Machine learning-based identification of a novel prognosis-related long noncoding RNA signature for gastric cancer

Abstract

Gastric cancer (GC) is one of the most common malignancies with a poor prognosis. Immunotherapy has attracted much attention as a treatment for a wide range of cancers, including GC. However, not all patients respond to immunotherapy. New models are urgently needed to accurately predict the prognosis and the efficacy of immunotherapy in patients with GC. Long noncoding RNAs (lncRNAs) play crucial roles in the occurrence and progression of cancers. Recent studies have identified a variety of prognosis-related lncRNA signatures in multiple cancers. However, these studies have some limitations. In the present study, we developed an integrative analysis to screen risk prediction models using various feature selection methods, such as univariate and multivariate Cox regression, least absolute shrinkage and selection operator (LASSO), stepwise selection techniques, subset selection, and a combination of the aforementioned methods. We constructed a 9-lncRNA signature for predicting the prognosis of GC patients in The Cancer Genome Atlas (TCGA) cohort using a machine learning algorithm. After obtaining a risk model from the training cohort, we further validated the model for predicting the prognosis in the test cohort, the entire dataset and two external GEO datasets. Then we explored the roles of the risk model in predicting immune cell infiltration, immunotherapeutic responses and genomic mutations. The results revealed that this risk model held promise for predicting the prognostic outcomes and immunotherapeutic responses of GC patients. Our findings provide ideas for integrating multiple screening methods for risk modeling through machine learning algorithms.

Keywords: gastric cancer; immunotherapy; long noncoding RNA; machine learning algorithm; prognostic signature.

FIGURE 1

Flow chart of this study.

FIGURE 2

The performances of different signatures selected using multiple feature selection methods (the numbers in parentheses represent the number of lncRNAs in the signature). (A) Univariate analysis (the red dotted line represents p < 0.05). (B) Multivariate analysis (the red dotted line represents p < 0.05). (C) Kaplan-Meier survival curve (the red dotted line represents p < 0.05). (D) AUC values of the ROC curves (the red and blue dotted lines represent 0.75 and 0.5, respectively). (E) Akaike’s information criterion (AIC). (F) C-index of the logistic regression model (the red and blue dotted lines represent 0.75 and 0.5, respectively).

FIGURE 3

Identification of the prognostic signature in the training cohort. (A) Decision curve analysis (DCA) was conducted to confirm the superiority of the risk score. (B) Multivariable analysis was conducted to validate the independent prognosis value of the model in the training cohort. (C) Kaplan-MeierKaplan-Meier curves of the signature for predicting the overall survival (OS) of GC patients. (D) The calibration curves were constructed to determine the accuracy of the nomogram for OS at 1, 3 and 5 years. (E) Time-dependent ROC curve analysis of the risk model in different years. (F,G) The distribution of the expression of the nine lncRNAs in the high- and low-risk groups.

FIGURE 4

Validation of the risk model in the test cohort and merged cohort. (A,B) Multivariable analysis was conducted to validate the independent prognostic value of the model in the test and merged cohorts. (C,E) Kaplan-MeierKaplan-Meier curves of the lncRNA signature for predicting OS in the test and merged cohorts. (D,F) Time-dependent ROC curve analysis of the risk model in the test and merged cohorts.

FIGURE 5

Functional enrichment analysis of the risk model. (A,B) Gene Ontology (GO) annotations and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis. (C) Gene set enrichment analysis (GSEA) in the MSigDB database of tumor hallmarks. (D) WikiPathways analysis.

FIGURE 6

The cnetplots show the relationship between tumor hallmarks and metabolic pathways (A,B).

FIGURE 7

Correlation between the risk model and the immune status of the tumor microenvironment (TME). (A) Heatmap showing the distribution of clinical features and immune cell infiltration. (B) Correlation between 10 tumor infiltrating immune cell types (red represents a negative correlation between two immune cell types and blue represents a positive correlation between two immune cell types; the larger the shape of the point, the stronger the correlation). (C) The distribution of 10 tumor infiltrating immune cell types in the high- and low-risk groups (**p < 0.01). Correlation between the risk score and infiltrating levels of M2-type macrophages (D) and Tregs (E). (F) The T-cell dysfunction scores in the high-risk and low-risk groups calculated using the TIDE algorithm. (G) The expression of PD-L1 in the high- and low-risk groups.

FIGURE 8

Correlation between the risk score and tumor mutation burden (TMB). Waterfall plot displaying the distribution of the top 20 mutated genes with the highest mutation frequency in the high- (A) and low-risk (B) groups. (C) Comparisons of the mutation status of TTN, TP53, PEG3, and SACS. (D) Kaplan-Meier curve analysis of the OS between the high-TMB group and the low-TMB group. (E) Kaplan-Meier curve analysis of OS of patients with low- or high-risk scores in the high-TMB group and low-TMB group.

FIGURE 9

External validation of the risk model. Kaplan-Meier curves of the lncRNA signature for predicting the OS of patients in the GSE15459 (A) and GSE62254 (C) datasets. Time-dependent ROC curves of the signature in the GSE15459 (B) and GSE62254 (D) datasets.

FIGURE 10

Evaluation of the expression of nine lncRNAs in the risk model. (A–I) Comparison of the expression levels of the nine lncRNAs between tumor tissues and normal tissues in TCGA datasets (Mann-Whitney U test). (J–R) The expression levels of the nine lncRNAs in 16 paired GC samples detected using qRT-PCR (paired t test). (S) The expression levels of the nine lncRNAs in two human gastric cancer cell lines (MKN-45 and AGS) and one human gastric epithelial cell (GES-1) (Mann-Whitney U test, ***: p < 0.001).