Patients with advanced NSCLC who were treated with PD-1 ICIs between June 2013 and June 2017 at the Erasmus University MC (Rotterdam, The Netherlands) and the Amphia Hospital (Breda, The Netherlands) were prospectively included. Patients with NSCLC stage IV who had been treated with nivolumab monotherapy (3 mg/kg intravenous, once every 2 weeks) were included, and patients who were treated with a prior line of immunotherapy were excluded from the study. The study was approved by the independent ethics committee board (MEC 02-1002 and MEC 16-011) and all patients provided written informed consent. Whole blood and baseline serum were available for DNA analysis and protein quantification, respectively. All assays were performed blinded to study endpoints; patients were assigned to a subject number. Patient characteristics included demographic and clinical data. BOR was assessed according to the Response Evaluation Criteria in Solid Tumors V.1.1 (RECIST V.1.1).14 A minimum duration of 90 days for stable disease (SD) was required. Confirmation of partial response (PR) or complete response (CR) was not required. After treatment initiation, radiological evaluation by CT was usually performed every 6 weeks. PFS was defined as the time from the first administration of nivolumab until progressive disease (PD), or death due to any cause, whichever occurred first. OS was defined as the time from the first administration of nivolumab until death due to any cause.

Quantitative analysis of granzyme B in serum was performed by a magnetic bead-based assay, using a 1×96-well microplate and magnetic antigranzyme-B-coated beads (Human Magnetic Luminex Assay, R&D Systems Inc, Minneapolis, Minnesota, USA). GZMB c.128T>C (rs8192917) was selected primarily based on its presumed contribution to resistance mechanisms of ICIs.9 A total of 12 single-nucleotide polymorphisms (SNPs) were selected in eight genes based on their contribution to T cell immunity and their correlation with overt T cell responses, such as the autoimmune diseases systemic lupus erythematosus and rheumatoid arthritis.15 16–25 A summary of the correlation of selected SNPs with overt T cell responses is displayed in online supplementary table 1. SNPs with a minor allele frequency (MAF) of <5% were excluded from the analysis. Linkage disequilibrium was tested using the application LDmatrix of LDlink V.5 (https://ldlink.nci.nih.gov/). Distribution of genotypes was tested for Hardy-Weinberg equilibrium (HWE) using the χ² test. Deviations from HWE may reflect genotyping errors, though, may also be a signal of disease association,26 for example, mutation of tumor suppressor genes. DNA isolation and genotyping for GZMB rs8192917, HLA-A rs60131261, IL10 rs3024493, IL2RA rs3024493, IFNG rs2430561/rs2069705/rs2069718, PDCD1 rs2227981/rs10204525/rs2227982, PTPN11 rs2301756 and ZAP70 rs13420683 has been performed using TaqMan analysis as previously described.27 Whole-blood specimens were used for extraction of DNA. Genotyping was performed using predesigned TaqMan allelic discrimination assays, consisting of two allele-specific minor groove-binding probes. PCRs were performed in a reaction volume containing DNA, primers, and the specific probes were labeled with fluorescent dye.

Peripheral blood samples were analyzed using multiplex flow cytometry as previously described by Kunert et al.6 In short, absolute cell counts of immune cell populations were determined on whole blood after lysis of red blood cells, and frequencies of particular subsets were determined in more detail on cryopreserved peripheral blood mononuclear cell samples that were stained with a mix of antibodies.

Expression of programmed death-ligand 1 (PD-L1) and tumor mutational burden (TMB) have previously been determined in a subset of patients of cohort 1 (n=26; patients with stage IV NSCLC) using archival formalin-fixed, paraffin-embedded tissue samples.28 Baseline expression of PD-L1 has been determined using antibody clone SP263 (Roche Diagnostics, Tucson, Arizona, USA), whereas TMB has been measured using the Oncomine TML assay (Thermofisher Scientific, Waltham, Massachusetts, USA).

To study the relationship between baseline protein levels or SNPs and survival (PFS and OS), Cox regression analysis was used. Patient survival was visualized by the Kaplan-Meier approach. Patients were dichotomized for low or high granzyme B serum levels based on the median value of all patients. Dominant, recessive, and additive models29 were used to test associations between SNPs and clinical outcomes including BOR, PFS, and OS. HRs and 95% CIs were calculated. For the association of SNPs with radiological response, BOR was used as an ordered categorical variable with three levels, namely CR/PR (pooled), SD, and PD, by means of ordinal logistic regression analysis. This method assumes that the relationship between each pair of outcome groups is the same. The proportional odds assumption was checked for all analyses. Univariate ordinal logistic regression was applied for both SNPs and patient factors, and OR and 95% CIs were calculated. Multivariable ordinal logistic analysis was performed including both single SNPs and patient factors, with p<0.1 according to the univariate analyses. All analyses, with a significant relation (p<0.05) between SNP and clinical outcome, were internally validated by bootstrapping,30 where 1000 bootstrap samples were generated (with replacement) and 95% CIs were calculated for either ORs or HRs. Furthermore, bias and bias-corrected 95% CIs were calculated. In general, a p value <0.05 was considered to be statistically significant. Analyses were performed using IBM SPSS Statistics V.24.0.0.1 (Chicago, Illinois, USA) and STATA V.15.1 (StataCorp LP, College Station, Texas, USA).

Baseline serum from a total of 78 (cohort 1) and DNA from a total of 322 patients with stage IV NSCLC (cohort 2) was available for analysis (table 1). There was a 16.5% subject overlap between the patient cohorts. All patients were treated with at least one dose of nivolumab at 3 mg/kg once every 2 weeks. Overall, median absolute starting dose was 222 mg (IQR 192–263 mg). The objective response rate was similar in cohort 1 and cohort 2, respectively 19.2% and 16.6%. A minority (7%) of all patients could not be evaluated for response by RECIST V.1.1 due to early death, loss to follow-up or incomplete or absent follow-up tumor assessments. In the univariate analysis of cohort 2, WHO performance status (0 vs ≥1) and sex (male vs female) were significantly associated with PFS (HR: 0.66; 95% CI: 0.46 to 0.95; p=0.027 and HR: 1.38; 95% CI: 1.06 to 1.80; p=0.18, respectively) and OS (HR: 0.52; 95% CI: 0.34 to 0.77; p=0.001 and HR: 1.40; 95% CI: 1.05 to 1.86; p=0.023, respectively). In the multivariable analysis, only WHO performance status and primary tumor type remained significant for their association with, respectively PFS or OS and BOR.

Serum levels of granzyme B were determined in cohort 1 (online supplementary figure 1), having a median concentration of 1.27 pg/mL, which was used as the cut-off to dichotomize patients with low and high levels of serum granzyme B (as described in the Method section). Patients with low baseline serum levels of granzyme B had a significant worse PFS (HR: 1.96; 95% CI: 1.12 to 3.43; p=0.018, figure 1A) than patients with high levels of baseline serum granzyme B. Likewise, low levels of granzyme B were significantly associated with worse OS (HR: 2.08; 95% CI: 1.12 to 3.87; p=0.021, figure 1B) compared with high levels of granzyme B.

To validate the findings in a second cohort by a different approach,31 genetic alterations of GZMB c.128T>C (rs8192917), together with a mechanism-of-action-based panel of other SNPs, were related to PFS, OS and BOR in cohort 2. Investigated SNPs with the corresponding haplotype frequencies, MAF and HWE statistics are shown in online supplementary table 2. After correction for WHO performance status and sex in multivariable analysis, patients with homozygous and heterozygous variants of GZMB rs8192917 had worse PFS (HR: 1.38; 95% CI: 1.02 to 1.87; p=0.036; table 2, figure 1C). This finding could be internally validated (HR 95% CI: 1.02 to 1.87; bias: 0.003; bias-corrected 95% CI: 1.03 to 1.88; p=0.036; online supplementary table 3). Remaining SNPs were not associated with PFS after correction for patient factors (table 3). No significant association of SNPs with OS was observed after correction of patient factors (table 3). Of note, homozygous and heterozygous variants of GZMB rs8192917 were not significantly associated with OS (HR: 1.19; 95% CI: 0.91 to 1.57; p=0.212; table 2, figure 1D). Homozygous and heterozygous variants of GZMB rs8192917 were significantly associated with worse BOR (OR: 1.60; 95% CI: 1.01 to 2.52; p=0.044; table 2) and could be internally validated for the association with BOR (OR 95% CI: 1.05 to 2.55, bias: 0.007; bias-corrected 95% CI: 1.09 to 2.57; p=0.030, online supplementary table 3). All the results from both univariate (p<0.1) and multivariate analyses of the association with PFS, OS or BOR are shown in table 3. Results from the internal validation of factors that were significantly associated after correction for patient factors (p value <0.05) are shown in online supplementary table 3. For a subset of patients, both serum levels of granzyme B and germline variation of GZMB c.128T>C (rs8192917) were available for analysis. Patients with homozygous and heterozygous variants of GZMB rs8192917 (n=19) had significant lower serum levels of granzyme B compared with patients with wild-type GZMB (n=34; mean difference 8.6 pg/mL, 95% CI 1.77 to 15.38, p=0.015).

Next, we assessed the relationship between serum granzyme B as well as genetic variations of GZMB and cytotoxic immune cell populations in blood. Overall, there was a tendency that patients with lower serum granzyme B levels had lower absolute numbers of peripheral CD56⁺ NK cells (mean 141 vs 205 cells/µL, p=0.16; online supplementary figure 2) and CD8⁺ T cells (346 vs 463 cells/µL, p=0.41; figure 2A) when compared with patients with higher serum granzyme B levels. In fact, patients with homozygous or heterozygous variants of GZMB rs8192917 had a higher number of CD8⁺ T cells but not CD56⁺ NK cells than patients with wild-type GZMB (figure 2B and online supplementary figure 2B). Further, we studied whether granzyme B levels or GZMB genotypes correlate with frequencies of those immune cell phenotypes that correspond to either terminal differentiation (CCR7^-CD45RA⁺) or exhaustion (PD1⁺ and TIM3⁺) of CD8⁺ T cells, as signs of T cell cytotoxicity. Interestingly, patients with lower serum granzyme B levels had a lower proportion of CCR7⁻CD45RA⁺CD8⁺ T cells (mean 44 vs 54 %, p=0.040) as well as PD1⁺TIM3⁺CD8⁺ T cells (mean 4.5 vs 7.8 %, p=0.001; figure 2A). No significant association nor trend could be observed between GZMB genotype and frequencies of T cells with these phenotypes (figure 2B).

Interaction analysis for PFS and OS was performed between serum granzyme B, and PD-L1 and TMB in a subset of evaluable patients. In the univariate analysis, TMB (HR: 7.31; 95% CI: 1.5 to 34.7; p=0.01) was significantly correlated with PFS, while expression of PD-L1 (HR: 4.30; 95% CI: 0.9 to 19.8; p=0.06) or serum granzyme B (HR: 2.34; 95% CI: 0.9 to 5.9; p=0.07) did not meet the threshold for significance (online supplementary table 4). Moreover, only serum granzyme B was significantly correlated with OS (HR: 2.95; 95% CI: 1.09 to 7.93; p=0.03; online supplementary table 5). Multivariate analysis of these biomarkers (namely, serum granzyme B, PD-L1 and TMB) for PFS and OS did not meet significance. Of note, the HRs were in favor of higher levels of PD-L1 (≥50%), TMB (≥10 mut/Mb) and serum granzyme B (≥1.27 pg/mL), yet expression of PD-L1 (%), TMB (mut/Mb) and serum granzyme B levels (pg/mL) were not significantly correlated with each other (online supplementary table 6).