Fig 2: A three-gene decision tree for classifying melanoma samples.a The tree trained on the 469 TCGA samples. Classification of a new sample into one of the four subtypes is done by traversing the tree from its root to one of its leaves (representing an assignment to a subtype). Three biomarkers (shown in black) are used to determine the route along the tree: overexpression of KLK8 distinguishes the “Keratin” subtype, overexpression of TIGIT distinguishes the “Immune” subtype, and finally, overexpression of TRIM63 distinguishes the “Melanogenesis-high” from the “Melanogenesis-low” subtype. b Threshold values and surrogate genes for the three decision tree predictors as identified by the algorithm. Threshold values are used to distinguish between high and low values (based on normalized expression values), and surrogates can be used as alternatives for the respective predictor gene.

Fig 3: Melanogenesis and immune characteristics of melanoma metastases in good and poor prognostic outcomes.a Hematoxylin and eosin (H&E) staining of lymph nodes containing melanoma metastases from six different patients taken at 20× magnification. Patients 1–3 had good survival, while patients 4–6 survived poorly. b Immunohistochemical staining of the three proteins of the decision tree on the lymph node samples of the six patients. Nuclei were stained blue using DAPI. Row 1: using KLK8 (pink) as a predictor for the Keratin subgroup. Row 2: using TIGIT (Green) as a predictor of the Immune subgroup. Row 3: using TRIM63 (Pink) as a predictor of the Melanogenesis-high subgroup. The assignments of the specimens from the six patients to subtypes based on the expression levels of the three predictor genes are summarized as a label at the bottom bar. c Immunohistochemical staining of additional biomarkers for general prognosis. Row 1: LCK, an immune protein indicative of good prognostic outcome. Row 2: melanogenesis protein OCA2. Row 3: melanogenesis protein GPNMB. d Color matrix quantifying the fluorescence intensity of immunohistochemistry across biomarkers and patients. For each protein, values were independently normalized across the samples. (The KLK staining of all samples is negative, as they are all nonprimary, so KLK is not included here).