Prognostic Value of Lipids Homeostasis and Identification of HGS
The expression pattern of lipid homeostasis based on the transcriptome profile from GSE62254 was calculated using ssGSEA. The association between lipid homeostasis and 5-year OS of GC was evaluated by using the Kaplan-Meier curve. We found lipid homeostasis was significantly associated with 5-year OS (Figure 1a). WGCNA was applied to construct a gene co-expression network to select the lipid homeostasis related module among the whole transcriptome profile (figure 1b). The optimal soft threshold was set to three to achieve a scale-free network (figure 1c). Seventeen modules were identified, while the yellow module was correlated with lipids homeostasis most (figure 1d,1e,1f). Genes within the yellow module could be enriched to oxidoreductase activity, hexosyltransferase activity, and UDP-glycosyltransferase activity (figure 2a). Univariate Cox (Supplementary Table S3) and LASSO-Cox (figure2b,2c) analyses were used to select the highly prognosis related genes. Finally, 18 genes and their coefficients were extracted for further study. We found TGFB2, NAB2, PJA1, ZNF382, GPX3, and PLOD2 were risky factors, while the others were protective factors for GC prognosis (figure 2d).
HGSscore for each case was calculated using mRNA expression level of genes and their LASSO coefficients. Then GSE62254 was divided into HGSscore-high and -low groups using the median HGSscore as the cutoff. Kaplan-Meier OS and DFS curves demonstrated that the two groups had significantly different prognosis (figure 2e,2f).
HGS and Clinical Characteristics in Training Cohort
Nomogram was developed based on the clinical characteristics and HGS to predict the 5-, 6-, and 7-year OS of GC (figure 3a). The area under ROC curves was 0.8305, 0.8296, and 0.7902, respectively (figure 3b). Calibration curves revealed a great consistency between Nomogram-predicted OS and actual OS (figure 3c, Supplementary Figure S1). Multivariate Cox regression demonstrated that HGS was an independent risk factor for GC prognosis (figure 3d). Subsequently, the correlations between HGS and pathological characteristics revealed HGS was significantly associated with advanced TNM and clinical stages (figure 3e). GSEA showed differentially expressed genes (DEGs) in the HGSscore-high group were related to cancer hallmark pathways, including Angiogenesis, EMT, FA Metabolism, HYPOXIA, and OXPHOS as well as lipids homeostasis pathway (figure 3f, 3g). Finally, we found cases with higher HGSscore often suffered from worse prognosis (figure 3h).
Validation of HGS in TCGA and GEO Cohort
To validate the prognostic value of HGS, we applied two independent cohorts (TCGA-STAD and GSE84437). In TCGA-STAD, lipid homeostasis was the riskiest pathway (HR:2.41, p=0.0051) among 12 lipid-related pathways (figure 4a). Kaplan-Meier curve demonstrated a significant difference in OS between cases in HGSscore-high and -low groups (figure 4b). Multivariate Cox regression also pointed out that HGS was an independent risk factor for patients’ survival in GC (figure 4c). Notably, we found cases with higher HGSscore were loaded with lower TMB (figure 4d), which indicated worse ICIs response in this population. To extend the relation of HGS and TMB, a comprehensive landscape of TMB in TCGA-STAD was illustrated. We found missense mutation dominated the variant subtypes in GC, whereas C>T single nucleotide polymorphism covered more than 50% (figure 4e). The top 20 mutated genes were as down, in which TTN, TP53, MUC16, ARID1A, and LRP1B were the most frequently mutated genes (figure 4e). Then we further analyzed the differences of TMB in HGSscore-high and -low groups, and we found TP53 was frequently mutated in the HGSscore-high cluster (figure 4f) while the TMB profile in HGSscore-low cluster was similar to the general population (figure 4g).
HGS kept a great prognostic value in the GSE84437 cohort as well (Supplementary Figure S2). The Kaplan-Meier curve showed that cases in different groups had absolutely different prognosis (p<0.001).
Mechanism of HGS in Tumor Microenvironment
Three immune/stromal cells infiltration predictive tools were deployed to reveal the potential mechanism of HGS within the TME. Specifically, TIMER was used to reveal the correlation between the single gene in HGS and immune cells infiltration. The riskiest and most predictive genes (TGFB2, NAB2, UBE2D1, ASPRV1) in HGS were analyzed. We found TGFB2 was positively related to macrophage infiltration, and NAB2 was associated with CD4+ T cell infiltration. Nevertheless, UBE2D1 was significantly correlated with DC infiltration, and ASPRV1 was relevant to CD4+ T cell infiltration (figure 5a).
Additionally, the Xcell algorithm was applied to outline the relationships between various immune/stromal cells and HGS (figure 5b). For immune cells infiltration, HGS was negatively associated with NKT, Th1 cells, Tgd cells, Basophils, pro B cells, CD4+ memory T cells, and pDC, while it was positively related with CD8+ naive T cells in TME. The correlations between HGS and stromal cells were much more complex, of which HGS was correlated with enormous stromal cells such as Epithelial cells and Fibroblasts. Moreover, Univariate Cox regression was used to evaluate the prognostic value of twelve immune/stromal cells (figure 5c). Among these selected cells, we found ly endothelial cells were the riskiest factor (HR:5.02) for GC, whereas pro B cells were the most protective factor (HR:-11.99). After that, Spearman analysis revealed that HGS was positively correlated with the StromalScore from Xcell (figure 5d). By comparing the PDL1 mRNA level, we found PDL1 was lower in the HGSscore-high group (figure 5e).
Finally, ESTIMATE was used to assess the tumor purity in TME. Wilcox test showed that tumor purity in the HGSscore-low group was slightly higher than that in the HGSscore-high group (figure 5f). Interestingly, the correlation analysis indicated that HGS was positively related to StromalScore resulting from ESTIMATE (figure 5g).
Relationship Between HGS and Chemotherapeutic Effect
The Cancer Therapeutics Response Portal (CTRP) and Genomics of Drug Sensitivity in Cancer (GDSC) databases were used to demonstrate the correlation between HGS and the chemotherapeutic effect. We found eight genes were significantly associated with chemotherapeutic effect of GC in CTRP, which were SLC10A7, ASPRV1, ZNF382, CDC42EP5, GPX3, STYK1, TGFB2, and PLOD2. Among these genes, SLC10A7, ASPRV1, and ZNF382 were related with chemotherapy response, while CDC42EP5, GPX3, STYK1, TGFB2, and PLOD2 were related with chemotherapy resistance (figure 6a). In GDSC, PLOD2 was considered to link with drug resistance, while SLC10A7 and ASPRV1 linked with drug response (Supplementary figure S3). GSEA of gene sets about Doxorubicin, Cisplatin, and Flurouracil resistance revealed that cases in HGSscore-high group had a satisfactory response to Doxorubicin, though keep resistant to Cisplatin and Flurouracil therapy (figure 6b).
Disruption of Lipid Homeostasis Pathway Mediated By Exogenous Lipids
To determine the effect of lipid homeostasis on GC, we detected relative cell viability and several mRNA expression levels in SGC7901 cell line. In cell viability assay, SGC7901 exposed to lipids demonstrated reduced cell viability compared to the ethanol group (Supplementary figure S4a). RT-qPCR showed that ASPRV1 was upregulated after treatment while the other five genes were downregulated (Supplementary figure S4b). Taken together, lipid homeostasis was considered to have an impact on GC cell.