Gene Signatures of Tumor Inflammation and Epithelial-to-Mesenchymal Transition (EMT) Predict Responses to Immune Checkpoint Blockade in Lung Cancer with High Accuracy

Abstract

Objectives

Treatment of non-small cell lung cancer (NSCLC) with immune checkpoint blockade (ICB) has resulted in striking clinical responses, but only in a subset of patients. The goal of this study was to evaluate transcriptional signatures previously reported in the literature in an independent cohort of NSCLC patients receiving ICB.

Materials and Methods

This retrospective study analyzed transcriptional profiles from pre-treatment tumor samples of 52 chemotherapy-refractory advanced NSCLC patients treated with anti-PD1/PD-L1 therapy. Gene signatures based on published reports were created and examined for their association with response to therapy and progression-free and overall survival (PFS, OS).

Results

Two signatures predicting response and outcomes were identified. One reflected the degree of immune infiltration and upregulation of interferon-gamma-induced genes. A second reflected the EMT status. Compared to those not responding to therapy, patients whose tumors responded to ICB had higher scores in an inflammatory gene signature (6.0±2.9 vs −5.5±3.4, p=0.014) or a more epithelial phenotype (−1.7±1.0 vs 2.1±1.2, p=0.016). Both signatures demonstrated a satisfactory predictive accuracy for response: AUC of 0.69 (95% CI: 0.54, 0.84) for the inflammatory and 0.70 (95% CI: 0.55, 0.85) for EMT signatures, respectively. A weighted score combining EMT and inflammatory signatures showed increased predictive value with AUC of 0.92 (95% CI: 0.85, 0.99). Kaplan-Meier curves for patients above and below the median combined score showed a significant separation for PFS and OS (all p<0.01, log rank test).

Conclusions

The EMT/Inflammation signature score may be useful in directing checkpoint inhibitor therapy in lung cancer and suggests that reversal of EMT might augment efficacy of ICB.

1. INTRODUCTION

Immune checkpoint blockade (ICB) using anti-PD1^1,2 or anti-PD-L1 antibodies^3,4 has demonstrated enhanced response rates, progression free survival (PFS), and overall survival (OS) in non-small cell lung cancer (NSCLC) patients. However, only a subset of patients achieve durable clinical benefit. Since these therapies are associated with immune-related adverse events and have significant financial implications,^5,6 a need exists for robust predictive biomarkers.

Measurement of PD-L1 protein expression on tumor cells by immunohistochemistry (IHC) has been the most commonly used predictive marker, however its utility has been limited by poor uniformity in the anti-PDL1 IHC antibodies, different thresholds for PD-L1 positivity across different trials, and inherent fluidity in PD-L1 expression.⁷ In addition, several trials have shown no correlation between PD-L1 expression and response rates.^1,8 Assessments of the tumor-immune cell infiltrate within the tumor microenvironment suggest that “hot” or more inflammatory tumors have better responses.^3,9–11

Other non-IHC-based biomarker strategies have shown promise. For example, in some studies, tumor somatic mutation burden (TMB) seems to predict response and PFS, presumably because these tumors have a higher neoantigen burden which makes them more sensitive to T cell activation therapy.^12,13

Another strategy is to use tumor gene-expression profiling to develop genomic signatures of response.^3,14–17 The most well-developed of these signatures reflects the inflammatory state of the tumor and includes genes that identify T cells, T cell activation, interferon-responsive genes, chemokine expression and adaptive immune resistance. This has been used to predict response to pembrolizumab in melanoma, gastric, and head and neck squamous cell carcinomas.¹⁴ Recently, the inflammation scores were combined with TMB scores and showed higher predictive value than either score alone in a similar range of cancers.^16,18

Other tumor factors may also regulate response to immunotherapy. Hugo et al. found a transcriptional signature referred to as IPRES (innate anti-PD1 resistance) that included genes involved in cell adhesion, extracellular matrix remodeling, angiogenesis, wound healing, and mesenchymal transition that predicted response to anti-PD1 antibody therapy in melanoma.¹⁵ Genes reflecting TGF-β expression and signaling of EMT/stroma-related genes have been reported to associate with resistance to PD-1 blockade in urothelial cancers.^19,20

There is also increasing evidence from preclinical studies to suggest that the level of epithelial-to-mesenchymal transition (EMT) of tumor cells contributes to the level of immunosuppression, with more mesenchymal tumors being more resistant to immunotherapy.^21,22 To our knowledge, the EMT hypothesis has not been rigorously examined in human cancers.

The goal of this study was to evaluate whether genomic data could predict response to ICB in NSCLC. To do so, we analyzed transcriptomic data from pre-treatment tumor specimens from 52 metastatic NSCLC patients treated with anti-PD1/PD-L1 monotherapy. Specifically, we assessed the predictive power of expression levels of macrophage, T cells, and TGF-β related genes, as well as a series of established genomic signatures reflecting inflammation, chronic interferon stimulation, the IPRES genes, and the EMT status of the tumors.

2. MATERIALS AND METHODS

2.1. Study Design and Patients

This single-center, retrospective, observational study was conducted at the Hospital of the University of Pennsylvania from September 2013 to August 2018 and was approved by the University’s Institutional Review Board. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement was followed to ensure the quality of data reported in this study.²³ Patients with metastatic NSCLC treated with anti-PD1 or anti-PD-L1 antibodies in the second-line setting that had sufficient residual formalin-fixed paraffin-embedded (FFPE) tumor material for detailed RNA analysis were included. A pathologist confirmed the presence of sufficient tumor material in FFPE slides. Initially, a total of 67 subjects were identified; 52 had RNA of sufficient quality for analysis. Baseline demographics and clinical variables were obtained from the electronic medical record.

2.2. Tissue Collection and RNA Sequencing

RNA was analyzed using the AmpliSeq™ Transcriptome Human Gene Expression Kit (ThermoFisher Scientific). RNA was extracted using the RNAstorm, RNA Isolation kit for FFPE (CELLDATA) from 3 FFPE slides per patient as per the manufacturer’s instructions. The quality of FFPE RNA was measured using the Agilent 2100 Bioanalyzer system (RNA 6000 Pico Kit). Sequencing libraries were prepared according to the AmpliSeq Library prep kit protocol as previously described.²⁴ Pooled libraries were amplified using emulsion PCR on an Ion Torrent OneTouch2 instrument and enriched following the manufacturer’s instructions. Libraries were then loaded onto an Ion P1 chip V2 and sequenced on the Ion Torrent Proton™ sequencing system, using the Ion PI sequencing 200 Kit v3 chemistry.

The data are raw read counts per gene. The primers used for each gene target regions of approximately the same length in all genes so do not need to be corrected for gene length. The data were normalized as “counts per million”, where the raw counts for each gene were divided by the total number of counts, multiplied by a constant (10⁶), and log2 transformed.

2.3. Gene Signature Generation

The goal of this study was to validate existing signatures from the literature in our independent patient population. By summing the log2 Z scores of genes, we calculated: (1) the IPRES score for innate anti-PD1 resistance using 16 of 17 (expression of CCL8 was too low) reported genes (see Supplemental Table 1);¹⁵ and (2) a chronic IFN activation score using STAT1, IFI44, IFIT1, IFIT3, OAS1,OAS2, MX1, IRF7, and ISG15 as described.^25,26 The TGF-β genes analyzed were those described by Mariathasan¹⁹ as TGFB1, TGFBR2, ACTA2, COL4A2, and TAGLN.

Our inflammation signature was based on a previously published gene score¹⁴ and created by summing (unweighted) log2 Z scores for 27 inflammation related genes (Supplemental Table 1).

There is not yet a validated lung cancer EMT signature, so we prospectively generated a gene list based on previous publications describing “classic” EMT genes in cancer^23,24 with the addition of some genes specifically mentioned in studies evaluating EMT in NSCLC.^27–29 Selected genes had levels of expression that were clearly above baseline (using a cutoff of 10 reads in our data set). Supplemental Table 2 shows the list of genes included along with their average expression level. Although the genes SNAI1, TWIST1, TWIST2, CDH2, and ZEB1 are classic mesenchymal markers, their expression levels were very low in our dataset and thus not included. Based on these criteria, we generated the EMT signature by adding the sum of the log2 Z scores of 6 established mesenchymal genes (AGER, FN1, MMP2, SNAI2, VIM, ZEB2) and subtracting the sum of the log2 Z scores of 6 established epithelial genes (CDH1, CDH3, CLDN4, EPCAM, MAL2, and ST14) (Supplemental Table 2). In this signature, the most mesenchymal tumors have the most positive EMT scores and the most epithelial tumors have the most negative scores.

2.4. Statistical Analyses

Comparisons among response groups was done using Pearson’s chi-squared test for categorical characteristics or Fisher’s exact test. Response variables were grouped into the binary categories of responders (including patients with partial responses and stable disease) versus non-responders. The responder group was defined by a radiologist reporting a decrease in size of overall disease burden or stability of disease ≥ 6 months. The association between a gene signature score and response was examined using a t-test after the normality of the gene scores were established. Descriptive statistics are displayed as mean±sem. The area under a receiving operating characteristics (ROC) curve (AUC) was obtained by fitting a logistic regression model with response status as the dependent variable and the signature score as an independent variable. The significance level of AUC against the null value of 0.5 was tested.

We created a combined score for inflammation and EMT using a simple unweighted score by taking the difference between the inflammation and EMT scores (higher scores are more inflammatory and more epithelial). We also generated a weighted score with the estimated regression weights from the multivariable logistic regression model that included both the inflammation and EMT scores. Based on the maximum likelihood estimation, the weighted score was computed as −0.60*EMT+0.19*inflammation.

The significance level for the change in AUC from the single signature model to 2-signature model was tested.³⁰ Kaplan-Meier curves for PFS and OS were plotted for patients above or below the median weighted combined score and compared by Cox proportional hazards models. PFS was defined as months since the start of ICB therapy until the date of progression or death, or censored at the cut-off date of August 31, 2018. All statistical analyses were two sided and performed using Graphpad Prism version 7.0 (La Jolla, CA) or Stata version 15.1 (College Station, TX).

3. RESULTS

3.1. Patient Characteristics

52 stage IV NSCLC patients were ultimately analyzed (Table 1). 63% had an adenocarcinoma histology and 85% were current or former smokers. 36 (69%) patients were treated with nivolumab, 14 (27%) with pembroluzimab, and 2 (4%) with atezoluzimab. 25 patients (48%) had a clinical response (partial response or stable disease) with 27 (52%) patients with progressive disease (non-responders). Response rates were not statistically different by histology, gender, age at diagnosis, age at treatment, pack years, or types of ICB treatment (all p>0.05, Table 1). The median PFS was 1.7 months in non-responders and 25 months in responders. The median OS was 6.0 months in non-responders and 25 months in responders.

Table 1.

Clinical Data Table

	Total	Non-responder	Responder	P-value
	52	27 (52%)	25(48%)
Immunotherapy Treatment
Nivolumab	36 (69%)	19 (70%)	17 (68%)	0.98
Pembrolizumab	14 (27%)	7 (26%)	7 (28%)
Atezolizumab	2 (4%)	1 (4%)	1 (4%)
Histology
Adenocarcinoma	33 (63%)	17 (63%)	16 (64%)	0.94
Squamous	19 (37%)	10 (37%)	9 (36%)
Driver Mutation*
Yes	9 (17%)	6 (22%)	3 (12%)	0.47
Gender
Male	23 (44%)	12 (44%)	11 (44%)	0.97
Female	29 (56%)	15 (56%)	14 (56%)
Age at Diagnosis
Median	62	62	63	0.59
Age at ICB Initiation
Median	64	64	65	0.79
Smoking History
Former/Current	44 (85%)	23 (85%)	21 (84%)	0.91
Never	8 (15%	4 (15%)	4 (16%)
Smoking (Pack Years)
Median	20	25	20	0.77
PFS (mo)
Events	42	27	15	<0.001
Median	3.3	1.7	25
OS (mo)
Events	41	25	16	<0.001
Median	9.6	6.0	25

Age at ICB: Age at time of starting immunotherapy; PFS (mo): Progression-free Survival in months; OS (mo): Overall survival in months

^*Driver mutations were defined as actionable EGFR mutations or EML4-ALK fusions

Footnote: Fisher’s exact test for categorical variables, t-test or median test for continuous variables, log-rank test for PFS and OS.

3.2. Gene Signatures and Response to ICB

PD-L1 gene expression levels were not significantly different between responder and non-responders (Supplemental Figure 1).

The inflammation signature scores were significantly higher in responders compared to non-responders. (Figure 1A, +6.0±2.9 vs −5.5±3.4, t test p=0.014). A ROC curve evaluating the inflammatory score had an AUC of 0.69 (Figure 1B, p =0.011). To assess the effects of tumor infiltration with macrophages and T cells, we examined the expression levels of a set of established macrophage genes (CD68, CD14, CD163, and CSF1R) and T cell genes (CD4, CD8A, and CD8B) and found no significant differences when comparing responder and non-responder groups (Supplemental Figure 2).

Figure 1. Analysis of the inflammatory and EMT signatures and response to checkpoint blockade.

(A) Comparison of the log2 z-scores of the 27-gene inflammatory gene signature between responders and non-responders. The mean inflammatory score was significantly higher in the responders compared to the non-responder group (6.0 vs. −5.5, p=0.014). (B) Receiver Operator Characteristics curve (ROC curve) utilizing the inflammatory signature to predict response to checkpoint blockade, AUC 0.69 (p=0.011). (C) Comparison of the log2 z-scores of the EMT signature between responders and non-responders. The mean EMT score was significantly lower (more epithelial) in the responders compared to non-responders (−1.7 vs. 2.1, p=0.016). (B) ROC curve utilizing the EMT signature to predict response to checkpoint blockade, AUC 0.70 (p=0.01).

The EMT signature scores were significantly lower (more epithelial) in the responder cohort (average score= −1.7±1.0) compared to non-responders (average score = +2.1±1.2) (Figure 1C; t test p=0.016). The area under the ROC curve was 0.70 (p=0.01) (Figure 1D). Similar to previous reports in melanoma,^15,31 (the epithelial marker E-cadherin (CDH1 was significantly higher in the responder group compared to the non-responders (Figure 2A; average reads: 53.0 ±5.6 vs 36.4 ±6.4, p=0.014). Interestingly, the non-responder group included a number of patients with very low levels of E-cadherin expression that had no overlap with the responding patients (Figure 2B). Analysis of EpCAM expression between responders and non-responders showed a near statistically significant difference (p=0.06) (Figure 2C), however, we did not observe a difference in the number of patients with very low expression as we did with E-cadherin (Figure 2D).

Figure 2. Analysis of E-cadherin expression levels and response to checkpoint blockade.

(A) Comparison of CDH1 gene expression levels in responders and non-responders. The mean CDH1 expression levels were significantly higher in the responder group compared to non-responders (53.0 ±5.6 vs 36.4 ±6.4, p=0.014). (B) Closer view of low-level CDH1 expression in responders and non-responders demonstrating enrichment of very low levels of CDH1 expression in the non-responder group. (C) Comparison of EpCAM gene expression levels in responders and non-responders. The mean EpCAM expression levels in responders compared to non-responders (125.8 vs 82.8, p=0.06). (D) Closer view of low-level EpCAM expression showing similar values between responder and non-responder groups.

We also assessed previously published gene signatures that have been associated with primary resistance to ICB. We did not observe a significant difference in the IPRES scores (p=0.1) (Supplemental Figure 3A), the chronic IFN scores (p=0.63) (Supplemental 3B), or the TGF-β signature scores (Supplemental Figure 4) between responders and non-responders.

3.3. Relationship of Inflammation and EMT Gene Scores

Because both the inflammatory and EMT gene signatures were significantly associated with response rates, we evaluated their relationship and noted a positive correlation between the two signatures (r²=0.32, p<0.0001) (Figure 3A). This relationship signifies that the more mesenchymal tumors (most positive EMT scores) tended to be the “hottest” (most inflammatory). Of interest, we noted (as has been previously reported),³² a significant positive correlation (r²=0.13, p=0.009) between PD-L1 expression levels and EMT scores, with the more mesenchymal tumors showing higher PD-L1 gene expression levels (Figure 3B).

Figure 3. Relationship between the EMT signature, Inflammatory signature and PD-L1 gene expression levels.

(A) Correlation between the log2 z-scores of the Inflammatory and EMT gene signatures (r²=0.32, p=0.0001). (B) Correlation between the log2 z-scores of the EMT gene signature and PD-L1 gene expression levels (r²=0.13, p=0.009).

3.4. Combined EMT and Inflammation Scores

Because the Inflammation and EMT signatures appear to assess different tumor programs, we generated a combined EMT/Inflammation score. The unweighted combination score was significantly higher in responders compared to non-responders (+7.7±2.3 vs −7.7±2.6) (p<0.0001) (Figure 4A). The weighted combination score was also significantly higher in responders compared to non-responders (2.2±0.4 vs −2.3±0.4) (p<0.0001) Figure 4C). A ROC analysis evaluating the ability of the combined scores to predict response had an AUC of 0.81 (p<0.001) for the unweighted score (Figure 4B) and AUC of 0.92 (p<0.001) for the weighted scores (Figure 4D). This value was higher than the AUCs using either the Inflammatory or EMT score alone (p=0.0007 and p=0.0051, respectively). Using the median value −0.4 for the weighted score, we could correctly classify 83% of patients’ response status with sensitivity of 84% and specificity of 81%.

Figure 4. Relationship between the combined EMT and Inflammatory gene signatures with response.

(A) Comparison of the log2 z-scores of the combined unweighted EMT/Inflammatory gene signature between responders and non-responders. The mean EMT/Inflammatory score was significantly higher in the responders compared to the non-responder group (7.7 vs. −7.7, p=0.0001). (B) ROC curve utilizing the unweighted EMT/Inflammatory signature to predict response to checkpoint blockade, AUC 0.81 (p=0.001). (C) Comparison of the log2 z-scores of the combined weighted EMT/Inflammatory gene signature between responders and non-responders. The mean EMT/Inflammatory score was significantly higher in the responders compared to the non-responders (2.2 vs. −2.3, p=0.0001). (B) ROC curve utilizing the weighted EMT/Inflammatory signature to predict response to checkpoint blockade, AUC 0.92 (p<0.001).

The weighted combined score was also highly predictive of PFS and OS. Kaplan-Meier curves for patients above and below the median showed a good separation for PFS and OS (Figure 5A and 5B, log rank test, all p<0.01) indicating a strong ability to discriminate patients with longer and shorter survival. The hazard ratio (HR) comparing high and low weighted combined scores was 0.32 (95% CI: 0.16, 0.60) for PFS and 0.40 (95% CI: 0.21, 0.76) for OS.

Figure 5. Association between the combined weighted EMT/Inflammatory score and outcome.

Patients were categorized as having an EMT/Inflammatory score above or below the median value for the full cohort and Progression-free Survival (PFS) and overall Survival (OS) were calculated. (A) Kaplan-Meier curves for PFS for patients above or below the median weighted combined score: Cox HR 0.32 (95% 0.16, 0.36, p<0.01). (B) OS for patients above or below the median weighted combined score: Cox HR 0.40 (95% 0.21, 0.76, p<0.01).

4. Discussion

We found that two distinct genomic signatures in lung cancer tumor samples obtained before the initiation of ICB can predict response to therapy. Higher scores of a 27 gene inflammation signature were associated with better responses in lung cancer, as has been shown in other tumors.¹⁴ Although the calculated IPRES scores, chronic interferon-gamma stimulation scores, and expression of TGF-β related genes failed to predict ICB responsiveness, the signature that specifically reflected the epithelial vs mesenchymal transition characteristics of the tumor was significantly associated with response to ICB. As predicted in preclinical models,^21,22 patients whose tumors were more mesenchymal had poorer responses and shorter PFS and OS than those with more epithelial tumors. Importantly, the EMT score was significantly positively correlated with the inflammation score, and a combined score predicted response with higher accuracy than either score alone suggesting these features drive independent programs within the tumor microenvironment. To our knowledge, the weighted EMT/Inflammation score is more predictive of response (an AUC of 0.92) and survival than any previously published score.^16,18

Our findings associating the mesenchymal phenotype with a poor response to ICB in lung cancer are novel, but consistent with previous studies in humans, as well as previous preclinical work, showing that forced mesenchymal transformation can lead to immunosuppression.^21,33 A proteomic analysis in a small cohort of melanoma patients (n=8) identified that mesenchymal protein features were associated with lack of response to ICB.³¹ Wang et al. found a higher gene score associated with stromal-related gene expression in a subgroup of patients with metastatic urothelial cancers who responded to nivolumab.²⁰ Chakravarthy et al. showed that a set of TGF-β associated extracellular matrix genes (many of which would qualify as “mesenchymal” ) was linked to a worse prognosis in two out of three published data sets of PD-1 blocked patients (two melanoma and one bladder).³⁴

Disagreement exists in the literature about the relationship of inflammatory genes to the mesenchymal phenotype. When examining TCGA data, Chae et al. found that a more mesenchymal signature was associated with lower T cell gene expression in NSCLC.³⁵ In contrast, our data (Figure 3A) showing that tumors with higher inflammation scores had higher (more mesenchymal) EMT scores is similar to reports from Lou et al.²⁹ and Chen et al. ³²

An interesting question raised by our findings is why mesenchymal differentiation was associated with decreased response to checkpoint blockade? One possibility is that EMT might simply reflect high levels of the immunosuppressive cytokine TGF-β in the tumor; however, we saw no genomic evidence of increased TGF-β related genes. We feel a more likely possibility is that the biological process of mesenchymal transition (which results in changes such as loss of HLA Class 1 antigen, upregulation of PD-L1, ^32,36 downregulation of tumor antigens, decreased expression of immunoproteosome subunits,³⁶ stimulation of autophagy,³⁷ and a well- recognized general resistance to any type of cell killing²⁸ makes cancer cells more resistant to T cell killing³⁸, as has been shown in a number of in vitro studies.^37,39,40 Our data also suggest that expression levels of E-cadherin might play a role. Although loss of both EPCAM and E-cadherin (classic epithelial to epithelial adhesion molecules) are hallmarks of mesenchymal transformation,⁴¹ only E-cadherin (in contrast to EPCAM) serves as the ligand for the adhesion molecule CD103 which is highly expressed on tissue resident memory CD8 T cells. A higher density of tissue resident memory cells in the tumor microenvironment has been associated with improved outcomes in lung cancer⁴² and other tumor types.⁴³ Mechanistically, the binding of CD103 to E-cadherin on target cells provides costimulatory signals and promotes effector T-cell function.⁴⁴ These findings suggest that the downregulation of E-cadherin may dampen the CD103-mediated immune attack thus serving as a potential mechanism of resistance to ICB. Intriguingly, we observed significantly higher levels of E-cadherin in the responder group compared to the non-responders. In addition, we observed that very low levels of E-cadherin expression were only found in the non-responders (Figure 2), suggesting that loss of E-cadherin might be a good predictor of non-response to checkpoint blockade.

An interesting implication of our study is that if EMT is a cause of resistance to anti-PD1 therapy, a possible therapeutic approach might be to first use an EMT-reversing agent and then follow this with checkpoint inhibition.⁴⁵ New drugs are being developed to reverse EMT⁴⁶ including inhibitors of the tyrosine kinase AXL.^47,48 It might also be possible to use drugs developed for other pathways that also affect EMT, such as ERK1–2 inhibitors⁴⁹ or the multi-kinase inhibitor drug nintedanib.⁵⁰

The strengths and weaknesses of this study should be considered. As strengths, we used samples from actual clinical practice rather than a formalized clinical trial. Our findings appear to be robust, in that we were able to obtain good data from small biopsy specimens (transbronchial biopsies and needle biopsies) that were not macrodissected and thus contained variable amounts of non-tumor tissue. We were also able to obtain highly congruent results using different analytic approaches (Nanostring®, HTGseq®, data not shown). This study also has limitations. Our data set was only moderate in size (52 patients) and derived from only one center. Thus, replication of these findings in other patient populations is needed. This study exclusively analyzed relapsed, second-line NSCLC patients who primarily received two different anti-PD1 inhibitor antibodies. Given the changing clinical standards of care, a goal for future studies will be to examine the value of these signatures in front-line patients receiving checkpoint antibodies alone or in combination with chemotherapy. The other major limitations to our study are that we did not consistently have PD-L1 immunohistochemical staining or measurements of TMB; our patients were treated in the second-line setting a number of years ago when neither of these tests were routinely performed. Although mRNA and protein levels do not always correlate perfectly, we did analyze the expression of the PD-L1 gene and found it did not correlate with response.

There are a number of possible extensions to our signatures that could make our findings more clinically applicable. This includes further optimization of weighting specific genes within each signature. As an alternative to genomic analysis, it may also be possible to obtain prognostic information about the “EMT status” of tumors by adding additional immunohistochemical stains to standard PD-L1 staining. Finally, one long-range goal would be to develop blood tests that might reflect the EMT status of tumors.

5. CONCLUSIONS

In summary, we have found that two genomic signatures, reflecting aspects of tumor inflammation and the EMT status of the tumor, when combined, predict response to immune checkpoint inhibitors in the second-line treatment of advanced NSCLC patients with very high accuracy. A higher combined EMT/Inflammatory score was independently associated with improved progression-free and overall survival. Application of these signatures to this population could provide a needed tool to direct checkpoint therapy to those patients most likely to benefit. Success in predicting responses in front-line therapy needs to be tested, but could extend this approach to a larger number of patients.

Highlights.

• Evaluation of gene signatures to predict response to immunotherapy in NSCLC.

• An EMT/Inflammation score was highly predictive of response and survival to ICB.

• Identifies potential role of EMT in primary resistance to immunotherapy in NSCLC.