IDDF2019-ABS-0302 Identification of DNA Methylation Signatures for Microvascular Invasion in Hepatocellular Carcinoma

Lixia Xu,Sui Pen,Qiye He,Zhihang Chen,Ming Kuang

DOI: https://doi.org/10.1136/gutjnl-2019-iddfabstracts.108

2019-01-01

Abstract:Background The presence of microvascular invasion (MVI) reduces overall survival of hepatocellular carcinoma (HCC). Recent studies showed DNA methylation markers could be applied to the diagnosis of cancers. However, it is unclear whether DNA methylation signatures could help diagnose MVI in HCC. Methods To identify DNA methylation markers for HCC and MVI diagnosis, we first generated genome-wide DNA methylation profiles from HCC tissues and adjacent normal liver tissues of 37 HCC patients (20 MVI+ and 17 MVI-). We also investigated DNA methylation profiles of plasma samples from these 37 HCC patients and 30 normal individuals as controls to minimize the interference of random background DNA methylation signals. Tissue and plasma samples were prepared into DNA methylation library and sequenced on Illumina Hiseq X10 platform. Using methylation haplotype load (MHL) and unmethylation haplotype load (UMHL) as metrics, we quantified DNA methylation profiles on methylation haplotype blocks (MHBs) by computing the degree of linkage between methylated or unmethylated CpGs in HCC and adjacent normal samples to identify discriminatory markers. Grouping samples from our cohort into training and validation sets respectively, we employed two supervised machine learning algorithms, random forest (RF) and support vector machine (SVM) to train and cross-validate binary predictive models. Results Supervised analyses identified 65 MHBs as classifiers for HCC tissues and adjacent normal liver tissues. Both RF- and SVM-built models were highly accurate in classifying HCC and normal liver tissues with AUC no less than 0.98 (AUC: 98%, CI: 97.3%∼98.8% for RF model; AUC: 99.9%, CI: 99.9 ∼ 99.9% for SVM model). We applied the RF-trained classification model to differentiate HCC plasma DNA from those of healthy controls, with AUC of 96% (CI: 95.1% - 96.9%). We further identified 6 MHL-quantified MHBs and 5 UMHL-quantified MHBs as classifiers for MVI- and MVI+ samples. Combining the MHBs’ methylation scores as classifiers and applying RF method, we trained and cross-validated MVI- and MVI+ classification models. We found the AUC of these models is 85.9% (CI: 83.5% - 88.3%). Conclusions These data showed that DNA methylation signatures can provide outstanding diagnostic accuracy for HCC and MVI.