Genome-wide DNA methylation markers associated with metabolic liver cancer

Abstract

Background and aims: Metabolic liver disease is the fastest rising cause of hepatocellular carcinoma (HCC) worldwide, but the underlying molecular processes that drive HCC development in the setting of metabolic perturbations are unclear. We investigated the role of aberrant DNA methylation in metabolic HCC development in a multicenter international study.

Methods: We used a case-control design, frequency-matched on age, sex, and study site. Genome-wide profiling of peripheral blood leukocyte DNA was performed using the 850k EPIC array. Cell type proportions were estimated from the methylation data. The study samples were split 80% and 20% for training and validation. Differential methylation analysis was performed with adjustment for cell type, and we generated area under the receiver-operating curves (ROC-AUC).

Results: We enrolled 272 metabolic HCC patients and 316 control patients with metabolic liver disease from six sites. Fifty-five differentially methylated CpGs were identified; 33 hypermethylated and 22 hypomethylated in cases versus controls. The panel of 55 CpGs discriminated between cases and controls with AUC=0.79 (95%CI=0.71-0.87), sensitivity=0.77 (95%CI=0.66-0.89), and specificity=0.74 (95%CI=0.64-0.85). The 55-CpG classifier panel performed better than a base model that comprised age, sex, race, and diabetes mellitus (AUC=0.65, 95%CI=0.55-0.75, sensitivity=0.62 (95%CI=0.49-0.75) and specificity=0.64 (95%CI=0.52-0.75). A multifactorial model that combined the 55 CpGs with age, sex, race, and diabetes, yielded AUC=0.78 (95%CI=0.70-0.86), sensitivity=0.81 (95%CI=0.71-0.92), and specificity=0.67 (95%CI=0.55-0.78).

Conclusions: A panel of 55 blood leukocyte DNA methylation markers differentiates patients with metabolic HCC from control patients with benign metabolic liver disease, with a slightly higher sensitivity when combined with demographic and clinical information.

Keywords: HCC; Liver cancer; MASLD; NAFLD; metabolic dysfunction-associated steatotic liver disease; metabolic liver disease.

Figure 1.. Epigenome-wide analysis for selection of differentially methylated CpGs associated with metabolic HCC.

The analysis was performed among 272 Metabolic HCC cases and 316 metabolic controls. (A) Manhattan plot with false discovery rate (FDR)-adjusted p-value threshold (red horizontal line) for selection of significant CpGs (q-value<0.05; n=164 CpGs) in the training data for further screening. (B) Q-Q plot of CpGs showing a lambda (λ) value that is closer to 1. (C) Volcano plot of the 164 FDR-significant CpGs, showing hypomethylated CpGs in red color and hypermethylated CpGs in green color among cases versus controls in the training data. (D) Results of a LASSO regression model with 10-fold cross validation, reducing the 164 FDR-significant CpGs to a parsimonious list of 55 CpGs with non-zero coefficients (33 hypermethylated and 22 hypomethylated) and scaling of absolute importance of each CpG in the presence of the other CpGs. This is the final set of CpGs used for the primary analysis.

Figure 2.. Distinguishing metabolic HCC from benign metabolic liver disease using demographic and clinical variables and differentially methylated CpGs.

The study sample comprised 272 Metabolic HCC cases and 316 metabolic controls. (A) Training and validation results from area under the receiver operating characteristic curve (AUC-ROC) analysis for a model that included age (continuous), sex, race (White, other), and type II diabetes mellitus (yes, no). (B) AUC-ROC analysis for a model that included only the 55 differentially methylated CpGs as shown in Table 2. (C) An elaborate multifactorial AUC-ROC analysis for a model that included age, sex, race, diabetes mellitus. and the 55 CpGs. Abbreviations: AUC, area under the receiver operating curve; HCC, hepatocellular carcinoma; sens., sensitivity; spec.: specificity.

Figure 3.. Discriminating between metabolic HCC and metabolic liver disease in a subgroup of participants with genetic data.

These analyses were performed among 75% of the study sample (n=439). (A) Training and validation results from area under the receiver operating characteristic curve (AUC-ROC) analysis for a model that included age (continuous), sex, race (White, other), diabetes mellitus (yes, no), and PNPLA3-rs738409 genotype. (B) Training and validation results for a model that included only the 55 differentially methylated CpGs as shown in Table 2. (C) Multifactorial AUC-ROC analysis for metabolic HCC combining the clinical and demographic variables with CpGs. This multifactorial model was built using LASSO regression with 10-fold cross validation and examining the clinical and demographic variables and the 55 CpGs. However, only 44 CpGs with non-zero coefficients were retained in addition to age, sex, race, diabetes mellitus, and rs738409 for prediction modeling. Abbreviations: AUC, area under the receiver operating curve; HCC, hepatocellular carcinoma; sens., sensitivity; spec.: specificity.

Figure 4.. Characterizing metabolic HCC using hypermethylated CpGs only, and in combination with clinical, demographic, and PNPLA3-rs738409.

The analysis was performed among 272 Metabolic HCC cases and 316 metabolic controls. (A) LASSO regression with scaled absolute importance of 42 hypermethylated CpGs used for the CpGs only model. (B) Differential distribution of the combined product of the 42 hypermethylated CpGs (estimated coefficients x beta values) between cases and controls. (C) Heatmap of 42 selected CpGs in the training data. (D) Modeling of area under the receiver operating characteristic curves (AUC-ROCs) for the hypermethylated CpGs only (n=42) in the training and validation samples. (E) A separate model that evaluated the combination of age (continuous), sex, race (White, other), type II diabetes mellitus (yes, no), and the hypermethylated CpGs in a distinct LASSO regression model with 10-fold cross validation, retaining 40 hypermethylated CpGs plus age, sex, race, and diabetes for prediction modeling. (F) A subgroup analysis modeling AUCs for the hypermethylated CpGs plus age, sex, race, diabetes, and PNPLA3-rs738409 among participants with genetic data (n=439) using a separate LASSO regression with 10-fold cross validation. This analysis retained 38 CpGs, age, sex, race, diabetes, and rs738409 for prediction modeling in the training (n=346) and validation (n=93) samples. Abbreviations: AUC, area under the receiver operating curve; HCC, hepatocellular carcinoma; sens., sensitivity; spec.: specificity.