The detailed information regarding the data being analyzed in this study is summarized in online supplemental table S1, and the flow diagram of this study is depicted as figure 1. In brief, we included patients treated with anti-PD-(L)1 antibodies at National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital and Chinese Academy of Medical Sciences and Peking Union Medical College and Sun Yat-sen University Cancer Center from December 2016 to December 2018 (named China cohort) as a discovery set, all patients were treated as part of clinical trials. Eligible patients for this study were determined mainly based on the following criteria: (i) >18 years old; (ii) Eastern Cooperative Oncology Group performance status: 0–1; (iii) have advanced or recurrent NSCLC; (iv) failure after first-line platinum-based doublets chemotherapy; (v) radiologically evaluable according to Response Evaluation Criteria in Solid Tumors (RECIST) V.1.1. CT or MRI scans were reviewed by the investigators. The comprehensive genomic profiling of 79 patients with NSCLC were implemented by WES. PD-L1 scoring was available in 49 out of 79 patients (online supplemental methods).

Multiple independent public cohorts were used to validate the association between EPHA mutation and immunotherapy efficacy. The first validation cohort (validation cohort 1, n=165) was a pooled cohort consisting three public datasets with patients with NSCLC treated with ICIs with available WES data, including the datasets of 75 patients treated with anti-PD-1 plus anticytotoxic T-lymphocyte antigen 4 (anti-CTLA-4) (Hellmann cohort),12 56 patients treated with anti-PD-(L)1 with or without anti-CTLA-4 (Miao cohort)30 and 34 patients treated with anti-PD-1 (Rizvi 34 cohort).31 The second validation cohort (validation cohort 2) consisted 1662 patients with a variety of cancer types who had received at least one dose of ICI therapy (Memorial Sloan Kettering Cancer Center [MSKCC] cohort), including 350 patients with NSCLC.7 The tumor tissues were subjected to MSKCC 468-gene panel (earlier versions included 341 or 410 genes) target sequencing. We also obtained the WES data of 2599 solid tumors in TCGA along with the corresponding mRNA expression data of 2541 solid tumors across six tumor types from cBioPortal (www.cbioportal.org) to study the mechanism underlying the association between EPHA mutation and immunotherapy (online supplemental methods).

In China cohort, baseline tumor assessments were performed within 1–28 days prior to the initiation of the anti-PD-(L)1 treatment, with the subsequent assessments being performed every 6–8 weeks until objective disease progression. The objective response rate (ORR) was defined as the percentage of patients with confirmed complete response (CR) or partial response (PR) by RECIST V.1.1. The disease control rate (DCR) was defined as the percentage of patients with confirmed complete response (CR), PR or stable disease (SD) by RECIST V.1.1. Durable clinical benefit (DCB) was defined as the percentage of patients who achieved CR, PR or SD lasted >6 months; all other patients were considered to have no durable benefit (NDB). Progression-free survival (PFS) was defined as the time from the beginning of ICI treatment to the date of PD or death from any cause. Patients who had not progressed were censored at the date of their last scan.

In the validation cohorts, tumor response was evaluated according to RECIST V.1.1 in Hellmann cohort,12 Miao cohort30 and MSKCC cohort.7 Objective response to anti-PD-1 treatment was assessed by investigator-assessed immune-related response criteria (irRC) in Rizvi 34 cohort.31 The definitions of DCB, NCB and PFS were consistent with those in the discovery cohort.

The detailed profiles of EPHA (EPHA1-8, EPHA10) mutation in each cohort are listed in the online supplemental figure S1. The non-synonymous mutations including TRUNC (Frameshift del, Frameshift ins, nonsense, nonstop, splice region, splice site), INFRAME (Inframe del and Inframe ins) and MISSENSE mutations of at least one EPHA subtype were defined as EPHA mutation (EPHA^mut ) in this study.

Continuous variables were compared by Mann-Whitney U test and categorical variables were compared by χ² test or Fisher’s exact test. Survival was estimated by Kaplan-Meier curves, with the p value determined by a log-rank test. HR was determined through the univariable and multivariable Cox regression. Variables with p<0.1 in the univariable regression and those which has been reported associated with the effect of immunotherapy in NSCLC were also included into multivariable Cox regression. Random-effect models were used to pool the effect sizes. The poor results of categorical variables were presented as relative risk (RR) and 95% CIs. Q-test and I² statistics were used to assess the heterogeneity. A result of p>0.1 and I² <50% indicated no significance between-study heterogeneity.

We used propensity-score matching with a ratio of 1:1 to analyze the association between EPHA status and PFS and OS in patients with NSCLC from TCGA. Propensity score was estimated by age, sex, smoking history, tumor stage, pathology, history of other malignancy and neoadjuvant therapy. False discovery rate (FDR) was used to estimate the significance of differences between the mRNA expression levels. All reported p values were two-tailed, FDR < 0.05 is considered statistically significant.

All analyses were performed using SPSS V.24.0 (SPSS, Chicago, Illinois, USA), and R V.3.5.2. Graphs in the present study were drawn by GraphPad Prism 8 and R V.3.5.2.

The discovery cohort included 79 patients with stage IV NSCLC who were treated with anti-PD-(L)1 treatment and had baseline tissue samples sequenced by WES (online supplemental table S4). The median age was 55 (IQR, 47–61) years and 54 patients (68%) were male. Nearly half of the patients had a history of smoking. The major histology was adenocarcinoma (47%). Most patients (74/79, 94%) received PD-1 inhibitor monotherapy and the rest received PD-L1 inhibitor monotherapy (5/79, 6%). The median follow-up time was 21 months.

In the discovery cohort, 22.8% patients with NSCLC harbored EPHA mutations, and most EPHA mutations were missense mutations (92.1%) (online supplemental figure S1). Patients harboring EPHA mutation (EPHA^mut ) had superior PFS (median PFS 6.87 months vs 2.10 months, HR 0. 38; 95% CI 0.21 to 0.68; p<0.001; figure 2A), a higher DCB (50% vs 16.4%; Fisher’s exact test, p=0.009; figure 2A), higher DCR (72.2% vs 36.1%; Fisher’s exact test, p=0.01; figure 2B) and numerically higher ORR (27.8% vs 16.4%; Fisher exact test, p=0.31; figure 2B) compared with those with wild-type EPHA (EPHA ^wt).

In the univariable analyses, besides EPHA mutation, several other indexed such as smoking history, sex and TMB (≥median vs <median) were also associated with the immunotherapeutic PFS with the HRs (95% CI) of 0.48 (0.30 to 0.76), 0.48 (0.29 to 0.79) and 0.46 (0.28 to 0.73), respectively (online supplemental table S5). In the multivariable Cox proportional hazards regression model adjusted by the smoking status, age, sex, PD-L1 expression and TMB (≥median vs <median), the association between EPHA mutation and PFS remained significant (HR 0.36; 95% CI 0.19 to 0.71; p=0.003; online supplemental table S5). Several factors which have been reported to be associated with the immunotherapeutic effectiveness, such as sex, age, histology, smoking and so on were also included simultaneously in the multivariable Cox regression model to exclude the potential confounding effects.5 32–34 These results indicated that EPHA mutations are associated with better clinical benefits of anti-PD-(L)1 therapy independent of PD-L1 expression and TMB.

In validation cohort 1 (online supplemental table S6), EPHA^mut was prevailed in 28.5% patients with NSCLC and it was associated with significantly longer PFS of ICIs (HR 0.48; 95% CI 0.31 to 0.74; p<0.001; figure 3A) and better DCB (74.5% vs 36.4%; Fisher’s exact test, p=0.02; figure 3A), DCR (91.2% vs 61.9%; Fisher’s exact test, p<0.001; figure 3B) and ORR (53.4% vs 25.7%; Fisher’s exact test, p=0.001; figure 3B), which was consistent with the discovery cohort. The trend of prolonged PFS in EPHA^mut patients was consistently observed across all three datasets included in the first validation cohort (figure 3C), and the pooled analyses showed a significantly longer PFS (HR 0.47; 95% CI 0.30 to 0.72; p<0.001; figure 3C), ORR (RR 2.12; 95% CI 1.39 to 3.22; p<0.001; online supplemental figure S2) and DCB (RR 2.04; 95% CI 1.52 to 2.74; p<0.001; online supplemental figure S2) in patients with EPHA ^mut versus EPHA ^wt . Statistical analyses for study heterogeneity did not present significant in all pooled estimates (p>0.10, I² <50%), indicating the consistency of the association between EPHA ^mut and favorable benefit to ICIs across the three datasets.

In the multivariable Cox proportional hazards regression model adjusted by PD-L1 expression and TMB (top 20% vs the rest) and other confounding factors, EPHA^mut remained an independent predictor for superior PFS (HR 0.59; 95% CI 0.37 to 0.96; p=0.03; online supplemental table S7).

In the validation cohort 2, which comprised 1662 patients with >10 types of tumors, we further validated the association between EPHA^mut and significantly longer OS in NSCLC (HR 0.48; 95% CI 0.33 to 0.71; p<0.001; figure 4A). However, no statistically significant association was observed between OS and EPHA status in other tumors (figure 4A). After adjusted for TMB (top 20% vs the rest), the association between EPHA^mut and OS remained significant in NSCLC (HR 0.60; 95% CI 0.39 to 0.93; p=0.02; figure 4B). EPHA^mut remained an independent predictor for ICIs in NSCLC in the multivariate Cox regression (HR 0.63; 95% CI 0.41 to 0.98; p=0.04; online supplemental table S8), as adjusted by histology, age, sex, treatment and TMB.

In addition, all the EPHA subtypes seemed to play uniform roles in predicting the clinical survival benefit with the HRs <1.0 in the pool analysis of the discovery and validation cohorts (online supplemental figure S3). Taken together, these data indicated that EPHA mutation might be predictive of the clinical benefit of ICIs independently in NSCLC.

We further explored whether the association between EPHA^mut and clinical benefit of immunotherapy may vary with the histological subtypes. In the discovery cohort, EPHA^mut was associated with significantly longer PFS (median PFS 9.82 vs 2.10 months; HR 0.30; 95% CI 0.13 to 0.67; p=0.002; figure 5A) than EPHA^wt in lung adenocarcinoma (LUAD), which was consistently observed in both validation cohorts (figure 5B,C). The pooled analysis further revealed that EPHA^mut was associated with significantly longer PFS (discovery cohort plus validation cohort 1; HR 0.38; 95% CI 0.25 to 0.58; p<0.001; online supplemental figure S4) and OS (validation cohort 2; HR 0.51; 95% CI 0.33 to 0.79; p=0.001; online supplemental figure S4) in LUAD.

After adjusted by TMB, PD-L1 and other confounding factors in the multivariate Cox regression, EPHA^mut remained an independent predictor for PFS and OS efficacy of ICIs in patients with LUAD based on either the discovery cohort (PFS HR 0.34; 95% CI 0.12 to 0.95; p=0.04; online supplemental table S9) or the validation cohorts (validation cohort 1, PFS HR 0.53; 95% CI 0.31 to 0.90; p=0.02; online supplemental table S10); validation cohort 2, OS HR 0.61; 95% CI 0.38 to 1.00; p=0.05; online supplemental table S11). No difference of PFS and OS was observed in patients with lung squamous cell carcinomas (LUSC) with EPHA^mut versus EPHA^wt (online supplemental figure S4).

To assess to potential prognostic value of EPHA mutation, survival analyses were further performed according to EPHA mutational status in the TCGA database. No significant difference was found in PFS or OS between EPHA^mut and EPHA^wt subsets in patients with NSCLC, LUAD and LUSC with standard treatment (figure 6), suggesting that EPHA mutation was not a prognostic factor.

Considering the previously identified correlations of PD-L1 expression and TMB with the clinical benefits of ICIs, the associations of EPHA mutations with PD-L1 expression, TMB as well as neoantigen burden were analyzed to investigate the possible mechanism. We first aimed to ascertain whether co-occurrence takes place between EPHA mutations with robust predictors, including PD-L1 expression, higher TMB and predicted neoantigens. As shown, EPHA ^mut was associated with significantly higher TMB in both the discovery cohort (median: 317 muts vs 105 muts, Mann-Whitney U test, p<0.001; online supplemental figure S5A) and validation cohort 1 (median: 332 muts vs 123 muts, Mann-Whitney U test, p<0.001; online supplemental figure S5C), along with significantly elevated predicted neoantigens (validation cohort 1, median: 476 vs 148, Mann-Whitney U test, p<0.001; online supplemental figure S5E), suggesting the co-occurrence between the EPHA ^mut and elevated TMB and predicted neoantigens. However, no association was observed between EPHA ^mut and PD-L1 expression (≥1% vs <1%) (discovery cohort: Fisher’s exact test, p=1.00; validation cohort: Fisher’s exact test, p=0.48; online supplemental figure S5B,S5D). PD-L1 expression was relatively balanced between EPHA ^mut group and EPHA ^wt group in the discovery and validation cohorts.

We further investigated other driver mutations co-mutated with EPHA. The co-occurrence of EPHA and EGFR, STK11, ALK, ROS1 seldom occurred (online supplemental figure S6A). The incidence rates of EPHA co-mutated with KRAS and TP53 were 2.1% and 14.9% in the discovery cohort and 15.5% and 19.3% in the validation cohort in LUAD, respectively (online supplemental figure S6A). However, few co-mutations of EPHA with KRAS and TP53 existed in LUSC. We further investigated the effect of co-mutation in predicting survival of immunotherapy in LUAD. The association between EPHA mutation and survival was not influenced by KRAS mutation or TP53 mutation in LUAD as shown in online supplemental figure S7B-7C, suggesting that EPHA mutation was an independent predictor of immunotherapy.

To further explore the underlying mechanism of the predictive values of EPHA mutations to ICI efficacy, 47 immune-related signatures and 43 TGF-β signaling genes based on the RNA and WES data from TCGA database (online supplemental tables S2 and S3) were analyzed. Gene Set Enrichment Analysis (GSEA) in NSCLCs revealed a prominent enrichment of signatures related to the downregulation of TGF-β signaling, while no difference of IFN-γ signaling was observed (data not shown). Further analyses in separate populations of LUAD or LUSC illustrated that LUAD rather than LUSC subset presented the prominent enrichments of signatures related to IFN-γ signaling upregulation and TGF-β signaling downregulation in EPHA^mut versus EPHA^wt groups (FDR adjusted p=0.03 for both; figure 7A,B).

Furthermore, compared with EPHA^wt patients, the mRNA expression levels of ten immune-related genes and five TGF-β-related genes were significantly increased and decreased, respectively, in patients with EPHA^mut (FDR adjusted p<0.05; figure 7C,D; online supplemental tables S12 and S13) in the LUAD subset, while such pattern was not observed in the LUSCs (figure 7E,F; online supplemental table S12).

No significant differences of TGF-β signaling or T cell gene signature between EPHA^mut and EPHA^wt were observed in the other types of tumor (online supplemental tables S12 and S13). GSEA results showed no enrichment in the IFN-γ or TGF-β signaling in patients with EPHA ^wt in bladder, esophageal carcinoma, skin cutaneous melanoma or head and neck carcinoma cohorts based on TCGA datasets (data not shown).

Collectively, the above results showed that the superior ICI benefits in NSCLCs with EPHA^mut might mainly be mainly attributed to LUAD. In terms of mechanism, the downregulation of TGF-β signaling and the increased T cell signatures mediated by EPHA^mut might be on account of the different susceptibility for ICIs between LUAD and other tumors.