A four-gene expression-based signature predicts the clinical outcome of melanoma

Purpose: Although clinical indicators provide effective prognostic information, the prognosis of melanoma is difficult due to its genomic and biological complexity. Our goal was to elucidate the impact of genes on survival. Methods: Public cohorts of melanoma gene expression and machine learning were used to develop a model for prognosis. A four-gene model was developed to predict the clinical outcome of melanoma in TCGA datasets. The performance was further validated in four independent cohorts. The relationship between clinical indicators and melanoma score was assayed and the correlated pathways were identified. Results: The samples with high melanoma scores had a significantly better survival rate than those with low melanoma scores in the training cohort. This observation was confirmed in four independent cohorts, GSE22138, GSE54467, GSE65904 and E-MTAB-4725. In addition, the melanoma score was independent of most clinically used indicators. Cox univariate regression showed that the melanoma score was significantly associated with survival. Multiple significantly enriched pathways were identified between the high-score and low-score groups. Conclusion: The melanoma score model was robust and effective for melanoma prognosis.