Machine learning modeling and prognostic value analysis of invasion-related genes in cutaneous melanoma

Abstract

In this study, we aimed to develop an invasion-related risk signature and prognostic model for personalized treatment and prognosis prediction in skin cutaneous melanoma (SKCM), as invasion plays a crucial role in this disease. We identified 124 differentially expressed invasion-associated genes (DE-IAGs) and selected 20 prognostic genes (TTYH3, NME1, ORC1, PLK1, MYO10, SPINT1, NUPR1, SERPINE2, HLA-DQB2, METTL7B, TIMP1, NOX4, DBI, ARL15, APOBEC3G, ARRB2, DRAM1, RNF213, C14orf28, and CPEB3) using Cox and LASSO regression to establish a risk score. Gene expression was validated through single-cell sequencing, protein expression, and transcriptome analysis. Negative correlations were discovered between risk score, immune score, and stromal score using ESTIMATE and CIBERSORT algorithms. High- and low-risk groups exhibited significant differences in immune cell infiltration and checkpoint molecule expression. The 20 prognostic genes effectively differentiated between SKCM and normal samples (AUCs >0.7). We identified 234 drugs targeting 6 genes from the DGIdb database. Our study provides potential biomarkers and a risk signature for personalized treatment and prognosis prediction in SKCM patients. We developed a nomogram and machine-learning prognostic model to predict 1-, 3-, and 5-year overall survival (OS) using risk signature and clinical factors. The best model, Extra Trees Classifier (AUC = 0.88), was derived from pycaret's comparison of 15 classifiers. The pipeline and app are accessible at https://github.com/EnyuY/IAGs-in-SKCM.

Keywords: Invasion-associated genes; Machine learning; Nomogram; Prognosis; Risk score; Skin cutaneous melanoma.