Employing machine learning using ferroptosis-related genes to construct a prognosis model for patients with osteosarcoma

Abstract

Identifying effective biomarkers in osteosarcoma (OS) is important for predicting prognosis. We investigated the prognostic value of ferroptosis-related genes (FRGs) in OS. Transcriptome and clinical data were obtained from The Cancer Genome Atlas and Gene Expression Omnibus. FRGs were obtained from the ferroptosis database. Univariate COX regression and LASSO regression screening were performed and an FRG-based prognostic model was constructed, which was validated using the Gene Expression Omnibus cohort. The predictive power of the model was assessed via a subgroup analysis. A nomogram was constructed using clinical markers with independent prognostic significance and risk score results. The CIBERSORT algorithm was used to detect the correlation between prognostic genes and 22 tumor-infiltrating lymphocytes. The expression of prognostic genes in erastin-treated OS cell lines was verified via real-time PCR. Six prognostic FRGs (ACSL5, ATF4, CBS, CDO1, SCD, and SLC3A2) were obtained and used to construct the risk prognosis model. Subjects were divided into high- and low-risk groups. Prognosis was worse in the high-risk group, and the model had satisfactory prediction performance for patients younger than 18 years, males, females, and those with non-metastatic disease. Univariate COX regression analysis showed that metastasis and risk score were independent risk factors for patients with OS. Nomogram was built on independent prognostic factors with superior predictive power and patient benefit. There was a significant correlation between prognostic genes and tumor immunity. Six prognostic genes were differentially expressed in ferroptosis inducer-treated OS cell lines. The identified prognostic genes can regulate tumor growth and progression by affecting the tumor microenvironment.

Keywords: ferroptosis; machine learning; osteosarcoma; prognostic model; tumor microenvironment.