DNA Methylation Signatures Correlate with Response to Immune Checkpoint Inhibitors in Metastatic Melanoma

A total of 71 patients (27 [38.0%] female and 44 [62.0%] male) with a median age of 65 years (range: 32–84 years) were retrospectively identified from the patient records of the Department of Dermatology, Medical University of Vienna, from 2015 to 2022. The inclusion criteria were as follows: (i) age ≥18 years; (ii) stage IVM1a–IVM1d metastatic melanoma according to the American Joint Committee on Cancer, 8th Edition [20]; (iii) sufficient formalin-fixed paraffin-embedded tumor tissue material prior to ICI initiation available; (iv) first-line therapy with ICIs including monotherapy with a programmed cell death protein 1 (PD-1) inhibitor (nivolumab, pembrolizumab), monotherapy with a cytotoxic T-lymphocyte associated protein 4 inhibitor (ipilimumab), or combination therapy of a PD-1 inhibitor (nivolumab) and cytotoxic T-lymphocyte associated protein 4 inhibitor (ipilimumab) and an ICI-based clinical trial study medication; (v) administration of a minimum of one cycle of ICI; (vi) radiologically measurable disease (lesion size ≥10×10 mm) prior to ICI therapy start; and (vii) availability of at least one radiological restaging under ICI therapy or the occurrence of death before the first performed radiological restaging. Patient and clinical characteristics, including age, sex, tumor mutation status (BRAF, NRAS, cKIT, or wild type), Eastern Cooperative Oncology Group (ECOG) performance status, baseline S100 and lactate dehydrogenase, and survival data (overall survival [OS], progression-free survival [PFS]) were obtained from patient records. Radiological assessment of response was defined according to the immune-based Response Evaluation Criteria in Solid Tumors (iRECIST) [38] and was performed by an independent radiologist blinded to patient outcomes. Objective response was determined by immune complete response or immune partial response, and disease control by immune complete response, immune partial response, or immune stable disease as best achieved response. Disease progression was defined as iPD and further subclassified as immune unconfirmed progressive disease and immune confirmed progressive disease. The date of immune unconfirmed progressive disease was classified as the timepoint of progression if no subsequent restaging was followed by a clear clinical progression or the occurrence of death. The date of the first radiologically confirmed progressive disease was defined as the immune confirmed progressive disease [38]. Patients with immune complete response and immune partial response were classified as responders, and patients with immune stable disease and iPD (immune unconfirmed progressive disease or immune confirmed progressive disease) were classified as non-responders. Patients who died due to early progression and received only one cycle of ICIs before the first restaging were classified as non-responders.

Tumor tissue samples from patients with melanoma before to the first cycle of ICI therapy were collected retrospectively. A total of 81 tumor tissue samples from 71 patients were included in the final methylation analysis. These were categorized into four groups: 28 lymph node metastases (34.6%), 20 organ metastases (24.7%), 24 cutaneous metastases (29.6%), and nine primary tumors (11.1%) (Table 1). Of note, in a subset of patients, 8/71 (11.3%) had more than one treatment-naïve tumor sample from different organ sites available. In detail, one patient had a lymph node and an organ metastatic lesion (12.5%), two patients had a lymph node metastatic lesion and a primary tumor tissue (25%), three patients had one lymph node and one cutaneous metastatic lesion (37.5%), one patient had a cutaneous metastatic lesion and a primary tumor tissue (12.5%), and one patient had a lymph node metastatic lesion, two cutaneous metastatic lesions, and a primary tumor tissue (12.5%) [Table S1 of the Electronic Supplementary Material (ESM)]. The 5×10-μm slides were obtained from formalin-fixed paraffin-embedded tumor tissue after macrodissection with ≥70% tumor cells, as evaluated by an independent certified dermatopathologist blinded to the patients’ clinical outcomes. Genomic DNA was extracted and treated with sodium bisulfite following standard procedures, and genome-wide DNA methylation analyses were performed using Infinium MethylationEPIC BeadChip microarrays (Illumina, San Diego, CA, USA) as previously described [39]. Raw Microarray Data (.idat files) were loaded into R (version R 4.0.4, R Foundation for Statistical Computing, Vienna, Austria) using the RnBeads package (.idat files). Single nucleotide polymorphism-associated, non-specific/cross-hybridizing, and sex chromosome-specific probes were excluded from further analysis. The SWAN algorithm was applied for data normalization [26]. Calculation of differential methylation between groups was conducted using RnBeads implementation of the limma package as well as computation of a combined rank score, which depends on the difference in mean methylation levels of two groups, the mean methylation quotient, and statistical significance. For subsequent analyses, the top 500 differentially methylated CpG sites (DMPs) were selected. To evaluate the robustness of the results, we have used a validation cohort [29]. Analyses of gene ontology (GO) were conducted using the geometh function from the missMethyl package [33] and heatmaps were generated using ClustVis [27].

Follow-up was calculated from the start date of ICI therapy until last contact or death, whatever occurred first. Descriptive statistics were used to describe patient characteristics, such as age, sex, and mutation status. For nominally scaled variables, absolute numbers and percentages were presented, and for metric variables, the mean, standard deviation, median, minimum, and maximum were applied. The Kaplan–Meier method was used to estimate PFS and OS from the first ICI application until the occurrence of disease progression or death. Patients without disease progression were censored on the date of their last visit. Kaplan–Meier curves were compared using the log-rank test and Cox’ proportional hazards models. To test the sensitivity-specificity trade-off, receiver operating characteristic classifier curve analyses were performed. The Kruskal–Wallis test and Mann–Whitney U test were used for comparisons between the groups. Statistical significance was determined by a p-value <0.05 (two-sided), multiple testing was corrected using the Bonferroni method. Statistical analyses were performed with Statistical Package for the Social Sciences (SPSS^®) 23.0 software (SPSS Inc., Chicago, IL, USA) and standard R functions (version R 4.0.4; R Foundation for Statistical Computing, Vienna, Austria). The data cut-off date was 30 August, 2022.

To investigate whether DNA methylation signatures can predict the response to ICIs, we performed Infinium MethylationEPIC microarray analyses of tumor specimens collected prior to therapy onset. Eighty-one treatment-naïve tumor tissue samples derived from 71 patients with melanoma were included in the analysis, of which 29 (40.8%) were classified as responders and 42 (59.2%) were non-responders. After quality control and exclusion of 192,085 probes owing to cross-reactivity, unreliable measurements, and sex chromosome specificity, 674,810 probes remained for a further statistical analysis. Testing for methylation differences between responders and non-responders of all tumor samples was performed using RnBeads. This resulted in a total of 58,831 DMPs with a false discovery rate <0.05 (Fig. 1a). Differentially methylated CpG sites were predominantly located within intergenic regions, gene bodies, and transcriptional start sites (Fig. 1b). Based on the top 500 DMPs, three clusters were identified revealing that hypomethylation mainly corresponded to the response to ICIs. Cluster 1 only consisted of responders (12/12), cluster 2 was evenly distributed between responders and non-responders (12/24), and cluster 3 contained mainly non-responders (39/45) (Fig. 1c). Patients in cluster 1 (12/12) and cluster 2 (23/24) had predominantly an ECOG performance status of zero, in contrast to cluster 3, where 7/45 patients had an ECOG performance status ≥1. The baseline lactate dehydrogenase (LDH) levels of most patients in cluster 1 (11/12) and cluster 2 (18/24) had baseline LDH levels within the normal range, and patients in cluster 3 had a comparably high number of patients with baseline LDH levels ≥1 upper limit of normal (13/45). A further exploratory analysis revealed that non-responders compared with responders showed a trend towards a higher ECOG status ≥1 (p = 0.0887), albeit not significant, and no differences between non-responders and responders in baseline LDH (p = 0.6646) and S100 levels (p = 0.7500) were observed. The other clusters did not differ according to factors such as the BRAF mutation status immunotherapy (IO) regimen, sex, localization of the utilized tumor tissue, mutation status, and baseline LDH and S100 values (Fig. 1c).

The predictive performance of the methylation signature was high with 80% sensitivity, 81% specificity, and an area under the curve of 0.829 (Fig. 1d). Progression-free survival and OS were significantly longer in patients from clusters 1 and 2 than in those from cluster 3 (p = 0.02 and p = 0.019, respectively) (Fig. 1e). Altogether, these results indicate that DNA methylation profiles in cutaneous metastatic melanoma correlate with the response to ICIs with high sensitivity and specificity.

A GO enrichment analysis including three sub-ontologies of biological processes (BP), cellular components (CC), and molecular functions (MF) of promotor differentially methylated regions analyzed with the missMethyl R package revealed that several pathways were significantly enriched. The most enriched pathways were related to olfactory pathways such as “detection of chemical stimulus involved in sensory perception of smell” with an enrichment false discovery rate of 6.13E−45, followed by “sensory perception of smell” and “detection of chemical stimulus involved in sensory perception” detected by GO:BP and “olfactory receptor activity” by GO:MF, which were recently demonstrated to be involved in the cell proliferation and migration processes of primary melanoma and melanoma metastasis [14]. Additionally, GO:BP and GO:MF revealed enrichments in “G protein-coupled receptor signaling pathways,” which were demonstrated to be involved in skin cancer development [32] and were shown to be engaged in melanogenesis including proliferation and migration [34]. Furthermore, GO:BP, GO:MF, and GO:CC displayed enrichments related to epidermal keratinocytes such as “keratinocyte differentiation,” “epidermis development,” “keratin filament,” and “intermediate filament cytoskeleton” (Table 2), indicative of an crosstalk between melanoma cells and the surrounding tumor microenvironment [21]. Finally, GO:BP enrichments related to B cells such as “humoral immune response” and “antimicrobial humoral immune response” were detected, which were demonstrated to promote a response to ICIs in metastatic melanoma [17].

Next, we performed differential methylation analyses of tissues derived from different organ sites. Based on the top 500 DMPs, lymph node metastasis-only samples revealed two clusters and clearly separated non-responders (17/18) in cluster 1 and responders (9/10) in cluster 2 (Fig. 2). Differential methylation analyses performed on different subgroups of patients (cutaneous metastases, primary tumors, and organ metastases) also identified two clusters containing mainly responders and non-responders (Fig. 3). In detail, an analysis of exclusively cutaneous metastases, based on 500 DMPs, identified two clusters, with the majority of responders in cluster 1 (9/10) and non-responders in cluster 2 (12/13) (Fig. 3a). A differential methylation analysis of only primary tumors, based on 500 DMPs, revealed a clear separation of non-responders (10/10) and responders (2/2); however, interpretation was limited by the low number of responders in this cohort (Fig. 3b). A differential analysis of the subset of organ metastases based on 500 mostly DMPs identified two clusters, with responders being mainly in cluster 1 (6/6) and non-responders preferentially in cluster 2 (10/12) (Fig. 3c). In a subset of patients, 8/71 (11.3%), more than one treatment-naïve tumor tissue sample was available (Table S1 of the ESM), and methylation profiling of all samples was performed accordingly. In 5/8 patients, the samples from different body sites clustered together (Fig. 4), whereas in 3/8 patients, an intra-patient different methylation pattern was observed. Irrespective, in-depth analysis per tissue methylation profiling showed that all samples correlated with response to ICI, irrespective of the tumor origin (Fig. 1c). These results support that tissues derived from different organ sites do not affect the DNA methylation-based prediction of ICI responses.

We aligned our findings to an independent data set. This validation cohort consisted of 43 patients with tumor tissue samples, of which 26 patients were classified as responders and 17 as non-responders [29]. Based on the 500 differentially methylated regions from our cohort, we identified similar trends (p = 0.146). In detail, we identified three clusters, showing responders mainly cluster 2 (15/20, 75%) and non-responders in cluster 3 (5/8, 62.5%), with hypomethylation being more prevalent in responders (Fig. S1 of the ESM).