In the first cohort, data of 73 GI cancer patients were retrospectively collected from all GI cancer patients treated with ICBs in the Department of GI Oncology, Peking University Cancer Hospital and Institute from August 1, 2015, to June 8, 2018. To explore genomic correlates of therapeutic efficacy, the 73 GI cancer patients were randomly assigned into the discovery (n=44) and validation (n=29) GI cohorts. Briefly, formalin-fixed paraffin-embedded (FFPE) samples from the 73 patients were subjected to whole-exome sequencing (WES) analysis. To characterize the different immunological features between CNA low and high GI patients, FFPE specimens from a subset of these GI patients (n=65) were subjected to an RNA IO panel sequencing (online supplementary materials).

In addition, 20 patients with advanced GC who were treated with ICBs between January 31, 2018 and May 24, 2019 in the Department of GI Oncology, Peking University Cancer Hospital and Institute were included in this study as an independent GC validation cohort (online supplementary table S1). Aiming to validate the predictive value of CNA burden, we performed WES analysis on these 20 GC samples.

Tumor burden was measured by imaging studies or physical examinations according to the Response Evaluation Criteria in Solid Tumors (RECIST) V.1.1 and iRECIST. Patients were stratified by clinical response. Briefly, the durable clinical benefit (DCB) group was defined as complete response, PR, and SD lasting for ≥24 weeks. No durable benefit (NDB) included patients with PD or SD that lasted <24 weeks.16

For the determination of TMB value, the number of somatic non-synonymous single nucleotide variants (SNVs) detected using next-generation sequencing was quantified, and the values were extrapolated to the WES analysis using a validated algorithm.17 TMB was measured in mutations per Mb.

For copy-number analysis, blood cell samples from patients were used as paired controls, and the CONTRA assay was used to call copy-number variations from the FFPE tumor samples for each patient.18 CNA burden analysis included measurements of the total CNA burden, CNgain, and CNloss. CNgain/CNloss was defined as the total number of genes with CNgain/CNloss per sample, as previously described.8 The CNA burden was calculated as the total number of genes with CNgains/CNlosses.15

RNA IO panel sequencing was conducted as previously described.19 Briefly, 10 ng of RNA that was extracted from the FFPE sample was reverse transcribed into cDNA, amplified with a primer pool, and ligated to unique barcode adapters. After purification, the libraries were quantified, pooled, and sequenced on the Ion S5 530 chip (Thermo Fisher Scientific, Waltham, MA). The detailed process of gene expression normalization and measurement of normalized reads per million (nRPM) values are described in the online supplementary materials.

CNA and gene expression data of stomach adenocarcinoma (STAD) and colon adenocarcinoma (COAD) types analyzed in TCGA project were obtained from the cBioPortal (http://www.cbioportal.org). The CNA burden was calculated as the total number of genes with CNgains or CNlosses.15 Gene expression levels of the IFN-γ signature and expanded immune signature (gene list was provided in the online supplementary materials) were calculated by using the RNA-Seq data. To quantify the relative infiltration of immune cell types into the tumor microenvironment (TME), single-sample gene set enrichment analysis (ssGSEA) was employed as previously reported.20–22

Statistical analyses were performed using SPSS V.22.0 software. Statistical tests included Fisher’s exact tests, Mann-Whitney U tests, and Student’s t-test. All statistical tests were two-sided.

This study included 93 metastatic GI patients treated with ICBs between August 1, 2015 and June 8, 2018 (n=73, comprising discovery and first validation GI cohorts, median follow-up time, 11.97 months, IQR: 6.03–17.3) and between January 31, 2018 and May 24, 2019 (n=20, second validation GC cohort, median follow-up time, 4.87 months, IQR: 2.83–10.47). To comprehensively explore the genomic correlates with immunotherapy efficacy, 73 GI patients were randomly assigned into discovery (n=44) and validation (n=29) cohorts. The baseline and treatment characteristics of all patients are shown in online supplementary table S1. Briefly, of 73 included GI patients, the cancer types were GC (50.7%), CRC (27.4%), and others types, including pancreatic neuroendocrine tumors, GI-NETs, and cholangiocellular carcinoma. In GI combined cohort, 47 patients (64.4%) were treated with anti-PD-1 therapy, 15 cases (20.5%) were treated with anti-PD-L1 therapy, and 11 cases (15.1%) received ICB combinational therapy. In the second validation GC cohort (n=20), the cancer types were GC (100%). Among the GC cases, 15 (75%) patients were treated with anti-PD-1 therapy, 3 cases (15%) were treated with anti-PD-L1 therapy, and 2 cases (10%) were treated with ICB combinational therapy (online supplementary table S1).

Tumor somatic copy-number changes have recently been associated with the clinical response to ICB in melanoma.8 11 We therefore investigated the predictive value of the three indices: CNA burden, CNgain, and CNloss in the GI cancer cohort (n=73), which was randomly assigned into the discovery and validation cohorts. First, we observed no remarkable differences in the number of copy-number changes among different GI cancer types (online supplementary figure S1 A-C).

Next, we identified that the CNA burden index revealed significantly decreased levels in DCB versus NDB patients receiving immunotherapy in the GI cohort, with AUCs of 0.712 (discovery, p=0.021), 0.864 (validation, p=0.001), and 0.775 (the combined GI cancer patients, p<0.0001) (figure 1A,B). Accordingly, a trend toward a lower burden of CNgain or CNloss was also observed in DCB group compared with NDB group with an AUC value of 0.728 (all GI cancer patients, p=0.001) and 0.757 (all GI cancer patients, p=0.0003) for CNgain and CNloss, respectively (online supplementary figure S2 A-B). Notably, the change in CNgain between DCB and NDB patients did not attain statistical significance in the discovery cohort (online supplementary figure S2 A).

Moreover, efficacy was further compared between groups with high (CNA>10) and low (CNA≤10) copy-number variation. Specifically, a drastically improved DCB rate was observed in CNA-low (69%) GI cancers compared with CNA-high (15.9%) cases (online supplementary table S2, p<0.001). A similar trend of responsiveness was identified in the CNgain-low (57.1%) and CNloss-low (57.5%) groups when compared with the CNgain-high (18.4%) and CNloss-high (12.1%) cases, respectively (p<0.01 for all comparisons, online supplementary table S2). Taken together, our findings strongly indicate that features for tumor genome alterations, including total CNA, CNgain, and CNloss, are potential predictive biomarkers for GI cancers with ICB treatment. Notably, the CNA burden demonstrated the highest AUC value, and therefore, the utility of the CNA burden was further evaluated.

Since MSI/MMR and PD-L1 expression have emerged as potential predictive biomarkers for PD-1/PD-L1 blockade, we examined the correlation between MSI/MMR or PD-L1 and the clinical benefit in GI cancer patients. Our data show that MSI-H/dMMR patients experienced a significantly higher DCB rate (59.1%) than MSI-L/MSS/pMMR patients (28.6%, p=0.022), while no association was found between PD-L1 positivity and the efficacy of ICB therapy in our cohort (online supplementary table S2).

As shown in online supplementary figure S1D, higher TMB was identified in CRC (median 38.7 mutations/Mb) than other cancer types (median 3.5 mutations/Mb). The median number of TMB in DCB patients was significantly higher than that of NDB patients, with AUCs of 0.696 (p=0.032) and 0.702 (p=0.004) in the discovery and combined cohorts, respectively (figure 2A,B). On the other hand, we did not observe a statistically significant AUC value (0.697) of TMB in the validation cohort (p=0.080) (figure 2B). In all GI cancer patients, 60% (18/30) of the patients with a higher mutation burden (designated as above 5, including intermediate and high TMB levels23) experienced a DCB compared with 20.9% (9/43) of those with a lower TMB (online supplementary table S2, p=0.001).

To assess whether the above-tested signatures, including CNA, TMB, MSI/MMR and PD-L1 status, influence the OS and progression-free survival (PFS), we used Kaplan-Meier analysis for the GI cancer cohort (figures 3 and 4 and online supplementary figure S3-4). As expected, CNA-low patients demonstrated a longer median OS (unreached for all three cohorts) than the CNA-high group (discovery, 7.63 months; validation, 5.6 months; and combined cohorts, 7.43 months) (figure 3A, log-rank test, p<0.05 for all comparisons). A similar trend was identified between CNgain/CNloss and OS in GI cancer patients, although statistical significance was not attained for CNgain in the validation cohort (online supplementary figure S3). In addition, patients with higher TMB levels revealed longer OS than patients with lower TMB in the validation and combined cohorts (figure 3B). In our cohort, neither MSI/MMR nor PD-L1 status was able to predict the OS outcome after ICB (figure 3C,D).

Consistently, longer PFS was identified in the lower burden groups for CNA, CNgain or CNloss than in the higher burden groups (figure 4A and online supplementary figure S4A–B, log-rank test, p<0.05 for all comparisons, except for CNgain in the validation cohort). In contrast, higher TMB levels were able to stratify patients with favorable PFS in all three cohorts (figure 4B). Moreover, the median PFS time of MSI-H/dMMR patients (7.24 months) was significantly longer than that of MSI-L/MSS/pMMR patients (2.67 months, p=0.0173) (figure 4C). However, no significant association was identified between PD-L1 status and PFS in the combined cohort (figure 4D).

Due to the predictive and prognostic value of CNA and TMB, we sought to explore the possibility of combining the two parameters in identifying DCB patients receiving immunotherapy. As revealed in figure 5A, no significant correlation was observed between CNA burden and TMB, suggesting no redundant feature between the two genomic determinants. All 73 patients can be classified into four subgroups based on TMB (high or low) and burden of CNA (high or low), and then the proportion of DCB patients was calculated in each subgroup (figure 5B–E). Notably, the proportion of DCB patients was remarkably higher in the TMB-high/CNA-low subgroup of patients (12 of 14) compared with the TMB-low/CNA-high subgroup (1 of 28) (figure 5C, Fisher’s exact test, p<0.05). Moreover, the median OS time of the TMB-high/CNA-low subgroup (unreached) was significantly improved compared with the other three subgroups (TMB-low/CNA-low, 17.3 months; TMB-high/CNA-high, 12.37 months; TMB-low/CNA-high, 6.23 months) (figure 5D, log-rank test p<0.05 for all comparisons).

Moreover, in order to validate the predictive value of CNA burden and its joint utility with TMB, an independent validation was performed using a cohort of 20 GC patients. Consistently, an increased DCB rate was observed in CNA-low (60%) GC patients compared with CNA-high (0%) cases (online supplementary figure S5A, Fisher’s exact test, p<0.01). However, there was no significant difference in DCB rates between TMB-high and TMB-low subgroups (online supplementary figure S5B). Intriguingly, a high proportion of DCB rate was also identified in the TMB-high/CNA-low subgroup of patients (five of six), as presented in online supplementary figure S5C-D. A larger sample size may help improve the performance of the predictive model.

Recent analyses have linked cancer genomic features, including TMB and CNA, with antitumor immunity. In particular, studies have proposed that high mutational load and low aneuploidy may correlate with increased T-cell responses.8 11 15 We therefore sought to determine the association between CNA burden and the inflammatory characteristics of the TME by using an RNA IO sequencing platform. IFN-γ–related (6-gene set) and expanded immune signature (18-gene set) mRNA profiles have recently been developed as pan-cancer biomarkers to predict the immune response to ICB in melanoma, head and neck squamous cell carcinoma and GC cohorts.24 Herein, we measured the IFN-γ and expanded immune signatures in 65 patient samples with an available RNA data set. Briefly, the CNA burden was partitioned into high or low at the upper quartile or lower quartile value for the 65 GI patients. As figure 6A demonstrates, the CNA-low subgroup revealed higher IFN-γ and expanded immune signatures than the CNA-high subgroup (p<0.05). Moreover, we observed 26 upregulated genes in the CNA-low samples versus the CNA-high group (figure 6B). These upregulated immune-related genes belong to the following pathways: lymphocyte markers, interferon signaling, lymphocyte regulation, and checkpoint pathways.

To confirm the correlation between tumor CNA burden and immune signature more broadly, we examined the IFN-γ and expanded immune signatures in the TCGA, STAD, and COAD cohorts. For each cohort, the burden of CNA was measured and then partitioned into high or low at the upper quartile or lower quartile value. Consistently, CNA-low samples showed higher immune signatures in both cohorts (figure 6C). Further ssGSEA analysis revealed that CNA-low GC and CRC samples were infiltrated with diverse immune cell types, including activated CD8+ T cells, activated CD4+ T cells, natural killer (NK) cells, and NK T cells (online supplementary figure S6A-B). Collectively, our data indicate that a lower CNA burden may correlate with an activated inflammatory response in the TME.