The ICI cohort comprised cutaneous metastases of AJCC stage IV (eighth edition) melanoma patients treated in three different dermato-oncology centers (sites I–III) in Germany. Formalin-fixed and paraffin-embedded (FFPE) tissue samples (deriving from the biobank of the Department of Dermatology, Goethe-University, Frankfurt am Main, Germany, from the Institute of Pathology, Charité—Universitätsmedizin Berlin, Berlin, Germany and from the Department of Dermatology, University Hospital Würzburg, Würzburg, Germany) treated with ICIs between October 2010 and December 2020 were included. Clinical outcome was measured as survival under ICI therapy (defined as the time from start of ICI treatment to date of decease) and OS (defined as the time from primary diagnosis of melanoma to date of decease). Additionally, radiological response to ICI treatment was classified by Response Evaluation Criteria in Solid Tumours for immune-based therapies (iRECIST: iCPD, confirmed progressive disease; iUPD, unconfirmed PD; iSD, stable disease; iPR, partial response and iCR, complete response). Hereby, patients with disease control (DC, defined as iCR, iPR or iSD in at least one ICI treatment approach) were discriminated from patients with progressive disease (PD, defined as iUPD or iCPD). We additionally included the following parameters for each patient: epidemiological characteristics (sex, age), BRAF and NRAS mutation status and brain metastases status (yes vs no) (detailed information on the patient cohort and study workflow is depicted in figure 1 and online supplemental table 1). Data not passing well-established quality controls were excluded. In case of multiple samples per patient at site I, we selected the representative sample per patient according to the minimum euclidean distance of its methylation profile to the median profile (figure 1). Finally, the ICI cohort consisted of 65 samples from 65 patients (from sites I–III) of which 61 patients received complete radiological workup according to iRECIST (figure 1). The study protocol was endorsed by the local ethical committee (SNO-5-2019).

Clinical data of initially 470 patients suffering from melanoma were retrieved from The Cancer Genome Atlas (TCGA) data archive (TCGA SKCM). Clinical annotations included sex, age, AJCC tumor stages and site of tumor resection/biopsy. As the information on exact AJCC tumor stages (A, B, C and D) was missing in 98 cases, we annotated the simplified AJCC stages I–IV. Comprehensive data on ICI treatment or radiological therapy response were not eligible for this cohort. Corresponding raw intensity data (IDAT) files generated by use of the 450 k Human Methylation Array (Illumina, San Diego, USA) were retrieved using the TCGA Genomic Data Commons tool. Data not passing well-established quality controls, patients with incomplete clinical annotations as well as melanoma in situ patients of AJCC stage 0 were excluded from the study resulting in 396 patients to set up the so-called TCGA melanoma cohort.

While TCGA data were available as 450 k Human Methylation array IDATs, for the ICI cohort of stage IV melanoma cutaneous metastases, we used the Human Methylation EPIC array (Illumina, San Diego, USA) comprising over 850 000 CpG sites by analysis of representative tissue punches or macrodissected 10 µm thick tumor slides from FFPE blocks. We followed standard protocols for tissue and DNA processing. Further processing of the microarrays and hybridization as well EPIC beadchip scanning were performed as indicated by the manufacturer.

DNA methylation data was obtained as IDAT files, which were used as input to the RnBeads software package. Quality control was performed using the built-in control probes on the EPIC array, and the data showed high overall quality. Furthermore, CpGs were filtered according to detection p-values, and annotated single nucleotide polymorphisms, sites on the sex chromosomes and potentially cross-reactive sites were discarded from the analysis.26 Methylation data were normalized using the “dasen” method from the wateRmelon R-package.26 TCGA methylation data were processed analogously.

In a first approach, all eligible CpG sites passing quality controls in 65 tumor bulk samples were enrolled in principal component analysis (PCA). In addition, differential methylation analysis for all eligible CpG sites using RnBeads was performed comparing patient groups with different clinical characteristics like the melanoma subtype, BRAF or NRAS mutation status, absence or presence of brain metastases or radiological DC versus PD according to iRECIST. Raw data IDAT files as well as processed data of the ICI cohort are accessible via Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/; GSE175699).

DNA methylation data of the bulk tumor samples of the TCGA melanoma cohort and the ICI cohort were investigated using the reference-free MeDeCom algorithm that dissects DNA methylation data into major components of variation, called latent methylation components (LMC).27 DNA methylation data of melanoma patients were processed according to a recently published protocol.28 The protocol selected the 5000 most variably methylated CpG sites across the samples as input to MeDeCom. Investigation of the cross-validation error and of the objective value for the parameter number of LMCs (kappa) and the regularization parameter (lambda) were performed, resulting in a set of LMCs. In order to prevent a strong dependence of the clustering on the LMC with the highest proportion across the samples, we decided to standardize the LMC proportions using z-scores. This allows for uncovering fine-grained changes in the LMC proportions and identification of subtle differences between the samples. LMC proportion values were standardized by subtracting the respective column mean and dividing by the column SD. Standardization was performed for LMCs 1–8 of all samples collectively (TCGA and ICI cohort), and separately in the TCGA and ICI cohort. Standardized LMC proportion-based clusters were further investigated regarding their prognostic and predictive significance. Hierarchical cluster analysis was performed using Ward’s minimum variance method.

To estimate the leukocyte content in the bulk tumor samples, the leukocyte unmethylation for purity (LUMP) algorithm as implemented within RnBeads was employed. LUMP provides leukocyte ratios from the DNA methylation data by screening for 44 CpG sites particularly hypomethylated in leukocytes.29 30

For a detailed deconvolution of the cellular composition of bulk melanoma samples, we applied the reference-based analysis MethylCIBERSORT. MethylCIBERSORT relies on DNA methylome-based reference data to infer distinct cellular contents (cancer cells, CD14-positive, CD19-positive, CD56-positive and CD8-positive cells, T regulatory cells, CD4-positive effector cells, eosinophils, fibroblasts and neutrophils). Methylation patterns of the sample of interest are compared with deposited cell-type-specific and determinating DNA methylomes.31 32 MethylCIBERSORT analysis was carried out according to the respective protocols.32 Briefly, EPIC array IDAT sets were imported in R’s “minfi” package to perform quality checks, Noob normalization and acquisition of beta values. Using the “MethylCIBERSORT” R package, a mixture file was built whose matrix consisted of beta values for comparison to a reference matrix (provided by TRF). This reference file contained signature methylation beta values of well-characterized cell types. After generating the mixture file, mixture and reference files were uploaded onto the CIBERSORT portal and deconvoluted (provided by the Alizadeh Lab, Stanford University, USA, developed by Newman et al31).

A patient classifier based on the 5000 most variable methylated CpGs selected for MeDeCom tumor deconvolution was developed for patient allocation to predictive LMC proportion-based clusters. Based on these CpG sites, we employed logistic regression using Lasso regularization (glmnet R-package)33 to predict whether a tumor sample would belong to the beneficial LMC proportion-based cluster or to the less beneficial cluster. Finally, this shrinkage led to 20 predictive CpGs. We employed nested 10-fold cross-validation to estimate model performance and the hyperparameter lambda simultaneously.

All statistical analyses were conducted using either JMP16 (SAS, Cary, USA) or R (R Core Team, 2019). The non-parametric Wilcoxon’s test was applied for comparisons of data distributions between two groups. Kaplan-Meier survival curves were compared by log-rank and Wilcoxon’s test. Likelihood ratios and risk ratios were computed in univariate proportional hazard models. Hierarchical clustering was performed by use of the Ward’s minimum variance method.

To identify prognostic DNA methylation signatures in melanoma patients, we first performed global approaches for dimensional reduction of large-scale DNA methylome data. As a next step, we used reference-free and reference-based tumor deconvolution analyses, which allowed for development of a patient classifier and a biological interpretation of results (figure 1). The TCGA melanoma cohort comprised 396 tumor samples located in skin, soft tissue, central nervous system, peripheral, non-central nervous system organs and lymph nodes, respectively. Tumors from all AJCC tumor stages I–IV were included (figure 1). For validation of the TCGA cohort statistics, we performed Kaplan-Meier survival analysis which revealed that the applied simplified AJCC stages I–IV were prognostic (figure 2A). The ICI cohort included tissue from cutaneous metastases of 65 patients with AJCC stage IV melanoma treated with ICI therapy exclusively. Regarding the mutational status, 45.3% of the primary tumors showed a BRAF mutation, while 25.6% were NRAS mutant. Brain metastases were observed in 40% of all patients (online supplemental table 1). We assessed OS, survival under ICI treatment as well as the radiological disease response assigning patients to either DC or PD. The Kaplan-Meier analysis of the survival under ICI therapy stresses a benefit in patients responding to ICI therapy according to the iRECIST criteria, thus proving cohort stringency (figure 2A).

In a first approach, we assessed global DNA methylation patterns including all eligible CpG sites that might differ between subgroups of the well-characterized cohort of stage IV melanoma patients treated with ICI. Using PCA neither BRAF/NRAS mutational status, brain metastases status, nor radiological response status showed a clear association to the principal components (online supplemental figure 1A). Accordingly, differential methylation analysis did not reveal major differentially methylated CpGs (online supplemental figure 1B). This emphasizes that there is no global DNA methylation state associated with any of the available clinicopathological sample annotations.

Since the initial global DNA methylation analyses (PCA and differential methylation) of the stage IV melanoma cohort did not reveal an association with ICI therapy, we aimed for a more refined yet profound and reference-free algorithm to assess potential prognostic and predictive methylation signatures. Therefore, we first applied the MeDeCom algorithm to the bulk tumor DNA methylation data of the combined TCGA and ICI cohort (parameter selection is depicted in online supplemental figures 2). Overall, the cross-validation error pointed to selecting eight components (parameter k) and the regularization parameter λ as 0.001. The proportions of the eight LMCs (LMC1-8) across the samples are visualized in figure 2B. Hierarchical cluster analysis of LMC proportion Z-scores of the entire cohort (TCGA+ICI) did not reveal clear clustering according to the cohort (TCGA vs ICI) and clusters did not show a clear association with AJCC tumor stages (figure 2B). In a next step, we analyzed the prognostic impact of LMC proportion-based clustering on OS of the TCGA cohort. Cluster analysis revealed two main clusters (1 vs 2, figure 3A). Although, patients of cluster 1 showed a trend toward a better OS in the total TCGA cohort as well as in AJCC stages II and IV, log-rank and Wilcoxon’s test did not indicate statistical significance (figure 3B). MeDeCom is a reference-free deconvolution algorithm of DNA methylation data, which generates LMCs and their proportions and follows dominant cell types. As we were interested in distinct cellular composition of LMC proportion-based clusters and as MeDeCom does not compare data to cell-type specific reference methylation profiles, we further investigated LMC proportion-based clusters 1 and 2 using the reference-based algorithms LUMP and MethylCIBERSORT.

Interestingly, the cancer cell fraction was significantly lower in the rather beneficial cluster 1, whereas immune and stromal cells were strongly increased in cluster 1 (figure 3D and online supplemental figure 3A,B). Taking a closer look at immune cell subtypes, especially CD8-positive T cells, B-cells, NK cells but also myeloid CD14-positive cells were enriched in cluster 1. On the contrary, CD4-positive T cells were significantly reduced in cluster 1. In summary, this indicates that LMC proportion-based clusters might be of prognostic relevance in certain subgroups of melanomas in a heterogenous melanoma cohort. However, a precise interpretation of LMCs with regard to immune biological processes remains elusive.

In search of predictive signatures of large-scale DNA methylation data, we further investigated cutaneous metastases of stage IV melanoma patients receiving immune checkpoint inhibition in three dermato-oncology centers in Germany. Based on our initial MeDeCom analysis which decomposes DNA methylation data of the total cohort into eight LMCs, we performed LMC proportion-based unsupervised hierarchical cluster analysis in the ICI cohort only (figure 4A). Hierarchical cluster analysis depicted two major clusters with an enrichment of patients with DC in cluster 2 (figure 4A,B). In line with this, survival calculated from the initiation of ICI treatment was significantly longer in cluster 2 patients (figure 4C) and cluster 2 was predictive for survival from initiation of ICI treatment in univariate analysis (figure 4D). The median duration of ICI treatment as an indirect marker for DC and treatment tolerability accounted for 183 days in the beneficial cluster 2 as compared with 90 days in cluster 1 (Wilcoxon’s non-parametric test p=0.03).

Additional MethylCIBERSORT facilitated a microenvironmental analysis of the tumor samples allocating to the predictive LMC-based clusters in the ICI cohort. While the cancer cell fraction, as well as most immune cell subsets did not show significant differences between the beneficial cluster 2 and cluster 1, only regulatory T cells (Tregs), CD56-positive (NK) cells and fibroblasts were depleted in cluster 2 (figure 4E and online supplemental figure 3). Timing of biopsy (before vs under ongoing ICI treatment) did not have a substantial impact on the cellular composition of tumor bulks with solely the fraction of CD56-positive NK cells being enriched after initiation of ICI treatment (online supplemental figure 4). In summary, MeDeCom analysis unraveled predictive signatures for DC in metastatic melanoma being treated with ICI.

Finally, the findings of reference-free methylome analyses in the ICI cohort were integrated to develop a patient classifier with the aim to allocate patients to the predictive LMC proportion-based cluster. Therefore, we used the 5000 CpGs computed by MeDeCom as an input for a logistic regression analysis. Lasso regularization revealed 20 of these 5000 CpG sites to predict sample allocation to the less favorable cluster 1 or the favorable cluster 2 of the ICI cohort (figure 4F). By means of this classifier model, we achieved a cross-validated accuracy of 89.5% as well as an area under the curve of the receiver operating characteristic curve of 0.9664 in our cohort of melanoma stage IV patients under ICI treatment (figure 4G).