A novel DNA methylation-based model that effectively predicts prognosis in hepatocellular carcinoma

Abstract

Purpose: To build a novel predictive model for hepatocellular carcinoma (HCC) patients based on DNA methylation data.

Methods: Four independent DNA methylation datasets for HCC were used to screen for common differentially methylated genes (CDMGs). Gene Ontology (GO) enrichment, and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis were used to explore the biological roles of CDMGs in HCC. Univariate Cox analysis and least absolute shrinkage and selection operator (LASSO) Cox analysis were performed to identify survival-related CDMGs (SR-CDMGs) and to build a predictive model. The importance of this model was assessed using Cox regression analysis, propensity score-matched (PSM) analysis and stratification analysis. A validation group from the Cancer Genome Atlas (TCGA) was constructed to further validate the model.

Results: Four SR-CDMGs were identified and used to build the predictive model. The risk score of this model was calculated as follows: risk score = (0.01489826 × methylation level of WDR69) + (0.15868618 × methylation level of HOXB4) + (0.16674959 × methylation level of CDKL2) + (0.16689301 × methylation level of HOXA10). Kaplan-Meier analysis demonstrated that patients in the low-risk group had a significantly longer overall survival (OS; log-rank P-value =0.00071). The Cox model multivariate analysis and PSM analysis identified the risk score as an independent prognostic factor (P<0.05). Stratified analysis results further confirmed this model performed well. By analyzing the validation group, the results of receiver operating characteristic (ROC) curve analysis and survival analysis further validated this model.

Conclusion: Our DNA methylation-based prognosis predictive model is effective and reliable in predicting prognosis for patients with HCC.

Keywords: DNA methylation; GEO; TCGA; hepatocellular carcinoma; prognosis.

Figure 1. CDMGs in four datasets

(A) Venn diagram showing 757 common hyper-methylated genes in HCC samples. (B) Venn diagram showing 21 common hypo-methylated genes in HCC samples.

Figure 2. GO enrichment analysis and KEGG pathway enrichment analysis of CDMGs

(A) Significant GO terms for biological processes. (B) Significant GO terms for cellular components. (C) Significant GO terms for molecular functions. (D) Significant KEGG pathways. Terms and pathways with P<0.05 were here considered significant.

Figure 3. PPI network of CDMGs

After removing the unconnected nodes, 180 nodes and 299 edges were retained. Red nodes represent common hyper-methylated genes, and blue nodes represent common hypo-methylated genes. Node size is proportional to the node’s degree. Line thickness indicates the strength of data support. The PPI score was set at 0.4.

Figure 4. Four SR-CDMGs selected through LASSO Cox analysis

(A) Ten-fold cross-validation for selection of the parameter λ. The solid vertical lines represent partial likelihood deviance ± standard error (SE). The two dotted vertical lines were drawn at the optimal values using minimum criteria (right) and 1-SE criteria (left). The parameter λ = 0.08905296 (log(λ) = −2.418524) was chosen using minimum criteria. (B) LASSO coefficients of the four SR-CDMGs. The dotted vertical line was drawn at the λ value chosen using minimum criteria. The upper axis represents the number of non-zero coefficients at each λ (the degrees of freedom for the LASSO model). The L1 norm represents the summation of absolute non-zero coefficients at each λ. The vertical axis represents the values of non-zero coefficients at each λ. The LASSO coefficients of the four SR-CDMGs HOXA10, CDKL2, HOXB4 and WDR69 were 0.16689301, 0.16674959, 0.15868618 and 0.01489826, respectively.

Figure 5. Kaplan–Meier survival analysis for the four SR-CDMGs

(A) Survival curve of HOXA10. (B) Survival curve of CDKL2. (C) Survival curve of WDR69. (D) Survival curve of HOXB4. Two-hundred and ninety-six patients from the TCGA were divided into a hyper-methylated and hypo-methylated groups according to the median value of the four SR-CDMGs. P-values were calculated with the log-rank test.

Figure 6. Survival analysis of the predictive model

(A) Distribution of the risk score of the predictive model, the survival status of patients and the methylation level of the four SR-CDMGs. (B) Kaplan–Meier survival curve of the predictive model. Patients were divided into low-risk and high-risk groups according to the median value of the risk score. P-values were calculated with the log-rank test.

Figure 7. Stratification analysis of the predictive model

(A) Univariate COX regression analysis of the predictive model in different subgroups stratified by important clinical variables. (B,C) Kaplan–Meier survival analysis of the predictive model in subsets of different AJCC stage patients with HCC (log-rank test). Other: Asian and black patients. Risk factors: hepatitis B, hepatitis C, hemochromatosis, cirrhosis, alcohol consumption, non-alcoholic fatty liver disease or α-1 antitrypsin deficiency.