Multiomics integration and machine learning reveal prognostic programmed cell death signatures in gastric cancer

Abstract

Gastric cancer (GC) is characterized by notable heterogeneity and the impact of molecular subtypes on treatment and prognosis. The role of programmed cell death (PCD) in cellular processes is critical, yet its specific function in GC is underexplored. This study applied multiomics approaches, integrating transcriptomic, epigenetic, and somatic mutation data, with consensus clustering algorithms to classify GC molecular subtypes and assess their biological and immunological features. A machine learning model was developed to create the Gastric Cancer Multi-Omics Programmed Cell Death Signature (GMPS), targeting PCD-related genes. We verified the expression of the GMPS hub genes using the RT-qPCR method. The prognostic influence of GMPS on GC was then evaluated. Single-cell analysis was performed to examine the heterogeneity of PCD characteristics in GC. Findings indicate that GMPS notably correlates with patient survival rates, tumor mutational burden (TMB), and copy number variations (CNV), demonstrating substantial prognostic predictive power. Moreover, GMPS is closely associated with the tumor microenvironment (TME) and immune therapy response. This research elucidates the molecular subtypes of GC, highlighting PCD's critical role in prognosis assessment. The relationship between GMPS and immune therapy response, alongside gastric cancer's microenvironmental features, provides insights for personalized treatment.

Keywords: Gastric cancer; Machine learning; Multiomics; Prognosis; Programmed cell death.

Fig. 1

Overview of the study design.

Fig. 2

GC integrated multimodal subtype analysis. (A) Extensive heatmap depicting consensus molecular subtypes, integrating mRNA, lncRNA, miRNA, DNA CpG methylation patterns, and gene mutations; (B) Stratification of GC cohorts via 10 diverse multimodal clustering methodologies; (C) Matrix representing consensus clustering of two emerging prognostic subtypes, derived from a synthesis of 10 distinct algorithms; (D) Kaplan–Meier survival plots delineating the prognostic trajectories of the identified subtypes; (E) TMB across the subtypes; (F) Mutated genes waterfall plot for each subtype; (G) Comprehensive evaluation of CNV within the subtypes.

Fig. 3

Molecular characterization and validation of GC CSs. (A) Enrichment of subtypes in GC treatment-related signatures, PCD features, and within TCGA classifications; (B) Activity spectrum of 23 transcription factors and potential regulators associated with chromatin remodeling; (C) Immune profile in the TCGA STAD cohort with heatmap annotations showing tumor samples’ immune and stromal scores, expression of immune checkpoint genes, and enrichment levels of 22 immune cells in the TME; (D) Validation of GC CSs in the nearest template analysis of the META-GC cohort; (E) Survival analysis of GC CSs in the META-GC cohort; (F) Consistency of CSs with NTP in TCGA-STAD; (G) Consistency of CSs with PAM in TCGA-STAD; (H) Consistency between NTP and PAM in TCGA-STAD cohort.

Fig. 4

Generation and predictive value of GMPS. (A) Combined results of 99 machine learning algorithms based on an integrated computational framework, with C-index calculated for each model through TCGA STAD and META-GC cohorts and ranked by average C-index; (B) GMPS hub genes selected through the RSF algorithm; (C) Univariate Cox regression analysis results of hub genes in training and validation cohorts; (D-E) Survival analysis of high GMPS and low GMPS in TCGA STAD and META-GC cohorts; (F) The expression level differences of hub genes were determined by RT-qPCR between gastric cancer cell lines and normal gastric mucosal epithelial cell lines; (G) Protein expression levels of hub genes in stomach tissues and gastric cancer in the HPA database.

Fig. 5

Clinical application value of GMPS. (A-B) Comparison of GMPS with other 10 published models in TCGA STAD and META-GC cohorts; (C-D) Comprehensive nomogram based on GMPS with calibration curves; (E) Net decision curve analyses of the nomogram in comparison with other clinical features; (F) Time-dependent C-index curves for the nomogram and GMPS.

Fig. 6

TME-related molecular characteristics in high and low GMPS patients. (A) Distribution of TME immune cell type characteristics in high and low GMPS patients; (B) Distribution of immune suppression features in high and low GMPS patients; (C) Distribution of immune exclusion features in high and low GMPS patients; (D) Distribution of immunotherapy biomarkers in high and low GMPS patients; (E) Distribution of TMB in high and low GMPS patients; (F) Survival analysis of GMPS combined with TMB.

Fig. 7

Value of GMPS in predicting immune therapy response in GC patients. (A) Difference curves of Restricted Mean Survival (RMS) time at 12 and 24 months after treatment between high and low GMPS groups; (B) Long-term Survival (LTS) difference curves after 3 months of treatment between high and low GMPS groups; (C) Distribution of GMPS across different immune therapy response groups; (D) Differences in activation levels between high and low GMPS groups at various stages of TIP; (E) Survival analysis for high and low GMPS groups in GSE78220; (F) Survival analysis for high and low GMPS groups in GSE135222.

Fig. 8

Potential therapeutics for high GMPS patients. (A) GSEA pathway enrichment analysis in the high CMLS group; (B) Sensitivity analysis of cisplatin; (C-D) Correlation and differential analysis of drug sensitivity for potential therapeutics screened from CTRP and PRISM datasets.

Fig. 9

PCD characteristics in GC single-cell transcriptomics. (A) Identification of cell types based on marker genes; (B) Heatmap of the top four marker genes in each cell cluster; (C) PCD activity scores in different cells; (D) Distribution of PCD scores among different cell types; (E) Expression of GMPS hub genes in various cell types.

Fig. 10

Correlation of GMPS with single-cell characteristics. (A) GSEA analysis in the high-risk group; (B-C) Ligand-receptor interactions of high and low-risk Fibroblast/Pericyte cells; (D-F) Full network diagrams of FN1, COLLAGEN and MIF signaling pathways, with heatmap of pathway network interactions across different cell types; (G) Heatmap of incoming and outgoing signaling patterns for each cell type.