Integration of machine learning to reveal the correlation between ferroptosis and M2 macrophages in head and neck squamous cell carcinoma.

Objective: To investigate the correlation between ferroptosis and M2 macrophages (M2Ms) in head and neck squamous cell carcinoma on the basis of multiomics data and machine learning methods.

Methods: M2M infiltration was assessed via the CIBERSORT algorithm, and Kaplan‒Meier (K‒M) survival analysis was conducted with the best cutoff value. The M2M-related genes (MRGs) were identified on the basis of the interactive results of weighted gene coexpression network analysis (WGCNA) and the Spearman test. The interactions between MRGs and ferroptosis genes were subsequently pooled to investigate their functions, and the hub genes were subsequently applied to establish a scoring system (MFRS) with 101 kinds of machine learning algorithms. The model with the highest concordance index was selected, and the predictive effect was assessed via the area under the curve (AUC) of the receiver operating characteristic (ROC) curves. The correlations of MFRS with immune infiltration, tumor mutation burden (TMB), copy number variation (CNV) and clinical treatment were analyzed, and the landscape of the model genes was displayed with multiomics data. Moreover, a pancancer analysis was conducted to reveal the roles of crucial model genes in different tumors.

Results: Patients with low M2 infiltration had a better prognosis. According to Spearman and WGNCA, a total of 1551 interactive MRGs were identified, 40 of which were also associated with ferroptosis. After the 13 hub genes were obtained from STRING, 101 kinds of machine learning algorithms were applied to establish the predictive model. Among them, the model concerning lasso combined with plsRcox had the best predictive effects, with the highest average C-index value of 0.645, consisting of ALOX12B, CYBB, DDR2, DRD4, NOX4, PRKCA, RGS4, SLC2A3, SLC3A2, TIMP1 and ENPP2. Patients with low MFRSs presented longer survival times, a more active immune microenvironment and greater sensitivity to immunotherapy; nevertheless, those with high MFRSs presented better chemotherapeutic responses. PRKCA was considered a hub model gene on the basis of external validation of multiomics data, and the pancancer analysis subsequently revealed that it performs important roles in tumors.

Conclusion: In this study, we constructed an MFRS model to predict patient prognosis and therapeutic response. This study also preliminarily reveals the roles of M2Ms and ferroptosis in HNSCC patients and provides potentially novel insight for treatment.