Protein Signatures of NK Cell–Mediated Melanoma Killing Predict Response to Immunotherapies

Melanoma is one of the most aggressive and deadly skin cancers (1, 2). Parameters predicting the survival rate of patients with melanoma include age, disease stage, mutational status, immune fitness, and environmental factors but also cellular characteristics such as organellar architecture and metabolic hotspots. In this context, we and others have shown that calcium and redox signaling as well as mitochondrial bioenergetics and dynamics determine melanoma pathobiology, drug sensitivity, and disease outcome (3–10).

Although the treatment of advanced melanoma is still challenging, the median survival of patients with advanced melanoma increased significantly in the last decade (11). This therapeutic improvement is mostly due to the clinical application of targeted and cytotoxic T cell (CTL)-based immune checkpoint therapies (12). While targeted therapies inhibit the frequently mutated and constitutively active RAS-RAF-MEK-ERK signaling cascade, the anti-melanoma checkpoint therapy is based on decreasing the CTL activation threshold by blocking the inhibitory CTLA4 (CTL-associated protein-4) and/or PD (programmed cell death protein)-1/PD-L1 receptors (13). Nevertheless, the outcomes of these therapeutic approaches show considerable variation and the tumor often relapses after an effective treatment (14). These limitations highlight the need for improved therapeutic strategies, but also for establishing novel algorithms and biomarkers that predict therapeutic responses.

In contrast to CTLs, natural killer (NK) cells exert their crucial antitumoral function without prior antigen-specific stimulation (15). Their principle of tumor cell recognition is controlled by a large repertoire of germline-encoded receptors with activating or inhibitory properties (16, 17). The sum of these activating and inhibiting signals determines the final NK-cell activation status and thus their cytotoxic capacity (18, 19). Hence, high NK-cell cytotoxicity correlates with a generally lower cancer risk and with spontaneous cancer regression (20). Moreover, the number and type of tumor-infiltrating NK cells are tightly linked with the overall survival of patients with melanoma or other cancers (20–24). NK cells are currently not used in anti-melanoma therapies. Yet, melanoma is highly immunogenic and its cells express ligands that are predicted to promote NK-cell recognition (25, 26). Accordingly, NK cells could facilitate the treatment of solid cancers in the future, as already demonstrated at different experimental settings (27–29). On the basis of this knowledge, several clinical trials are currently evaluating the potential of NK cells in treating solid tumors (see https://clinicaltrials.gov/).

In addition to using NK cells for direct elimination of transformed cells, NK cell–related parameters such as tumor infiltration may help to predict the disease course in patients with melanoma (23). Moreover, NK cells promote checkpoint-directed immunotherapy by recruiting dendritic cells into the tumor (23, 30). In light of the heterogeneity and plasticity of cancer cells, intrinsic cellular properties that modulate the susceptibility toward NK-cell cytotoxicity could also open up hitherto unexplored avenues for antitumoral immunotherapy.

Using real-time two-dimensional (2D) and three-dimensional (3D) NK cell–based killing assays in a panel of genetically distinct melanoma cell lines, protein microarray screens, melanoma xenografts, and bioinformatic analyses, we investigated the cytotoxic potential of human NK cells against melanoma. Analyses of patient-individual transcriptome datasets identified a “protein-killing-signature” and allowed us to develop a robust algorithm that predicts the intrinsic sensitivity of melanomas to PD-1-directed immunotherapies. Our results reveal novel biomarkers and expression signatures that might be used to predict outcomes of immune cell–based therapies.

All chemicals were purchased from Sigma-Aldrich, unless otherwise indicated.

Cells were isolated from leukoreduction system chambers of the local blood bank (UMG Ethics approval 2/3/18). Human peripheral blood mononuclear cells (PBMC) of healthy thrombocyte donors were isolated by density gradient centrifugation using Lymphocyte Separation Medium 1077 (PromoCell, # C-44010). Primary human NK cells (pNK) were isolated from PBMCs by negative bead isolation (Dynabeads Untouched Human NK cell Kit, #11349D). Isolated pNKs were maintained in AIMV medium (Life Technologies, #12055-091) supplemented with 10% fetal calf serum (FCS) (Sigma-Aldrich, #A9418) and 0.05 μg/mL IL2 (Thermo Fisher Scientific, #15596-026) for at least one day (Supplementary Fig. S1A–S1D). NK-92 cells (DSMZ #ACC 488), were kept in culture for 2–3 months (24–30 passages) in MEM α (Thermo Fisher Scientific, #11900-016) supplemented with 12.5% FCS, 12.5% horse serum (Thermo Fisher Scientific, #16050122), 2 mmol/L l-glutamine (Sigma-Aldrich, #67513), and 10 ng/mL IL2.

Target cells including K562 and melanoma cell lines were kept in culture for 2–3 months (24–30 passages). The leukemia cell line K562 (ATCC, #CCL-243) was cultured in RPMI1640 medium (Thermo Fisher Scientific, #21875-034) supplemented with 10% FCS. Melanoma cell lines were maintained in melanoma growth medium [TU2%, consisting of 80% MCDB153-Basalmedium (Biochrom AG, #F 8105), 20% L-15, Leibowitz-Medium (PromoCell, #C-24300), 2% FCS, 1.68 mmol/L CaCl₂ (Sigma-Aldrich, #21115), 5 ng/mL human insulin (Sigma-Aldrich, #I9278)]. The 1205Lu cell line was cultured in TU2% with 200 mmol/L l-glutamine and without insulin. The 15 melanoma cell lines used in this study have been provided, authenticated, and documented by the Herlyn laboratory (Wistar Institute, Philadelphia, PA; see Supplementary Table S1 and Resources at https://wistar.org/our-scientists/meenhard-herlyn). All cell lines were regularly tested for Mycoplasma contamination by using PCR Mycoplasma Test Kit I/C (PromoCell #PK-CA91-1024) according to the manufacturer's manual. Protein expression in melanoma cells was manipulated by using siRNA or DNA transfection or by CRISPR-Cas9. Immunoblotting, IHC, immunofluorescence, and flow cytometry were applied to determine protein abundance. For full description of the methods used, please see Supplementary Data.

Kinetic cytotoxicity assays were performed as described previously in ref. 31 (graphic presentation in Fig. 1A). In brief, target (cancer) cells (25,000 cells/well) were loaded with 0.5 μmol/L calcein-AM (Life Technologies, #C1430) and seeded in black clear-bottom 96-well plates (BD, #353219). For optimal NK cell–mediated killing (NKmK), an effector to target (E:T) ratio of 5:1 (if not indicated differently) was used (Supplementary Fig. S1E–S1H). NKmK of melanoma cells (Y) was recorded by measuring the decrease of the fluorescence signal over time (Ex.: 485 nm; Em.: 535 nm). Living target cells (NC) and Triton X-100 lysed cells (PC) were used as negative and positive controls, respectively. Each condition was pipetted in triplicates to avoid unspecific effects. The mean fluorescence was calculated for every 10-minute intervals (t). NKmK (%) was quantified using the equation depicted in Fig. 1A. As the values of the “killing” condition and the negative control inevitably display different starting fluorescence signals, the difference was corrected with the following index I: I_{t = 0} = Y_{t = o} / NC_{t = 0}. Fluorescence was recorded with Infinite M200Pro (Tecan) or CLARIOstar (BMG LABTECH) plate readers. The kinetic measurement (cycles = 25; interval time = 10 minutes; bottom reading mode) was performed at 37°C and 5% CO₂.

To block NKG2D, pNKs were incubated with 25 μg/mL anti-NKG2D REAfinity (Miltenyi, #130-122-332) for 15 minutes and with 5 μg/mL during the cytotoxicity assay.

Spheroids were generated as described previously (32). Shortly, 5,000 WM983B melanoma cells/well were seeded in a 96-well plate on top of a layer of 1.5% (w/v) Difco Noble Agar (BD Biosciences, #214220). Melanoma cells were allowed to form spheroids for 96 hours. Then they were treated for 48 hours with 2.5 nmol/L rapamycin or DMSO (control). The spheroids were harvested and mixed in 2 mg/mL bovine collagen I (Thermo Fisher Scientific, #A1064401). Embedded spheroids were transferred into a new 24-well plate and coated on top of a layer of collagen I. The cell-collagen-media mix was subsequently covered with fresh AIMV medium or with AIMV medium containing prestained pNKs (using 5 μmol/L CellTracker Green CMFDA, Invitrogen, #C2925) in a 10:1 ratio for 48 hours. Spheroids were stained with 4 μmol/L EthD-1 (Live/Dead Viability/Cytotoxicity Kit, Invitrogen, #L3224) and imaged using a Zeiss Axiovert S100TV microscope equipped with a Fluar 10×/0.5 objective, CMOSpco.edge camera and pE-340fura (CoolLED) LED light source. Images were acquired using the VisiView 4.2.0.0 software (Visitron Systems GmbH) using GFP (green signal; Ex. 470/40, Dichroic mirror T495 LPXR, Em. 525/50) and RFP (red signal; Ex. 545/25, Dichroic mirror T565 LPXR, Em. 605/70) filters and processed using ImageJ (NIH, Bethesda, MD).

Reverse phase protein array (RPPA) is a high-throughput approach that allows simultaneous detection and quantification of hundreds of proteins (33). In short, melanoma cell protein lysates were immobilized on nitrocellulose-coated slides and probed with primary antibodies detecting cancer-related (total and phospho) proteins. Subsequently, biotinylated secondary antibodies were used for detection and quantification of the protein abundance or their posttranslational modifications. Melanoma cell lysates were prepared as described previously (34). The RPPA assays and the data processing were performed by the MD Anderson Center RPPA core facility (Houston, TX; see ref. bib33) and (https://www.mdanderson.org/research/research-resources/core-facilities/functional-proteomics-rppa-core/rppa-process.html). The readouts (log₂ scale) were normalized by using median-centering across antibodies. Unsupervised hierarchical clustering using centered correlation and complete linkage was performed on normalized log₂ median-centered protein values using Cluster 3.0 software (http://bonsai.hgc.jp/~mdehoon/software/cluster/software.htm#ctv).

Linear regression analyses were applied to determine the correlation between NKmK and RPPA datasets. To this end, the slope, the intercept point and the Pearson correlation coefficient were calculated for each protein. The slope was interpreted as the extent of change in NKmK relative to the abundance/phosphorylation of the respective protein. Accordingly, proteins, which are relevant for NKmK, display higher absolute slope values. The Pearson correlation coefficient was used to calculate the P value using a t-distribution with n-2 degrees of freedom. For the volcano plot, −log₁₀(P) was plotted as a function of the slope for each protein in the RPPA panel. Proteins with P < 0.05 were considered as “hits.”

Proteins that showed a Pearson correlation coefficient above 0.7 (positive hits) or below −0.7 (negative hits) were used to derive the protein-killing-signature. To this end, the median centered log₂ (log₂MedCen) values of all positive hits and the inverse of the log₂MedCen values of all negative hits were averaged to yield the average signature protein abundance for each melanoma cell line in the training dataset. A linear regression of NKmK vs. the average abundance yielded the prediction formula that was used to calculate the theoretical NKmK values of the validation cell line panel. Predicted and experimentally determined NKmK of the validation dataset were plotted and correlated.

Oxygen consumption rate (OCR) and extracellular acidification rate (ECAR) were measured as described in ref. 5 using an XF96 Extracellular Flux Analyzer (Seahorse Bioscience). Briefly, 100,000 melanoma cells were seeded on the day of the measurement into a Seahorse 96-well plate. Cells were prepared for the measurement 3 hours after seeding by exchanging medium to Seahorse XF DMEM (Agilent, #103575-100) supplemented with 25 mmol/L glucose, 1 mmol/L pyruvate, and 2 mmol/L glutamine and incubated at 37°C in an incubator without CO₂ for 1 hour. Periodic measurements were performed at basal state and after the administration of 3 μmol/L oligomycin, 1 μmol/L CCCP (carbonylcyanid-3-chlorophenylhydrazone), and 2 μmol/L rotenone plus 1 μmol/L antimycin A. Basal (basal state) and maximal (after addition of CCCP) OCR and basal ECAR were quantified and are presented as mean ± SEM.

Melanoma cells (7 × 10⁵) were seeded into 60 mm dishes 4 hours before treatment. NVP-BKM120 (Selleckchem, #S2247, BKM120, 1 μmol/L), MK2206 2HCl (Selleckchem, #S1078, MK2206, 1 μmol/L), and rapamycin (Selleckchem, #S1039, 2.5 μmol/L) were diluted in DMSO (Sigma-Aldrich, #D8418). As control, DMSO without drugs was used. During treatment (24 or 96 hours), cells were incubated at 37°C and in 5% CO₂.

Melanoma cells (1.2 × 10⁶) were seeded into 100 mm cell culture dishes 4 hours before rapamycin treatment (2.5 nmol/L). After 96 hours, RNA was isolated with NucleoSpin RNA Plus kit (Macherey-Nagel, #740984.250) according to manufacturer's instructions.

Quality and integrity of RNA was assessed using a Fragment Analyzer. All samples exhibited an RNA integrity number over 8. RNA sequencing (RNA-seq) libraries were performed using 200 ng total RNA of a non-stranded mRNA Seq (TruSeq RNA Library Preparation, #RS-122-2001). Libraries were sequenced on the Illumina HiSeq 4000 (SE; 1 × 50 bp; 30–35 Mio reads/sample).

Sequence images were transformed with Illumina software BaseCaller to BCL files, which was demultiplexed to fastq files with bcl2fastq v2.20. The sequencing quality was asserted using FastQC.

Sequences were aligned to the reference genome Homo sapiens (GRCh38.p13) using the RNA-seq alignment tool (version 2.7.8a) allowing for two mismatches within 50 bases. Subsequently, read counting was performed using featureCounts. Read counts were analyzed in the R/Bioconductor environment (version 4.0.5) using the DESeq2 package version 1.31.5. Candidate genes were filtered using an absolute log₂-fold change >1 and FDR-corrected P value <0.05. Gene annotation was performed using Homo Sapiens entries via biomaRt R package version 2.46.3.

Melanoma cells were xenografted onto NOD-SCID gamma (NSG) mice and NOD/SCID (NS) mice (The Jackson Laboratory). All animals were housed in ventilated cages. Animal experiments were authorized by the Landesamt für Natur, Umwelt und Verbraucherschutz (LANUV) Nordrhein-Westfalen and in accordance with the German law for animal protection and/or according to institutional guidelines at the Ontario Cancer Institute of the University Health Network (Toronto, Canada; G1605/17 AZ84-02.04.2017.A053). Six to 8-week-old sex-matched mice were injected subcutaneously with 2 × 10⁶ cells (1205Lu or WM88) per mouse in the left flank and tumor growth was monitored twice a week.

Initially, the RPPA datasets were filtered by the correlation of the relative signal intensity (log₂MedCen) with the determined NKmK/NK-92mK of the indicated panel of melanoma cell lines.

Genes displaying either a positive (>0.6) or a negative (<−0.6) Pearson correlation coefficient were selected for further analysis. For the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis, David 6.8 application (35) was used. The selected proteins were matched to the human background gene list to identify the biological pathways that display an enrichment in our protein-killing-signature. Signaling pathways were ordered on the basis of their fold enrichment. KEGG pathways referring to specific cancers and other diseases (pathway maps 6.2–6.12) are only shown in Supplementary Data S3, S4, and S10. The P values of each protein were used as a second criterion. To generate proteomaps, uniprot IDs of gene marks that are either positively or negatively correlated to NKmK were submitted to the proteomaps (version 2.0) application (36). All proteomaps were generated considering all human genes as a background. The size of each polygon represents protein abundance, weighted by protein size. Functionally related proteins are summarized in bigger distinct colored areas and indicate metabolic and signaling pathways involved in NKmK.

The transcriptome datasets generated by Hugo and colleagues (37), containing gene expression data of 27 patients with melanoma characterized as responders (14) and nonresponders (13) to anti-PD-1 therapy, was used to train a linear support vector machine (SVM) with the scikit-learn Python library (38). Of 25,268 genes in the dataset, the 50 genes with the highest absolute correlation to the NK-92mK signature were used as an initial set of features. Because of their high correlation (0.98), the average expression of datasets 27A and 27B was calculated and considered as one dataset. A classification pipeline was set up consisting of four steps. First, the gene expression values were scaled to the interval [0,1]. For each split of the data, the constants for the transformation of the values were derived from the training set and applied to the test set. Next, the training set was used to select the k best features, where k is an integer between 1 and 20. Features were ranked with a χ² test. After standardizing the remaining features based on the data distribution of the training set, a linear SVM (Supplementary Data S11) was trained on the training set, and then used to predict the responsiveness toward PD-1 therapy for each sample in the test set. Hyperparameters of the SVM were optimized by performing 5-fold cross-validations (5FCV) for different values of the cost parameter C. The 5FCV was repeated three times with randomly selected splits to reduce bias. The mean accuracy from all 3 × 5 splits was used as the final score for any given parameter set, and the best-performing value of C was then chosen for the model. Finally, the classification performance of the pipeline was evaluated again by using the average score from three randomized runs of 5FCV for validation.

To visualize the SVM, a principal component analysis (PCA) was applied to the dataset. A scatter plot projection of the data and of the SVM boundary on the first two principal components was created with the Python-library Matplotlib. The ability of the SVM boundary to distinguish between classes in the full dataset was evaluated by calculating the sensitivity and specificity of the model after the PCA was applied.

The genes that led to the highest classification performance of the machine learning pipeline were plotted in a clustered heatmap, using the Seaborn library. A genewise Z-score was applied, and Ward hierarchical clustering was performed for the genes.

Data obtained from experiments were analyzed or processed using ImageJ, GraphPad Prism 8, Image Studio Lite, and Microsoft Excel. The data were tested for normal distribution using Shapiro–Wilk test and statistical significance was determined with unpaired, two-tailed Student t test unless otherwise specified. The significant differences are indicated by asterisks *, P < 0.05; **, P < 0.01; ***, P < 0.001.

To evaluate the susceptibility of melanoma cells to NK cell–mediated killing (NKmK) we used a well-established real-time killing assay (Fig. 1A; ref. 31) and determined the killing efficiency in eight genetically distinct melanoma cell lines (Supplementary Table S1). Interestingly, our results showed that the susceptibility toward NK-cell killing between the different melanoma cell lines is highly variable. Moreover, our findings indicated that these intrinsic differences are consistent and characteristic for each particular melanoma cell line (Fig. 1B and C). For example, the 1205Lu melanoma cells (blue) were killed to more than 90% after 4 hours, while WM88 cells (orange) were almost completely resistant to the primary NK cells and were killed to only around 15%. The values for all other cell lines lay between these two extremes, with WM3734 (green) displaying NK-cell susceptibility of around 50%. Notably, the primary human NK cells isolated from healthy blood donors displayed a rather small variability despite the fact that for some melanoma cell lines, we used NK cells from more than seventy healthy individuals (Fig. 1B and C). On the basis of these findings, we concluded that the differences in NKmK are controlled to a much higher degree by the intrinsic properties of the different melanoma cell lines. To further explore this hypothesis, we used the NK cell line NK-92. These cells originate from a non–Hodgkin lymphoma patient and their antitumor capacity has already been thoroughly investigated and used in treating hematologic malignancies (39, 40). We performed NK-92 killing assays with three melanoma lines, 1205Lu (highly susceptible to primary NKs), WM3734 (killed to around 50% by primary NKs), and WM88 (almost completely resistant to primary NKs). In general, NK-92 cells showed a lower cytotoxicity against melanoma cells than primary NK cells (Fig. 1D and E). Nevertheless, the relative melanoma cell line susceptibility was comparable with the one observed when using primary NK cells (Fig. 1F); again, 1205Lu cells were highly susceptible, WM88 cells were almost completely resistant, and WM3734 cells showed medium susceptibility to NK-cell killing. These findings thus strengthened the hypothesis that the overall heterogeneous susceptibility to NKmK resided with the melanoma cells themselves, and not with the NK cells. On the basis of these results, we next sought to identify molecular determinants that were responsible for the differences in NKmK between the tested melanoma cell lines.

MHC class I (HLA A, B, C) proteins have been suggested to be involved in the regulation of NK cell–based antitumor responses (18). We therefore evaluated the general HLA status of the three exemplary melanoma cell lines and found that HLA abundance does not correlate with the NKmK of the examined melanoma cell lines (Supplementary Fig. S2A and S2B), confirming previous findings (41). Thus, we concluded that the HLA status does not play a major role in determining the melanoma cell susceptibility to primary NK cell–mediated and NK-92 cell–mediated killing.

To explore alternative regulatory mechanisms involved in NKmK, we performed RPPA screens. In contrast to conventional proteomics, RPPA determines not only abundance but also activation status of proteins using high-throughput antibody-based detection (Fig. 1G; ref. 33). For this study, we investigated the abundance and activation status of around 300 cancer-related proteins (see Supplementary Data S1, S2, and S9). Initially, we performed RPPA for eight melanoma cell lines (Supplementary Data S1). To assess the potential involvement of all analyzed proteins in NKmK, we correlated their relative abundance and/or activation status with the killing efficiency (extrapolated from the results shown in Fig. 1B). The RPPA values showing a higher Pearson correlation coefficient than the set threshold were selected and further evaluated (Fig. 1G). The “hits” that correlated in a positive manner with the killing efficiency (high expression = high killing) were marked as dark yellow circles in the volcano plot, and those with negative correlation (high expression = low killing) were marked with purple circles. The heatmap displayed as Fig. 2A provides additional information regarding expression scores, expression-killing correlations, and the killing rates for all eight melanoma cell lines (see also Supplementary Data S1). To test the robustness and specificity of our findings, we performed additional NK-92 cell killing assays (Supplementary Fig. S2C and S2D) and evaluated the results as described for the primary NK cells. The heatmap identified positively and negatively correlating hits (Fig. 2B; Supplementary Data S2). Not all identified hits for NKmK and NK-92mK matched; however, we observed a considerable overlap. To identify molecules and pathways that control the differences in NKmK, we performed a bioinformatic processing of the identified “protein hits.” A “proteomap”-based pathway analysis indicated high contribution of metabolic signals as well as environmental and genetic factors in both positively and negatively correlating protein hits (Fig. 2C). These findings were confirmed for NK-92–mediated killing (Fig. 2D). To identify the most relevant signaling pathways, we performed additional KEGG-based pathway analyses and identified mitochondrial bioenergetics (i.e., central carbon metabolism in cancer pathway) and mTOR signaling pathway as major determinants of NKmK (Fig. 2E; Supplementary Data S3). Similar results were obtained when melanoma cell killing by NK-92 cells was analyzed (Fig. 2F; Supplementary Data S4).

In summary, these findings further supported the notion that specific intrinsic traits of melanoma cells profoundly contribute to NKmK efficacy. Moreover, our RPPA screens suggested that the metabolic status of melanoma cells plays a central role in this context.

As seen in Fig. 2A and B and in Supplementary Data S1 and S2, RPPA identified several mitochondrial proteins as important regulators of NKmK. On the basis of these findings and also on the results from our pathway analyses, we sought to evaluate the role of mitochondrial metabolism, that is, bioenergetics on NKmK. For this purpose, we determined OCR and ECAR in melanoma cells with high (1205Lu), median (WM3734), and low (WM88 and WM1366) NKmK. Measurements of basal OCR indicated that 1205Lu cells have the lowest respiration rate of all four cell lines. Compared with 1205Lu, the OCR of WM3734 was higher and the OCRs of WM88 and WM1366 were the highest (Fig. 3A and B). The strong correlation between OCR and NKmK (Fig. 3C) confirmed that mitochondria and their metabolic state play an important role in determining melanoma cell sensitivity toward NK-cell cytotoxicity. A similar trend and a comparable NKmK-OCR correlation was observed when we calculated maximal OCRs (Fig. 3A, D, and E). Importantly, the acidification measurements indicated no significant correlation between ECAR and NKmK (Fig. 3F–H).

Out of all mitochondrial proteins that we identified as regulators of NKmK, DIABLO (SMAC) was the most prominent negative regulator of NKmK (see Fig. 2A and B) as its expression strongly correlated with NKmK (Supplementary Fig. S3A). To test this connection and to validate the RPPA data, we assessed DIABLO abundance by immunoblotting and IHC (IHC). The resistant WM88 cells had higher DIABLO expression than the highly susceptible 1205Lu cells (Supplementary Fig. S3B and S3C), a finding that confirmed the data obtained by RPPA. Next, we tested whether the manipulation of DIABLO expression can affect NKmK. We thus transiently silenced DIABLO in WM3734 cells (Supplementary Fig. S3D and S3E) and measured NKmK (Supplementary Fig. S3F). The quantification of these results showed that silencing DIABLO elevated the killing efficiency by around 10% (Supplementary Fig. S3G and S3H), thus supporting the hypothesis that DIABLO is a negative regulator (high DIABLO = low killing) of NKmK.

SNAI1, among other functions, promotes epithelial–mesenchymal transition (EMT) by repressing E-cadherin expression (42) and is a positive regulator of NKmK of melanoma cells according to our RPPA analysis (Fig. 2A and B). The correlation of RPPA-obtained expression levels with the killing efficiency yielded a Pearson correlation coefficient of around 0.9 (Supplementary Fig. S4A). As for DIABLO, immunoblotting and IHC supported and confirmed the RPPA data regarding SNAI1 abundance (Supplementary Fig. S4B and S4C). To test the contribution of SNAI1 in NKmK, we transiently overexpressed SNAI1 in WM88 cells. We confirmed the overexpression (Supplementary Fig. S4D and S4E) and proceeded to evaluate the killing efficiency. As predicted (high SNAI1 = high killing), the killing assays showed an increase of almost 2-fold for NKmK of SNAI1-overexpressing WM88 cells (Supplementary Fig. S4F–S4H).

In summary, the evaluation of the RPPA data suggested and confirmed that mitochondrial metabolism and respiration as well as EMT (via SNAI1) might play an important role in NKmK. Moreover, these data suggested that this regulation can be confirmed by manipulating single proteins such as DIABLO and SNAI1. Yet, the relatively small effects for DIABLO, implied that targeting single proteins might have only a limited influence on NKmK. Therefore, we next examined the contribution of the most relevant signaling pathway(s).

Our RPPA revealed several components of the PI₃K/AKT/mTOR pathway (Fig. 2E and F), as positive regulators of melanoma susceptibility to NKmK. Interestingly, no members of the PI₃K/AKT/mTOR axis were identified as negative regulators. In addition to elevated expression levels, based on the phosphorylation status of RPS6KB1 (p70S6K) and other PI₃K/AKT/mTOR-relevant proteins, RPPA also indicated an enhanced pathway activity. Using the abundance and phosphorylation status of proteins such as PRAS40, RPS6K, and RPS6 we also characterized our melanoma cell lines as mTOR^high (1205Lu) or mTOR_low (WM88; Fig. 4A; Supplementary Data S1, S2, and S9). On the basis of these results and given that RPS6KB1 (p70S6K) is downstream of mTORC1 and can be thus used to evaluate both pathway abundance and activity, we exploited this kinase for further validation of the RPPA data. Indeed, our IHC analyses confirmed the RPPA findings by showing that the phosphorylation status of RPS6KB1 (p70S6K) is elevated in 1205Lu versus WM88 cells (Fig. 4B).

The PI₃K/AKT/mTOR pathway is crucially involved in the pathobiology of cancer, particularly in melanoma (43, 44). Consequently, it has been targeted therapeutically by several inhibitors in clinical trials. Rapamycin blocks mTORC1 together with novel drugs such as BKM120 and MK2206 targeting PI₃K and AKT, respectively, is one of the best established inhibitors of PI₃K/AKT/mTOR (43). However, the role of the PI₃K/AKT/mTOR pathway in melanoma cell susceptibility toward anti-melanoma immunity is not yet fully clarified.

Given that RPPA identified PI₃K/AKT/mTOR components as positive regulators of NKmK, we reasoned that inhibitors of this pathway should cause a decrease in NKmK. To test this hypothesis, we treated WM3734 and 1205Lu cells with rapamycin and determined NK killing efficiency after 24 and 96 hours. Indeed, rapamycin significantly inhibited NKmK in WM3734 cells (Fig. 4C–F). Moreover, and as predicted, rapamycin had an even more profound effect on NKmK in the mTOR^high 1205Lu cells, particularly after 96 hours (Fig. 4G–J). The efficacy of rapamycin (2.5 nmol/L) was confirmed by immunocytochemistry of p70S6K phosphorylation (T389) status. As depicted in Fig. 4K–N, rapamycin treatment reduced p70S6K phosphorylation (i.e., mTOR activity) in a time-dependent manner in both cell lines. Furthermore, these data suggested that the low rapamycin concentration, did not induce formation of stress granules marked by G3BP1, which occur upon acute translation inhibition triggered by various stresses. The slight decrease in G3BP1 intensity after 96 hours, also indicated that the treatment affects melanoma cell biology and is inducing a mild translational inhibition. The BKM120 and MK2206 inhibitors, which act upstream of rapamycin and inhibit PI₃K and the serine and threonine kinase AKT, respectively, had similar effects as rapamycin and decreased NKmK after 24 and 96 hours of treatment (Supplementary Fig. S5A–S5D).

These results confirmed the identification of PI₃K/AKT/mTOR signaling as an essential regulator of NKmK. However, the molecular mechanism and players involved in this regulation were not understood. To answer this important question, we performed RNA-seq of WM3734 and 1205Lu cells treated with rapamycin (2.5 nmol/L) for 96 hours. The processed data (Fig. 5A; Supplementary Fig. S6A and Supplementary Data S5 and S6) demonstrated that rapamycin significantly affects the expression of a number of genes. Notably, the pathway analysis showed that many of these genes are involved in the metabolic regulation of melanoma pathobiology (Fig. 5B; Supplementary Fig. S6B and Supplementary Data S7 and S8). However, a direct link between the genes affected by rapamycin with NKmK was still missing and warranted further investigation. Hence, we expanded and focused our analysis to examine expression levels of genes known to control the expression of NK cell–relevant ligands and receptors, as described in Shimasaki and colleagues (45). Our analysis showed that in 1205Lu cells, seven out of eight significantly downregulated receptors and ligands, are classified as activating (white). Moreover, five out of six upregulated genes, are classified as inhibiting (black) or partially inhibiting (gray; Fig. 5C). In the less-mTOR–dependent WM3734 cells, the correlation between activating and inhibiting ligands/receptors was six to two in the downregulated fraction and two to six in the upregulated fraction (Supplementary Fig. S6C).

To validate these findings, we determined the effect of rapamycin on MICA and MICB (MICA/B) receptor expression. MICA/B are prominent regulators of NK-cell cytotoxicity, particularly in melanoma cells (25) and were downregulated on an mRNA level in both cell lines following rapamycin treatment. Antibody-based flow cytometry indicated that both MICA and MICB were downregulated following rapamycin also on a protein level (Fig. 5D–F; Supplementary Fig. S6D–F). The importance of MICA/B for NKmK was further validated by treating the NK cells with antibodies against NKG2D, a receptor that binds to MICA/B and promotes NKmK (26). As seen in Fig. 5G and H; Supplementary Fig. S6G and S6H, anti-NKG2D significantly reduced the NKmK of 1205Lu and WM3734 cells.

In summary, we could conclude that the PI₃K/AKT/mTOR axis controls NKmK via transcriptional regulation of surface receptors and ligands that determine melanoma cell sensitivity to NK-cell killing.

To test the role of PI₃K/AKT/mTOR axis independently of rapamycin, we generated RAPTOR-depleted 1205Lu cells. As depicted in Fig. 6A, RAPTOR is an important regulator of the mTOR signaling cascade and its downregulation should mimic the effect of rapamycin. Indeed, immunofluorescence of RAPTOR and p70S6K phosphorylation confirmed the decrease in RAPTOR abundance as well as its inhibitory effect on the mTOR activity (Fig. 6B–D). The reduction in NKmK and downregulation of MICA/B expression in the RAPTOR-depleted cells further strengthened the link between mTOR signaling and NK-cell killing (Fig. 6E–H).

To elucidate the role of PI₃K/AKT/mTOR in a more realistic setting, we determined NKmK in 3D melanoma cell culture. To this end, the viability of the melanoma cells was determined in melanoma spheroids exposed to primary NK cells. Addition of NK cells significantly increased the number of dead melanoma cells compared with spheroids not exposed to NK cells (Fig. 6I and J, compare gray vs. green). Pretreatment of the spheroids with rapamycin completely abrogated this cytotoxic activity and caused a significant reduction of NKmK (gray vs. red), thus confirming our findings obtained with the real-time 2D killing assay. According to current literature, PI₃K/AKT/mTOR inhibitors are in clinical trials and should suppress tumor growth (46). However, our findings suggested that inhibition of PI₃K-AKT-mTOR also decreases NK cell–mediated melanoma-directed immunity. To understand this complex role of PI₃K/AKT/mTOR, we performed melanoma xenografts in two different immunosuppressed mouse strains, namely NOD SCID (NS, which have functional NK cells) and NOD SCID gamma (NSG, which are devoid of NK cells). Moreover, we used two cell lines with distinct PI₃K/AKT/mTOR profiles (the mTOR_low WM88 and the mTOR^high 1205Lu cells) for these experiments. By means of this experimental setup, we could simultaneously evaluate the contribution of NK-cell killing potential and the role of PI₃K/AKT/mTOR pathway in NKmK. The tumor volumes of 1205Lu xenografts in NSG mice were much larger compared with those in NS mice (Fig. 6K). Moreover, WM88 xenografts grew significantly slower and no difference could be observed between NS and NSG mice (Fig. 6L).

In summary, because the major difference between the immune cell profiles of NSG and NS mice is the presence/absence of NK cells, we concluded that the PI₃K-AKT-mTOR pathway is a critical regulator of NK cell–mediated anti-melanoma immunity.

At higher concentrations, rapamycin, BKM120 and MK2206 may affect melanoma cell viability (47, 48) and thus “falsely” increase NKmK. Our results, however, showed a decrease of NKmK in cells treated with mTOR inhibitors, suggesting that unspecific effects related to melanoma cell viability do not play an important role in our experimental setting. The killing assays did not contain mTOR inhibitors, however, the dead cells could in theory release some of the inhibitors and thus affect the cytotoxic potential of NK cells as shown previously (49). To exclude unwanted effects on NK-cell cytotoxicity by mTOR inhibitors, we compared the susceptibility of nontreated melanoma cells with drug-pretreated melanoma cells in the presence of mTOR inhibitors or DMSO during the killing assay. Rapamycin and MK2206 did not affect NKmK, while BKM120 slightly decreased NKmK (Supplementary Fig. S7A–S7M). The latter was therefore not used in subsequent experiments.

In summary, our findings demonstrated that the PI₃K-AKT-mTOR signaling is a major, positive regulator of NK cell–mediated melanoma killing.

Our bioinformatic analyses and experimental validations suggested that predicting NKmK of any given melanoma cell line by using the abundance and/or activation status of several highly correlating proteins, might be possible. We tested this possibility and found that the relative protein abundance and/or activity (i.e., phosphorylation status) strongly correlated with the NKmK of primary NK cells (Fig. 7A). To further explore the predictive relevance of NKmK, we developed a prediction model, that is, “protein-killing-signature” based on the RPPA data. The model is based on a simple linear regression analysis that considers the average abundance of all NKmK-correlating proteins (Pearson correlation coefficient ± 0.7) for the eight melanoma lines (training set, Figs. 1B, C, and 2A; Supplementary Data S1) and the respective primary NK killing rates. To test the power of our prediction tool, we performed RPPA analysis of seven additional, randomly selected and genetically diverse melanoma cell lines (Supplementary Data S9) and used the prediction model to calculate the theoretical NKmK for all seven lines (Fig. 7B). Next, we performed killing assays with the new seven melanoma lines (validation panel) and again observed a heterogeneous NKmK (Fig. 7C). To evaluate the predicting power of our model, we correlated the validated NKmK with the predicted NKmK and obtained a relatively high correlation between the predicted and the real NKmK values (Fig. 7D). The validation of our predictive calculations, was extended by pathway analyses of positively and negatively correlating hits for the NKmK of the seven melanoma cell lines (validation panel). Notably, this analysis again identified metabolic factors and PI₃K-AKT-mTOR as highly relevant regulators of NKmK (Fig. 7E and F; Supplementary Data S10).

NK cells have been suggested to play a role in determining efficiency of immune checkpoint therapies in patients with cancer (23, 30). Hence, we sought to utilize our NK cell–based “protein-killing-signature” to predict melanoma patient sensitivity to anti-PD-1 melanoma immunotherapy. We thus processed transcriptomes of melanoma samples from patients categorized as responders and nonresponders as reported by Hugo and colleagues (37). First, we aligned our protein-killing-signatures to the transcriptomic datasets and selected the 50 genes with the highest absolute correlation to NKmK. By training a SVM, we created a list of 17 genes with the highest categorization power (Supplementary Data S11). Of note, for the SVM, we used the protein signature obtained with the NK-92 cells (Supplementary Data S2), due to the marginal difference between experiments. A heatmap that depicts transcriptomes of these 17 genes in all patients displays two loose clusters (top left and bottom right corner) and thus suggests a possible correlation between NKmK and patient sensitivity to anti-PD-1 checkpoint therapy (Fig. 7G). To examine this correlation in more detail, we asked whether SVM classification can be used for predicting anti-PD-1 efficacy and applied a 5FCV. Indeed, the prediction power of the SVM trained on our “protein-killing-signature” yielded a sensitivity of 81% and a specificity of 78% (see Supplementary Data S11 for all other parameters). In order to visualize the performance of the SVM classifier, we applied a PCA to the full RPPA dataset. Figure 7H shows a projection of the patient data on the first two principal components. The linear decision boundary of the SVM demonstrated a relatively high sensitivity of 86% and a specificity of 85% on this two-dimensional dataset representation. Given that four of the 17 proteins/genes selected in Fig. 7G were identified based on their phosphorylation/activation status, we repeated the analysis with proteins that were identified as relevant for NKmK based only on their abundance. The sensitivity of 74% and specificity of 83% confirmed that the combination of NKmK and RPPA can be used as a prognostic tool for anti-PD1–based therapies (Supplementary Fig. S8A and S8B; Supplementary Data S11).

In summary, we demonstrated that our protein-killing-signature–based prediction tools can robustly estimate NKmK for any given melanoma cell subtype as well as predict patient sensitivity to immune checkpoint therapy.

Melanoma is arguably one of the best examples of how cancer cells control and evade immune responses. Being a cancer with one of the highest mutational loads, melanoma is very immunogenic but also extremely heterogeneous (50). Indeed, melanoma cells undergo very dynamic and diverse phenotypic changes; a trait that profoundly affects both their aggressiveness as well as their drug responsiveness (6, 51). How this dynamic phenotypic switching may affect CTL and, in particular, NK cell–mediated anti-melanoma immune responses, was the focus of this study.

Anti-PD-1 (nivolumab, pembrolizumab) and anti-CTLA4 (ipilimumab) immune checkpoint therapies are effective but have serious side effects (52). Moreover, less than approximately 40% of all patients with melanoma respond to the treatment, and the percentage of complete responses is even lower (53). Accordingly, predicting possible responders and nonresponders is very important. Some studies predict strategies for CTL-based immunotherapies based on genomic and transcriptomic screens of melanoma patient samples and melanoma cell lines (37, 54–57). In contrast, platforms that predict melanoma sensitivity to NK cell–based immunotherapy or that utilize NK-cell killing for predicting sensitivity to immune checkpoint inhibition are not available. In this context, our study identified molecular players, pathways, and mechanisms that control melanoma cell susceptibility to NK cell–mediated killing and predicted possible therapeutic outcomes.

We used RPPA screens to determine protein abundance but also enzymatic activity (through their phosphorylation status) in melanoma cells for this purpose. Given that our main goal was to identify regulators of antitumor immunity in melanoma rather than NK cells, we minimized the influence of NK cell–intrinsic traits by using a high number of healthy human donors (more than 70 for some melanoma cell lines) and the NK-92 cell line. In addition, stimulation with IL2 further reduced the differences in intrinsic NK cell cytotoxicity between donors as well as between biological replicates. By applying a standardized real-time cytotoxicity assay, we evaluated the NK killing efficiency in a training panel of genetically distinct melanoma cell lines and established a protein signature, which was used to determine signaling pathways as well as to predict NKmK of additional melanoma cell lines. Our bioinformatic analysis identified single protein hits (SNAI1, DIABLO), biological processes (mitochondrial respiration) as well as signaling pathways (PI₃K/AKT/mTOR) that are critically involved in controlling NKmK. Given that SNAI1 is one of the major regulators of EMT (42), this finding suggested that NKmK is higher in dedifferentiated melanoma cells. We also demonstrated that mitochondrial respiration is tightly linked with NKmK; a finding that warrants further exploration, in particular because a recent study identified mitochondria as major determinants of responsiveness to immune checkpoint therapy in patients with melanoma (58). Moreover, we showed that high PI₃K/AKT/mTOR activity correlates with high NKmK. This was an intriguing finding given that the PI₃K/AKT/mTOR signaling axis is an important regulator of cancer cell biology (43, 59) and its drug inhibitors have been and are being tested as anti-melanoma therapies (60–63). However, if the regulation of susceptibility to NKmK is also included, the overall picture becomes more complex than could be discerned from previous studies.

While NK cell–based therapies are not in clinical use for treating melanoma yet, it is tempting to speculate on the possible clinical impact of our findings. Microarray screens of protein abundance and activity as well as in vitro cytotoxicity assays with patient-derived material might be helpful tools for personalized cancer therapies of the future. For example, they could provide rapid information about the cytotoxic potential of the patients' immune cells as well as the protein abundance/activity pattern of the individual's cancer. Ultimately, this information (i.e., tools) could be used in categorizing patients with cancer as “probable responders” and “probable non-responders” and become a central decision-making aid toward personalized therapies. In any case, we believe that our approach can basically be feasible in clinical situations, although some aspects of routine performance might have to be refined and optimized. From a practical point of view, the approach could be realized in a cost- and time-effective manner with readily available cells and reagents. Thus, clinical trials could be started within a reasonable timeframe.

In summary, our results contribute to a paradigm shift in melanoma research in several ways: First, they broaden our understanding regarding the critical role of mitochondria, metabolic factors, and the PI₃K/AKT/mTOR signaling network as regulators of anti-cancer immunity. The influence of these parameters on immunologic responses within the complex tumor biology, with a special focus on NK cell sensitivity, must obviously be considered when dealing with tumor biology in its entirety. Second, they expand our understanding of why the therapeutic response to specific inhibitors of immunologic regulations is very heterogeneous. Our perspective of the complex pathophysiology of melanoma is thus broadened. From a practical point of view, our data potentially provide a basis to use protein-based signatures for the prediction of therapeutic responses in melanoma and other malignancies.

A. Denger reports grants from DFG during the conduct of the study. N. Künzel reports grants from DFG during the conduct of the study. M.P. Schön reports grants and personal fees from pharmaceutical companies: AbbVie, BMS, Biogen, Leo, UCB, Janssen, Novartis, Allmirall outside the submitted work. V. Helms reports grants from DFG during the conduct of the study. M. Hoth reports grants from German Research Foundation during the conduct of the study. No disclosures were reported by the other authors.

S. Cappello: Investigation, visualization, writing–original draft. H.-M. Sung: Investigation, visualization. C. Ickes: Investigation, visualization. C.S. Gibhardt: Investigation, visualization. A. Vultur: Supervision. H. Bhat: Investigation, visualization. Z. Hu: Investigation. P. Brafford: Investigation, methodology. A. Denger: Investigation, visualization. I. Stejerean-Todoran: Investigation, visualization. R.-M. Köhn: Investigation. V. Lorenz: Investigation, visualization. N. Künzel: Investigation. G. Salinas: Investigation, methodology. H. Stanisz: Supervision, methodology. T.J. Legler: Resources, supervision. P. Rehling: Supervision, methodology. M.P. Schön: Supervision, funding acquisition, writing–review and editing. K.S. Lang: Supervision, funding acquisition, methodology. V. Helms: Supervision, funding acquisition, methodology. M. Herlyn: Supervision, funding acquisition, methodology. M. Hoth: Supervision, funding acquisition. C. Kummerow: Conceptualization, supervision, visualization, methodology. I. Bogeski: Conceptualization, supervision, funding acquisition, writing–original draft, writing–review and editing.

The authors thank Andrea Paluschkiwitz, Sandra Janku, Anette Benemann, and Ulrike Fischer for their technical assistance. Many thanks to Magdalena Shumanska for her help with PBMC isolation and to Gijsbert van Belle for his help with fluorescence microscopy. The authors are grateful to Alexander Flügel and Lutz Walter for helpful discussions and comments. This work was supported by the German Research Foundation (DFG) through SFB1027 Projects C4 (to I. Bogeski), C3 (to V. Helms), and A2 (to M. Hoth), IRTG1816 (to I. Bogeski), SFB1190 project P13 (to P. Rehling), and P17 (to I. Bogeski).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.