Un modelo integrado de aprendizaje automático basado en quimiocinas predice el pronóstico y guía la inmunoterapia en...

Hepatocellular carcinoma (HCC), the second leading cause of cancer-related death globally, has limited clinical benefits from immune checkpoint inhibitors. This study aimed to address the critical challenges of prognostic heterogeneity and low immunotherapy response rates in patients with HCC by integrating chemokine-related gene expression profiles using Machine learning (ML) algorithms. Using the cancer genome atlas hepatocellular carcinoma and GSE14520 datasets, we performed unsupervised consensus clustering of 227 chemokine-related genes to define HCC subtypes. Prognostic genes were screened using univariate Cox regression, and 10 ML algorithms were integrated under 10-fold cross-validation to construct a predictive model. Immune infiltration, pathway enrichment, and immunotherapy sensitivity were analyzed using ESTIMATE, single-sample gene set enrichment analysis, and immunophenoscore. Two chemokine-related subtypes (A/B) were identified, and subtype B demonstrated a significantly prolonged overall survival (P < .05). The StepCox[both] + SuperPC model, constructed from 92 prognostic genes, exhibited robust performance in both training and validation sets (C-index: 0.719/0.653). High-risk patients were characterized by metabolic reprogramming and an immunosuppressive tumor microenvironment, whereas the low-risk group displayed immune-stromal synergy. Among the 11 core genes screened, high expression of SPP1 and SLC1A2 was significantly associated with poor prognosis (P < .05), whereas ITGAM/HILPDA may serve as a predictor of programmed cell death protein 1/cytotoxic T-lymphocyte-associated protein 4 inhibitor sensitivity. In this study, we developed a chemokine-based ML model that classifies HCC patients into 2 subtypes with distinct survival outcomes and immune-metabolic features. We identified an 11-gene prognostic signature, including SPP1 and SLC2A1 associated with poor prognosis, and genes such as ITGAM and P2RY6 predictive of immunotherapy response. These findings provide insights into the chemokine-immune-metabolic network and support further validation toward personalized HCC treatment.