Study participants were enrolled in a clinical study of lung cancer that has been ongoing since 1991 at The University of Texas MD Anderson Cancer Center. The recruitment method was described previously.9 Briefly, the subjects were incident cases of lung cancer diagnosed and histologically confirmed at MD Anderson between 1995 and 2013. The schematic of study design involving discovery and validation sets for 941 early-stage NSCLC patients (discovery set: n=536, validation set: n=405) as well as bioinformatic and functional analyzes are shown in online supplementary figure S1 and table 1. Subjects in the discovery and validation sets were recruited for a genome-wide association study (GWAS) of lung cancer and the OncoArray study, respectively. Clinical data were abstracted from chart review, and epidemiologic data were collected from each participant during an in-person interview. The peripheral blood was collected from the antecubital area of arm after the interview. Participants were considered never-smokers if they had smoked less than 100 cigarettes in a lifetime. Former smokers were those who had quit smoking more than 1 year before lung cancer diagnosis. Current smokers were those who were currently smoking or had quit smoking within 1 year from the date of lung cancer diagnosis (cases). To avoid confounding by race/ethnicity and to minimize heterogeneity of participants, this study was restricted to non-Hispanic white patients with stage I or II NSCLC who were treated at MD Anderson Cancer Center.

NSCLC (lung adenocarcinoma (LUAD) and LUSC (lung squamous cell carcinoma)) datasets from The Cancer Genome Atlas (TCGA) database were accessed for the analysis of gene expression in tumors.10 11 Clinical data and level-3 RNA-seq datasets were downloaded from http://firebrowse.org/. The RNA-seq data of 1016 NSCLC patients (512 cases from LUAD dataset and 504 cases from LUSC dataset) were included in the analysis, and 110 patients among them had corresponding normal tissue data available.

Genomic DNA was isolated from peripheral blood samples using the QIAamp DNA Blood Maxi Kit according to the manufacturer’s instructions. Discovery cohort genotypes were generated using the HumanHap300 BeadChip (Illumina, San Diego, California, USA) for the 421 patients included in our previous GWAS of lung cancer,12 and the HumanHap660 BeadChip (Illumina) for additional 115 patients in this study. The analysis focused on 307 260 SNPs that were included and passed quality control filters, including a call rate of at least 95% and minor allele frequency of at least 0.01. Genotyping for the validation cohort (405 patients) was performed using the Custom Infinium OncoArray-500K beadchip according to the manufacturer’s instructions. The assay was run on the iScan system (Illumina). Genotyping data were analyzed and exported using the Genome Studio software program (Illumina). Due to variation in SNP coverage between the GWAS and OncoArray Beadchips, we identified linked SNPs to replace missing loci from the discovery phase using data from the 1000 Genome Project (European population, Phase 3 data, r² >0.8; http://www.internationalgenome.org/data). All genotyping data were analyzed and exported using the Genome Studio software program (Illumina). For quality control, we included 3% of the samples as replicates, and the call rate and concordance of the different beadchips were similar at 99% or greater. Finally, 412 487 SNPs that passed quality control filter using the same criteria as GWAS were included in the analysis.

On the basis of the stepwise cancer immune response of T cells,4 we generated T cell cancer immune response pathways and genes by extensively searching the keywords of each step (antigen presentation, T cell priming/activation, T cell trafficking, T cell infiltration, T cell recognition and T cell cytotoxicity) in the Kyoto Encyclopedia of Genes and Genomes (KEGG, http://www.genome.jp/kegg/), Biocarta (https://cgap.nci.nih.gov/Pathways/BioCarta_Pathways) and Reactome (http://www.reactome.org/) databases. Furthermore, previously published related studies and gene lists from commercially customized gene panels (Nanostring nCounter PanCancer Immune Profiling Panel and HTG EdgeSeq Immuno-Oncology Assay) were referenced.4 11 Genes involved in at least two pathways or mentioned in two databases were selected for the final gene list. For each selected gene, tagging SNPs from 10 kb flanking and within gene regions were included. A total of 314 T cell cancer immune response genes from 25 pathways were selected (online supplementary table S1), and corresponding genotyping data were extracted from the GWAS and OncoArray data. After quality control, 280 genes were considered resulting in the selection of 2450 SNPs.

For the in vitro T cell assays, 19 pairs of healthy donors matched by age, sex and smoking status from our lung cancer GWAS and OncoArray studies, whose genotypes corresponded to low and high risk, were selected. To minimize the impact of lung cancer on T cell phenotypes, only healthy controls were analyzed. Low-risk and high-risk groups were defined by the number of unfavorable genotypes (UFGs) of expression Quantitative Trait Loci (eQTL)-significant SNPs with the low-risk group having 0–1 UFG and high risk group having 3–4 UFGs (online supplementary table S2).

Peripheral blood mononuclear cells (PBMCs) of healthy donors were isolated from 5 to 20 mL of whole blood, counted and resuspend into freezing medium (90% FBS and 10% dimethyl sulfoxide (Sigma-Aldrich, St. Louis, Missouri, USA)) at the density of 5×10E6/mL. The cells were placed in Cryo-SafeTM Cooler (Bel-Art, Wayne NJ) in −80°C overnight and transferred to a liquid nitrogen freezer until use. Cryopreserved PBMCs (5×10E6/mL) were thawed in a 37°C water bath without shaking. The cells were cultured overnight according to previous literature.13

Rested PBMCs were diluted to a density of 2×10E6 cell/mL in medium containing 20% FBS and 50 IU/mL interleukin-2 (IL-2) (BioLegend, San Diego, California, USA). One hundred and fifty microliter of cell suspension was transferred into 96-well round-bottom plate. Cells were stimulated with 4 µg/mL OKT3 (BioLegend) and monensin (MEDICAL &BIOLOGICAL LABORATORIES, MBL, Japan) then stained with CD107a antibody (MBL) for 5 hours at 37°C in the dark. After stimulation, cells were washed twice with Fluorescence-activated Cell Sorting (FACS) buffer and stained with PE-CD8 antibody (BioLegend) for 15 min at 4°C in the dark. The cells were analyzed using a BD LSRFortessa X-20 analyzer (BD Bioscience, San Jose, California, USA). The gating strategy and representative data are shown in online supplementary figure S2. Gating strategy consisted on isolating single cells by gating on FSC-A vs FSC-H, followed by SSC-A vs SSC-H. Dead cells were excluded by Sytox Blue. Degranulated CD8 T cells were identified as a per cent of CD107a+CD8+T cell population. Data were analyzed using FlowJo (V.10.0.8) software.

To assess the cytotoxic potential of T cells, we performed an in vitro killing assay using NSCLC A549 and H460 cell lines as target cells. The entire experimental workflow is depicted in online supplementary figure S3. Briefly, rested PBMCs were cultured and expanded with Roswell Park Memorial Institute (RPMI) 1640 medium enriched with 2 mM L-glutamine and supplemented with 20% FBS and 8 µg/mL phytohemagglutinin (Remel, Thermo Fisher, Waltham, Massachusetts, USA) for 48 hours. The viable CD3 + lymphocytes were labeled with an anti-CD3 antibody (BioLegend), counterstained with Sytox blue (Thermo Fisher), and sorted using a BD FACS Aria II cell sorter (BD Bioscience) (gating strategy in online supplementary figure S4A). A549 and H460 were purchased from ATCC (Manassas, Virginia, USA) and cultured with F­12K medium and RPMI­1640 medium supplemented with 10% FBS, respectively. The two cell lines have been verified by short tandem repeat profiling of 14 known loci and tested for Mycoplasma contamination showing negative result by the MD Anderson Characterized Cell Line Core facility on January 23 2018. The HLA genotypes of these two target cells were identified previously.14 A549 and H460 cells were plated in a 96-well plate (Greiner, Austria) the day before the assay. The target cells were labeled with calcein AM (green fluorescence) and ethidium homodimer-1 (red fluorescence) using LIVE/DEAD Viability/Cytotoxicity kit (Thermo Fisher). Meanwhile, purified T cells were stimulated with 4 µg/mL OKT3 (BioLegend) for 5 hours. To distinguish from target cells, T cells were labeled with anti-CD3 fluorescence antibody (Alexa Fluor 647) and added to the tumor cells at a ratio of 10:1 and incubated in an IN Cell Analyzer 2200 (GE, USA). The interactions between T cells and tumor cells were scanned every 30 min under fluorescence microscopy using the IN Cell Analyzer 2200. All images were further processed with the Developer Tool Box 1.9.2 (GE, USA); all dead cell signals were counted. T cell cytotoxicity was calculated from triplicate samples as (experimental dead cell count – negative control dead cell count (spontaneous)) / (positive control dead cell count (maximal) – negative control dead cell count (spontaneous)) and expressed as a percentage. The positive control was assessed by complete lysis of target cells using 70% ethanol. The negative control was assessed by adding only medium to the target cells. Wilcoxon signed rank test were applied in the comparison of UFG and Non-UFG carriers. The inhibitory rate by effector:target cell ratio over time is shown in online supplementary figure S4B.

RNA was extracted from purified viable CD3 + T cells before and 6 hours after co-culture with the target tumor cells using Trizol Reagent (Thermo Fisher) according to the manufacturer’s instructions. RNA quantity and quality were determined using a NanoDrop 1000 Spectrophotometer (Thermo Fisher). Reverse transcription was performed using the High-Capacity cDNA Reverse Transcription Kit (Thermo Fisher) according to the manufacturer’s instructions. The T cell activation-related genes IL2, IFNG, PRF1, GZMB and TNFA; T regulatory cell genes FOXP3 and IL4 (data not shown due to undetected expression); T cell trafficking gene EOMES; T cell checkpoint genes HAVCR2, PDCD1, CTLA4, LAG3, CD137, VISTA, IDO1 and ICOS; 1and lineage markers CD4 and CD8 were selected and incorporated into the panel. The gene expressions were determined using TaqMan probes (Applied Biosystems, Thermo Fisher) and a 48.48 Dynamic Array (Fluidigm, San Francisco, CA), according to the manufacturer’s instructions. Probes used are listed in online supplementary table S3. Transcript abundance was calculated by comparison with a standard curve. Each gene expression assay was tested in duplicate, and the mean Ct value was normalized to the averaged expression of CD3E and then subjected to analysis using the 2^-ΔΔCt method.

Analysis of eQTL effects of validated SNPs associated with recurrence and survival was carried out using HaploReg v4.1 from Broad Institute (http://archive.broadinstitute.org/mammals/haploreg/haploreg.php).15 Only cis-eQTLs (acting on local genes) were considered. Variants showing cis-eQTL effects in TRA and TRB loci were not considered due to highly variable transcription of these genes.16

Primary endpoints of the study were OS and recurrence. The OS rate was defined as the number of living patients after diagnosis divided by the total number of living and deceased patients after diagnosis. Survival time was defined as duration from diagnosis to death of any cause or the last follow-up, Time to recurrence was computed from the date of pathological diagnosis to the date of first documented recurrence or last follow-up. Patients who were lost to follow-up were censored. The risk of death or recurrence for each SNP in patients in the discovery and validation cohorts was estimated as HR and 95% CI values using the multivariate Cox proportional hazards model with adjustment for sex, age, smoking status, tumor stage, performance status and treatment. We assessed three genetic models of inheritance (dominant, recessive and additive) for each SNP using the discovery dataset and multivariable Cox proportional hazard regression analysis. The model with the smallest p value was used to measure the statistical significance of the association between each SNP and recurrence-free survival (RFS) or OS in the genotyping data. Only the dominant model was considered when the rare homozygous genotype was <5% in both living and deceased patients. A meta-analysis was used to estimate the HR and 95% CI of the combined discovery and validation populations. For the integration of genotype data, SNPs identified from GWAS not found in the OncoArray panel were replaced with linked SNPs (r² ≥0.8). Kaplan-Meier analyzes and log-rank tests were used to calculate the survival difference associated with individual genotypes. To evaluate the cumulative effects of the genetic variants, we combined the UFG (genotypes associated with significantly increased risk in the main effects analysis) for each participant. If multiple SNPs within a haplotype block showed significant main effects, only the SNP most strongly associated with the smallest P value was selected for the analysis. If the HR <1, the reciprocal value was applied in the cumulative analysis of UFG. The RNA-seq data from TCGA database were analyzed with R software (V.3.4.2), and the Wilcoxon rank-sum test was used to compare the difference in gene expression between tumor and normal tissues. All statistical tests were two-sided, with P values less than 0.05 considered statistically significant.

The characteristics of the discovery and validation cohorts are given in table 1. The discovery group included 536 patients with early-stage NSCLC, 54.1% of whom were women. The patients’ mean age at diagnosis was 65.4 years. The median survival time (MST) was 71.7 months, and the median follow-up time was 59.9 months. Within the cohort, there were 108 (20.1%) never smokers, 243 (45.3%) former smokers, and 185 (34.5%) current smokers. Among the patients, 58.0% received surgery only, 24.6% received surgery plus adjuvant chemotherapy, and 3.6% received surgery plus radiotherapy. At the time of the current study, 230 (42.9%) of the patients had died. The validation cohort included 405 patients with early-stage NSCLC (51.1% women). The patients’ mean age was 67.0 years. The MST was 94.7 months, and the median follow-up time was 33.5 months. The cohort included 17 (4.2%) never smokers, 237 (58.5%) former smokers, and 151 (37.3%) current smokers. Among the patients, 51.4% received surgery only, 24.2% received surgery plus chemotherapy, and 1.2% received surgery plus radiotherapy. At the time of the study, 280 (69.1%) of the patients had died. The details of chemotherapy were provided in online supplementary table S4. No significant differences were found comparing the host and clinical characteristics of the discovery and validation groups except for smoking status, treatment, and vital status. The discovery cohort had higher percentage of never smokers, while the validation cohort had higher percentage of non-surgery treated patients and higher number of deaths, however, discrepancy in smoking status between two cohorts have no significant impact on the overall results according to our results in multivariable models with/without adjusting the covariate (table 2, online supplementary table S5).

Among 2450 selected SNPs from 280 genes, 285 SNPs were associated with recurrence risk (p<0.05). Of these SNPs, 7 SNPs in six genes showed consistent results in both discovery and validation sets (table 2). The SNP most strongly associated with recurrence risk was rs1964986 in TRB. This intronic SNP was associated with 1.6- to 2.4-fold increased risk of recurrence in the discovery and validation sets (meta-analysis: HR 1.84, 95% CI 1.35 to 2.52, p=1.15E-04). Patients with AA genotype showed lower median RFS than those with CC/CA genotypes in both discovery (log-rank p=0.02) and validation sets (log-rank p=0.001) (figure 1A–B). IL2RB:rs3218339 also was associated with increased recurrence risk in both sets (table 2). In Kaplan-Meier survival analysis, rs3218339 CT/TT genotype carriers had shorter median RFS compared with CC carriers for both sets (log-rank p<0.05) (figure 1C–D). Another five variants were also associated with recurrence risk in both sets. Although with similar trends for both sets, the results were not consistently significant during the Kaplan-Meier analyzes (online supplementary figure S5).

Among all SNPs analyzed, 258 were associated with risk of death during the discovery set (p<0.05). However, 7 variants were associated with death risk in both datasets (table 2). The SNP most strongly associated with OS was rs10108662 of IDO1. The variant genotype was associated with 1.5-fold to 2.5-fold increased risk of death in the discovery and validation sets (meta-analysis HR 1.87, 95% CI 1.40 to 2.51, p=2.17E-05). Patients with variant AA genotype demonstrated decreased survival compared with those with CC/CA genotypes during the Kaplan-Meier analysis; however, the result was significant only in the validation set (figure 1E–F). Similarly, variant alleles of GRB2:rs959260, GRB2:rs4789182, and JAK1:rs4915675 were associated with increased risk of death in both sets, whereas CUL1:rs122571, CUL1:rs243538, and TRB:rs1573618 were associated with reduced death risk (table 2). Again, Kaplan-Meier analyzes of these SNPs were not significant in both sets (online supplementary figure S6).

To assess the joint effects of identified SNPs on NSCLC outcomes, we conducted the UFG analysis for risks of recurrence and death. The SNPs associated with recurrence or survival demonstrated cumulative effect on recurrence or death risk and RFS or OS in both sets (online supplementary table S6 and figure S7).

Subgroup analyzes were performed to identify SNPs that are predictive of NSCLC outcomes in surgery-only or surgery-plus-chemotherapy patients for both phases (online supplementary table S7). Results showed 4 SNPs were validated for recurrence and 9 SNPs for survival.

To further explore the predictors of lung cancer treatment, we performed subgroup analysis in pooled population (table 3). We identified 5 SNPs associated with recurrence risk in surgery-only and surgery-plus-chemotherapy patients. Interestingly, two variants were associated with recurrence risk in opposing directions (groups 1 and 2) for the treatment groups, whereas the remaining three variants conferred altered recurrence risk in similar fashion (groups 3 and 4). Also, we identified 6 SNPs correlated with death risk in both treatment groups. Five of the variants displayed opposite death risks (groups 1 and 2), whereas the remaining variant conferred similarly decreased death risk (group 4). To rule out the impact of chemotherapy type, we added the covariate in the subgroup analysis and found that it had no significant impact on the subgroup analysis (online supplementary table S8).

We conducted UFG analysis to assess the cumulative effect of two SNPs that predicted opposite effects on recurrence risk in the two treatment groups. Among surgery-only patients, intermediate-risk (1 UFG) and high-risk (2 UFG) patients displayed recurrence risks that were 1.9-fold and 3.7-fold higher, respectively, and shorter RFS than that of low-risk (0 UFG) group (p for trend=3.16E-03, log-rank p=4.29E-03). Conversely, among surgery-plus-chemotherapy patients, intermediate-risk and high-risk patients showed reduced recurrence risks that were 0.4-fold and 0.1-fold lower, respectively, and longer RFS than that of the low-risk group (p for trend=2.57E-03, log-rank p=0.049, figure 2A,B). The cumulative effect of 5 SNPs that predicted opposite effects on OS in the treatment groups was also analyzed. UFGs were arbitrarily defined from the vantage point of surgery-only patients with three or more UFGs as high risk; 2 UFGs as intermediate risk; and 0 or 1 UFG as low risk. Among the surgery-only patients, high-risk and intermediate-risk patients had death risks that were 3.9-fold higher (95% CI 1.91 to 7.77, p=0.013) and 2.6-fold higher (95% CI 1.22 to 5.46, p=1.7E-04), respectively, than that of the low-risk group, and had shorter MST than the low-risk group (log-rank p=3.91E-04). In contrast, among the surgery-plus-chemotherapy patients, intermediate-risk and high-risk patients had death risks that were 0.3-fold lower (95% CI; p=0.007) and 0.1-fold lower (95% CI 0.03 to 0.26; p=1.44E-05), respectively, than that of the low-risk group and had significantly longer MST than the low-risk group (log-rank p=1.82E-05). The combined UFG and Kaplan-Meier survival analyzes of the treatment-specific SNPs are shown in figure 2C,D and online supplementary table S9.

Analysis of cis-eQTL revealed that 6 SNPs correlated with gene expression in PBMCs (online supplementary table S10). We analyzed the gene expression data for these eQTL-associated genes in 1016 lung cancer tissues and 110 normal lung tissues from TCGA database. The result showed that compared with normal tissues, tumor tissues had altered expression of GRB2, JAK1 and PSMD3 (p<0.01), but not IDO1 (p=0.11) and GSK3B (p=0.13) (online supplementary figure S8).

To investigate the association between specific genotypes and T cell phenotypes, we implemented T cell CD107a degranulation and in vitro T cell killing assays. The definitions of favorable and UFGs for SNPs showing eQTL effects are listed in online supplementary table S11. In CD107a degranulation assay, we evaluated the association between UFG carrier status and CD8 + T cell degranulation. In 19 pairs of donors who belonged to low-risk group (0 or 1 UFG) or high-risk group (≥3 UFGs), we found that the percentage of CD107a+CD8+T cells was significantly higher in low-risk group (mean=7.3, 95% CI 4.7 to 9.8) than high-risk group (mean=4.5, 95% CI 2.9 to 6.2) (paired t-test p=0.02) (figure 3A,B, online supplementary table S2). Furthermore, in eight pairs of patients (the remaining pairs were not tested for cytotoxicity due to insufficient samples in storage), our in vitro T cell killing assay indicated that donors in low-risk group showed higher T cell cytotoxicity against NSCLC cells than donors of high-risk group in both target cell models (p<0.05, figure 3C, online supplementary figure S9 and table S2). We also captured images showing process of T cell engaging an NSCLC cell (figure 3E).

Furthermore, we evaluated the expression of a panel of T cell function-related genes in isolated CD3 +T cells before (baseline) and after (activated) coculture with NSCLC cells. In available samples, T cells from high-risk group displayed lower expression of IL2, GZMB, TNFA, LAG3, and VISTA, and higher expression of IDO1 than low-risk group (p<0.05, figure 3D).