A retrospective review of pathology records at the Netherlands Cancer Institute (NKI) identified 50 patients who underwent a liver biopsy or partial liver resection with a diagnosis of metastatic cutaneous or UM between 2002 and 2017 (online supplemental figure 1). As the number of liver biopsies/resections of patients with CM were low, we included an additional 16 samples from patients that underwent a liver biopsy or partial liver resection for metastatic CM between 1999 and 2017 at the Melanoma Institute Australia (MIA). The CM cohort consisted of 38 patients. The majority of patients (80%) did not receive any treatment prior to the liver biopsy. Some patients received treatment prior to the liver biopsy (online supplemental table 1), but all patients had progressive disease at the moment of the biopsy. In total, 22 patients underwent a biopsy/resection at the NKI and 16 patients at MIA. The UM cohort included 28 untreated patients who underwent a biopsy of a liver metastasis between 2012 and 2016 at the NKI within the SECIRA-UM study, in which all patients were treated with a combination of ipilimumab and radiofrequency ablation,31 following written informed consent. The study was approved by the Institutional Review Board of the NKI.

Tumor DNA and RNA was isolated from formalin-fixed, paraffin-embedded (FFPE) sections. A pathologist scored the tumor percentage and indicated the most tumor-dense region on a H&E stain slide for subsequent DNA/RNA isolation. A total of 5–10 FFPE slides (10 µm) were used for simultaneous isolation of DNA and RNA using the AllPrep DNA/RNA FFPE isolation kit (Qiagen, 80234) and the QIAcube, according to the manufacturer’s protocol.

Germline DNA was isolated from peripheral blood mononuclear cells using AllPrep DNA/RNA/miRNA Universal isolation kit (Qiagen, 80224) and the QIAcube, according to the manufacturer’s protocol.

Strand-specific libraries were generated using the TruSeq RNA Exome Library Prep Kit (Illumina), according to the manufacturer’s instructions. In brief, total RNA was randomly primed and reverse transcribed using SuperScript II Reverse Transcriptase (Invitrogen) with the addition of Actinomycin D. Second strand synthesis was performed using Polymerase I and RNaseH, replacing dTTP with dUTP. Resulting cDNA fragments were 3′ end adenylated and ligated to Illumina Paired-end sequencing adapters and subsequently amplified by PCR. Libraries were validated on a 2100 Bioanalyzer using a 7500 chip (Agilent) and pooled. The pooled libraries were enriched for target regions using the probe Coding Exome Oligos set (CEX, 45MB), according to the manufacturer’s instructions (Illumina). The enriched libraries were subjected to a second round of enrichment followed by a 10-cycle PCR amplification and cleanup using AMPure XP beads (Beckman). Resulting target enriched pools were analyzed on a 2100 Bioanalyzer using a 7500 chip (Agilent), diluted and pooled equimolar into a multiplexed sequencing pool. Next, libraries were sequenced with 65 base pair (bp) single-end reads on a HiSeq2500 using V4 chemistry (Illumina). Fastq files were mapped to the human reference genome (Homo.sapiens.GRCh38.v82) using STAR(2.6.0 c)32 with default settings. Count data generated with HTseq-count33 was analyzed with DESeq2.34 Centering of the normalized gene expression data per dataset was performed by subtracting the row means and scaling by dividing the columns by the SD. Next, previously defined gene expression signatures (IFN-γ,35 microenvironment cell populations-counter (MCP-counter)36 and Danaher immune cell37) were analyzed. The average IFN-γ signature was calculated by the average z-score of all genes within the IFN-γ signature. Immune cell populations of the Danaher immune cell signature were evaluated by comparing the z-score of the immune subset. The distance among samples was computed using the Euclidean distance. DNA was fragmented up to 160–180 bp fragments by Covaris DNA shearing and purified using 2X Agencourt AMPure XP PCR Purification beads, according to the manufacturer’s instructions (Beckman Coulter). Library preparation for Illumina sequencing of the sheared DNA was performed using the KAPA Hyper Prep Kit (KAPA Biosystems, KK8504). DNA libraries were cleaned up using 1X AMPure XP beads. The libraries were sequenced with 100 bp paired-end reads on a HiSeq2500 System, according to the manufacturer’s instructions.

Fastq files were aligned to GRCh38 using Burrows-Wheeler Aligner,38 followed by marking of duplicate reads by Picard MarkDuplicates. Subsequently, base quality scores were recalibrated using GATK BaseRecalibrator and single nucleotide variants were called using GATK MuTect2.39 All identified mutations were required to have passed all Mutect2 tests (FILTER field equals “PASS”). Variants were subsequently annotated using Variant Effect Predictor (VEP).40 TMB was calculated by summarizing the total number of non-synonymous, somatic mutations per sample with minimal variant allele frequency (VAF) of 0.05 (5%).

The COSMIC mutational signatures (V2—March 2015)14 41–44 were assessed using MutationalPatterns.45 Both non-synonymous as well as synonymous somatic mutations with a minimal VAF of 0.05 were used to calculate the relative contribution of each of the 30 COSMIC signatures in each sample.

DNA copy number profiles were generated from the exome sequence data using CopywriteR46 with default settings and a 500 kb bins size. Segmentation was performed using circular binary segmentation47 after which gains and losses were identified using CGHcall.48

For neoantigen prediction, variants were annotated using SnpEff 4.3t.49 Based on these variants, candidate tumor-specific neoepitopes were predicted and annotated using an in-house epitope prediction pipeline, which uses a random forest model to score the probability of surface expression of candidate neoepitopes based on the major prerequisites for (neo-)antigen presentation: RNA expression level (Salmon V.0.9.1),50 proteasomal processing (NetChop V.3.1)51 52 and human leukocyte antigen binding (netMHCpan V.4).53 Candidate neoepitopes that have a model prediction score below 0.02 were filtered out.

Immunohistochemistry (IHC) of FFPE tumor samples was performed on a BenchMark Ultra autostainer (Ventana Medical Systems). In brief, paraffin sections were cut in 3 µm, heated at 75°C for 28 min and deparaffinized in the instrument with EZ prep solution (Ventana Medical Systems). Heat-induced antigen retrieval was carried out using Cell Conditioning 1 (CC1, Ventana Medical Systems) for 48 min at 95°C.

PD-L1 was detected using clone 22C3 (1/40 dilution, 1 hour at RT, Agilent/DAKO). Bound antibody was detected using the OptiView DAB Detection Kit (Ventana Medical Systems). Slides were counterstained with Hematoxylin and Bluing Reagent (Ventana Medical Systems). A scoring system of three levels of PD-L1 expression by IHC stain was used: less than 1%, between 1% and 50% and more than 50%. FFPE sections were stained for CD163 (clone: ERP14643-36, Abcam). Pigmentation was scored by a pathologist as 0–3 (0=no melanin pigment, 1=pigmentation visible at high power, 2=moderate pigmentation visible at low power, 3=high pigmentation readily visible at low power with dense melanin content).54

Using NanoString’s GeoMx digital spatial profiling (DSP) we performed multiplexed and spatially resolved profiling analysis on pretreatment FFPE liver metastasis samples. The DSP technology uses a cocktail of primary antibodies conjugated to unique oligonucleotide tags with an ultraviolet (UV) photocleavable linker. Here, 4 µm-thick FFPE tissues were incubated with an antibody cocktail of 44 unique oligonucleotide-labeled antibodies (online supplemental table 2). Melanoma cells, leukocytes and T cells were defined by fluorescence imaging with antibodies for S100B/PMEL17, CD45 and CD3, respectively (online supplemental table 1). Based on fluorescence imaging, regions of interest (ROIs) (200 µm–600 µm in diameter) within the tumor-enriched tissue areas (S100B/PMEL17 positive) were chosen for multiplex profiling. Tumor-infiltrating lymphocyte (TIL) high and low ROIs were selected based on CD3 positive staining within these tumor-enriched tissue areas. Photocleaved oligos were transferred into a microwell and quantified using optical barcodes in the nCounter platform. For analysis, digital counts were first normalized with internal spike-in controls (External RNA Control Consortium; ERCCs) to account for technical variation, then normalized to the geometric mean of housekeeping controls of their defined ROIs, and subsequently background was subtracted using the IgG controls (rabbit, mouse).

Statistical analysis was performed using GraphPad Prism 7 software (GraphPad Software, San Diego, California, USA) or with the R programing language. Measures of spread in PD-L1 and pigmentation were calculated using χ² test. Measured of spread in immune infiltration, mean centered DSP data, or ratios in CM and UM patients were calculated using an independent t-test. For DSP analysis of individual data points, a linear mixed effect model was used to control for multiple sampling within a slide, using the Satterthwaite’s approximation for df for p value calculation. All analyzes were two-sided and used a significant level of p value ≤0.05.

CM and UM have previously been shown to have a different genetic profile at the primary tumor site. Genome characterization has largely been restricted to primary tumors and a wide variety of metastatic sites, however, genome profiling of these melanomas at the same metastatic site is lacking. We performed whole exome sequencing on liver tumor samples and paired germline DNA from 16 CM patients and 15 UM patients to determine mutational burden and oncogenic drivers. As expected, a strong difference in TMB and oncogenic drivers between CM and UM liver metastases was observed (figure 1). UM liver metastases have significantly lower non-synonymous mutational load than CM liver metastases (p<0.0001, figure 1A,B), and consequently displayed fewer predicted neoepitopes (figure 1C).

Previous analyzes indicated that CM patients most frequently harbor alterations in BRAF, RAS, NF1, TP53 and CDKN2A, resulting in deregulation of the MAPK/ERK pathway in the majority of the CM patients.20 In UM, the most commonly mutated genes are GNA11, GNAQ, BAP1, EIF1AX and SF3B1, which leads to upregulated signaling via MEK, Akt and protein kinase C.21 22 These CM-specific and UM-specific mutations were also observed in the liver metastases samples (figure 1A). BRAF kinase mutations were present in 81% of the CM patients, 46% were located at codon 600 (figure 1D). In addition, BRAF mutations at other codons were observed, which could be a consequence of a high TMB and are therefore most likely passenger mutations (figure 1A,D). Most UM liver metastases carried mutations in GNA11 (67%), BAP1 (80%) and/or GNAQ (27%). Only one CM tumor carried a GNA11 mutation, which has previously been described in a rare subgroup of CM which is also characterized by a lower mutational burden.55 The UM-specific mutations in EIF1AX and SF3B1 were observed less frequently than the UM-specific GNAQ/11 and BAP1 mutations, which may potentially be explained by a lower metastatic rate of UM tumors with EIF1AX and SF3B1 mutations.56

Chromosomal anomalies that are often found in primary tumors of CM and UM patients were also detected in the liver metastases of these patients.23 57 Specifically, loss of chromosome 6q, 8p, 9 and 10, and copy number gains in chromosome 6p, 7, 8, 20 were observed in CM metastases (figure 1E). Copy number aberrations on chromosome 1 (1p loss, 1q gain), chromosome 3 (loss), chromosome 6 (6p gain, 6q loss) and chromosome 8 (8p loss, 8q gain) were observed in UM metastases (figure 1E). Overall, the genetic differences between CM and UM are maintained at the metastatic site.

Both CM and UM arise from the transformation of melanocytes. Regardless of the fact that they have common lineage from melanin-producing cells, CM and UM showed a significant difference in the level of pigmentation. Whereas 58% of the UM metastases displayed pigmentation (score 1, 2 or 3), only 27% of CM metastases showed presence of melanin pigmentation (p=0.0265) (online supplemental figure 2A,B). These pigment-producing cells can express the same MDA (PMEL, MelanA, tyrosinase), however there was only significant difference for MelanA (MLANA) (p=0.0062) (online supplemental figure 2C,D); UM patients had a significantly higher expression of MelanA compared with CM patients.

PD-L1 expression was analyzed by conventional IHC and DSP, in order to determine if this predictor could explain the difference in response rate of CM and UM patients to ICB. Both techniques identified higher and more frequent PD-L1 expression on CM liver samples compared with UM samples (figure 2A–C). Specifically, although the majority of both UM and CM liver metastases lacked PD-L1 expression, the percentage of PD-L1 positive CM metastases, as determined by IHC, was significantly higher (11/41, 27%) as compared with UM metastases (2/31, 6%) (p=0.0260, figure 2A). This observation was confirmed by DSP analyzes, showing a significantly higher expression of PD-L1 (p=0.0068) in CM as compared with UM liver metastases (figure 2C). CM patients that received treatment prior to biopsy did not show higher PD-L1 expression to untreated CM patients (online supplemental figure 2E).

Loss of MHC has previously been described as an escape mechanism from ICB therapy.25 58 Using DSP analysis, we assessed expression of MHC class I and class II proteins by staining for β₂ microglobulin (B2M) and HLA-DR, respectively (figure 2D,E). No significant difference in B2M and HLA-DR expression between CM and UM was found.

PD-L1 expression has been reported to be positively correlated with immune cell infiltration,59 and the observed difference in PD-L1 expression between CM and UM could therefore potentially be coupled to a difference in immune cell infiltrate. To address this, immune-related signatures that have previously been correlated with clinical benefit to ICB therapy in late stage malignancies35–37 were assessed. In this analysis, we focused on the previously defined IFN-γ,35 MCP-counter36 and Danaher immune cell37 signature. The IFN-γ response signature showed no apparent difference in expression between CM and UM samples (figure 3A). As PD-L1 forms one of the IFN-γ responsive genes,60 we also evaluated the association between PD-L1 expression and IFN-γ gene expression profile of CM and UM metastases. A positive correlation between IFN-γ (defined as the average expression of genes included in the IFN-γ signature) and PD-L1 expression was observed (figure 3B).

The MCP-counter signature, which quantifies the absolute abundance of eight immune and two stromal cell populations, showed also no difference in immune cell infiltrates between CM and UM samples (figure 3C). When applying the Danaher immune cell signature evaluating a set of marker genes for different immune cell subpopulations, no distinct difference between CM and UM patients was found (figure 3D). Comparing the different immune subsets of the Danaher signature, showed no significant difference in expression between CM and UM patients (figure 3E). These findings on RNA level were also confirmed by DSP data, in which no difference in protein level of CD4 and CD8 was observed between CM and UM patient samples (figure 3F). However, a significantly lower expression of CD3 in UM liver metastases was found compared with CM liver metastases (p=0.0433). This could potentially indicate a higher expression of CD3⁺CD4⁻CD8⁻ double-negative T cells61 in CM liver metastases compared with UM liver metastasis, however, a single cell characterization (eg, multiplex IHC or flow cytometry) for CD3, CD4 and CD8 expression would be required in order to prove this.

Next, we examined the ratio between different immune subsets identified by the Danaher immune cell signature. Strikingly, the ratio of exhausted CD8⁺ T cells (defined by the RNA expression of LAG3, CD244, EOMES, PTGER4) to cytotoxic T cells (defined by the RNA expression of Perforin 1, Granzyme A/B/H, NKG7, KLRK1, KLRB1, KLRD1, CTSW, GNLY), CD8⁺ T cells (defined by the RNA expression of CD8A and CD8B) and Th1 cells (defined by the RNA expression of TBX21) was significantly higher in metastases of UM patients compared with CM patients (p=0.0061, p=0.0058, p=0.0312, respectively, figure 3G), while the individual subsets of these T cells (exhausted CD8⁺ T cells, cytotoxic T cells, CD8⁺ T cells and Th1 cells) in both tumors was similar (figure 3E). In addition, the ratio of exhausted CD8⁺ T cells to cytotoxic T cells, to CD8⁺ T cells and to Th1 cells were also associated with TMB (figure 3H), in which a higher TMB was associated with a lower ratio of exhausted CD8⁺ T cells to the other immune populations. Furthermore, a high IFN-γ score (defined as average expression IFN-γ signature) was likewise associated with a lower ratio of exhausted CD8⁺ T cells to CD8⁺ T cells in UM metastases (p=0.0386) (figure 3I). These data are both consistent with a model in which the relative level of T cell exhaustion is associated with melanoma subtype and with a model in which low TMB—and consequently low immunogenicity—is associated with increased T cell exhaustion.

To further explore the dissimilarities between CM and UM liver metastases, we performed additional transcriptional analyzes of CM and UM liver metastasis samples. First, we evaluated the distance among samples. An apparent clustering of CM and UM samples was detected (figure 4A). In order to further investigate which genes contribute to this difference, we analyzed the genes that were differentially expressed genes (DEGs) between CM and UM liver metastases (figure 4B), and identified 6523 DEGs (p<0.05). The genes with the lowest p values that were found to be more highly expressed in CM liver metastases were hardly expressed in UM liver metastases, and vice versa (figure 4C). The stark difference between CM and UM liver metastases was also highlighted when CM and UM were analyzed for disparities by principal component analysis. Comparing PC1 to PC2 showed a close clustering of UM samples, more closely to PC2, while CM samples cluster less together (figure 4D).

Gene set enrichment analyzes (GSEA) on DEGs and PC1/PC2 (figure 4E, online supplemental figure 3) demonstrated increased expression of hallmark gene sets involved in metabolic pathways, such as oxidative phosphorylation (OXPHOS), xenobiotic metabolism and fatty acid metabolism in UM. It has previously been described that UM is ranked among the tumors with the highest OXPHOS signature,62 and this property has been associated with invasiveness and drug resistance.63 Pathways upregulated in CM involved proliferation (mitotic spindle, E2F targets, MYC targets and G2M checkpoint) and immune pathway (tumornecrosefactor-alfa (TNFA) signaling via NFKB) (figure 4E). GSEA on PC1 showed a diversity of enriched biological states or processes, and revealed significant enrichment of genes upregulated in response to UV radiation. An estimated 60%–70% of CM are thought to be caused by UV radiation exposure. This fits with the observation that a part of the CM samples tend to correlate to PC1, which was not observed for the UM samples. In addition, analyzes for the presence of COSMIC mutational signatures showed a contribution of UV-related mutations (COSMIC signature 7) in the CM samples (online supplemental figure 4). GSEA on PC2 also revealed in the highest ranked pathways (based on false discovery rate) an enrichment for immune associated pathways (TNFA signaling via NFKB, IL6 JAK STAT3 signaling, interferon alpha and gamma response) (online supplemental figure 3).

Unbiased comparison of CM and UM by DSP analysis revealed significantly higher expression of phosphatase and tensin homolog (PTEN) and β-catenin (CTNNB1) in UM liver metastases compared with CM liver metastases when taking both areas with high and low TIL infiltration into consideration (figure 5A,B, online supplemental figure 5). The higher expression of PTEN in UM liver metastases is in line with the previous observation that a part of CM liver samples have a missense mutation in the tumor suppressor gene Pten and a loss of chromosome 10 (location Pten) (figure 1A,E), and as a consequence lower or absent protein expression (figure 5B). In addition, a significantly higher expression of β-catenin is observed in UM liver metastases compared with CM liver metastases.

DSP analysis also revealed significantly higher expression of CD163 (marker for monocytes/macrophage lineage) (p<0.0001) and CD66b/CEACAM8 (marker for granulocytes) in CM liver metastases compared with UM (figure 5B,C). To confirm our DSP analysis observation, an IHC stain for CD163 was performed on the FFPE slides, and a clear difference in CD163 expression was observed (figure 5D). This implies that CM liver metastases have a higher infiltration of CD163+tumor associated macrophages (TAMs). TAMs in the tumor microenvironment (TME) could release anti-inflammatory cytokines, such as interleukin (IL)-10 and transforming growth factor beta (TGF-β), which inhibit other immune cell functions.64 In order to determine whether there was an associated higher expression of TGF-β in CM liver metastases compared with UM liver metastases, we examined for the TGF-β pathway associated gene signature.65 However, we observed that liver metastases of CM patients exhibit a lower (average) expression of the TGF-β signature compared with UM patients in general (p=0.0005, figure 5E,F).