Multi-omics-based subtyping of melanoma suggests distinct immune and targeted therapy strategies

Abstract

Background: Melanoma is a highly heterogeneous malignancy with diverse molecular and clinical behaviors. A precise molecular classification is critical for improving prognostic assessment and guiding personalized therapy.

Methods: We performed an integrative multi-omics analysis of skin cutaneous melanoma using data from The Cancer Genome Atlas (TCGA) and validated our findings in independent cohorts. Multi-layered data, including transcriptomic, genomic, epigenetic, and immune landscape profiles, were analyzed using unsupervised clustering and machine learning approaches to define molecular subtypes. Functional assays and in silico drug screening were employed to explore subtype-specific vulnerabilities.

Results: Three robust molecular subtypes (CS1, CS2, CS3) were identified, each with distinct genomic alterations, tumor microenvironment characteristics, and clinical outcomes. The CS2 subtype was immunologically "hot," characterized by high tumor mutational burden (TMB), elevated neoantigen load, strong immune infiltration, and activated IFN-γ signaling. CS2 tumors showed significant enrichment of immune checkpoint gene expression and were associated with favorable response to anti-PD-1 therapy in external validation cohorts. In contrast, CS1 and CS3 were immunologically "cold" with immune exclusion, high chromosomal instability, and activation of oncogenic pathways linked to immune evasion. Transcriptomic drug sensitivity modeling suggested that CS1 and CS3 may benefit from HSP90 or MEK inhibitors. Moreover, COL11A2 was identified as a subtype-enriched oncogenic driver predominantly expressed in CS1/CS3, and its silencing impaired tumor cell proliferation, invasion, and epithelial-mesenchymal transition (EMT) features.

Conclusions: This study presents a refined multi-omics classification of melanoma that reveals biologically and clinically distinct subtypes with divergent immune and therapeutic profiles. It offers a framework for subtype-specific treatment strategies, and identifies COL11A2 as a potential target in immune-cold melanomas.

Keywords: immune checkpoint therapy; in-silico drug screening; melanoma; molecular subtypes; multi-omics integration.

Figure 1

Multi-omics integrative molecular subtyping of melanoma. (A) Determination of the optimal clustering number based on two clustering statistics. (B) Integrative clustering using a fully Bayesian latent variable model identified three distinct clusters, showing overlap with previous TCGA classifications. (C) Distinct molecular patterns across different omics platforms: transcriptome expression, DNA methylation, CNA, and somatic mutation. (D, E) Kaplan–Meier survival plots showing the association of our classification with progression-free survival (PFS) and overall survival (OS).

Figure 2

Genetic delineation of integrative subtypes. (A) Mutational landscape indicating 65 mutations with significantly different frequencies among the subtypes. (B) Box plot showing that the CS2 subtype has a significantly higher TMB. (C) Box plot demonstrating that CS2 has more neoantigens than CS1 and CS3.

Figure 3

Chromosomal Instability Analysis. (A) Focal-level CNA profiling for each subtype. (B) Broad-level copy number alteration (CNA) profiling across the genome. (C) Quantification of chromosomal stability using FGA, FGG, and FGL metrics. (D) Comparative analysis showing significantly lower focal-level amplifications/deletions in CS2. (E) Validation of subtype-specific drug sensitivity: estimated IC₅₀ distributions for five compounds across CS1–CS3 subtypes.

Figure 4

Immune Profiles across Cutaneous Melanoma Subtypes in TCGA Cohorts. (A) Differential immunocyte infiltration across subtypes. (B, C) Upregulation of immune-related genes (e.g., CD274, PDCD1, CTLA4) in CS2. (D) Illustration of CS2’s immune/stromal cell infiltration. (E, F) Bar plots showing CS2’s higher enrichment of immune and stromal cells. (G) Higher tumor-infiltrating lymphocyte methylation score in CS2, indicating a lower proportion of tumor-infiltrating leukocytes based on methylation. ** P<0.01, *** P < 0.001.

Figure 5

Validation in Gide’s Cohort. (A) Schematic of the 90-gene classifier (see Supplementary Table S5 ). (B) Application of the classifier in the TCGA cohort using NTP. (C) Classification of Gide’s cohort using the 90-gene signature. (D) Comparison of predicted versus actual subtype labels. (E) Differential immunotherapy response among predicted subtypes. (F, G) Kaplan–Meier plots of PFS and OS for predicted subtypes in Gide’s cohort.

Figure 6

Immune Profiles across Cutaneous Melanoma Subtypes in Gide’s Cohorts. (A) Differential immunocyte infiltration across subtypes. (B) Upregulation of immune-related genes (e.g., CD274, PDCD1, CTLA4) in CS2. (C) Illustration of CS2’s immune/stromal cell infiltration. (D) Bar plots showing CS2’s higher enrichment of immune and stromal cells. **** P < 0.0001.

Figure 7

Epigenetic Validation in Conway’s Cohort and Therapeutic Drug Prediction. (A) 90-promoter classifier based on genes in promoter CpG islands (see Supplementary Table S6 ). (B) Reproduction of the three subtypes in Conway’s cohort using methylation profiles. (C) CS2 in Conway’s cohort shows significantly higher tumor-infiltrating lymphocyte methylation scores. (D, E) In silico drug sensitivity prediction using ridge regression analysis (see Supplementary Tables S7, S8 ). * P<0.05, ** P<0.01, *** P<0.001.

Figure 8

Expression and functional analysis of COL11A2 in melanoma cells. (A) COL11A2 expression in melanoma vs. adjacent normal tissues. (B) COL11A2 expression in melanoma cell lines. (C) qRT-PCR of COL11A2 after siRNA knockdown. (D, E) CCK-8 assay showing proliferation in knockdown vs. control. (F, G) Flow cytometry of apoptosis in A-375 cells post-knockdown. (H, I) Transwell assays of migration/invasion in A-375 cells. (J) Western blot for cleaved caspase-3, Bcl-2, E-cadherin, and Vimentin. (K) Relative levels of cleaved caspase-3, Bcl-2, E-cadherin, and vimentin in A-375 cells transfected with si-NC or si-COL11A2, normalized to β-actin. Data are mean ± SD (n = 3). * P<0.05, ** P<0.01, *** P<0.001.