Classification of node-positive melanomas into prognostic subgroups using keratin, immune, and melanogenesis expression patterns

Cutaneous melanoma tumors are heterogeneous and show diverse responses to treatment. Identification of robust molecular biomarkers for classifying melanoma tumors into clinically distinct and homogenous subtypes is crucial for improving the diagnosis and treatment of the disease. In this study, we present a classification of melanoma tumors into four subtypes with different survival profiles based on three distinct gene expression signatures: keratin, immune, and melanogenesis. The melanogenesis expression pattern includes several genes that are characteristic of the melanosome organelle and correlates with worse survival, suggesting the involvement of melanosomes in melanoma aggression. We experimentally validated the secretion of melanosomes into surrounding tissues by melanoma tumors, which potentially affects the lethality of metastasis. We propose a simple molecular decision tree classifier for predicting a tumor’s subtype based on representative genes from the three identified signatures. Key predictor genes were experimentally validated on melanoma samples taken from patients with varying survival outcomes. Our three-pattern approach for classifying melanoma tumors can contribute to advancing the understanding of melanoma variability and promote accurate diagnosis, prognostication, and treatment.


Supplementary Information -Section 1 Supplementary Figures and Tables
: Characterization of the four melanoma subgroups. Figure S1: Gene Ontology enrichments on the five gene clusters. The analysis was performed in PROMO 1 .

Analysis of the topology and predictor biological function of decision trees for subsampled datasets
The results demonstrate the hierarchy of the biological functions by which melanoma samples can be partitioned into distinct subgroups, and also show that the final tree presentd in Fig. 4 is a representative of a stable tree topology and is using predictor genes that are biomarkers of the above three biological functions.  Table S5: Clinical details for the six patients selected for Immunohistochemical staining. Patients 1-3 survived for more than 60 months after diagnosis and were therefore labeled as "Good Survival", whereas patients 4-6 survived for less than 24 months and were therefore labeled as "Poor Survival". Each sample was assigned to a melanoma subgroup based on the decision tree's logic: The KLK8 expression levels distinguished sample 6, the only primary sample, from all other samples, and assigned it to the Keratin subgroup. The high expression levels of the immune marker TIGIT on samples 1-3 and 5 assigned them to the Immune subgroup. Finally, the high levels of the Melanogenesis marker TRIM63 assigned sample 4 to the Melanogenesis-high subgroup. The prediction procedure correctly identified sample 6 to keratin, and samples 1-4 to melanoma subgroups that are in agreement with their evaulated survival category . However, sample 5 was assigned to the Immune subgroup, which does not match its evaluated survival category.

5-year survival 5-year recurrence
Melanoma subgroups Lund subtypes Figure S14: Comparison of five-year survival and recurrence between the melanoma subgroups identified in this study and the Lund subgroups. Only samples for which Lund subtype is available were included in the analysis (n=327).

5-year survival 5-year recurrence
Melanoma subgroups Lund subtypes Figure S15: Five-year survival and recurrence plots for metastasis samples using the melanoma subgroups identified in this study and the Lund subgroups. Only metastasis samples for which Lund subtype and followup information are available were included in the analysis. Analysis was performed after removing 104 non-metastasis samples and 100 samples for which Lund subgroup was not available. Of the remaining 260 samples, 260 had survival data, and 135 had recurrence data available. A small group with < 15 samples representing the non-metastatic (Normal or Keratin) subtype was excluded from each analysis, (n=12 for our classification and n=13 for Lund in the survival analysis and n=5 for our classification and n=4 for Lund in the recurrence analysis).

Supplementary Information -Section 5
Validation of the tree on the dataset of

Cirenajwis et al., 2015
To validate the three-gene decision tree described in Fig. 4, we downloaded the Lund dataset by Cirenajwis et al. 15 from GEO 17 (GSE65904). The dataset contains expression profiles of 214 stage III melanoma samples, measured by Illumina Human-HT12v4.0 BeadChip arrays. We applied a log2 transformation to the data, and mapped each of the three predictor genes to a corresponding probeset ID in the Lund dataset: A single probeset was available for the TRIM63 gene (ILMN_1702489), where two probesets were available for both TIGIT and KLK8, of which we manually selected one (ILMN_2125017 and ILMN_1735700 respectively).
Before applying the decision tree on the samples, we manually calibrated the tree threshold values. This step was required because the new dataset was generated using a different platform (microarrays) and therefore had different expression value distributions for the three genes compared with the TCGA dataset (RNA-Seq), on which the decision tree was trained. Furthermore, the distributions differed also due to the different characteristics of the samples here compared to TCGA samples, where samples identified as primary had relatively large tumors and thus were likely more advanced 2 .
When executed, the decision tree assigned a subgroup label to each one of the melanoma samples (See Table S7). Samples showing high expression of KLK8 (above 7.44) were assigned to the Keratin subgroup (n=13). Out of the remaining samples, those with high expression of TIGIT (above 7.36) were assigned to the Immune subgroup (n=80). Out of the remaining samples, those with high expression of TRIM63 (above 9.45) were assigned to the Melanogenesis-high subgroup (n=41), and lastly, all remaining samples were assigned to the Melanogenesis-low subgroup (See Fig. S16). Remarkably, the Immune, Melanogenesis-low, and Melanogenesis-high subgroups showed distinct survival curves with relative risk that is in agreement with their relative risk on the TCGA dataset. These results demonstrate the ability of the decision tree to identify prognostic subgroups of melanoma (Fig. S17).