Abstract
BACKGROUND: Laryngeal cancer (LCA) is the second most common type of head and neck malignancy, characterized by high recurrence rates and poor overall survival (OS). However, progress in curing LCA through molecular-targeted diagnostics and therapies is slow and limited. The occurrence and progression of cancer are closely associated with metabolic reprogramming. Therefore, this study aimed to identify metabolism-related LCA subtypes through a comprehensive analysis of transcriptomic, mutational, methylation, and single-cell RNA sequencing, in hopes of finding factors which influences the prognosis of LCA. METHODS: First, to identify metabolism-related LCA subtypes, data from 114 patients with LCA from The Cancer Genome Atlas (TCGA) dataset were collected for an unsupervised clustering analysis, which focused on the expression characteristics of survival-related metabolic genes. Subsequently, prognostic and diagnostic models have been developed using machine learning techniques. Specifically, the prognostic model utilized the least absolute shrinkage and selection operator (LASSO) Cox regression, whereas the diagnostic model was built using the Random Forest (RF) algorithm. Furthermore, to ensure the reproducibility, the results of the subtypes and models were validated using three independent bulk RNA datasets and a scRNA-seq dataset. RESULTS: Two robust subtypes were identified and independently validated. Each subtype has a distinct prognostic outcomes and molecular features. Specifically, the LCA1 subtype exhibited better prognosis, enriched metabolic pathways, and higher mutation frequencies. Notably, significant damaging mutations in the methyltransferases NSD1 were observed in this subtype. In contrast, the LCA2 subtype was associated with poorer prognosis, higher immune infiltration, and elevated methylation levels. Moreover, in LCA2 tumors, higher levels of T cell/APC co-inhibition and inhibitory checkpoints were observed. In addition, the diagnostic model demonstrated strong performance, achieving an area under the curve (AUC) values of 1.000 in the training group and 0.947 in the validation group. The prognostic model effectively predicted patient outcomes, with the RiskScore emerging as an independent prognostic factor. CONCLUSION: This study offers new perspectives for patient stratification and presents opportunities for therapeutic development in LCA. Furthermore, we explored the potentials of several key tumor markers for both diagnosis and prognosis prediction.