to structure subclonal populations of cancer cells across two time points and to stochastically model the order of clonal driver events and cell population dynamics (34). As expected, the total size of mutant clones was decreased in the post-pembrolizumab treatment sample of a responding patient (EP-72; Fig. 3C). Conversely, a patient with a poor response showed clonal evolution (with a subclone harboring EGFR and HLA-B mutations) between the two time points, consistent with a previous report (ref. 35; Fig. 3D). We then applied this approach to multi-region biopsies. By mapping sequencing data to visually annotated endoscopic tumor sampling regions, we simulated how subclones were populated in the spatiotemporal context for each patient posttreatment. While responsive patients showed intra- and interpatient genetic heterogeneity, negative selection with lower dN/dS ratios was dominant, with a decreased cell population size after two cycles of therapy (Fig. 3E and F). Notably, we found that the dN/dS ratio was strongly predictive of the overall mutant clone size (P = 9.57 × 10⁻⁸, Pearson χ² test; Supplementary Fig. S2F). In a nonresponding patient EP-71, intratumoral heterogeneity was present and mutational selection was preserved, but the overall mutant clone size was maintained after therapy, reflecting an intrinsically resistant tumor (Fig. 3G).

In an effort to elucidate potential transcriptional signatures or molecular determinants that are associated with pembrolizumab response, we compared the gene-expression profiles for pretreatment MSI-H gastric cancer specimens with those for normal adjacent tissue (NAT). Gene signature enrichment analysis showed that the pathways related to T-cell receptor (TCR) signaling were significantly increased in patients who showed a good response to pembrolizumab, reflecting preexisting immune recognition (Fig. 4A). IL2-mediated signaling events were also enriched in the responder group, and IL2 is crucial for T-cell survival and activation and notably lost during T-cell exhaustion (36). From the analysis of intratumoral TCR clonality (1-Pielou evenness index), which is inversely correlated with TCR repertoire diversity, the patients with decreased TCRβ clonality in pretreatment (i.e., patients with a diverse pretreatment TCR repertoire) showed superior tumor reduction (Fig. 4B). Moreover, the patients with more diverse intratumor TCRβ repertoire in pretreatment (higher than median of Pielou evenness index; ≥0.92, n = 6) demonstrated significantly longer PFS (not reached vs. 4.2 months; HR, 0.090; 95% CI, 0.012–0.67; P = 0.0209) than those with less diverse intratumor TCRβ repertoire in pretreatment tumor biopsies (lower than median of Pielou evenness index; <0.92, n = 6; Fig. 4C).

To decipher the complexity of cellular diversity and dynamic immunologic changes in the TME in response to pembrolizumab therapy, we performed single-cell RNA sequencing (scRNA-seq) in paired tumor tissue specimens from five patients with MSI-H gastric cancer [one responder (EP-72); four nonresponders (EP-75-EP-78)] before and after two cycles of pembrolizumab treatment (Supplementary Fig. S3A). Dimensional reduction analysis via Uniform Manifold Approximation and Projection (UMAP) revealed six major global cell types, including epithelial cells, stromal cells, myeloid cells, B cells, and T cells, and natural killer (NK) cells, as well as several subtypes (Fig. 4B–F; Supplementary Fig. S3B). Responders had a higher number of T and NK cells and a lower number of stromal cells in tumor tissues at baseline than nonresponders (Fig. 4E). Following two cycles of pembrolizumab, responders had expanded T- and NK-cell populations and a decrease in the epithelial cell population, whereas nonresponders exhibited a decrease in the T- and NK-cell populations, and unchanged epithelial cell population (i.e., minimal tumor cell death), an increased stromal cell population, and increased cancer-associated fibroblast abundancy (Fig. 4E–H). Further analysis of the T-cell populations (Fig. 4I; Supplementary Figs. S3B and S4A) revealed that effector CD8⁺ T cells, exhausted CD8⁺ T cells, and γδ T cells were increased and that regulatory T cells (Tregs) were decreased in responders after two cycles of pembrolizumab. In nonresponders, the populations of CD4⁺ T cells and Tregs were decreased, while that of exhausted CD8⁺ T cells was increased during pembrolizumab therapy (Fig. 4I), consistent with the failure to reinvigorate an antitumor immune response.

Exhausted CD8⁺ T cells are divided into functionally heterogeneous subpopulations: stem-like cells, transitory cells, and terminally differentiated (TD) cells (37). Stem-like cells can reproduce by themselves, differentiate into transitory cells and TD cells (which are highly cytotoxic), and only exhibit proliferative bursts when activated by PD-1 inhibitors (30, 37, 38). Responding patients demonstrated a dominant proportion of stem-like exhausted cells before pembrolizumab treatment; however, nonresponding patients demonstrated a higher proportion of TD-exhausted cells and a lower proportion of stem-like cells (Fig. 4J; Supplementary Fig. S4B). Moreover, responding patients exhibited increased proportions of proliferating, transitory, and TD cells during pembrolizumab therapy, suggesting improved cytolytic function and cytokine production of exhausted T cells. Taking our bulk RNA-seq and scRNA-seq data together, patients with MSI-H gastric cancer who showed a good response to pembrolizumab demonstrated abundant preexisting tumor-infiltrating lymphocytes (TIL), a diverse pretreatment TCR repertoire, and a high proportion of stem-like exhausted cells in dysfunctional CD8⁺ TILs.

Given that the expansion of effector T cells in tissue was associated with a pembrolizumab response, we hypothesized that dynamic changes in peripheral blood (PB) T cells may be a noninvasive marker for therapeutic benefit in MSI-H gastric cancer, as has been suggested in patients with lung cancer (39). Using multi-color flow cytometry, we analyzed PB T-cell populations before and after two cycles of pembrolizumab. Considering that the targets of pembrolizumab are PD-1–expressing cells, we focused on the characteristics and dynamics of PD-1⁺ T cells in PB. Among all enrolled patients (n = 19), the percentage of Ki-67⁺ cells among PB PD-1⁺CD8⁺ T cells was significantly increased during pembrolizumab therapy, suggesting that PD-1–expressing cells proliferated in response to pembrolizumab (Fig. 5A). Conversely, the percentage of PB Ki-67⁺ PD-1⁺ Treg cells was significantly decreased during pembrolizumab therapy. The ratio of the number of CD8⁺ T cells to the number of Tregs in PB was not significantly different during pembrolizumab treatment and was not correlated with the response to pembrolizumab. Thus, the proportional changes in PB CD8⁺ T cells and Tregs did not reflect the changes in immune cells observed in the tumor tissue. In addition, the intensity of thymocyte selection-associated high mobility group box protein (TOX), a master regulatory factor for T-cell exhaustion (40–42), did not exhibit any differences during therapy or in association with pembrolizumab response in our cohort (Fig. 5A; Supplementary Fig. S5A). Interestingly, the percentage of PD-1⁺ cells among peripheral CD8⁺ T cells at the baseline was significantly associated with the durable clinical benefit (more than 6 months of PFS) to pembrolizumab in patients with MSI-H gastric cancer (P = 0.0435, Mann–Whitney test; Fig. 5B). The patients who harbored among the highest levels of PD-1 expression among PB CD8⁺ T cells before pembrolizumab treatment demonstrated complete remission (EP-67) or dramatic response (EP-80) in tumor assessment (Fig. 5C and D). Conversely, patients who did not respond to pembrolizumab and progressed within 6 months were likely to demonstrate low PD-1 expression in CD8⁺ T cells before pembrolizumab treatment (Fig. 5B and E). Taken together, the higher percentage of baseline PD-1⁺ in PB CD8⁺ T cells was associated with a clinical benefit to pembrolizumab in MSI-H gastric cancer.

In this study, we used a suite of tissue and PB approaches to prospectively examine genomic and immunologic determinants of clinical response to pembrolizumab in MSI-H gastric cancer in a phase II clinical trial. The clinical activity (ORR of 55.6% and DCR of 88.9%) in our 19-patient cohort was well aligned with published MSI-H gastric cancer data, including Keynote-059 (ORR 57.1%, 4 among 7 patients, third-line; ref. 13), Keynote-061 (ORR 46.7%, 7 of 15 patients, second-line; ref. 14), Keynote-062 (ORR 57.1%, 8 among 14 patients, first-line; ref. 12), and Keynote-158 (ORR 45.8% 11 among 24 patients, more than second-line; ref. 19). Our dataset is the largest prospective PD-1 monotherapy study in MSI-H gastric cancer and provided a unique group within which to explore the variable response to PD-1 blockade.

Using an engineered mouse system, Mandal and colleagues recently reported that the insertion–deletion mutational load accumulated by MSI-H tumors is strongly associated with the response to the PD-1 blockade (26). Moreover, the predictive value of the TMB was reported in patients with MSI-H colon cancers who were treated with PD-1 or PD-L1 inhibitors (30); the MSIsensor score, a quantitative measure of MSI (MSI burden), was shown to be significantly correlated with the TMB. Our data provided further support for this observation and suggested a strong relationship between nonsynonymous single-nucleotide variations (SNV) and PD-1 response even within MSI-H gastric cancer. Specifically, the median TMB was 24 Mut/Mb among 14 evaluable patients, 37 Mut/Mb among responders (n = 6), and 23 Mut/Mb among nonresponders (n = 8). From the ROC curve analysis, 26 Mut/Mb was optimal cutoff to discriminate response of pembrolizumab in this report, and further investigation will be needed to generalize this threshold in MSI-H gastric cancer. We extended this analysis and observed that the depth of clinical response (% tumor decrease) was also significantly related to the TMB (Fig. 2D; P = 0.0057).

The genomic consequences of MSI-H biology are the accumulation of thousands of frameshifts and SNV mutations. While MSI testing is a principal component in the molecular subtyping of gastric cancer, we sought to explore whether specific recurrent genomic alterations differed among MSI-H gastric cancer cases. Although limited by sample size, we identified genomic features of the TCGA genomically stable (GS) or mesenchymal subtypes (CDH1, FGFR, and RHOA mutations) in a poor responder (EP-63). Alterations in Wnt/β-catenin pathway members were also observed in nonresponders and have been implicated in intrinsic PD-1 resistance (43, 44). Beyond interpatient genomic baseline heterogeneity, we found preliminary evidence that intrapatient heterogeneity could influence the response to pembrolizumab. Importantly, within our MSI-H gastric cancer cohort there was no clear association between PD-L1 expression and outcomes.

As of June 2020, pembrolizumab is FDA-approved for solid tumors with an elevated tumor mutation burden, defined as greater than or equal to 10 Mut/Mb as determined by the FoundationOne hybrid capture next-generation sequencing assay. Unsurprisingly, all patients in our cohort would meet this TMB-high definition, although patients with TMB greater than 26 Mut/Mb achieved significant clinical benefits suggesting a potential role for TMB even within MSI-H gastric cancer. To move beyond TMB, we utilized tissue- and plasma-based platforms to explore features underlying the variable response in a responder-enriched population. We identified that the presence of effector T cells in the TME, and the status of T-cell exhaustion were associated with response to pembrolizumab. Furthermore, increased TCR-related signaling and a diverse TCR repertoire in whole transcriptome analysis, as well as a higher proportion of T cells and a lower baseline proportion of stromal cells, especially cancer-associated fibroblast, were preferentially identified in responding patients. These findings suggested that both quantity and functional state of tumor-infiltrating T cells in the TME are required for an optimal pembrolizumab-induced antitumor response. Despite the small number of patients, we also confirmed that the existence of stem-like exhausted T cells in tumor tissues was associated with a response to anti–PD-1 (30). Although several of the pretreatment features we identified have been variably associated with ICI benefit, this is the first report of these features in gastric cancer.

Our cohort was designed to extend to on-treatment tissue and peripheral analyses, something largely lacking from existing data sets. Using on-treatment biopsy samples, we provided evidence that cancer immunoediting occurred during pembrolizumab treatment and that the collapse of clonal architecture during therapy was associated with pembrolizumab response. Our results also suggested that the inherently high mutation rate in MSI-H gastric cancer likely spontaneously generated subclonal somatic alterations, engendered immune escape, and underly the lesser response and/or primary progression seen in some patients on our trial. Whether serial analysis of circulating tumor DNA could guide adaptive strategies to boost response remains to be seen. Beyond tumor genomic changes, there is a clear interplay between TME and tumor clones in gastric cancer (39, 45, 46). Our scRNA-seq data supported clear differences in both baseline and adaptive TME composition between responders and nonresponders. We observed that nonresponders had frequent mutations and also upregulations in Wnt/β-catenin pathway and cancer-associated fibroblast abundancies. It is interesting to note that decreased T-cell infiltration and a lower NK-cell composition were observed in nonresponders. Activated Wnt/β-catenin signaling in cancer has been designated as intrinsic resistance mechanism for immunotherapy, and selective inhibitors for Wnt/β-catenin pathway could be a promising combination partner for ICIs (43). Whether NK cell–based immunotherapies could be a complementary treatment option for MSI-H PD-1 nonresponders, or those with baseline predictors of poor response, warrants further exploration.

Noninvasive pretreatment and on-treatment predictors of therapeutic benefit are clinically attractive tools, and several studies have explored potential biomarkers (39, 47–50). Currently, there are no peripheral biomarkers for immunotherapy response used in routine clinical practice. We confirmed that patients who demonstrated high percentages of peripheral PD-1⁺ CD8⁺ T cells achieved durable clinical benefit with pembrolizumab monotherapy. Although PD-1⁺ cells have been reported to be correlated with PB CD8⁺ T cells and tumor-infiltrating CD8⁺ T cells, the proportions of cancer antigen–specific T cells among peripheral PD-1⁺ CD8⁺ T cells and the roles in antitumor activity of PB PD-1⁺ cells require further investigation.

Our study is not without limitations, and although this is the largest prospective clinical data in MSI-H gastric cancer, the relatively small number of patients and lack of racial diversity may limit generalization. Although Asian patients with non–small cell lung cancer had better clinical outcome for ICIs than did white patients by pooled analysis (median OS, 18.7 vs. 11.1 months P = 0.005; ref. 51), we could not find ethnic differences with the response of pembrolizumab in MSI-H gastric cancer from available data. Furthermore, prior genomic analyses have not observed significant differences in mutational landscape between Asian and non-Asian gastric cancers.

Taken together our data point toward a subset of MSI-H gastric cancer with particular immune responsiveness and characterized by a composite of more elevated TMB, abundant infiltration of T cells, greater TCR clonal diversity, and less stem-like exhausted T cells at baseline. Patients with these features may not require anything beyond PD-1 monotherapy to achieve optimal outcomes. Of equal importance clinically is our observation of baseline unfavorable genomic and immunologic features, identifying a group of MSI-H gastric cancers that may require additional therapies to derive benefit from PD-1 blockade. In this subset, our data may point toward combination therapies geared toward reducing Treg populations (cytotoxic agents, ICOS, etc.) and/or enhancing and increasing NK-cell populations. Engineered model systems recapitulating MSI-H biology and broad genomic and immunologic screens for therapeutic vulnerabilities will be important to nominate candidate combinations for further testing. However, our findings suggest that detailed pretreatment and on-treatment characterizations are feasible and will likely be required to stratify MSI-H cancers for treatment with either a PD-1 blockade alone or novel combination approaches.

Patients enrolled in this study had measurable, histologically confirmed metastatic and/or recurrent gastric adenocarcinoma. The trial was conducted in accordance with the Declaration of Helsinki and the Guidelines for Good Clinical Practice (ClinicalTrial.gov identifier: NCT02589496). The trial protocol was approved by the Institutional Review Board of Samsung Medical Center (Seoul, Korea), and all patients provided written informed consent before enrollment (IRB#: 2015–09–053). To be eligible to participate in this study, patients were required to meet the following criteria: (i) histologically confirmed diagnosis of gastric or gastroesophageal junctional adenocarcinoma, (ii) MSI-H subtype detected by PCR; (iii) age of at least 19 years, (iv) prior failure of at least one line of chemotherapy that included platinum/fluoropyrimidine, (v) willingness to undergo a procedure to obtain fresh-frozen tissue within 42 days of treatment initiation for biomarker analysis, (vi) adequate organ function per protocol, (vii) at least one measurable lesion according to RECIST 1.1, and (viii) Eastern Cooperative Oncology Group performance status of 0 or 1 (7). All patients were naïve to anti–PD-1, anti–PD-L1, or anti–PD-L2 antibodies.

This prospective open-label phase II trial was designed as a single-arm phase II study at an academic cancer center. We previously reported a phase II trial of pembrolizumab in gastric cancer (Cohort A; ref. 28), and expanded an independent cohort for the patients with MSI-H only (Cohort B). All the patients in this report (Cohort B) were not included in cohort A and have not been previously reported. Pembrolizumab (200 mg) was administered via 30-minute intravenous infusion every 3 weeks until documented disease progression or unacceptable toxicity occurred, or up to 24 months. Tumor responses were evaluated every two cycles according to the RECIST 1.1 criteria. Toxicity was graded on the basis of the NCI Common Terminology Criteria for Adverse Events 4.0.

Tumor tissues were obtained at any time between day -42 and day 1 prior to initiation of the study treatment. Matched PB was also collected prior to initiation of the study treatment. After two cycles of pembrolizumab therapy, blood and tissues were obtained when sampling was available. Tumor tissues were obtained using multi-region biopsy with endoscopic mapping (52). If the tumor content was estimated to be more than 40% after thorough pathologic review, tumor DNA and RNA were extracted from freshly obtained tissues using a QIAamp Mini Kit (Qiagen) according to the manufacturer's instructions. In cases with DNA, we used RNaseA (catalog no. 19101; Qiagen). We measured concentrations and 260/280 and 260/230 nm ratios with an ND1000 spectrophotometer (Nanodrop Technologies, Thermo-Fisher Scientific) and then further quantified DNA/RNA using a Qubit fluorometer (Life Technologies.

Tissues were freshly cut into 4-μm sections, 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 a Dako Autostainer Link 48 system (Agilent Technologies) using the Dako PD-L1 IHC 22C3 pharmDx Kit (Agilent Technologies) with the 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–stained cells (tumor cells, lymphocytes, and macrophages) divided by the total number of viable tumor cells and multiplied by 100. The specimen was considered to have PD-L1 expression if CPS ≥ 1.

Tumor tissue MSI status was determined using PCR analysis of five markers with mononucleotide repeats (BAT-25, BAT-26, NR-21, NR-24, and NR-27), as described previously (53). Briefly, each sense primer was end-labeled with FAM, HEX, or NED. Pentaplex PCR was performed, and the PCR products were run on an Applied Biosystems PRISM 3130 automated genetic analyzer. Allelic sizes were estimated using Genescan 2.1 software (Applied Biosystems). Samples with allelic size variations in more than two microsatellites were considered MSI-H.

Genomic DNA was extracted from tissues (described above) and matched blood using the QIAamp DNA Blood Kit (Qiagen). Purified DNA was sheared to an average size of 150 bp in a Covaris Adaptive Focused Acoustics (AFA) sonication device (S2, Covaris Inc.) and indexed with unique barcode tags using PCR. Prepared libraries were assessed for quality and quantity using a Qubit Bioanalyzer (Agilent Inc.) and Mx3005P 1PCR (Agilent Inc.). Exomes were targeted using the SureSelect XT Human All Exon V6 kit (Agilent). Samples were multiplexed and flow-cell clusters were created using the TruSeq Rapid Cluster Kit and TruSeq Rapid SBS Kit (Illumina). Capture exomes were sequenced using the HiSeq 2500 platform (Illumina), and paired-end 100-bp sequence data were generated.

Total RNA concentration and quality were estimated using Quant-IT RiboGreen (Invitrogen). To determine the percentage of fragments with a size greater than 200 nucleotides (DV200), samples were run on the TapeStation RNA ScreenTape (Agilent). Overall, 100 ng of total RNA was used for sequencing library construction, using the TruSeq RNA Access Library Prep Kit (Illumina) according to the manufacturer's protocol. Briefly, total RNA was first fragmented into small pieces using divalent cations at elevated temperatures. The cleaved RNA fragments were copied into first-strand cDNA using SuperScript II reverse transcriptase (Invitrogen) and random primers. This was followed by second-strand cDNA synthesis using DNA polymerase I, RNase H, and dUTP. These cDNA fragments were subjected to an end-repair process, addition of a single “A” base, and subsequently, ligation of the adapters. The products were then purified and enriched using PCR to create the cDNA library. All libraries were normalized, and six were pooled into a single hybridization/capture reaction. Pooled libraries were incubated with a cocktail of biotinylated oligos, corresponding to the coding regions of the genome. Targeted library molecules were captured via hybridized biotinylated oligo probes using streptavidin-conjugated beads. After two rounds of hybridization/capture reactions, the enriched library molecules were subjected to a second round of PCR amplification. The captured libraries were quantified using KAPA Library Quantification kits for Illumina Sequencing platforms according to the qPCR Quantification Protocol Guide (KAPA BIOSYSTEMS, #KK4854) and assessed using the TapeStation D1000 ScreenTape (Agilent Technologies, # 5067–5582). Indexed libraries were then submitted to an Illumina HiSeq2500 (Illumina, Inc.), and paired-end (2 × 100 bp) sequencing was performed by Macrogen Inc. (South Korea).

For single-cell preparation, tumor tissue was dissociated with the gentleMACS Dissociator and Tumor infiltrating lymphocyte Kit (Miltenyi Biotec) according to the manufacturer's protocol. The cells were then cryopreserved in liquid nitrogen until use. All samples showed a viability of around 90% on average after thawing. scRNA-seq libraries were generated using the Chromium Single Cell 3 Library & Gel Bead Kit v3 (10 × Genomics) following the manufacturer's instructions. Briefly, the Chromium instrument was used to separate single cells into gel bead emulsions that facilitated the addition of cell-specific barcodes to all cDNAs generated during oligo-dT–primed reverse transcription. As a result, a cell barcoding sequence and UMI were added to each cDNA molecule. Libraries were constructed and sequenced at a depth of approximately 50,000 reads per cell using the HiSeq2500 platform (Illumina).

WES reads were aligned to the reference human genome GRCh37 using BWA-MEM (54) followed by preprocessing steps, including duplicate marking, indel realignment, and base recalibration using the Genome Analysis Toolkit (GATK; version 4.1.1.0), (55) generating analysis-ready BAM files. To increase the sensitivity for identifying both the lower and higher allele frequencies of somatic variants in the given tumor and paired normal BAM files at the genomic locus, we used the union variant callsets from two tools: MuTect2 (56) and Strelka2 (57). Default parameters were applied, and both variant callers were run with dbSNP (version 138; ref. 58), 1000G (phase I; ref. 59) and HapMap (phase III; ref. 60) data for known polymorphic sites. Filtered variants with minimum depth ≥ 5 and minimum alternative alleles ≥ 2 were annotated using the Ensembl Variant Effect Predictor (VEP; release version 87; ref. 61) with the GRCh37 database.

Allele-specific copy-number variation was estimated from WES using the FACETS (version 0.6.0; ref. 62) in default quantification mode with the given tumor and paired normal BAM files. The outputs, together with somatic point mutations, were used to compute purity, ploidy, and cancer cell fraction (CCF) as a function of the clonal composition in the given tumor by running ABSOLUTE (version 1.0.6; ref. 63), allowing us to calculate the proportion of the genome with an aberrant copy number relative to tumor ploidy, weighted on a per autosomal chromosome basis (wGII, weighted genomic instability index).

We explored replication slippage variants at microsatellite regions using MSIsensor (version 0.6; ref. 64), and somatically occurring MSI events were applied to represent an MSI score from a paired tumor and from normal BAM files.

We used Combined Annotation–Dependent Depletion (CADD) score to estimate the relative pathogenicity in genetic variations (65). According to the author's recommendation, the variants with CADD score higher than 10 points were regarded as pathogenic variant.

Mutational signature analysis was performed using the deconstructSigs package (version 1.6.0) in R, which selects combinations of known mutational signatures that account for the observed mutational profile in each sample (66). Exome regions were defined by the Agilent Sureselect V5 target region. Only somatic mutations in exome regions were considered, and trinucleotide counts were normalized by the number of times each trinucleotide context was observed in the exome region. Mutational signatures as defined by Alexandrov and colleagues, (67) named “signatures”, were utilized as the target signature set to be assessed in this sample set and were represented by the following terms: age (SBS1 and SBS5), apolipoprotein B mRNA editing enzyme, catalytic polypeptide-like (APOBEC; SBS2 and SBS13), UV (SBS7a, SBS7b, SBS7c, and SBS7d), temozolomide (SBS11), smoking (SBS4), immunoglobulin gene hypermutation (SBS9), homologous recombination deficiency (HRD; SBS3), mismatch repair deficiency (MMRD; SBS6, SBS15, SBS20, and SBS26), nucleotide excision repair deficiency (NERD; SBS8; ref. 68), DNA proofreading deficiency (DPD; SBS10a and SBS10b), and base excision repair deficiency (BERD; SBS18).

RNA sequence reads were annotated with ENSEMBL (version 98), aligned to the human reference genome (GRCh38) using STAR (version 2.6.1; ref. 69), and quantified in units of transcript per million (TPM) as a function of gene expression using RSEM (version 1.3.1; ref. 70), applying the option parameters suggested by the GTEx project (https://github.com/broadinstitute/gtex-pipeline/blob/master/TOPMed_RNAseq_pipeline.md). TPM values less than one were considered unreliable and substituted with zero.

To explore the activity of signaling pathways, we used gene set variation analysis (GSVA; ref. 71) in default mode. Together with the TCR-related signaling pathways of interest, we applied all the other gene sets, which were previously applied in TCGA's study of gastric adenocarcinoma (ref: TCGA GC, 2014, Nature, PMID 25079317), from the NCI Pathway Interaction Database, http://www.ndexbio.org/).

We profiled the TCR repertoire from RNA sequences using TRUST4, same as updated version of TRUST (72) with default option parameters. To compare CDR3 clonotypes across tumor samples, we quantified clonal abundance and diversity from the TRUST4 output table. To elevate the accuracy and get rid of ambiguous results, we filtered samples out, below 5 diverse TCR clonotypes (responder: 5 patients, nonresponder: 8 patients in pretreatment data). To calculate the TCR clonality, we adopt an equation of clonality score derived from Shannon entropy clonality (1-normalized entropy; ref. 73).

(X_i, the frequency of TCR sequence; n, the number or unique TCR sequences in the repertoire)

scRNA-seq reads were aligned to the GRCh38 human genome reference and quantified using a Cellranger (version 3.1.0). Further analyses were conducted using Seurat (version 3.1.4). Putative doublets were predicted using Scrublet (74) and filtered out with poor-quality cells that had less than 200 detected genes. Afterwards, cells with a mitochondrial genome content greater than 15% were further excluded from the analysis. Raw feature counts were then log-normalized, scaled, and subjected to linear dimensional reduction using principal component analysis (PCA). We applied the Harmony algorithm (75) using PCs embedded into a space with reduced dimensionality to adjust for different batch-derived potential biases across data sets. Following the application of the UMAP algorithm to visualize cells in a two-dimensional space (76), we then identified clusters of cells using the “FindClusters” function from Seurat, which applied a shared nearest neighbor modularity optimization-based clustering algorithm. Different cell-type clusters were identified by performing differentially expressed gene analysis for each cluster and annotated on the basis of the expression of representative markers.

PBMCs were isolated by Ficoll (GE Healthcare, 17–5442–02). After processing, PBMCs were resuspended in freezing media (Recovery cell culture freezing medium, Gibco) and stored in liquid nitrogen. For flow cytometry analysis, 1 × 10⁶ cells were thawed and stained using the Live/Dead Fixable Cell Stain Kit (Invitrogen: L34975) to exclude dead cells from the analysis. After being washed with FACS staining buffer, the cells were stained with fluorochrome-conjugated antibodies for 20 minutes at room temperature. For intracellular staining, surface-stained cells were fixed and permeabilized using a Foxp3 Staining Buffer Kit (eBioscience: 00–5523–00) according to the manufacturer's instructions. As the therapeutic binding of pembrolizumab (humanized IgG4) to PD-1⁺ cells interfered with the PD-1 staining of posttreatment specimens, anti-human IgG4 Fc staining in addition to anti–PD-1 staining was used to define PD-1⁺ cells (Supplementary Fig. S5B). The fluorochrome-conjugated mAbs used in multi-color flow cytometry are listed in Supplementary Table S4. Multi-color flow cytometry was performed using a Northern Lights flow cytometer (Cytek), and the data were analyzed using FlowJo V10.6 software (Treestar).

The primary endpoint of the trial was the ORR per RECIST 1.1. The secondary endpoints included RR, DCR, PFS, overall survival (OS), safety profile, and exploratory biomarker analysis. PFS was defined as the time from the start of treatment until the date of disease progression or death resulting from any cause. OS was measured from the start of treatment to the date of death from any cause. The RR was calculated as the percentage of patients experiencing a confirmed CR or PR, and DCR was calculated as RR + stable disease (SD) per the RECIST 1.1 guidelines. All statistical analyses were performed using R (higher than 3.3.3), Matlab R2010b, or Prism software version 8.4 (GraphPad). Statistical associations between continuous variables were evaluated using Spearman correlations, and associations between continuous and categorical variables were evaluated using rank sum statistics. The nonparametric Mann–Whitney U test was used to compare the two groups. Paired values were compared using the nonparametric Wilcoxon matched-pairs signed rank test.

All raw sequencing data were deposited in the European Nucleotide Archive (ENA; accession number: PRJEB40416).

S.J. Klempner reports personal fees from Merck, BMS, Eli Lilly, Daiichi-Sankyo, Astellas, Natera, and Pieris, and other support from Turning Point Therapeutics outside the submitted work. W. Park reports other support from Geninus Inc. outside the submitted work. J. Lee reports grants from Merck during the conduct of the study, grants from Mirati, and grants from Oncologie outside the submitted work. No disclosures were reported by the other authors.

M. Kwon: Conceptualization, resources, formal analysis, funding acquisition, investigation, methodology, writing–original draft, writing–review and editing. M. An: Conceptualization, resources, data curation, validation, investigation, methodology, writing–original draft, writing–review and editing. S.J. Klempner: Conceptualization, data curation, formal analysis, funding acquisition, validation, investigation, methodology, writing–original draft, writing–review and editing. H. Lee: Conceptualization, resources, formal analysis, investigation, methodology, writing–review and editing. K. Kim: Resources, formal analysis, supervision, investigation, methodology, writing–review and editing. J.K. Sa: Data curation, formal analysis, methodology. H. Cho: Data curation, formal analysis, methodology, writing–review and editing. J. Hong: Investigation, writing–review and editing. T. Lee: Data curation, formal analysis, methodology, writing–review and editing. Y. Min: Resources, investigation, methodology, writing–review and editing. T. Kim: Resources, investigation, methodology, writing–review and editing. B. Min: Resources, investigation, writing–review and editing. W. Park: Resources, data curation, formal analysis, methodology, writing–review and editing. W. Kang: Resources, formal analysis, investigation, methodology, writing–review and editing. K. Kim: Conceptualization, resources, formal analysis, investigation, writing–review and editing. S. Kim:Conceptualization, formal analysis, supervision, investigation, writing–original draft, writing–review and editing. J. Lee: Conceptualization, resources, formal analysis, supervision, funding acquisition, investigation, methodology, writing–original draft, writing–review and editing.

This work was supported by AGA Research Foundation's AGA-Gastric Cancer Foundation Ben Feinstein Memorial Research Scholar Award in Gastric Cancer – AGA2020–13–02 (to S.J. Klempner), and Stand Up To Cancer Gastric Cancer Interception Award (to S.J. Klempner, H. Lee, J. Lee). This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: HR20C0025; to S.T. Klempner, K. Kim). Pembrolizumab was donated through Merck MISP grant. This article was supported by SKKU Excellence in Research Award Research Fund, Sungkyungkwan University, 2020 (to J. Lee).