UCLA

Background & aim

Microvascular invasion (MVI) is a critical prognostic risk factor in hepatocellular carcinoma (HCC). This study evaluated the performance of 5-hydroxymethylcytosine (5hmC) modifications in circulating cell-free DNA (cfDNA) in preoperative assessment of MVI.

Methods

A total of 907 patients with HCC were enrolled from two centers, including 671 in the training cohort, 152 in the internal validation cohort, and 84 in the external validation cohort. Preoperative clinical data, laboratory parameters, and cfDNA-derived 5hmC profiles were collected. Feature selection was performed using XGBoost, and modeling was conducted using a multilayer perceptron (MLP) neural network. Survival analyses were performed to evaluate the prognostic significance of the MVI prediction model. RNA sequencing analysis was performed to explore the potential mechanism underlying the proposed model.

Results

The 181-5hmC-modification signature demonstrated strong discriminatory performance, achieving an area under curve (AUC) of 0.852 in the training cohort, 0.862 in the internal validation cohort, and 0.864 in the external validation cohort, respectively. Univariate and multivariate analyses identified the α-fetoprotein (AFP) level (odds ratio [OR] 1.576, P = 0.039), Barcelona Clinical Liver Cancer (BCLC) stage (OR 3.051, P < 0.001), and the 5hmC signature (OR 46.891, P < 0.001) as independent predictors of MVI. The 5hmC signature demonstrated significantly higher predictive accuracy than AFP levels or BCLC stage alone. Survival analysis showed that the 5hmC signature significantly stratified both recurrence-free and overall survival in resectable HCC patients. Additionally, interpretability analysis based on RNA sequencing revealed that lower MVI prediction scores were associated with immune-related pathways and immune infiltration levels.

Conclusions

We developed and validated a circulating cfDNA-derived 5hmC signature that non-invasively predicts preoperative MVI status, with potential clinical utility in the management of resectable HCC.

Impact and implications

This study presents the first integration of cfDNA-derived 5hmC profiling with machine learning for preoperative MVI prediction in resectable HCC. The proposed 5hmC signature demonstrates potential for predicting MVI status and prognosis before surgery. Integration of RNA sequencing analysis provides biological support for the model's predictions, strengthening its clinical relevance. As a blood-based assay, this approach offers practical advantages for potential routine clinical implementation.