Systematic Evaluation of the Predictive Gene Expression Signatures of Immune Checkpoint Inhibitors in Metastatic Melanoma

Abstract

Advances in immunotherapy, including immune checkpoint inhibitors (ICIs), have transformed the standard of care of many types of cancer including melanoma. ICIs have improved the overall outcome of melanoma patients; however, a significant proportion of patients suffer from primary or secondary tumor resistance. Therefore, there is an urgent need to develop predictive biomarkers to better select patients for ICI therapy. Numerous biomarkers that predict response of melanoma to ICIs have been investigated, including biomarker signatures based on genomics or transcriptomics. Most of these predictive biomarkers have not been systematically evaluated across different cohorts to determine the reproducibility of these signatures in metastatic melanoma. We evaluated 28 previously published predictive biomarkers of ICIs based on gene expression signatures in 8 previously published studies with available RNA-sequencing data in public repositories. We found that signatures related to IFN-γ-responsive genes, T and B cell markers and chemokines in the tumor immune microenvironment are generally predictive of response to ICIs in these patients. In addition, we identified that these predictive biomarkers have higher predictive values in on-treatment samples as compared to pre-treatment samples in metastatic melanoma. The most frequently overlapping genes among the top 18 predictive signatures were CXCL10, CXCL9, PRF1, RANTES, IFNG, HLA-DRA, GZMB and CD8A. From gene set enrichment analysis and cell type deconvolution, we estimated that the tumors of responders were enriched with infiltrating cytotoxic T-cells and other immune cells and the upregulation of genes related to interferon-γ signaling. Conversely, the tumors of non-responders were enriched with stromal related cell types such as fibroblasts and myofibroblasts, as well as enrichment with T helper 17 cell types across all cohorts. In summary, our approach of validating and integrating multi-omics data can help guide future biomarker development in the field of ICIs and serve the quest for a more personalized therapeutic approach for melanoma patients.

INTRODUCTION

Despite the significant recent advances in the systemic therapy of melanoma, it remains a lethal disease for a large proportion of patients. It is estimated that there will be about 99,780 newly diagnosed invasive melanoma cases and 7,650 projected deaths in the United States in 2022¹. For the past decade, several immune checkpoint inhibitors (ICIs) have transformed the standard of care of many types of cancer including melanoma. Ipilimumab, a human anti-cytotoxic T-lymphocyte-associated antigen 4 antibody (anti-CTLA-4) represents the first ICI approved by the U.S. FDA for the treatment of melanoma in 2011². Following this approval, monoclonal antibodies against programmed death 1 (anti-PD-1) pembrolizumab³ and nivolumab⁴ were subsequently approved in 2014 for melanoma. Similarly, the combination of ipilimumab and nivolumab was approved for melanoma in 2015⁵. Most recently, the fixed dose combination of relatlimab (anti-Lymphocyte Activating 3, anti-LAG3) and nivolumab was approved⁶. These ICIs have demonstrated substantial benefits as systemic treatment options in the advanced, neoadjuvant and adjuvant disease settings of melanoma. However, only a subset of patients achieves long-term survival when treated with these new agents. For example, at 6.5 years the overall survival rates with ipilimumab plus nivolumab, nivolumab, and ipilimumab were 49%, 42% and 23%, respectively, as reported in the CheckMate-067 pivotal trial⁷. Furthermore, ICIs are associated with significant toxicities that are primarily immune mediated and have high economic costs. However, we have no clinical grade reliable biomarkers in the clinic to accurately predict benefits and risks. Therefore, it is paramount to develop biomarkers that can guide patients to a specific optimal therapeutic treatment based on predicted ability to respond and save patients the unwanted toxicities and costs for those with no predicted benefits.

Numerous biomarkers that predict response of melanoma to ICIs have been investigated, including potentially predictive immune related biomarker signatures generated within the tumor microenvironment (TME) and the peripheral circulation⁸. Biomarker signatures based on genomic or transcriptomic data have been especially promising. For genomics data, these predictive biomarkers include tumor mutational burden (TMB)⁹, neoantigen load¹⁰, genotypes of HLA-I^10,11, T-cell repertoire¹², aneuploidy (also known as somatic copy number alterations, SCNAs)¹³, and germline variations¹⁴. On the other hand, predictive biomarkers derived from transcriptomics data included tumor oncogene expression signatures such as genes related to MYC¹⁵, WNT/ß-catenin ^16,17 or RAS¹⁸ signaling; or gene expression profiles within the tumor immune microenvironment (TIME) such as interferon-γ (IFN-γ) responsive genes¹⁹, chemokines^20,21, major histocompatibility complex (MHC) class I and II²², cytotoxic T-cell and T-cell effector^23,24 gene expression markers that have been reported to be predictive of ICI response in metastatic melanoma. However, most of these predictive biomarkers, especially the gene expression-based signatures, have not been systematically evaluated across different cohorts to determine the reproducibility of these signatures in metastatic melanoma.

In this study, we focused our evaluation on 28 previously published predictive biomarkers of ICIs based on gene expression signatures. We included 531 patients with advanced melanoma treated with ICIs and belonging to 15 cohorts across 8 previously published studies with available RNA-sequencing data in public repositories. We further compared the performance of these predictive biomarkers between pre-treatment and on-treatment tumor biopsies in 5 patient cohorts. We found that gene expression-based signatures developed from IFN-γ-responsive genes and T-cell markers in the TME are generally predictive of ICIs responders in these patient cohorts. Conversely, gene signatures involved in tumor oncogene signaling or stroma primarily relate to tumor suppressive elements and tend to be associated with non-responders to ICIs. Finally, we also identified that these predictive biomarkers have higher predictive values, as measured by area under the receiver operating curves, in on-treatment samples as compared to pre-treatment samples in metastatic melanoma.

MATERIALS AND METHODS

Data Sets, Treatment Cohorts and RNA-seq Data

We collected eight previously published RNA-sequencing data sets including 15 cohorts of melanoma patients treated with immunotherapy from public repositories. These data sets included: Gide et al²⁵ (n = 91, four cohorts), Hugo et al²⁶ (n = 26, one cohort), Lee et al²⁷ (n = 78, two cohorts), Liu et al²⁸ (n = 122, two cohorts), Riaz et al²⁹ (n = 98, two cohorts), Van Allen et al³⁰ (n = 41, one cohort), Du et al³¹ (n = 50, two cohorts) and Lauss et al³² (n = 25, one cohort). We considered as responders, patients who had a RECIST criteria Complete Response (CR) or Partial Response (PR). Patients with RECIST criteria Stable Disease (SD) or Progressive Disease (PD) were classified as non-responders. Previously processed RNA-seq data were downloaded from the repositories (Table 1), and the quantified values were transformed as log₂(CPM+1), log₂(TPM+1), or log₂(FPKM + 1). Supplementary Table 1 provides the link to the public repositories.

Table 1:

Summary of the 15 cohorts across 8 studies used in this study.

Data Set	Cohort	Description	Treatment(s)	Pre / On	# Patients	Responders	Non- Responders	REF
Gide	Gide_Pre_PD1_CTLA4	91 Samples treated with anti PD-1 or anti PD-1 CTLA4 31 Female Samples : 60 Male samples Median age = 61	Ipilimumab Nivolumab Pembrolizumab	Pre	32	21	11	25
	Gide_Pre_PD1		Pembrolizumab Nivolumab	Pre	41	19	22
	Gide_On_PD1_CTLA4		Ipilimumab Nivolumab Pembrolizumab	On	9	5	4
	Gide_On_PD1		Nivolumab Pembrolizumab	On	9	4	5
Hugo	HugoLo_IPRES_2016	26 Samples 25 Pre-anti PD1 treatment 1 On anti PD1 treatment 8 Female Samples : 18 Male Samples Median Age = 61 M Stage 21 M1c : 3 M1b : 1 M1a : 1 M0	Pembrolizumab	Pre /On	26	13	13	26
Lee	Lee_Pre	RNA seq analysis of 78 tumors from 55 patients. 28 paired pre and on treatment biopsies and 50 unpaired biopsies.	Pembrolizumab Nivolumab	Pre	43	22	21	27
	Lee_On		Pembrolizumab Nivolumab	On	35	6	29
Liu	Liu_Naive	122 Samples 51 Female : 71 Male M Stage 91 M1c : 14 M1b : 10 M0 : 7 M1a	Pembrolizumab Nivolumab	Pre	74	31	43	28
	Liu_Prog		Pembrolizumab Nivolumab	On	48	17	31
Riaz	Riaz_Pre	Patients were stratified based on whether they had received previous ICI therapy with Ipilimumab. 98 samples were then treated with Nivolumab to characterize the genomic changes of pre and post therapy.	Nivolumab	Pre	49	10	39	29
	Riaz_On		Nivolumab	On	49	10	39
Van Allen	VanAllen_antiCTLA4_2015	Patient Details Median Age = 61 13 Female : 28 Male M Stage 29 M1c : 8 M1b : 3 M1a : 1 M0	Ipilimumab	Pre	41	8	33	30
Du	Du_Pre	42 paired pre and on treatment tumor samples. 8 unpaired pre and on treatment tumor samples.	Ipilimumab Nivolumab Pembrolizumab	Pre	19	6	13	31
	Du_On		Ipilimumab Nivolumab Pembrolizumab	On	31	5	26
Lauss	Lauss	Majority of samples obtained from lymph node or subcutaneous metastases. All patients were previously treated with other immunotherapies like IL-2 and anti-CTLA4.	Adoptive T-cell therapy	Pre	25	10	15	32

Generating Gene Expression Signature Scores

We surveyed the previously published predictive biomarkers based on gene expression signatures in the literature (Table 2 and Supplementary Table 2). For each published gene signature, we collected the gene names and harmonized using the NCBI Entrez gene number. To quantify the published gene expression score, we first transformed the gene expressions across samples in a study into a Z-score. Next, we averaged the standardized Z-score across the number of genes in the signature as previously described³³. This score is used to compare responders and non-responders of ICIs in a particular study based on the area under the receiver operating curve (AUROC) as previously described³⁴. A perfect classifier will have an AUROC = 1, whereas a random classifier will have an AUROC = 0.5. Mann-Whitney U test was performed in comparing the two groups and p<0.05 was deemed as statistically significant.

Table 2:

Summary of the 28 predictive gene signatures evaluated in this study.

Gene Signature Name	Description	Number of Genes	Reference
Angiogenesis	A signature is constructed from genes which are highly co-expressed in the angiogenesis pathway.	16	34
Chaurio	A set of genes that reflect T and B cell responses in human cancer.	7	39
Chemokine	Chemokine gene set.	12	20,21
Davoli	A signature based on scoring the immune infiltration based on gene expression profile.	7	13
effector_t	Effector T cell gene set	6	24
Glycolysis	This signature is related to the expression of glucose transporters and genes related to the metabolism of glucose.	30	34
gMDSC	The granulocytic myeloid-derived suppressor cell signature contains markers for neutrophils and chemokines which are ligands for the neutrophils.	43	34
Hypoxia	This signature was developed based on the transcriptional response of various cell types to hypoxia.	20	34
ifng-effector	IFNγ and T effector gene set	8	23
ifng18	18 gene IFNγ immune gene set	18	19
ifng6	6 gene IFNγ associated gene set	6	19
Impres	IMPRES score based on logical comparison of quantile-normalized expression of pre-defined checkpoint gene pairs.	15	48
Ipi-neo	A gene signature that reflects a proinflammatory tumor microenvironment and elements of T and B cell interplay.	32	38
MHC-I	6 Major histocompatibility complex (MHC) class I genes.	6	22
MHC-II	13 Major histocompatibility complex (MHC) class II genes.	13	22
mMDSC	The monocytic myeloid-derived suppressor cell signature contains markers of the monocytic morphology.	209	34
MYC	The Myc signature is an oncogenic signature focused on MYC pathway activation.	32	34
NRS	The neoadjuvant response signature (NRS) including genes involved in T cell activation, adaptive immune response, and T cell migration.	68	49
Ock	A gene signature predictor based around Bayesian compound covariate predictor algorithm. Stratifies patients into a responder/non-responder.	105	50
Pan	A gene panel of 18 DNA damage repair genes, developed to predict response to ICI.	18	51
Proliferation	A marker that emphasizes genes related to the cell cycle and proliferation rates in cell lines.	227	34
Ras	The Ras signature focuses on prediction of the response against PI3K and RAS inhibitors.	11	34
Roh	A gene signature that combines cytolytic markers, HLA molecules, IFNγ pathway genes, chemokines, and adhesions molecules.	41	52
Rooney	A gene signature based around the expression of two key cytolytic effectors.	2	53
Stroma	This signature is based on the infiltration of stromal cells into the tumor.	51	34
TIP Hot	Genes that serve as a marker of 'hot' (inflamed) tumors.	12	54
TLS	Tertiary lymphoid structures gene signature.	9	55
WNT	A signature developed to predict WNT mutations within human tumor datasets.	13	34

Gene Set Enrichment Analysis

Gene set enrichment analysis (GSEA)³⁵ was performed by comparing the responders versus non-responders in each data set. We used two gene sets in this study, the Hallmark gene sets obtained from MSigDB³⁶ to interrogate the enrichment of pathways in the responders versus non-responders data. In order to further deconvolute the cell types in the bulk transcriptomics, we used gene sets obtained from TIMEx³⁷ in comparing the responder vs non-responder samples in each data set. TIMEx is a novel deconvolution method for bulk transcriptomics based on scRNA-seq signatures. TIMEx currently contains 37 cell type signatures, which are classified into 4 main cell groups (malignant, immune, stromal and other), 14 major myeloid and lymphoid immune cell lineages (e.g. conventional CD4+ T cells), 15 minor immune cell lineages (e.g. naïve CD4, effector CD4) and 4 stromal cell types. We used TIMEx cell type signatures to deconvolute the bulk RNA-seq in this study. To estimate the p-value of the gene sets and cell types, we performed permutation analysis of 1000 gene sets. Gene sets with a false discovery rate q-value <0.1 were deemed as significant. We also performed GSEA using KEGG and REACTOME gene sets.

RESULTS

Melanoma Patient Cohorts treated with Immune Checkpoint Inhibitors

In this study, we analyzed 531 melanoma patients collected from 15 cohorts across 7 studies^25-31 with RNA-seq data available for patients treated with monotherapy anti-PD-1/PD-L1/anti-CTLA4 or combination of anti-PD-1 plus anti-CTLA-4. We also included one Adoptive T-cell therapy in this study³² as another immunotherapy treated cohort to test the general predictive value of the gene signatures. Among these data sets, five cohorts from four studies^25,27,29,31 have pre-treatment and on-treatment cohorts, which were investigated separately in this study. Table 1 summarizes the patient cohorts used in this study.

Performance of the Published Predictive Gene Expression Signatures in the Data Sets

Here, we collected 28 previously published predictive gene signatures that have been associated with responses to ICI immunotherapy. Table 2 summarizes the predictive gene signatures evaluated in this study. To evaluate the previously published gene expression signatures in predicting ICI responses, we tested these signatures on the 15 cohorts. We used area under the receiver operating curve (AUROC) as the metric to evaluate the predictive power of these signatures. Figure 1 summarizes the performance of the 28 published gene expression signatures on the 15 cohorts. In this evaluation, 18 predictive gene signatures achieved AUROC > 0.6; 3 gene signatures have AUROC between 0.5 to 0.6, and 7 predictive gene signatures have AUROC < 0.5.

Figure 1: AUROC performance of the 28 predictive gene signatures across 15 melanoma ICIs- cohorts.

Heatmap depicting the AUROC of the 28 predictive gene signatures evaluated across the 15 cohorts across 8 studies where columns and rows represent cohorts and signatures, respectively. The “average” column represents the average AUROC for a particular gene signature across all cohorts. The signatures are ranked by the average AUROC from the highest to the lowest predictive value.

There are seven gene signatures that performed equally well in these cohorts with average AUROC > 0.69. The IFNγ-18 gene signature¹⁹ achieved an average AUROC of 0.70 (range: 0.48 – 0.95). The remaining six gene signatures achieved the same average AUROC of 0.69 and include the IFNγ/effector T-cell signature²³ (range: 0.48 – 1), the Ipi-neo signature³⁸ (range: 0.5 – 0.95), the Davoli cytotoxic immune signature¹³ (range:0.48 - 0.95), the T-cell effector signature (range: 0.47 – 1), the Chaurio immune signature³⁹ (range: 0.47 – 0.95) and the MHC II signature²² (range: 0.47 – 0.95). No statistical significance different between these seven gene signatures in AUROC across the data sets.

From the results, we noticed that there is no correlation between the number of genes in the predictive gene expression signatures to the average AUROC in these cohorts (Supplementary Figure 1). This suggests that the content of the genes that can capture the tumor immune microenvironment is more important in the predictive power of these signatures than the number of genes included in the signature. In addition, on average, we observed that the evaluated gene expression signatures performed poorly in two cohorts: Hugo²⁶ (average AUROC = 0.44) and Liu_naive²⁸ (average AUROC = 0.50). This may be related to the heterogeneity of the patients included in these two cohorts.

Correlations of Gene Signatures

To interrogate the common themes across these predictive gene signatures, we performed Pearson’s correlation analysis of the average AUROC across the data sets. Notably, there are two main clusters from the pairwise correlation plot as illustrated in Figure 2. One cluster contains the predictive gene signatures with average AUROC ≥ 0.6 (n = 18 gene signatures), which are enriched for gene signatures involved in immune activation (e.g. cytotoxic T cells, IFNγ, chemokines and MHC I/II). The other cluster is made up by gene signatures with average AUROC < 0.6 (n = 10 gene signatures), enriched with gene signatures related to the intra-signaling of cellular cells (e.g. MYC, RAS, WNT signaling or DNA Damage Repair genes) or related to stroma (e.g. hypoxia, angiogenesis, stroma/EMT/TGF-ß signaling). The latter cluster suggests that these gene signatures predict non-responders of ICIs, potentially representing the immune suppressive elements in the tumor samples.

Figure 2: Correlation of the 28 predictive gene signatures.

Pearson’s correlation of the 28 predictive gene signatures across all 15 cohorts based on AUROC. The heatmap illustrated the two clusters, the “immune active” and the “immune suppressive” clusters.

Overlapping genes in pro-inflammatory/cytotoxic T-cell/IFNγ activation

We next investigated the overlapping genes found in the 18 gene signatures related to immune activation. There were 91 genes present in at least two predictive gene signatures (Figure 3). The most frequent gene was CXCL10 which was present in 9 predictive gene signatures. CXCL10, together with CXCL9 (present in 8 signatures) and CXCL11 (present in 4 signatures) are commonly known as the “inflammatory” chemokines, which are induced by interferon (IFN)-γ and bind to CXC chemokine receptor 3 (CXCR3) to mediate the immune response through the activation and recruitment of leukocytes such as T cells, natural killer (NK) cells and monocytes. PRF1 (perforin 1), another important player in leukocyte activation and recruitment is found in 8 of the immune activation predictive signatures. CCL5 (C-C chemokine ligand 5), also known as Regulated upon Activation, Normal T-cell Expressed, and Secreted (RANTES), another “inflammatory” chemokine in the CCL5/CCR5 axis is found in 7 predictive gene signatures. The following four genes: IFNG (interferon-γ), HLA-DRA (major histocompatibility complex, class II, DR alpha), GZMB (Granzyme B) and CD8A (Cluster of Differentiation 8A), were included in six predictive gene signatures. Notably, two cytotoxic T lymphocyte marker genes, GZMB and CD8A, were included in the predictive gene signatures. Taken together, the most common genes shared between the 18 predictive gene signatures involved genes related to inflammatory chemokines, activation of T-cells and interferon-γ signaling.

Figure 3: Common genes in the 18 immune activation gene signatures.

The top panel represents the occurrence of a gene in more than two times in the 18 immune activation gene signatures. The bottom panel shows the genes present in a particular gene signature.

Gene signatures have higher predictive power in On-treatment data sets.

Next, we evaluated the predictive power of the 18 immune activation gene signatures in five cohorts of data with pre-treatment and on-treatment patients. Interestingly, as illustrated in Figure 4, we observed that these immune activation gene signatures on average have higher predictive power for responders to ICIs in the on-treatment as compared to the pre-treatment patients. This may suggest the dynamic changes of the tumor immune microenvironment in patients after receiving therapy could assist in better predicting responses to ICIs.

Figure 4: Comparison of 18 immune activation gene signatures in pre-treatment vs on-treatment cohorts.

The top panel shows the 18 immune activation gene signatures in the five on-treatment cohorts. The bottom panel shows the same 18 immune activation gene signatures in the five pre-treatment cohorts. As illustrated in the figure, the gene signatures have higher predictive value (based on AUROC) in the on-treatment as compared to the pre-treatment cohort.

Tumors of responders to ICIs were enriched with interferon, inflammatory and immune-related gene sets

To further dissect the immune pathways involved in predicting responses to ICIs in treated melanoma patients, we performed gene set enrichment analysis (GSEA) using the Hallmark gene sets in responders versus non-responders for the five pre-treatment and on-treatment cohorts. In both the pre-treatment and on-treatment cohorts, the tumors of responders were enriched with genes related to interferon, inflammation, and immune activation. In contrast, tumors of non-responders to ICIs were enriched with genes related to MYC, oxidative phosphorylation, myogenesis, and cell cycle related genes. This is consistent with our earlier findings related to the gene signatures enriched with immune suppressive elements. Similar GSEA results were also observed by using KEGG (Supplementary Table 2) and REACTOME (Supplementary Table 3) gene sets.

Responders to ICIs were enriched with infiltrated Immune cell types

Finally, we used the TIMEx cell type signatures to deconvolute the tumor microenvironments of the melanoma samples from responders as compared to non-responders to ICIs. As illustrated in Figure 6, in both pre-treatment and on-treatment cohorts, the responders were enriched with infiltrating immune cell types in the bulk transcriptomics, especially CD8 T cells, B cells, monocytes, macrophages, dendritic cells and natural killer cells. Notably, exhausted CD8 T cells (CD8T_Ex), CD8 T cells, effector CD8 T cells (CD8T_EFF) and central memory CD8 T cells (CD8T_CM) were significantly enriched (FDR < 0.1) in all on-treatment data sets of the responders to ICIs. The enrichment of immune cell types in the responders are corroborating with the immune active gene signatures are more predictive to ICIs responses. Conversely, the non-responders were enriched with stromal related cell types such as fibroblasts and myofibroblasts. Interestingly, we observed T helper 17 (T_H17) cell type was significantly enriched (FDR < 0.1) in non-responders to ICIs across all cohorts.

Figure 6: TIMEx cell types analysis.

GSEA of the pre-treatment and post-treatment cohorts based on the TIMEx cell types. Asterisk (*) indicates gene set with FDR < 0.1. The Average column represents the average enrichment scores across the cohorts.

DISCUSSION

In this study, we evaluated 28 previously published predictive gene expression biomarker signatures for ICI benefits in 15 cohorts across 8 melanoma studies. Using AUROC as the performance metric, we found that 18 predictive gene signatures achieved AUROC > 0.6 in the tested patient cohorts. We noticed that these 18 gene signatures were enriched with genes related to immune activation including pro-inflammatory chemokines and immune cell receptors, cytotoxic T cells, IFN-γ-responsive genes and MHC I and II genes. Conversely, the other 10 gene signatures that had lower AUROC in this evaluation were enriched with tumor cell intrinsic signaling pathways such as glycolysis, MYC, RAS and WNT pathways. In addition, stroma related gene signatures also performed poorly in this study, which may suggest that these signatures are predicting the “immune suppressive” elements in the melanoma samples. For example, there is a close interaction between angiogenesis and immune suppression^40,41. Pro-angiogenic factors such as VEGF, NRP-1, PlGF are immune suppressive and negatively impact antigen presentation and effector T cell function through a variety of mechanisms including the stimulation and recruitment of regulatory T cells (Treg), myeloid-derived suppressor cells (MDSC) and tumor-associated macrophages (TAM). These suppressive immune cells can in turn drive angiogenesis and create a vicious cycle of immune deactivation^40,41. In fact, combinations of ICI and anti-angiogenic agents are being evaluated in clinical trials⁴², including in melanoma.

Upon closer examination of the genes involved in the 18 predictive gene signatures with AUROC > 0.6, we found that the “inflammatory” chemokines such as CXCL9, CXCL10, CXCL11, CCL2, CCL3, CCL4, CCL5 were common genes in these signatures. These pro-inflammatory gene expressions have been determined to play an important role in the activation and recruitment of CD8+ cytotoxic T cells in melanoma⁴³. The presence of lymphocytes was previously demonstrated to correlate with defined chemokine gene expressions, and a subset of six chemokines (CCL2, CCL3, CCL4, CCL5, CXCL9, and CXCL10) was previously validated by protein array or quantitative reverse transcription-PCR to be preferentially expressed in tumors that contained T cells, supporting the value of monitoring chemokines as potential predictors of ICI benefit⁴³. In addition, other important players in leukocyte activation and recruitment such as PRF1 were also found to be common among the predictive gene signatures. IFNγ, GZMB and CD8 reflecting cytotoxic T cell activation were also commonly present among the predictive gene signatures. Taken together, the common genes shared among the 18 predictive gene signatures were related to inflammatory chemokines, T-cell activation and interferon-γ signaling.

Due to the complexity and the dynamic nature of the TME, static biomarkers at baseline may not be sufficient to accurately predict response to ICIs. Several studies have indicated that specific immune-related changes in the TME following treatment initiation could provide an early indication of outcomes with immunotherapy^8,38. For instance, Tarhini et al. reported that ipilimumab induces massive CD8+ T cell infiltration in the TME early-on following systemic therapy. Furthermore, an immune related gene expression signature was associated with clinical benefits when tested in pre-treatment and early on-treatment specimens and across the clinical endpoints evaluated^8,38. To investigate this further, we compared the 18 immune active gene signatures in the five patient cohorts where the pre-treatment and on-treatment gene expression data were available. Overall, we observed that the predictive gene signatures have more predictive power (based on AUROC) in the on-treatment cohorts as compared to the pre-treatment cohorts. This suggests that ICIs induce immune cellular activation and recruitment to the TME and result in dynamic immune changes that could be captured and prove to be more predictive of ICI benefits than gene expression profiling at baseline. Testing early-on treatment samples could be relevant to identify patients who may have primary resistance to ICI therapy and could be saved the toxicities and costs of continued systemic therapy. Along the same lines, a recent study derived predictive pathway-based signatures in pre-treatment and on-treatment melanoma samples based on transcriptomic data and clinical information. They demonstrated that the on-treatment pathway-based signatures were more predictive than the pre-treatment pathway-based signatures in predicting response to ICIs in melanoma³¹. Similarly, there is an ongoing interest in the development of on-treatment prognostic biomarkers such as the monitoring of circulating tumor cells utilizing liquid biopsies for response to treatment and early detection of disease progression^44,45.

To further dissect the molecular pathways and tumor microenvironment in the responders and non-responders to ICIs, we performed gene set enrichment analyses in these cohorts. Similar to the predictive gene signatures results, we observed that the tumors of ICI responders were enriched with genes related to interferon-γ, inflammation and immune activation. In contrast, tumors of non-responders were enriched with genes related to MYC, oxidative phosphorylation, myogenesis and cell cycle regulation. In the immune cell type deconvolution analysis, we observed that CD8+ T cells, B cells, monocytes, macrophages, dendritic cells and natural killer cells were enriched in the tumors of responders. Among the top enriched immune cell types, exhausted CD8+ T cells (CD8T_Ex), CD8+ T cells, effector CD8+ T cells (CD8T_EFF) and central memory CD8+ T cells (CD8T_CM) were significantly enriched in all on-treatment data sets of the tumors of responders. This may suggest that ICIs may be reversing effector T cell dysfunction and exhaustion in on-treatment cohorts, thereby enhancing their anti-tumor activities in the responders. On the other hand, the tumors of non-responders were enriched with stroma related cell types such as fibroblasts and myofibroblasts, which have been previously indicated to be associated with resistance in various immunotherapies. Notably, we observed that T-helper 17 (T_H17) cell type was significantly enriched in non-responders to ICIs across all on-treatment cohorts. T_H17 cells are a subset of helper T cells whose dysregulation has been implicated in the pathogenesis of autoimmune inflammation. In addition, T_H17 cells have been shown to trans-differentiate into a more immunosuppressive phenotype that plays a role in tumor immune evasion⁴⁶. T_H17 cells have recently been demonstrated to contribute to resistance to the combination of MEK inhibitor and anti-PD-L1 in pre-clinical lung cancer models⁴⁷. Furthermore, increasing T_H17 associated genes (RORγt, IL17RA, TGFB1, and CCR5) were correlated with poor overall survival and predictive of non-response to ICIs in melanoma patients⁴⁷. This observation proposes that targeting T_H17 could improve responses to ICIs and warrant further investigation in melanoma.

CONCLUSIONS

The goal of developing a reliable and robust gene signature to predict response to ICIs remains a challenging task. Various gene expression signatures have been published, but the reproducibility and predictive power of these signatures have not been independently evaluated. In this study, we evaluated 28 predictive gene signatures across 15 cohorts of patients with melanoma treated with ICIs. We found that predictive gene expression signatures related to IFN-γ signaling, chemokines and immune cell infiltration and activation have higher predictive power for ICI response in this patient population. We also observed that on-treatment gene expression profiles have better predictive value thus capturing the dynamic changes in the TME induced by ICIs. Future biomarker development in the field of ICIs could benefit from integrating multi-omics data and has the potential to move treatment toward a more personalized therapeutic approach for melanoma patients.

Supplementary Material

tE1

Supplementary Table 1: Link to public repositories for the data.

tE2

Supplementary Table 2: Detailed gene expression signatures evaluated in this study.

tE3

Supplementary Table 3: GSEA results using KEGG gene sets in pre- and on-treatment cohorts.

tE4

Supplementary Table 4: GSEA results using REACTOME gene sets in pre- and on-treatment cohorts.

5

Supplementary Figure 1: Correlation of the number of genes in the predictive gene expression signatures and the average AUROC. No correlation was observed between the number of genes in the predictive gene expression signatures and the average AUROC evaluated in this study (p > 0.05).

Figure 5: Gene sets enriched in ICI responders.

Gene set enrichment results of the pre-treatment and post-treatment cohorts based on the Hallmark gene sets. Asterisk (*) indicates gene set with FDR < 0.1. The Average column represents the average enrichment scores across the cohorts.