A Dysfunctional T-cell Gene Signature for Predicting Nonresponse to PD-1 Blockade in Non–small Cell Lung Cancer That Is Suitable for Routine Clinical Diagnostics

Abstract

Purpose:

Because PD-1 blockade is only effective in a minority of patients with advanced-stage non–small cell lung cancer (NSCLC), biomarkers are needed to guide treatment decisions. Tumor infiltration by PD-1^T tumor-infiltrating lymphocytes (TIL), a dysfunctional TIL pool with tumor-reactive capacity, can be detected by digital quantitative IHC and has been established as a novel predictive biomarker in NSCLC. To facilitate translation of this biomarker to the clinic, we aimed to develop a robust RNA signature reflecting a tumor's PD-1^T TIL status.

Experimental Design:

mRNA expression analysis using the NanoString nCounter platform was performed in baseline tumor samples from 41 patients with advanced-stage NSCLC treated with nivolumab that were selected on the basis of PD-1^T TIL infiltration by IHC. Samples were included as a training cohort (n = 41) to develop a predictive gene signature. This signature was independently validated in a second cohort (n = 42). Primary outcome was disease control at 12 months (DC 12 m), and secondary outcome was progression-free and overall survival.

Results:

Regularized regression analysis yielded a signature using 12 out of 56 differentially expressed genes between PD-1^T IHC-high tumors from patients with DC 12 m and PD-1^T IHC-low tumors from patients with progressive disease (PD). In the validation cohort, 6/6 (100%) patients with DC 12 m and 23/36 (64%) with PD were correctly classified with a negative predictive value (NPV) of 100% and a positive predictive value of 32%.

Conclusions:

The PD-1^T mRNA signature showed a similar high sensitivity and high NPV as the digital IHC quantification of PD-1^T TIL. This finding provides a straightforward approach allowing for easy implementation in a routine diagnostic clinical setting.

Translational Relevance.

PD-(L)1 blockade therapies have significantly improved the survival of a subset of patients with advanced-stage non–small cell lung cancer. However, most patients do not benefit, but still are at risk of adverse effects associated. This asks for biomarkers to better predict benefit. PD-1^T TIL, a tumor-reactive T-cell population, constitute such a biomarker, and can be accurately quantified by digital image analysis in formalin-fixed paraffin-embedded (FFPE) tumor tissue. Yet, this method is complex, limiting its application in routine patient care. As an alternative, a PD-1^T mRNA expression signature measured in FFPE tissue samples with industry-standard technology is more suitable for implementation in routine diagnostics. The PD-1^T signature reached equally high sensitivity and negative predictive value as the digital image analysis–based IHC quantification of PD-1^T TIL, making it particularly suited for reliably identifying patients with advanced-stage non–small cell lung cancer who have a low chance of benefitting from PD-(L)1 blockade therapy.

Introduction

Pharmacologic blockade of the inhibitory immune receptor programmed cell death protein 1 (PD-1) or its ligand programmed death-ligand 1 (PD-L1) has improved the clinical outcome of many cancers, including non–small cell lung cancer (NSCLC; refs. 1–5). High tumor PD-L1 expression has been associated with clinical benefit in NSCLC treated with PD-1 blockade (1, 2, 6), and these results have led to the implementation of PD-L1 IHC as a biomarker in clinical practice. However, other studies have shown suboptimal correlation of PD-L1 expression to clinical outcome (3–5). Therefore, robust biomarkers that more accurately predict who will benefit, and who not, are needed. In particular, biomarkers with a high negative predictive value (NPV) that reliably predict lack of clinical benefit are important to offer patients alternative treatments.

Previously, we reported high levels of PD-1^T tumor-infiltrating lymphocytes (TIL) as a novel predictive biomarker for long-term benefit to PD-1 blockade with a high NPV (7). PD-1^T TIL are a subset of PD-1+ T cells with a dysfunctional phenotype, high tumor-specific expression of PD-1, and high capacity for tumor recognition (8, 9). PD-1^T TIL can be measured via algorithm-based quantitative analysis of PD-1 IHC in formalin-fixed paraffin-embedded (FFPE) tumor tissue (7, 8). Although digital image analysis can yield accurate quantitative PD-1 protein expression data, this method is challenging to validate across centers, affecting routine clinical care.

We therefore aimed at developing a reliable method that can detect the signal represented by PD-1^T TIL and that can easily be applied in routine clinical care. The NanoString nCounter is a robust platform that allows for measuring very low input RNA amounts isolated from FFPE tissue (10). Several mRNA signatures have already been developed for this platform, including the tumor inflammation signature (TIS) that has demonstrated predictive potential for clinical benefit to PD-1 blockade in multiple cancer types (11–13).

In the current study, we used the NanoString nCounter platform to develop and validate an RNA expression signature that reflects a tumor's PD-1^T TIL status and predicts clinical outcome of patients with NSCLC treated with PD-1 blockade.

Materials and Methods

Patient cohorts and study endpoints

In this study, 152 patients with stage IV NSCLC were included who all started second- or later-line monotherapy with nivolumab. Patients with tumors harboring known sensitizing EGFR mutations or ALK translocations were excluded. A total of 115 patients received PD-1 blockade therapy since March 2015 at the Netherlands Cancer Institute/Antoni van Leeuwenhoek hospital (NKI-AVL), The Netherlands, and were split in a training (n = 94) and validation cohort (n = 21). Thirty-seven patients from the CERTIM (Immunomodulatory Therapies Multidisciplinary Study Group) treated since July 2015 at Cochin University Hospital, France (13) were pooled with the validation cohort (Fig. 1). All patients received nivolumab, administered per label as a single agent.

Figure 1.

Study design for the development and validation of the PD-1^T signature as biomarker for nonresponse to PD-1 blockade in NSCLC. Overall workflow for the development of the PD-1^T signature using PD-1^T IHC-high (≥90 per mm²) patients with DC 12 m (n = 12) and PD-1^T IHC-low (<90 per mm²) patients with PD (n = 29). An independent cohort of patients was used for validation (n = 42).

RECIST version 1.1 was used to determine the objective response. Patients treated at NKI-AVL, who were not evaluable for response assessment according to RECIST, were determined as having progressive disease (PD) by the treating physician. Disease control (DC) status (complete response/partial response or stable disease) at 12 months after treatment initiation was used as the primary clinical outcome measure. Progression-free survival (PFS) and overall survival (OS) were used as secondary outcome measures. PFS and OS were defined as the time from the date of initiation of PD-1 blockade treatment to the date of progression (for PFS) or death (for OS). Patients who had not progressed or died were censored at the date of their last follow-up.

RNA or gene expression data derived from pretreatment archival FFPE tumor tissue samples were collected from the cohorts. Written informed consent was obtained from all patients treated at NKI-AVL for research usage of material not required for diagnostic use by institutionally implemented opt-out procedure. The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board (IRB) of NKI (CFMPB586). Gene expression data of the CERTIM cohort was used from Damotte and colleagues, a study that was previously approved by the ethics committee (CPP Ile de France II, No. 2008-33, 2012 06-12, 2018 MS1) in agreement with article L.1121-1 of the French law (13). In the NKI-AVL training cohort (n = 94), 28 samples were excluded on the basis of low RNA yield and/or low RNA quality. Eight additional samples were excluded because of quality control failure in the NanoString nCounter profiling assay (NanoString; Fig. 1). Eight samples in the NKI-AVL validation cohort and six samples in the CERTIM cohort were excluded because these were obtained 2 years or more before start of PD-1 blockade. Two additional samples in the NKI-AVL validation cohort were excluded because these contained abundant normal lymphoid tissue (n = 2; Fig. 1). We excluded lymph node resections, as these contain PD-1 high-expressing T cells in normal lymphoid tissue, which could potentially lead to false-positive test results. Lymph node biopsies were included, as these are usually targeted biopsies in the tumor region.

IHC

PD-1 and PD-L1 immunostaining of samples from the NKI-AVL cohorts were performed on fresh-cut slides from FFPE blocks using an anti–PD-1 antibody (NAT105, Roche Diagnostics) and anti–PD-L1 antibody (22C3 DAKO, Agilent) on a BenchMark Ultra autostainer Instrument (Ventana Medical Systems) as described previously (7). PD-L1 immunostaining of samples from the CERTIM cohort was performed using a different anti–PD-L1 antibody (E1L3N, Cell Signaling Technology, catalog No. AB_2687655) on a Bond automat (Leica Biosystems) as described previously by Adam and colleagues (14).

PD-1 IHC slides were scanned at 20× magnification with a resolution of 0.50 per μm² using an Aperio slide AT2 scanner (Leica Biosystems).

Digital quantification of PD-1^T TIL

In 11/94 (12%) pretreatment samples, automated quantification of PD-1^T TIL was performed using an image analysis algorithm with a cutoff of 0.25 optical density (OD) of PD-1 staining as described previously (7). In 83/94 (88%) samples, PD-1^T TIL numbers were used from a previously published cohort and quantified using the same approach (7). PD-1^T TIL numbers of 58 samples were used to develop the signature and are provided in Supplementary Table S1.

PD-L1 scoring

Tumor PD-L1 expression in pretreatment FFPE samples was assessed in the NKI-AVL training (n = 58) and validation cohort (n = 21) using the clinical grade laboratory-developed test IHC assay with the 22C3 DAKO clone (Agilent) as described previously (7). For 35/37 (95%) samples in the CERTIM cohort, PD-L1 tumor proportion score (TPS) data have been reported previously (13). In 2/37 (5%), the PD-L1 status was unknown. The expression levels were scored using a different anti–PD-L1 antibody (E1L3N, Cell Signaling Technology, catalog No. AB_2687655) as previously described and validated by the PATTERN French thoracic pathologists’ group (14). PD-L1 TPS data are provided in Supplementary Table S1.

RNA extraction and hybridization to nCounter tagset

RNA of pretreatment FFPE samples from the NKI-AVL cohorts were isolated with the AllPrep DNA/RNA FFPE isolation kit (No. 80234, Qiagen) according to the manufacturer's recommendations and quantified by Tapestation (Agilent). RNA from the CERTIM cohort was extracted with High Pure FFPE RNA Isolation Kit (Roche Diagnostics) according to the manufacturer's recommendations and quantified using fluorimetry with Qubit RNA XR Assay Kit (Invitrogen, Thermo Fisher Scientific). A total of 200- to 300-ng RNA from the NKI-AVL cohorts and 30- to 100-ng RNA from the CERTIM cohort were hybridized to NanoString PanCancer IO 360 Panel code set (NanoString), according to the manufacturer's recommendations. After hybridization, nonbound probes were washed off, and the RNA–probe complex was bound to the cartridge on the NanoString Flex Prep Station (NanoString) according to manufacturing protocol. The cartridge was sealed and transferred to the digital analyzer for imaging.

Statistical analysis

The Mann–Whitney, Fisher exact, and linear-by-linear association tests, respectively, were used to assess differences in patient characteristics between training and validation cohorts. Differences were considered statistically significant if ^*, P < 0.05. Correlations between the PD-1^T signature and PD-1^T TIL assessed by IHC or the PD-1^T signature and the TIS, respectively, were evaluated using linear regression analysis.

A two-level batch effect correction on the mRNA expression data was performed on all NKI-AVL and CERTIM patients using an empirical Bayes linear regression. This was performed to correct for batch effects between the NKI-AVL and the CERTIM cohort and between the different NKI-AVL cohorts. Both batch effect correction and gene expression analysis were performed with R 4.1.0 and the package limma 3.48.0. Differential gene expression analysis was performed using linear regression on the gene log-expression. Separate models were fitted for each gene, and the computation of moderated t-statistics and log-odds of differential expression was performed via empirical Bayes moderation. Analysis of main biological processes involved in the gene signature was performed by gene ontology analysis using the Fisher exact test with the R package topGO 2.44.0 (SCR_014798). P values were adjusted via Benjamini–Hochberg.

A prediction model was built using logistic regression combined with regularized regression for variable selection using LASSO (least absolute shrinkage and selection operator). By adding a penalization term on the coefficients of the model, the coefficients of the models are constrained to zero leading to variable selection. Because of the limited sample size and unequal distribution over the DC 12 m and PD patient groups, cross-validation was limited to threefold. Thus, a threefold cross-validation for the selection of the optimal penalization term of the regularized regression, based on the deviance, was performed. This is a goodness-of-fit statistics commonly used for generalized linear models. The results of the regression were transformed to obtain probability scores using the formula: with K being the results of the logistic regression. K was computed as K = c₁ * STAT1 + c₂ * OAS1 + c₃ * TAP1 + … + c₁₂ * LAG3. The coefficients of the prediction model are provided in Supplementary Table S2.

The cross-validation and prediction model building were performed with R and the package glmnet 4.1-2. On the basis of the NKI-AVL training cohort, a threshold was chosen from the probability scores that were provided by the prediction model to classify a patient as predicted to achieve DC upon therapy. This threshold was set at the best sensitivity (detection of DC), while keeping a satisfactory specificity. A prediction model including only the PDCD1 gene was built using logistic regression with samples from the NKI-AVL training cohort. All prediction models aimed to predict DC 12 m, and therefore the reference category used in the logistic regression was PD. Signature probability scores are provided in Supplementary Table S1. ROC curves were produced for each prediction model using the package pROC 1.18.0. The area under the ROC curves (AUC) was compared using the DeLong test with the pROC 1.18.0 R package.

Genes in the TIS are normalized using a ratio of the expression value to the geometric mean of the housekeeper genes that are used only for the TIS. This is then followed by a log₂ transformation. The TIS score was calculated in the NKI-AVL validation cohort (n = 21) as a weighted linear combination of the 18 gene expression values (11, 15). This analysis was performed by NanoString as part of their intellectual property. For the 37 samples in the CERTIM cohort, TIS scores have been reported before (13). TIS scores of 42 samples were used for analysis and are provided in Supplementary Table S1.

Data availability

The raw mRNA expression data generated in this study have been deposited in Zenodo, under accession codes https://doi.org/10.5281/zenodo.10213261 and are available under restricted access. Access to these data can be obtained upon a scientifically sound request with the corresponding author. All requests will be reviewed by the IRB of the NKI and will require the requesting researcher to sign a data access agreement with the NKI.

Results

PD-1^T signature development

To develop a predictive mRNA signature that reflects tumor infiltration by PD-1^T TIL, we first selected 94 pretreatment samples from patients with advanced-stage NSCLC treated with nivolumab (training cohort). Sixty-six of 94 (70%) of these samples had sufficient RNA available for gene expression analysis using the NanoString nCounter platform. Only archival samples were used, and therefore the majority with insufficient RNA were blocks that were already (partially) used for previous molecular analysis. Fifty-eight of 94 (62%) samples were successful in obtaining high-quality mRNA expression data to be used for signature development, as eight samples did not meet RNA quality criteria (Fig. 1).

Of these 58 patients, 12 showed DC at 12 months (DC 12 m; n = 12) and 46 developed PD within 12 months of treatment (Fig. 1). DC12 was chosen as the primary clinical outcome measure based on previous observations that patients with DC 12 m were more accurately identified by the biomarker as compared with DC 6 m (i.e., higher sensitivity). This was also found for a group with no long-term benefit (i.e., higher NPV; ref. 7). Clinicopathologic characteristics and treatment outcomes are summarized in Supplementary Table S3. In line with our previous work, the number of PD-1^T TIL per mm² was significantly higher in patients with DC 12 m compared with patients with PD (P < 0.01; Fig. 2A). PD-1^T IHC-high versus IHC-low status was called using a previously established cutoff of 90 PD-1^T TIL per mm² (Supplementary Fig. S1A and S1B; ref. 7). Twelve of 12 (100%) patients with DC 12 m were classified as PD-1^T IHC-high. Twenty-nine of 46 (63%) patients with PD were classified as PD-1^T IHC-low, and 17/46 (37%) as PD-1^T IHC-high (Figs. 1 and 2A).

Figure 2.

PD-1^T signature development for prediction of nonresponse to PD-1 blockade. A, PD-1^T TIL per mm², as measured by digital IHC algorithm-based quantification, in pretreatment samples from patients with DC 12 m (n = 12) and PD (n = 46) in the training cohort (n = 58). Only patients with available PD-1^T IHC and gene expression data are included. Dashed line indicates a cutoff of 90 PD1^T TIL per mm². Medians, interquartile ranges, and minimum/maximum shown in boxplots, **, P < 0.01 by Mann–Whitney U test. The red dots indicate PD-1^T IHC-high (≥90 per mm²) patients with DC 12 m (n = 12), and the blue dots indicate PD-1^T IHC-low (<90 per mm²) patients with PD (n = 29). B, Gene set enrichment analysis displaying gene sets that are significantly enriched in PD-1^T IHC-high tumors from patients with DC 12 m (n = 12). Pathways are ordered by P value (log-transformed), P values were calculated by Fisher exact test. C, Volcano plot showing the differentially expressed genes between PD-1^T IHC-high tumors from patients with DC 12 m and PD-1^T IHC-low tumors from patients with PD (n = 29) in the training cohort (n = 41). A total of 56 genes reached statistical significance. The red line indicates a P value <0.05. P values were computed by moderated t-statistics. D, Heat map showing the expression of the 12-gene PD-1^T signature in pretreatment samples from PD-1^T IHC-high patients with DC 12 m (n = 12) and PD-1^T IHC-low patients with PD (n = 29) in the training cohort (n = 41) ordered by probability score. Each column represents one patient [blue: PD-1^T signature high (score of ≥0.35) and light gray: PD-1^T signature low (score of <0.35), green: PD-1^T IHC-high patients with DC 12 m, dark gray: PD-1^T IHC-low patients with PD] and rows display genes. Positive values (red) indicate higher gene expression, and negative values (blue) indicate lower gene expression. E, Correlation of PD-1^T TIL and the PD-1^T signature in the training cohort (n = 41), R² = 0.603, ****, P < 0.0001. R² and P values were calculated using linear regression analysis. F, PD-1^T signature scores in pretreatment samples from PD-1^T IHC-high patients with DC 12 m (n = 12) and PD-1^T IHC-low patients with PD (n = 29) in the training cohort (n = 41). The dashed line indicates a cutoff score of 0.35. Medians, interquartile ranges, and minimum/maximum shown in boxplots, ****, P < 0.0001 by Mann–Whitney U test. G, ROC curve for predictive value of the PD-1^T signature for DC 12 m (AUC, 0.97; 95% CI, 0.93–1.00) in the training cohort (n = 41).

To obtain a distinctive mRNA expression gene set with maximum contrast, we performed differential gene expression analysis between PD-1^T IHC-high patients with DC 12 m (n = 12) and PD-1^T IHC-low patients with PD (n = 29; Figs. 1 and 2A). After correction for multiple testing, 54 genes were significantly higher expressed, and two genes were significantly lower expressed in the PD-1^T IHC-high DC 12 m group. Some of the top-ranked genes included LAG3, CTLA4, CXCR6, and CXCL13, which have previously shown to be highly expressed in PD-1^T TIL (ref. 8; Supplementary Table S4). In addition, various pathways related to active immune responses in the tumor microenvironment (TME), including type I IFN signaling, regulation of lymphocyte chemotaxis, and natural killer cell–mediated cytotoxicity were significantly upregulated in the PD-1^T IHC-high DC 12-m group (Fig 2B).

Regularized regression analysis (LASSO) yielded a 12-gene PD-1^T signature (STAT1, OAS1, TAP1, HEY1, CXCL13, IFIT2, IL6, TDO2, CD6, CTLA4, CD274, LAG3) as being most predictive (Table 1). All genes in the signature were positively associated with DC 12 m, except for HEY1, which showed a negative association (Fig. 2C and D). In line with the selection strategy of the samples, we observed a high correlation between the number of PD-1^T TIL assessed by IHC and the PD-1^T signature score within samples (R² = 0.603; P < 0.0001; Fig. 2E). The PD-1^T signature was able to separate the preselected PD-1^T IHC-high DC 12-m group from the preselected PD-1^T IHC-low PD group with high significance (P < 0.0001; Fig. 2D–F). The AUC was 0.97 [95% confidence interval (CI), 0.93–1.00] (Fig. 2G). We aimed for a sensitivity of ≥90% to minimize the risk of undertreatment and a specificity of ≥50%, a strategy that was previously used for the PD-1^T TIL IHC biomarker (7). A probability score of 0.35 was selected as optimal cutoff (Fig. 2F). This cutoff resulted in a sensitivity of 92%, specificity of 93%, positive predictive value (PPV) of 85%, and NPV of 96% (Table 2).

Table 1.

Overview of the upregulated and downregulated PD-1^T signature genes.

PD-1T signature genes
STAT1	Up
OAS1	Up
TAP1	Up
HEY1	Down
CXCL13	Up
IFIT2	Up
IL6	Up
TDO2	Up
CD6	Up
CTLA4	Up
CD274	Up
LAG3	Up

Table 2.

Predictive accuracy of the PD-1^T signature and PD-L1 TPS, summary of training and validation results.

	Biomarker	AUC	Cutoff	Sensitivity	Specificity	PPV	NPV
Training cohort (n = 41)	PD-1T signature	0.97; 95% CI, 0.93–1.00	<0.35 vs. ≥0.35	92%	93%	85%	96%
Validation cohort (n = 42)	PD-1T signature	0.87; 95% CI, 0.74–0.99	<0.35 vs. ≥0.35	100%	64%	32%	100%
Validation cohort (n = 40)	% PD-L1 TPS	0.63; 95% CI, 0.34–0.91	<50% vs. ≥50%	50%	82%	33%	90%
			<1% vs. ≥1%	50%	62%	19%	88%

PD-1^T signature validation

Next, the predictive performance of this PD-1^T signature was validated in an independent cohort of 42 patients with advanced-stage NSCLC treated with nivolumab. Six of 42 (14%) patients showed DC 12 m and 36/42 (86%) showed PD (Fig. 1). In contrast to the training cohort, tumor samples were not preselected. None of the clinicopathologic characteristics differed significantly between training and validation set, except for the performance score as more patients in the validation cohort showed a higher performance score (P < 0.01; Supplementary Table S3).

In the validation cohort, in line with previously observed patterns of PD-1^T TIL quantified by IHC, PD-1^T signature scores were significantly higher in the DC 12-m group than in the PD group (P < 0.01; Fig. 3A). A high PD-1^T signature score (≥0.35) correctly identified six out of six patients in the DC 12-m group (sensitivity for treatment benefit 100%), and a low score (<0.35) identified 23/36 patients with PD (specificity for no treatment benefit 64%), yielding a PPV of 32% and an NPV of 100% with an AUC of 0.87 (95% CI, 0.74–0.99; Fig. 3A–B; Table 2). Similar to the training cohort, most signature genes were overexpressed in the PD-1^T signature-high DC 12-m group (n = 6), compared with the PD-1^T signature-low PD group (n = 23), and the HEY1 gene was expressed at lower levels. The PD-1^T signature-high patients with PD (n = 13) showed a similar gene expression profile as the PD-1^T signature-high DC 12-m group (Fig. 3C).

Figure 3.

PD-1^T signature validation for prediction of nonresponse to PD-1 blockade. A, PD-1^T signature scores in pretreatment samples from patients with DC 12 m (n = 6) and patients with PD (n = 36) in the validation set (n = 42). The dashed line indicates a cutoff score of 0.35. Medians, interquartile ranges, and minimum/maximum shown in boxplots, **, P < 0.01 by Mann–Whitney U test. B, ROC curve for predictive value of the PD-1^T signature for DC 12 m (AUC, 0.87; 95% CI, 0.74–0.99) in the validation cohort (n = 42). C, Heat map showing the expression of the 12-gene PD-1^T signature in pretreatment samples from patients with DC 12 m (n = 6) and patients with PD (n = 36) in the validation cohort (n = 42) ordered by probability score. Each column represents one patient [blue: PD-1^T signature high (score of ≥0.35), light gray: PD-1^T signature low (score of <0.35), green: patients with DC 12 m, dark gray: patients with PD] and rows display genes. Positive values (red) indicate higher gene expression, and negative values (blue) indicate lower gene expression. PFS [D; median 8.3 vs. 1.8 months; HR, 0.36 (95% CI, 0.18–0.69), ***, P < 0.001] and OS [E; median 7.0 vs. 5.6 months; HR, 0.34 (95% CI, 0.17–0.68), ***, P < 0.001] of patients with PD-1^T signature-high (n = 6) and PD-1^T signature-low (n = 36) pretreatment samples in the validation set (n = 42). Tick marks represent data censored at the last time the patient was known to be alive and without disease progression or death. P value was determined by log-rank test.

In addition, in the validation cohort, PFS was significantly longer in PD-1^T signature-high patients (median, 8.3 months) versus PD-1^T signature-low patients (median, 1.8 months) with HR 0.36 (95% CI, 0.18–0.69), P < 0.001. The median OS was 7.0 months versus 5.6 months, respectively, with HR 0.34 (95% CI, 0.17–0.68), P < 0.001 (Fig. 3D and E).

Comparison with PD-L1 IHC and PDCD1 gene expression

In previous work, we showed that the predictive performance of PD-1^T TIL was superior to the PD-L1 TPS (7). Therefore, we compared the predictive performance of PD-L1 TPS with the PD-1^T signature in the validation cohort. Nine of 42 (21%) pretreatment samples showed a PD-L1 TPS of ≥50%, 7/42 (17%) between 1% and 50%, and 24/42 (57%) showed no PD-L1 expression. In 2/42 (5%), the PD-L1 status was unknown, and these patients were excluded from further analysis (Supplementary Tables S1 and S3). PD-L1 TPS was not significantly higher in patients in the DC 12-m group compared with patients with PD (P = 0.30; Fig. 4A). The AUC was substantially lower compared with the PD-1^T signature (AUC, 0.63; 95% CI, 0.34–0.91; P = 0.13), indicating a lower discriminatory ability (Fig. 4B). At 50% cutoff, the sensitivity was also lower (50%), as well as were specificity (82%), PPV (33%), and NPV (90%; Table 2). Furthermore, we observed that a PD-L1 TPS of ≥50% was not associated with significantly better PFS [HR, 0.77 (95% CI, 0.38–1.54), median PFS 5.7 vs. 2.3 months], but did show borderline significance for improved OS [HR, 0.40 (95% CI, 0.20–0.81), median OS 11.4 vs. 3.0 months] (Fig 4C and D). Next, we assessed the predictive accuracy at 1% cutoff, as this cutoff has also been previously studied, though with contradictory results (3–5). Here, sensitivity (50%) and NPV (88%) were similar to the performance of 50% PD-L1 TPS, with only a slightly lower specificity (62%) and PPV (19%; Table 2). PD-L1 TPS ≥1% showed similar survival outcomes as PD-L1 TPS ≥50% (Fig. 4E and F).

Figure 4.

Association of PD-L1 with treatment benefit and survival. A, PD-L1 TPS in pretreatment samples from patients with DC 12 m (n = 6) and patients with PD (n = 34) in the validation cohort (n = 40). Mean shown as dashed line, P = 0.30 by Mann–Whitney U test. Note that for two patients, PD-L1 TPS was unknown. B, ROC curve for predictive value of PD-L1 TPS for DC 12 m (AUC, 0.63; 95% CI, 0.34–0.91) in the validation cohort (n = 40). PFS [C; median 5.7 vs. 2.3 months; HR, 0.77 (95% CI, 0.38–1.54), P = 0.47] and OS [D; median 11.4 vs. 3.0 months; HR, 0.40 (95% CI, 0.20–0.81), ^*, P = 0.03] of patients with PD-L1 TPS ≥50% (n = 9) and PD-L1 TPS <50% (n = 31) pretreatment samples in the validation cohort (n = 40). PFS [E; median 4.2 vs. 2.1 months; HR, 0.87 (95% CI, 0.46–1.63), P = 0.65] and OS [F; median 8.9 vs. 2.5 months; HR, 0.56 (95% CI, 0.30–1.07), *, P = 0.0498] of patients with PD-L1 TPS ≥1% (n = 16) and PD-L1 TPS <1% (n = 24) pretreatment samples in the validation cohort (n = 40). Tick marks represent data censored at the last time the patient was known to be alive and without disease progression or death. P value was determined by log-rank test.

Next, we explored the predictive value of PDCD1 gene expression, encoding PD-1, and compared this with the PD1^T signature. Using all 58 pretreatment samples with available gene expression data from the first cohort for training, we observed that PDCD1 scores of the DC 12-m group were slightly higher compared with the PD group (P = 0.03), with an AUC of 0.71 (95% CI, 0.54–0.88) and a score of 0.20 as the optimal cutoff (Supplementary Fig. S2A and S2B). In the validation cohort, signature scores did not significantly differ between the DC 12-m group and the PD group (P = 0.06; Supplementary Fig. S2C). The AUC was lower compared with the PD-1^T signature (0.74; 95% CI, 0.50–0.99, P = 0.36; Supplementary Fig. S2D), as well as the sensitivity (83%), NPV (93%), and specificity (36%) when using the predefined cutoff of 0.20 (Supplementary Table S5). Patients with high (≥0.20) PDCD1 scores did not show significantly prolonged PFS and OS compared with those with low (<0.20) PDCD1 scores in the validation cohort (Supplementary Fig. S2E and S2F).

Taken together, these findings show that the PD-1^T signature had a higher accuracy for predicting DC 12 m and survival compared with PD-L1 TPS. The predictive performance of PDCD1 alone was lower as compared with the PD-1^T signature, further highlighting that only the presence of a T-cell subset with high expression of PD-1, which is reflected by the PD-1^T signature, and not total PD-1 expression is predictive for response to PD-1 blockade.

Discussion

In spite of the success of PD-(L)1 blockade therapies, the majority of patients do not benefit from these agents. Therefore, biomarkers that reliably allow to identify patients without clinical response are urgently needed to guide alternative treatment decisions beyond PD-1 blockade. In addition, such biomarkers should be developed using robust clinical-grade platforms that can easily be implemented in a clinical setting. We previously established PD-1^T TIL, a functionally and transcriptionally distinct intratumoral T-cell population with enriched tumor reactivity, as a novel predictive biomarker for long-term benefit to PD-1 blockade (7, 8). However, the measurement of PD-1^T TIL status using advanced digital image analysis–based quantification of IHC stainings is complex, and multiple—predominantly technical—sources of potential bias are challenging to cope with in a routine diagnostic setting. Therefore, we here developed a clinically applicable mRNA signature reflecting the presence of PD-1^T TIL by using the NanoString nCounter platform. This study shows that the PD-1^T TIL IHC biomarker could successfully be translated into a gene expression signature, as the latter had a similar predictive performance to the digital IHC quantification approach (7). Importantly, a high sensitivity and NPV (100%) was reached, which should allow patients to be reliably identified without clinical benefit to PD-1 blockade alone. In this relatively small number of patients, the PD-1^T signature performed superior to PD-L1 TPS, which is similar to previous work (7). The clinical applicability of the NanoString nCounter platform has been demonstrated previously (16), and this platform has been shown to have a high analytic sensitivity, technical reproducibility, and to generate robust data from routine FFPE samples (10). Therefore, we expect that by using this approach, the PD-1^T signature can now easily be applied in a clinical setting and that a similar approach can be exploited for other promising biomarker candidates.

Genes in the PD-1^T signature are related to, for instance, coinhibitory signaling (CD274, CTLA4, LAG3), cytokines and chemokines (CXCL13, IL6), IFN signaling (IFIT2, OAS1, STAT1), antigen presentation (TAP1), and angiogenesis (HEY1). This is in line with features of a TME with a preexisting adaptive immune response as well as with ongoing immunosuppressive stimuli associated with T-cell dysfunction (17, 18). Of note, LAG3, CTLA4, and CXCL13 correspond to the dysfunctional phenotype that characterizes PD-1^T TIL (8), indicating that the gene signature captures the presence of PD-1^T TIL in the TME. Intriguingly, the PDCD1 gene was not among the signature genes, probably due to a partial overlap in expression levels between PD-1^T TIL high and low tumors. Notably, the predictive performance of the PD-1^T signature was better than PDCD1 gene expression alone. This is in line with the notion that the predictive capacity is specifically driven by the PD-1^T TIL subset, which reflects a tumor-reactive population that is likely crucial for response to PD-1 blockade and not simply by the presence of PD-1–positive T cells (8).

The development of predictive mRNA signatures that characterize a broader transcriptomic immune profile of the TME has gained interest. For example, the commercially available TIS has shown to predict clinical benefit in different cancer types, including NSCLC (11–13). We correlated the PD-1^T signature scores to the TIS scores in samples from the validation cohort (n = 42). We observed a good correlation between the signatures (R² = 0.705, P < 0.0001), in line with a partial overlap in three genes (STAT1, CD274, LAG3; Supplementary Fig. S3). While the small sample size limits conclusions that can be made from this analysis, it may further support the notion that these types of mRNA signatures can robustly detect tumor immune environments that are responsive to immune checkpoint inhibitors (ICI). In the future, it will be interesting to compare the performance of the signatures in a larger patient cohort to understand whether they report on similar features of the TME or whether there may be possible additive value for a subgroup of patients.

While the analysis of samples from two distinct expert centers on NSCLC ICI therapy strengthens the results of the current study, a number of limitations should be noted: First, the sample size was low in both the training and the validation cohorts. Thus, additional studies with higher patient numbers are needed. Second, this is a retrospective study, which makes further validation in prospective studies necessary. Third, different PD-L1 IHC antibodies were used in the validation cohort, which could potentially have introduced analytic differences (19, 20). Finally, an important emerging question in the clinical treatment of NSCLC is how to preselect patients for therapy with either single-agent PD-1 blockade or in combination with chemotherapy. As we developed the PD-1^T biomarker aimed at high sensitivity and NPV, the signature is currently not designed to make such predictions. Nevertheless, retraining of the biomarker for high specificity and PPV in a suitable patient cohort could be considered for this purpose. Such an approach, however, may be potentially limited by the number of false positives driven by a relatively large fraction of mixed responses in the biomarker-high, PD group as shown previously (7).

Taken together, in the current study, we developed a PD-1^T gene expression signature with high sensitivity and NPV that can be used as a predictive biomarker for nonresponse to PD-1 blockade in NSCLC. Our data demonstrate that our digital image–based IHC assay can reliably be replaced by a matching gene expression signature with comparable predictive performance. This could provide a straightforward FFPE-based approach that allows for easy implementation in a routine diagnostic clinical setting. Moreover, the strategy used has the potential to bring other expression-level–based biomarkers to routine clinical diagnostics where these can support shared decision-making for therapeutic strategies.

Supplementary Material

Supplementary Fig. S1

PD-1T quantification by an algorithm-based immunohistochemical (IHC) imaging approach.

Supplementary Fig. S2

Predictive accuracy of PDCD1 gene expression.

Supplementary Fig. S3

Correlation of the PD-1T signature with the Tumor Inflammation Signature (TIS). Correlation of the PD-1T signature and the TIS in the validation cohort (n=42).

Supplementary Table S1

Overview of all analyzed biomarkers per patient.

Supplementary Table S2

Coefficients of the PD-1T signature genes used in the prediction model.

Supplementary Table S3

Patient characteristics and treatment outcomes for training and validation cohorts. P-values were calculated by Mann-Whitney, Fisher exact or linear-by-linear association tests.

Supplementary Table S4

Overview of significantly expressed genes between tumors from PD-1T IHC high patients (n=12) and PD-1T IHC low patients (n=29).

Supplementary Table S5

Predictive accuracy of PDCD1 gene expression signature, summary of training and validation results.

Authors' Disclosures

K. Hummelink reports grants from Dutch Cancer Society and Dutch Ministry of Health, Welfare and Sport during the conduct of the study. W.S.M.E. Theelen reports grants from MSD, AZ, and Sanofi/Regeneron outside the submitted work. K. Leroy reports personal fees and nonfinancial support from Roche and personal fees from AstraZeneca, Janssen, Amgen, GSK, and MSD outside the submitted work. E.F. Smit reports other support from AstraZeneca, Bristol Myers Squibb, Bayer, DSI, Eli Lilly, MSD, Merck, Novartis, Pfizer, Takeda, Regeneron, Roche Genentech, Roche Diagnostics, Boehringer Ingelheim, and Sanofi outside the submitted work. G.A. Meijer reports other support from Hartwig Medical Foundation, Sysmex, and Exact Sciences and grants from CZ Health Insurance outside the submitted work; in addition, G.A. Meijer has a patent for Progression Markers for Colorectal Cancer issued and licensed to CRCbioscreen BV, a patent for Protein Biomarkers for Detection of Colorectal Cancer issued and licensed to CRCbioscreen BV, a patent for Protein Biomarkers (II) for Detection of Colorectal Cancer in Stool issued and licensed to CRCbioscreen BV, and a patent for Methods and Compositions for Analyses of Cancer, pending. G.A. Meijer is also co-founder and board member (CSO) of CRCbioscreen BV, CSO of Health-RI (Dutch National Health Data infrastructure for research and innovation), and a member of the supervisory board of IKNL (Netherlands Comprehensive Cancer Organisation). G.A. Meijer has research collaborations with Exact Sciences, Sysmex, Sentinel CH. S.p.A., Personal Genome Diagnostics (PGDX), DELFi, and Hartwig Medical Foundation. These companies provided materials, equipment and/or sample/genomic analyses. D.S. Thommen reports grants from Bristol Myers Squibb and Asher Bio outside the submitted work. K. Monkhorst reports personal fees from Lilly, Bayer, and Amgen outside the submitted work. No disclosures were reported by the other authors.

Authors' Contributions

K. Hummelink: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. R. Tissier: Data curation, formal analysis, supervision, validation, investigation, visualization, methodology, writing–review and editing. L.J.W. Bosch: Resources, Data curation, funding acquisition, validation, methodology, project administration. O. Krijgsman: Formal analysis, investigation, methodology. M.M. Van den Heuvel: Resources, data curation, writing–review and editing. W.S.M.E. Theelen: Resources, Data curation, writing–review and editing. D. Damotte: Resources, data curation, validation, investigation, writing–review and editing. F. Goldwasser: Resources, data curation, validation, investigation, writing–review and editing. K. Leroy: Resources, Data curation, validation, investigation, methodology, writing–review and editing. E.F. Smit: Conceptualization, resources, data curation, supervision, funding acquisition, investigation, methodology, project administration, writing–review and editing. G.A. Meijer: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. D.S. Thommen: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. K. Monkhorst: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, Project administration, writing–review and editing.