Machine Learning-based Development and Validation of a Cell Senescence Predictive and Prognostic Signature in Intrahepatic Cholangiocarcinoma

Abstract

Background: Previous studies have shown that cellular senescence is strongly associated with tumorigenesis and the tumor microenvironment. Accordingly, we developed a novel prognostic signature for intrahepatic cholangiocarcinoma (ICCA) based on senescence-associated long non-coding RNAs (SR-lncRNAs) and identified a lncRNA-miRNA-mRNA axis involving in ICCA. Methods: Based on the 197 senescence-associated genes (SRGs) from Genacards and their expression in Fu-ICCA cohort, we identified 20 lncRNAs as senescence-associated lncRNAs (SR-lncRNAs) through co-expression and cox-regression analysis. According to 20 SR-lncRNAs, patients with ICCA were classified into 2 molecular subtypes using unsupervised clustering machine learning approach and to explore the prognostic and functional heterogeneity between these two subtypes. Subsequently, we integrated 113 machine learning algorithms to develop senescence-related lncRNA signature, ultimately identifying 11 lncRNAs and constructing prognostic models and risk stratification. The correlation between the signature and the immune landscape, immunotherapy response as well as drug sensitivity are explored too. Results: We developed a novel senescence related signature. The predictive model and risk score calculated by the signature exhibited favorable prognostic predictive performance, which is a suitable independent risk factor for the prognosis of patients with ICCA based on Kaplan-Meier plotter, nomogram and receiving operating characteristic (ROC) curves. The results were validated using external datasets. Estimate, ssGSEA (single sample gene set enrichment analysis), IPS (immunophenotype score) and TIDE (tumor immune dysfunction and exclusion) algorithms revealed higher immune infiltration, higher immune scores, lower immune escape potential and better response to immunotherapy in the high-risk group. In addition, signature identifies eight chemotherapeutic agents, including cisplatin for patients with different risk levels, providing guidance for clinical treatment. Finally, we identified a set of lncRNA-miRNA-mRNA axes involved in ICCA through regulation of senescence. Conclusion: SR-lncRNAs signature can favorably predict the prognosis, risk stratification, immune landscape and immunotherapy response of patients with ICCA and consequently guide individualized treatment.

Keywords: cellular senescence; cholangiocarcinoma; long non-coding RNAs (lncRNAs); machine learning; signature.

Figure 1

The graphical abstract of study.

Figure 2

Identification of SR-lncRNAs and integration of 113 machine learning algorithms. (A) Cox regression identified 20 SR-lncRNAs. (B) Mutational landscape of the Fu-ICCA cohort. (C) Interaction network diagram of SRG and SR-lncRNAs. (D) Integration of 113 machine learning algorithms to output optimal SRLS. SR-lncRNAs: senescence-related lncRNAs; SRG: senescence-related gene; SRLS: senescence-related lncRNA signature.

Figure 3

Construction of distinct molecular clusters. (A) Based on unsupervised clustering 244 patients with ICCA were stratified into 2 distinct subgroups. (B, C and D) The consensus CDF and the relative change in the area below the CDF curve confirmed that the consensus matrix represents a value of k = 2. (E) OS of different clusters. (F) Clinical characterization and heterogeneity of SR-lncRNAs expression between different clusters. OS, overall survival.

Figure 4

Characteristic of immune and biological function within two clusters. (A) GSVA illustrated the up-regulated and down-regulated pathways in cluster1 and cluster2. (B) CIBERSORT analysis to examine the relative ratios of 22 immune cells. (C) Estimate algorithm of ICCA patients was conducted to calculate estimate score. (D) A significant difference in three immune checkpoint markers between two clusters. (E) The associations between the mRNAsi and the clusters. (F) The proportion of each cell is shown in CIBERSORT. (G) The differences of immune cells between clusters using the Xcell algorithm.

Figure 5

Construction of prognostic model and validation of SRLS in internal and external validation cohort. (A, B) The most optimal model exported 11 lncRNAs identified as SRLS. (C) The correspondence between the different phenotypes. (D, G, J) The expression characteristics of the 11 crucial lncRNAs in our risk model among the low- and high-risk groups in all cohorts. (E, H, K) The risk plots in all cohorts. (F, I, L) The high-risk group had considerably poorer overall survival (OS) when compared to the low-risk group in all cohorts.

Figure 6

Characteristic of Clinical utility and predictive applicability for SRLS. (A, B, C and D) The occurrence of various clinicopathological factors in both low and high-risk groups. (E, F) Univariate and multivariate Cox regression analyses of risk scores in test and train cohorts were performed to exclude the interference of other clinical factors. (G, I) The calibration curve in demonstrated a strong consistence between the predicted survival probabilities and the actual survival rates for the respective time intervals in both cohorts. (H, K) DCA (Decision Curve Analysis) demonstrated the strong predictive performance of SRLS compared to other clinical predictors in both cohorts.

Figure 7

Prediction performance for immune landscape, immune therapy, stemness of SRLS in ICCA. (A) Estimate algorithm reveals higher proportion of immune cells and lower proportion of stromal cells in high-risk group. (B) A strong positive correlation between risk score and tumor stemness score. (C) A significant positive correlation between SRLS score and the expression of PD-L1. (D) The high-risk group was characterized by greater immune suppression and the low-risk group by greater immunogenicity based on immunophenotype score (IPS). (E) Significant differences observed in the expression of immune checkpoint genes among the two risk subgroups. (F) Lower TIDE prediction score suggests that high risk patients are more likely to benefit from immunotherapy. (G) Results of ssGSEA suggest an increased infiltration of almost all immune cells in the TME of the high-risk patients (H, I, J) Predictions of response to immunotherapy also showed higher rates in the high-risk group.

Figure 8

Identification of hub lncRNA-miRNA-mRNA axis involving in senescence in ICCA. (A, B) Characterization of the difference of biological function between high and low risk groups via GSEA. (C) The lncRNA-miRNA regulatory network was constructed to discover the target miRNA of ADAMTSL4-AS1. (D)ADAMTSL4-AS1 and miR-214-3P were significantly decreased in tumor samples in the TCGA-CHOL cohort. (E) A positive correlation between the expression levels of ADAMTSL4-AS1 and miR-214-3p; The predicted binding sites for both are shown. (F) A series of target mRNAs closely related to senescence, such as CDKN1A(p21), HACD3, CDK6. (G) Risk score were significantly and positively correlated with SASP score. (H) The expression of senescence-related proteins that is targeted by miR-214 was significantly different in tumor tissues and normal tissues.

Figure 9

Predicting chemotherapeutic drug sensitivity in ICCA using SRLS. (A, B, C, D, E, F, G, H) Common chemotherapeutic drug sensitivity in ICCA between high and low risk groups including Axitinib, AZD7762, Cisplatin, Cytarabine, Dactolisib, Docetaxel, MK2206 and Palbciclib.