Mapping heterogeneity in patient-derived melanoma cultures by single-cell RNA-seq

Recent technological advances in single-cell genomics make it possible to analyze cellular heterogeneity of tumor samples. Here, we applied single-cell RNA-seq to measure the transcriptomes of 307 single cells cultured from three biopsies of three different patients with a BRAF/NRAS wild type, BRAF mutant/NRAS wild type and BRAF wild type/NRAS mutant melanoma metastasis, respectively. Analysis based on self-organizing maps identified sub-populations defined by multiple gene expression modules involved in proliferation, oxidative phosphorylation, pigmentation and cellular stroma. Gene expression modules had prognostic relevance when compared with gene expression data from published melanoma samples and patient survival data. We surveyed kinome expression patterns across sub-populations of the BRAF/NRAS wild type sample and found that CDK4 and CDK2 were consistently highly expressed in the majority of cells, suggesting that these kinases might be involved in melanoma progression. Treatment of cells with the CDK4 inhibitor palbociclib restricted cell proliferation to a similar, and in some cases greater, extent than MAPK inhibitors. Finally, we identified a low abundant sub-population in this sample that highly expressed a module containing ABC transporter ABCB5, surface markers CD271 and CD133, and multiple aldehyde dehydrogenases (ALDHs). Patient-derived cultures of the BRAF mutant/NRAS wild type and BRAF wild type/NRAS mutant metastases showed more homogeneous single-cell gene expression patterns with gene expression modules for proliferation and ABC transporters. Taken together, our results describe an intertumor and intratumor heterogeneity in melanoma short-term cultures which might be relevant for patient survival, and suggest promising targets for new treatment approaches in melanoma therapy.


INTRODUCTION
Large-scale mutation analyses of malignant melanoma in the past identified mutations in the BRAF oncogene in 50% of melanoma samples [1,2]. The most prevalent BRAF V600E missense mutation leads to an activation of the classical mitogen-activated protein kinase (MAPK) pathway. Targeted treatment of metastatic melanoma patients using small molecule inhibitors such as vemurafenib, dabrafenib and encorafenib directed

Research Paper
Oncotarget 847 www.impactjournals.com/oncotarget against activated (mutated) BRAF kinase has shown promising results in recent years, significantly improving overall survival of affected patients [3]. However, a significant number of patients show primary resistance, and recurrences under inhibitor treatment occur as secondary resistance in the vast majority of cases. Recent studies have shown that combination treatments of BRAF and MEK1/2 inhibitors are significantly more effective than BRAFinhibitor treatment alone [4]. However, 50% of patients develop a secondary resistance after 6-9 months [5].
There are a series of mechanisms described that underlie the secondary resistance of BRAF-mutant melanomas that occur after BRAF inhibitor treatment, including NRAS mutations, aberrant BRAF splicing, BRAF amplifications, MAP2K1 (MEK1) mutations, PTEN and PIK3CA mutations, and COT1 overexpression [6,7]. In addition, mechanisms of primary treatment resistance of BRAF-mutant melanoma cells may be due to a MITF low/NF-κB high phenotype, which could be linked to a specific gene expression profile [8]. These results suggest that primary and secondary resistance mechanisms may be either due to genetic changes (mutations, amplifications) or changes in gene expression of specific pathways.
It has been suggested that recurrences and treatment failures may derive from intratumor heterogeneity [9]. That is, multiple subclonal mutations, gene expression patterns or epigenetic mechanisms may be present in tumor lesions and create a genetically heterogeneous population of tumor cells. Here, we analyzed the intratumoral heterogeneity in three short-term cultures derived from three different patients with metastatic malignant melanoma using single-cell RNA-seq. We used a comprehensive analysis and visualization strategy based on self-organizing maps (SOM) machine learning which is called 'high-dimensional data portrayal' because it visualizes the gene expression landscape of each individual cell. As a clustering method, SOMs offer several advantages compared with alternative methods such as non-negative matrix factorization, K-means, hierarchical clustering or correlation clustering [10]. By this means we identified gene expression patterns that may be useful for designing new treatments targeting tumor sub-populations.

Gene expression portraits of single-cell transcriptome heterogeneity in a BRAF/NRAS wild type melanoma sample
We applied microfluidic single-cell RNA-seq to measure the transcriptome of 92 single cells obtained from a BRAF/NRAS wild type melanoma short-term culture (Ma-Mel-123). In order to rule out intermixture of benign non-melanoma cells, we inferred largescale copy number variations (CNVs) from expression profiles by averaging gene expression over stretches of 50 genes on their respective chromosomes (Supplementary Figure S1). Data are shown as heatmap and revealed extensive copy number variations as a typical feature of cancer cells, basically ruling out an intermixture of benign cells such as fibroblasts.
For analysis of subpopulations, we used selforganizing map (SOM) machine learning which bundles a series of sophisticated downstream analysis tasks such as gene module selection, sample diversity clustering and functional knowledge discovery [11]. Its performance was previously demonstrated in different studies on cancer heterogeneity [12,13].
SOM classified the cells into three major groups as proliferation, pigmentation and stromal type ( Figure 1A; Supplementary Figure S2) according to the major gene categories represented in each group. The majority of the 92 cells (n = 42) were defined by genes involved in processes of cellular proliferation such as DNA replication, DNA repair, chromosome segregation and mitosis [14]. The pairwise correlation map shows that the expression landscapes of group 1 virtually anti-correlates with those of groups 2 and 3 ( Figure 1B). We identified four main clusters of co-expressed genes which were called spot-modules A-D ( Figure 1C, 1D; Table 1;  Supplementary Table S1).
Spot A, upregulated in cells of group 1, is enriched with genes related to cell cycle and cellular proliferation ( Figure 1C). Spot B is enriched with genes involved in energy metabolism and oxidative phosphorylation (Oxphos) which are concertedly upregulated in cells of group 1 and group 2. Spot C is enriched with genes related to pigmentation and melanocyte differentiation. Spot D is enriched with genes related to tumor stroma. For further details, see also Supplementary Figures S2, S3. A closer inspection of the single-cell expression portraits in the different groups revealed a fine structure ( Figure 1C, Supplementary Figure S4). E.g., expression portraits of group 1 can be separated into cells showing characteristic patterns of the G2 phase of cell cycle (subgroup SG1a, 13 cells) and G1S-phase (SG1c, 7 cells).
To obtain a more detailed view on the biological context we mapped selected gene sets of different categories into the SOM and characterized them in terms of gene set-maps and GSZ-profiles (Supplementary Figure S5). Interestingly, genes from G1 and G2 phase of the cell cycle accumulate in close proximity (Supplementary Figure S5A). The gene sets oxydative phosphorylation (Oxphos) and melanocyte differentiation accumulate in spots B and C (Supplementary Figure S5B). In addition to the stromal signature in group 3 cells, we identified a cell population with enriched ABC-cassette membrane transporters which is described in detail below (Supplementary Figure S5C). In order to demonstrate that different cellular subpopulations identified in the single-cell analysis are also present in the original tumor sample, immunofluorescence stainings were performed of the original frozen tumor material (Ma-Mel-123 tissue). Antibodies used were directed against proliferation markers TOP2A and Ki-67 (spot module Proliferation; Table 1 and Supplementary  Table S1), ITGA1 (spot module Stroma; Table 1;  Supplementary Table S1), and CDK4 (highly expressed in all cells). As shown in Supplementary Figure S6, the mentioned proliferation markers showed disseminated staining of individual cells within the tumor tissue with 30-40% of positively staining tumor cells. Staining of ITGA1 showed a lower percentage of staining of 15-20%  up(+)/down(-) regulated in group. b) Gene sets refer to the categories assigned in the methodical part. Only sets with p values < 10-7 (Fishers exact test) were taken into account. Literature sets were taken from Ref. 31,62-68. c) Genes are ranked with decreasing correlation coefficient with respect to mean spot expression profile; full gene lists for each spot are provided as Supplementary Tables (Supplementary Tables S1, S2). ABC transporter and aldehyde dehydrgenases were grouped together.

d)
Kinases are selected using the criteria‚ high and almost invariant expression as described in the Supplementary material. MMIC markers were taken from independent studies (summarized in [23]). Melanoma key genes (MKG) collect genes directly related to melanoma. Oncotarget 850 www.impactjournals.com/oncotarget of individual cells but also stained endothelial cells lining microvasculature. CDK4 expression was found in the cytoplasm and nucleus and was homogeneously distributed over the whole sample (Supplementary Figure S6). These findings are in agreement with the percentages of cell subpopulations of the single-cell analysis (45% of cells with the proliferation signature; 27% of cells with the stromal signature). Thus, by analysis of a selected set of marker proteins we could validate the single-cell data in vivo, which argues for a preservation of cell subpopulations in tumor tissues in tumor-derived short-term cultures.
An independent single-cell RNA-seq study on different melanoma tissues confirms cellular subpopulations found in a melanoma short-term culture Gene sets as determined in a recent single-cell transcriptome analysis on 19 melanoma tissues were mapped into our single-cell RNAseq data to study whether expression characteristics of these experiments may be identified in our data sets [15]. As shown in Supplementary Figure S7, the heatmap of the GSZ score of the different melanoma gene sets of the mentioned study revealed strong correlations with our data in terms of spot enrichment, expression profiles and group assignments. The group of proliferative cells (group 1 of our study) showed high expression of cell cycle signature genes (principal component 2 (PC2)) from the study of Tirosh and co-workers [15] (Supplementary Figure S7A). Group 2 cells of our study show high expression of MITFand melanoma-signature genes (PC4 of the study of Tirosh and co-workers) [15]. AXL signature genes were highly expressed in our group 3 (stromal genes). The right part of the Supplementary Figure S7A shows selected GSZprofiles and gene set maps which illustrate accumulation of the genes in spots A-D in a set-specific manner showing that the melanoma cells studied by us and by Tirosh and co-workers show similar activation patterns of different cellular programs. Moreover, cells with an activated KDM5B (JARID1B) signature refer to group 2 or 3 of our study and thus to low cycling cells in agreement with their study [15]. AXL high expression co-regulates with our stromal signature (Supplementary Figure S7A). Slowcycling cells of our study divided into two populations according to the activity of MITF-and AXL-programs (Supplementary Figure S7C, S7D). A similar division was also shown for individual samples in the study of Tirosh and co-workers [15]. Overall, gene sets referring to PC1 to PC5 of the study by Tirosh and co-workers agree with our spot and group characteristics and thus support our classification scheme (Supplementary Figure S7B, S7C). Taken together, the cellular subpopulations in the present study show large overlaps with the mentioned study on gene signatures active in vivo.

Single-cell transcriptome data reflect prognostic expression signatures of melanomas in situ
Gene expression patterns in the spot modules of the present study were compared with prognostic gene patterns observed in melanoma samples of independent studies [16][17][18][19][20] (Figure 2). In the study of Jönsson and co-workers on metastatic lesions, the proliferative subtype had the shortest overall survival compared to all other groups [16]. In the study of Harbst and co-workers on primary melanomas the gene signatures of proliferation and pigmentation combined defined a high-grade class of melanomas with poorer overall survival compared to other subgroups [17]. The study of Kaufmann and coworkers and Winnepenninckx and co-workers found different gene signatures in patient material of patients with metastasis (bad prognosis) or no metastasis in a 4-year follow-up period [18,19]. The study by Gerami and co-workers provided a meta-analysis of the mentioned studies [20]. Each of the signature gene sets associated with poor prognosis were strongly upregulated in cells of either group 1 or 2 ( Figure 2A, 2B). Group 1 genes are upregulated in early relapse melanomas, highly metastatic and high-grade melanomas. Interestingly, the gene set maps reveal that the signature sets shift with respect to their cell cycle characteristics from the G2 to the G1S pattern between the different prognostic melanoma types ( Figure 2C). These findings show that single-cell gene expression patterns from an in vitro analysis may identify cell subsets with impact on the patients' prognosis.
Expression of top 100 genes from each spot-cluster (A-D) was analyzed in a further independent data set [21] ( Figure 2D). Our cell signatures of spot B (Oxphos) and C (Pigmentation) show maximum expression in high grade primary melanomas whereas spot signature A (Proliferation) shows strongest expression in highgrade and metastatic melanomas ( Figure 2D). Spot D (Stroma) shows downregulation during tumor progression [21]. Hence, high similarities of gene activity of our proliferation signature were found in high grade primary and metastatic tumors.

Mapping of the kinome identifies signaling kinases as potential targets for targeted treatment of the BRAF/NRAS wild type melanoma
We next explored the cellular kinome heterogeneity across the cell groups [22]. Genes of different kinase families were mapped to the different spot-modules (Supplementary Figure S8). Most of the candidate genes identified were co-expressed with group-specific spot signatures/modules (Table 1). Interestingly, the cell cycledependent kinases CDK4 and CDK2 were consistently highly expressed in the vast majority of cells analyzed with CDK4 showing the most prominent expression in the Oncotarget 851 www.impactjournals.com/oncotarget  Figure S8). In addition, KIT was found to be among the top ranked kinases (Supplementary Figure S8). CDK4 was also part of the PC1, and CDK2 was part of the MITF and melanoma signature in the study of Tirosh and co-workers which is suggestive for a role of both kinases in this study [15].

CDK4 inhibitor palbociclib reduces melanoma cell proliferation
To test the impact of CDK4 expression on treatment response of the BRAF/NRAS wild type melanoma cells, different small molecule inhibitors were used targeting MAPK, CDK2 and CDK4 pathways. After 72 h of exposure, CDK4 inhibitor palbociclib reduced melanoma cell proliferation of Ma-Mel-123 cells with significantly higher efficiency than BRAF inhibitor vemurafenib (p = 0.0017) ( Figure 3B). The activity of MEK1/2 inhibitor cobimetinib was comparable to that of palbociclib (p = 0.80). Significant higher activity was also obtained for palbociclib in comparison to CDK2 inhibitor purvanolol but not against ERK1/2 inhibitor SCH772984 (p = 0.14).
Control experiments were carried out with benign fibroblasts (48 h Table 1; Supplementary Table S2). The co-expression of ABC-transporters with ALDHs supports their possible role in resistance against chemotherapy and tumor stem cell activity [23][24][25]. Mapping of this signature into the melanoma data set of Raskin and coworkers shows activation in early stages of melanoma progression (Supplementary Figure S9B) [21].

MMIC signatures associate with melanoma single-cell expression
Another MMIC-marker is KDM5B (JARID1B), a demethylase of H3K4me3. KDM5B (JARID1B) is often found in association with weakly transcribed genes which demethylate their gene bodies during embryonic stem cell self-renewal [26][27][28]. In melanoma, KDM5B (JARID1B) is regarded as a marker of slow-cycling melanoma cells which are essential for long-term tumor growth [29] Our KDM5B (JARID1B) signature clearly reflects decreased gene activity in group 1 cells (Proliferation) ( Figure 4C). KDM5B (JARID1B)-positive cells are assumed to form a side population that displays characteristics of enriched chemoresistance and tumorigenesis [30]. About 7% of our cells mainly of group 3 overexpress this KDM5B-signature ( Figure 4C).
Taken together, cells which could possibly be candidates for MMICs are those with the ABC transporter and ALDH gene expression signatures and those with KDM5B expression.

Modes of epigenetic regulation
Next, we analyzed a series of embryonic stem cell (ESC) and neural progenitor cell gene expression signatures characterizing poorly differentiated tumors [31][32][33] and signatures of 15 chromatin states in fibroblasts, melanocytes (MCs) and neural progenitor cells (NPCs) [34] (Supplementary Figure S10). So-called poised and repressed genes are regulators for fate decisions of embryonic stem cells ensuring their pluripotency. We find that the expression signatures of poised and repressed genes in ESCs, MCs, NPCs and partly also in fibroblasts show similar profiles as the ABC-and the KDM5Bsignatures identified above (Figure 4, Supplementary  Figures S10, S11).
Deregulation of these states due to epigenetic mechanisms is associated with cellular reprogramming and cancer stemness [35,36]. Changes of expression-inactive chromatin states (poised, repressed and silent chromatin) were especially observed for the ABC-transporter and KDM5B-positive cells.
Finally, we systematically studied the gene expression levels of more than fifty chromatin modifying enzymes in the single cell data (Supplementary Figure S12). Cells of the stromal type (group 3) show a high activity level of enzymes that remove methylation marks from DNA and histone lysine side chains. Among them is a high fraction of demethylases of H3K9me3. This result together with the finding that the DNA demethylases TET1 and TET3 are upregulated in ABCtransporter type cells suggests a deregulation of the DNA-methylation machinery with consequences for DNA-methylation. In summary, we find that epigenetic modes of transcriptional regulation such as chromatin www.impactjournals.com/oncotarget remodeling, DNA methylation changes and modulation of the expression of chromatin modifying enzymes are potential regulators of different functional states in melanoma.

Single-cell transcriptome patterns of BRAF mutant/NRAS wild type and BRAF wild type/ NRAS mutant short-term cultures
In order to further substantiate our findings, a SOM analysis of an experimental replicate of the BRAF/NRAS wild type culture was performed, together with a t-distributed stochastic neighbour embedding (tSNE) analysis. As shown in Figure 5A, 5B and Supplementary Figure S13A, the cells of the replicate analysis clustered together with those of the first analysis revealing no major transcriptional differences between independent experiments of the same short-term-culture after an additional passaging. The cells showed similar percentages of cells within each group (group 1: 46% vs 45%; group 2: 27% vs 23%; group 3: 27% vs 32%) ( Figure 5B; Supplementary Figure S13; Table 2; Supplementary Table S3). Oncotarget 854 www.impactjournals.com/oncotarget Furthermore, two short-term cultures of two additional patients were analyzed ( Figure 5). These two cultures consisted of BRAF mutant/NRAS wild type and BRAF wild type/NRAS mutant cells, respectively. As shown in Figure 5A, 5B and Supplementary Figure S13 Table 2). BRAF/NRAS wild type cells were characterized by a gene signature which combines the oxphos and pigmentation signatures. A subset of signature genes of the stromal cells found in the wild type short term culture (Group 3) was also uniquely expressed in these cells ( Figure 5D). In comparison to both others, the expression signature of BRAF mutant/ NRAS wild type cells was governed by large-scale copy number variations evident in chromosomal mapping of the expression data ( Figure 5C, Table 2; Supplementary Figure S1). The BRAF wild type/NRAS mutant cells predominantly up-regulated genes associated with the stromal signature in a functional context different to the one found in the wildtype short-term culture ( Table 2; Supplementary Figure S13). An overlapping gene expression signature between different cultures was found for genes involved in cellular proliferation. A fraction of 30% to 45% (45% in BRAF wt/NRAS wt, 30% in BRAF mut/ NRAS wt and 36% in BRAF wt/NRAS mut) of cells from each culture was highly proliferative and showed a common proliferation signature independent of their culturespecific expression patterns ( Figure 5D). This common proliferative signature was almost identical between the mutant cells and double wild type cells and also with the cell cycle signatures described by Tirosh and co-workers [15]. Additionally, it provided expression profiles that closely resembled those of metastatic, relapsed and highgrade melanomas ( Table S3).
The single-cell study of Tirosh and co-workers detected an antagonism of MITF-and AXL-related transcriptional programs. In the present analysis, NRASmutant cells activated the AXL program and BRAF/NRAS wild type cells the MITF program which also resembles the melanoma pigmentation subtype described earlier [16] (Supplementary Figure S7D; Supplementary Figure S13). Detailed analysis of the BRAF/NRAS wild type data revealed differential activation of the MITF-and AXL-programs in groups 2 and 3 of our data (Supplementary Figure S7).
Taken together, we could show that the same expression profiles were identified when performing

DISCUSSION
Intratumor heterogeneity based on a complex subclonal molecular structure in malignant tumors has been regarded as a major factor that influences therapeutic response and secondary treatment resistance [9,37]. In order to further estimate intratumor heterogeneity and the subclonal structure of malignant tumors, single-cell sequencing has been used for the analysis of a series of tumor cell lines, solid tumors and haematological malignancies in recent years [37][38][39]. In the present study, three different patient-derived melanoma cell cultures Oncotarget 857 www.impactjournals.com/oncotarget were analyzed. Intercellular (intratumor) heterogeneity was largely driven by genes involved in proliferation, stroma and MITF/AXL programs. The proliferation signature showed strong overlaps between the different cultures.
It is known that cell cycle control, cellular proliferation and oncogenic signaling are largely dependent on intracellular mechanisms such as kinasedirected signaling. We therefore more closely investigated the differential expression of the cellular kinome in the BRAF/NRAS wild type culture [22]. Among the most differentially expressed kinases and kinase-related genes are CDK1, NEK2, CLK3, MET, MOK, KDR, RAF1 and KIT. Interestingly, CDK2 and CDK4 were the only kinases that were almost homogeneously expressed in the vast majority of single cells, with CDK4 being more prominent than CDK2. Similar findings regarding kinases and kinase-related genes have been reported by several other studies [38][39][40][41][42].
The identified gene expression pattern was highly suggestive for a targeted treatment approach directed against CDK4 and potentially also CDK2, which were the only signaling kinases expressed in the vast majority of cells. Interestingly, in a recent study, the cyclinD-CDK4/6 axis has been emphasized as an interesting target for tumor treatment [43]. Indeed, CDK4 inhibitor palbociclib was more effective than BRAF inhibitor vemurafenib for treatment of the bulk short-term culture. This finding is not surprising as the short-term culture is derived from a BRAF wild type tumor, which is known to not respond to BRAF inhibition [3]. However palbociclib activity equaled the activity of MEK1/2 and ERK1/2 inhibitors in this culture. MEK1/2 inhibitors such as trametinib have been shown to be active in BRAF/NRAS wild type melanomas and induced at least a partial response in 10% of melanoma patients [44]. Although we observed some effects of palbociclib on fibroblasts which might argue against its widespread clinical use, palbociclib is well tolerated in treatment of other tumor entities [45].
Despite the dominating proliferation signature found in all three cultures, also other functional groups such as a pigmentation and a stromal signature were identified in the double wild type melanoma short-term culture. This raised the question if the heterogeneity of patient-derived short-term cultures revealed by single-cell RNA-seq might reflect in part different gene expression patterns found in primary or metastatic melanomas. In a direct comparison of the single-cell data and an immunofluorescence analysis of the original tumor of the BRAF/NRAS wild type tumor sample, we provided strong evidence that the cell composition revealed by single-cell RNA-seq reflects the actual situation in the original primary tissue. Moreover, some gene expression signatures found in our single-cell data overlapped with patterns of prognostic relevance in a series of independent studies on melanoma patients [16][17][18][19]. Genes associated with high proliferative activity and genes of the pigmentation signature were found in the gene signatures related to negative patients' prognosis [16,17,19]. Moreover, the proliferation signature of our study was also highly expressed in high-grade and metastatic melanomas in another independent study [21]. Notably, a recent single-cell analysis of 114 selected genes of invasive and proliferative signatures in two melanoma cell lines with different metastatic capacity is in line with our results as the single-cell gene expression differed significantly between both cell lines [46]. Thus, in vitro gene signatures may at least in part predict in vivo behavior of melanoma cells.
In a more recent study, and in accordance with our data, a significant intertumor heterogeneity was observed in a single-cell RNA-seq analysis of 19 melanoma tissues [15]. Major subgroups of transcriptional heterogeneity were associated with cell cycle, spatial context of cells, and a drug-resistance program (MITF low/AXL high signature). The majority of the different gene signatures of tumor cells of the study of Tirosh and co-workers could be mapped to our cell subpopulations, arguing for a conservation of in vivo gene signatures in in vitro cultures at least in those with low passage numbers [15]. Among these overlapping gene signatures were the proliferation signatures and the MITF/AXL signatures. Furthermore, the KDM5B (JARID1B) signature in this study showed a strong association with non-cycling cells in our study. It is noteworthy that KDM5B (JARID1B) is a known marker of slow-cycling cells [29]. Together, these findings supported the notion that in vivo gene signatures are conserved in short-term cultures.
Recently, single-cell analyses have been used to identify candidates for tumor stem cells [38,40]. Similar to these studies, we analyzed a set of genes recently described as markers for malignant melanoma stem or initiating cells (MMIC) [23]. No single stem-like subpopulation could be defined by the presence of all these genes. However, a subpopulation of about 10% of cells was defined by an increased expression of ABC transporter and ALDH genes in all three cultures. Overexpression of ABC transporters is found in cancer stem cells of different tumors and, similarly, ALDH genes have been described as a characteristic feature of cancer stem cells [47]. Thus, the ABC/ALDH subgroup presents interesting gene candidates for MMICs in our study [23][24][25]. In addition, our analyses of epigenetic marks of gene expression further favor the presence of a small subpopulation of cells with stemness character. Genes poised by histone marks and by polycomb genes, both hallmarks of stemness, were found in our ABC transporter and ALDH subpopulation.
Interestingly, another small group of cells concordantly expressed KDM5B (JARID1B), a gene known to be involved in epigenetic regulation of gene expression. As mentioned above, a KDM5B (JARID1B) signature has also been described in the study of Tirosh and co-workers [15]. In accordance with this latter study, www.impactjournals.com/oncotarget we could show that KDM5B is associated with nonproliferating cells, which is consistent with other previous studies where it was shown to be a marker for slow-cycling melanoma cells as well as involved in long-term melanoma growth [15,24]. Strikingly, KDM5B was also identified by Patel and co-workers to be significantly higher expressed in non-cycling glioblastoma cells [38]. Since it is known that stem-like cells divide with lowest rates, cells of the identified KDM5B subgroup are interesting candidates for stem-like melanoma cells [29,48].
Taken together, we identified marked intertumor and intratumor (intercellular) heterogeneity in three different patient-derived melanoma cultures including a small subpopulation of stem-like cells. Specifically targeting these stem-like cells or cells with negative impact on the patients' prognosis may in the future provide promising treatment strategies.

Melanoma cell culture
Three melanoma short-term cultures were used: Ma-Mel-123 (23 passages), Ma-Mel-108 (29 passages), and Ma-Mel-93 (15 passages). All three are derived from subcutaneous metastases. Ma-Mel-123 is BRAF and NRAS wild type, Ma-Mel-108 is BRAF V600E mutant and NRAS wild type, Ma-Mel-93 is BRAF wild type and NRAS G13R mutant [49]. Cell cultures were authenticated immediately before the experiments. Clinical information of Ma-Mel-123, Ma-Mel-108, Ma-Mel-93: age at biopsy in years (67,35,75), gender (female, female, male), localization of primary tumor (mucosa, occult, occult), type of primary tumor (mucosal, occult, occult), stage at biopsy (IV, IV, IV), biopsy origin (all cutaneous metastases), and survival in months after biopsy (6.3+, 8.8, 9.2). Melanoma cells were cultured under standard culture conditions [49]. All clinical investigations have been conducted according to the principles expressed in the Declaration of Helsinki and have been approved by the Ethics committee of the Medical Faculty of the University of Leipzig (Az 023-16-01022016).

Single-cell RNA-seq
Cell suspensions was generated by washing melanoma cells from different cultures with PBS followed by trypsinizing with TripLE Xpress (Invitrogen, Darmstadt, Germany) for 5 minutes. The reaction was stopped by adding media containing FCS. Single melanoma cells from short-term cultures were captured on an integrated fluidic circuit RNA-seq chip (Fluidigm, Hamburg, Germany) using the Fluidigm C1 system. Cells were loaded onto the chip at a concentration of 200 cells/μL and imaged by phase-contrast. Cell capture, cell lysis, reverse transcription, and cDNA amplification were performed on the chip as described [50]. Illumina libraries were constructed using the Illumina Nextera XT DNA Sample Preparation kit using the protocol supplied by Fluidigm. Sequencing libraries were pooled (3 µL each) and purified with 18% SPRI beads. Library concentration and size distribution were assessed on an Agilent Bioanalyzer and with Qubit dsDNA HS Assay kits and a Qubit 2.0 Fluorometer (Thermo Fisher Scientific, Invitrogen, Darmstadt). Each cell was pairedend sequenced (100 base reads) on an Illumina HiSeq 2500 to a depth of 2-5 million reads and base-calling, adaptor trimming, and de-multiplexing were performed as described [51,52].

Analysis of single-cell data by SOM portrayals and tSNE
Raw reads were processed using a custom script and aligned to a Bowtie2 [53] indexed human genome (grch38 sourced from ENSEMBL) using TopHat [54] with default settings. Transcript levels were quantified as fragments per kilobase of mapped reads (FPKM) generated by Cufflinks [55] using gencode protein coding genes (grch38 v22 Havana). We excluded cells that did express neither of two housekeeping genes ACTB and GAPDH. Ninety-two single cells remained for the transcriptome analysis of the BRAF/NRAS wt/wt culture, 84 cells for the replicate analysis of this culture, 54 cells for the BRAF mutant/NRAS wild type, and 77 cells for BRAF wild type/NRAS mutant culture, respectively. The mean number of genes detected was 6014 in the BRAF/NRAS wild type culture (4078 in the replicate) and 4655 in the mutant cell cultures. The higher gene detection rate in the first analysis of the shortterm culture Ma-Mel-123 compared to the replicate might explain minor batch effects in the subsequent analysis.
Gene-centric expression data were quantile normalized, then centralized and clustered using selforganizing map (SOM) machine learning. The SOM method translates the approximately 15,000 gene expression profiles into 2,500 metagene expression profiles and visualizes their expression levels in each sample using a two-dimensional quadratic 50x50 grid. SOM size and topology were chosen to allow robust identification of expression modules in terms of so-called spots [10,56]. For the functional interpretation we applied gene set analysis using Fisher's exact test and the gene set Z-score (GSZ) [57]. We considered gene sets related to biological processes (BP), cellular components (CC) or molecular function (MF) of the gene ontology (GO) classification, and standard literature sets [56,58] and literature sets curated by our group. Sample diversity analysis, class discovery and presentation were performed as described previously [13]. For estimation of the quality of sample clustering we made use of the silhouette-plot [13]. All downstream methods have been described previously [10,56] and are implemented in the R-package 'oposSOM' [11]. www.impactjournals.com/oncotarget In order to demonstrate the similarity of replicate samples, a t-distributed stochastic neighbour embedding analysis (tSNE) [59]) was performed. The R-package 'Seurat' was used for running tSNE analysis with default settings [60]. Only small but systematic differences between the two replicates of wild type/wild type cells were interpreted as batch effects and were not considered in downstream analyses. Largescale copy number variations (CNVs) for the three short-term cultures were inferred from expression profiles by averaging gene expression over stretches of 50 genes on their respective chromosomes, as described in a recent study [38].

ACKNOWLEDGMENTS AND FUNDING
We thank B. Höber and A. Weihmann of MPI-EVA for DNA sequencing and S. Vorberg for immunofluorescence analysis. B.T. and T.G. were

CONFLICTS OF INTEREST
M. Kunz has received honoraria from the Speakers Bureau of Roche Pharma and travel support from Novartis Pharma GmbH and Bristol-Myers Squibb GmbH. D. Schadendorf has received personal fees from Roche, Novartis, GSK, Merck and Bristol-Myers Squibb, Amgen, Boehringer Ingelheim and Leo Pharma. All other authors declare no conflict of interest.