Molecular Subtypes of Mixed Gastric Cancer Defined by Machine Learning for Predicting Prognosis and Treatment
Minchao Rao1,2, Ruiwen Ruan1,3, Jianping Xiong1,3
1Jiangxi Medical College, Nanchang University, Nanchang, Jiangxi Province, 330006, China.
Background:
Gastric cancer (GC) is traditionally classified into intestinal (IGC), diffuse (DGC), and mixed (MGC) types based on pathological features, with each subtype exhibiting distinct clinical outcomes. Among these, DGC is associated with poor prognosis, characterized by low cell adhesion and a high stromal component. Recent proteomic studies have revealed significant differences in extracellular matrix (ECM) composition between DGC and IGC, highlighting the critical role of ECM in tumor biology. MGC, which combines both intestinal and diffuse characteristics, presents substantial heterogeneity, complicating prognosis and personalized treatment approaches. This study reclassifies MGC using extracellular matrix receptor (ECMR) and cell adhesion (CA)-related genes (ECRGs), closely linked to the biological behavior of DGC, to provide insights into prognosis and treatment response.
Methods:
RNA sequencing data and clinical information from GC patients were collected from the TCGA and GEO databases, excluding cases of pure IGC and DGC. Based on ECMR and CA-related genes, supervised clustering via non-negative Matrix Factorization (NMF) was used to identify molecular subtypes in MGC. Differential expression and Cox regression analyses were performed to identify prognostic genes, and an ECMR and CA-based gene signature (ECRS) was developed using machine learning techniques. Gene Set Variation Analysis (GSVA) was conducted to assess functional differences between risk groups, while TIDE and pRRophetic analyses were used to predict responses to immunotherapy and chemotherapy.
Results:
A total of 239 MGC patients were classified into two molecular subtypes with significant differences in prognosis. Subtype 2 displayed distinct ECM interactions and connective tissue development pathways. To refine the ECRS model, we tested 117 model combinations across 10 machine learning algorithms, selecting the configuration with the best predictive accuracy. This optimized model distinguished biological and immune characteristics between high- and low-risk groups, with low-risk patients showing greater sensitivity to immunotherapy and standard chemotherapy.
Conclusion:
This study identifies novel molecular subtypes of MGC based on ECMR and CA-related genes and establishes an effective ECRS model to predict prognosis, immunotherapy response, and chemotherapy sensitivity. This model supports personalized treatment strategies for MGC.
