Identification of Multiple Prognostic Biomarkers sets for Risk stratification in SKCM

Title:
SKCM Prognostic Biomarker Sets – Multiple independent gene expression signatures for risk stratification in skin cutaneous melanoma

Description:

Project: SKCM Prognostic Biomarker – Identification of multiple prognostic biomarker sets for risk stratification in skin cutaneous melanoma

Publication: Malik, S., Tomer, R., Arora, A., & Raghava, G.P.S. (2026). Identification of multiple prognostic biomarkers sets for risk stratification in SKCM. Frontiers in Bioinformatics, 5, 1624329. https://doi.org/10.3389/fbinf.2025.1624329

Overview:
This repository accompanies the SKCM prognostic biomarker publication and provides seven independent, non‑overlapping sets of prognostic biomarkers for predicting overall survival (OS) in skin cutaneous melanoma (SKCM). Unlike existing studies that identify only a single biomarker signature, this work systematically generates multiple distinct gene sets (20 genes each) with no overlap, enabling flexible and robust risk stratification. SKCM is the most lethal form of skin cancer, with rising global incidence driven primarily by UV radiation exposure. Early detection and accurate risk assessment are critical for timely intervention and improved survival rates.

Dataset summary:

• Source: TCGA (The Cancer Genome Atlas) – SKCM cohort

• Samples: 287 patients with complete survival data

• Original feature space: ~20,000 genes (transcriptomic profiling)

• Survival classes (based on overall survival time):

  • Class 0: 0–1 year (29 patients – high risk)

  • Class 1: 1–3 years (110 patients – intermediate risk)

  • Class 2: 3–5 years (50 patients – intermediate risk)

  • Class 3: >5 years (98 patients – low risk/long‑term survivors)

• Class balancing: SMOTE applied to address severe imbalance (final: 110 patients per class; 440 total for ML training)

Key Gene Biomarkers Identified (Primary Set – 20 genes):

TEKT5, ZNF154, H2AC14, BX284668.6, MYCNOS, STUM, SERTM2, RPSAP18, REG4, PSCA, PAEP, ACTR3C, MSLN, MRPS18AP1, ISLR, IL37, IGLV3.16, H2BC11, GPR25, MTND4P35

Notable biomarkers with literature support in SKCM:

• PSCA (Prostate Stem Cell Antigen): Upregulated in SKCM; positively correlated with RNA modification genes

• PAEP (Progestogen‑Associated Endometrial Protein): Upregulated; correlates with poor survival outcomes

• ACTR3C: Mutations (p.Gly58Arg, p.Gly58Glu) may alter cell signaling/motility in melanoma progression

• MSLN: Downregulated in SKCM

• ISLR: Downregulated; negatively correlated with tumor mutation load (TMB)

• STUM: Biomarker for response to Nivolumab (immune checkpoint inhibitor)

• SERTM2: Gene alterations frequently observed in melanoma (tumor progression, immune evasion)

LASSO Cox regression – 17 prognosis‑related genes:

Drug‑gene interactions (DGIdb): 17 drugs identified targeting 5 of the selected genes (mostly inhibitors). Detailed in Supplementary Table S9.

Correlation analysis (Pearson) – top correlated genes with OS:

Univariate Cox regression: 4,324 genes significantly associated with OS (p < 0.01).

• 1,264 genes – HR > 1 (risk factors)

• 3,060 genes – HR < 1 (protective factors)

GO enrichment (top 50 positively correlated genes):

• Biological processes: lipid translocation (p = 0.00002), phospholipid translocation (p = 0.00006)

• Cellular components: mitochondrial outer membrane (p = 0.0002), organelle outer membrane (p = 0.0004)

• Molecular functions: cysteine‑type endopeptidase inhibitor activity (p = 0.001), protein serine/threonine kinase activity (p = 0.001)

Reactome pathways: Ion Transport By P‑type ATPases (p = 0.0003), Ion Channel Transport (p = 0.0009), plus RNA Polymerase I Promoter Opening, Packaging of Telomere Ends, Post‑translational Protein Modification, Metabolism of Proteins.

Model Performance (CatBoost – best performing classifier):

Primary biomarker set (20 genes):

Class‑wise AUC (Primary set – Test):

Performance across all seven biomarker sets (Test – AUC):

Best overall: Fifth biomarker set – AUC 0.91, Kappa 0.64, MCC 0.64

Class‑wise AUC across all seven sets (Test):

Key observations:

• Class 0 (high‑risk patients): Exceptional prediction across all sets (AUC 0.90–0.99) – clinically valuable for identifying patients needing immediate intervention

• Class 2 (3–5 years): Consistently high AUC (0.88–0.93) – robust intermediate survival prediction

• Class 1 (1–3 years): Moderate performance (AUC 0.75–0.85) – challenging mid‑range group

• Class 3 (>5 years): Variable performance (AUC 0.73–0.90) – requires further refinement

Ensemble Models (Primary set – 20 genes):

External validation (GEO dataset GSE65904):

Class‑wise AUC on GSE65904 (Third set):

Comparison with existing methods:

Our approach (multiple independent biomarker sets) outperforms existing single‑signature methods, with 4 out of 7 sets achieving AUC > 0.88.

Feature Selection Methods Evaluated:

Machine Learning Algorithms Evaluated:

CatBoost (best), Extra Trees (ET), Random Forest (RF), XGBoost (XGB), LightGBM, SVM, KNN, MLP, AdaBoost, Gradient Boosting, Logistic Regression

Data Curation & Quality Control:

• Source: TCGA SKCM cohort via TCGAbiolinks (R) / Xena Browser

• Samples: 287 patients with complete survival data

• Normalization: Z‑score scaling (mean = 0, SD = 1)

• Class definition: Based on overall survival time (months)

• Class balancing: SMOTE (Synthetic Minority Over‑sampling Technique)

• Validation split: Stratified 80/20 (5‑fold CV) – each class represented equally

• External validation: GEO GSE65904 (independent melanoma cohort)

• Multiple testing correction: Benjamini‑Hochberg FDR (p < 0.05)

Usage:
These datasets and models are designed for:

• Identifying multiple independent prognostic biomarker sets for SKCM

• Risk stratification of SKCM patients (high‑risk vs. long‑term survivors)

• Training and benchmarking machine learning classifiers (CatBoost, RF, ET, XGB) for survival prediction

• Drug target discovery (DGIdb analysis of prognostic genes)

• External validation of biomarker signatures on independent melanoma cohorts

• Understanding molecular determinants (lipid translocation, ion transport, mitochondrial function) of melanoma prognosis

Key Biological Insights:

• CREG1, PCGF5, VPS13C – high expression associated with longer survival (positive correlation)

• TTYH2, AC008687.4 – high expression associated with shorter survival (negative correlation)

• Lipid/phospholipid translocation and mitochondrial outer membrane pathways enriched in positively correlated genes

• Ion transport pathways (P‑type ATPases, ion channels) significantly enriched

Related Resources:

• TCGA SKCM data: https://xenabrowser.net (GDC TCGA Melanoma)

• GitHub: https://github.com/raghavagps/skcm_prognostic_biomarker

• DGIdb (drug‑gene interactions): https://www.dgidb.org

License: CC BY 4.0 (as per Frontiers open access license)

Contact:
Prof. Gajendra P. S. Raghava