Identification of multiple prognostic biomarker sets for risk stratification in SKCM.

Introduction: The majority of available transcriptomics-related cancer prognosis studies strive to define one collection of biomarkers that can be used to predict high-risk patients. However, using a single biomarker profile could restrict its strength and applicability to diverse groups of patients. In order to fill this gap, we discuss the prospect of determining several, discrete sets of prognostic biomarkers in Skin Cutaneous Melanoma (SKCM). Our search identifies various genes including CREG1, PCGF5 and VPS13C whose expression pattern depicts significant correlations with overall survival (OS) in SKCM patients. Methods: We developed machine learning-based prognostic models using SKCM gene expression data to predict 1-, 3-, and 5-year overall survival. Advanced feature selection approaches were applied to identify prognostic biomarkers. The primary biomarker set consisted of 20 genes selected using state-of-the-art feature selection techniques. Machine learning classifiers were trained to distinguish high-risk from low-risk patients using these biomarkers. The process was systematically repeated to identify seven independent biomarker sets, each containing 20 unique genes without overlap. Model performance was evaluated using AUC and Cohen's Kappa metrics on an independent test dataset. Validation was further performed using the GEO dataset GSE65904, employing subsets of biomarkers from the primary and third sets. Results: The primary biomarker-based prognostic model demonstrated strong predictive ability, achieving an AUC of 0.90 and a Kappa of 0.58 in identifying high-risk SKCM patients. A second independent 20-gene set, with no overlap with the first, produced an AUC of 0.89 and Kappa of 0.56. Across all seven biomarker sets, performance ranged from 0.84 to 0.91 (AUC) and 0.48 to 0.64 (Kappa). Notably, the fifth biomarker set yielded the highest performance with an AUC of 0.91 and Kappa of 0.64. External validation confirmed the predictive utility of selected biomarkers where genes from the primary set achieved an AUC of 0.83 on GSE65904. While genes from the third set achieved an AUC of 0.86 on the same dataset. Discussion: Our results show that only one gene-expression signature is not sufficient to predict SKCM prognosis. Alternatively, high-risk patients can be accurately predicted using multiple independent biomarker sets providing flexibility in both clinical and computational practices. The high similarity in the results of all seven sets (AUC 0.84-0.91; Kappa 0.48-0.64) signifies the stability and strength of the method. The external validation of these biomarkers with GEO data also helps to confirm the reliability of these biomarkers and hints at their potential wider applicability. This work facilitates transparency by ensuring that all the data and code is publicly accessible (https://github.com/raghavagps/skcm_prognostic_biomarker), which also promotes future developments in creating multi-signature prognostic tools in melanoma.