Patient demographic and clinical characteristics

A total of 35 patients and 24 healthy individuals were included (Fig. 1). We grouped 35 patients into 2 categories, including the MVI- group with 17 tissue and plasma samples, and MVI + group with 18 tissue and plasma samples. There were no significant differences in all oncology indicators and liver function indicators between the MVI + group and MVI- group except for AFP, indicating that the baseline data between the two groups were basically balanced (Table 1).

Fig. 1

Study flowchart. HCC hepatocellular carcinoma, MHB methylation haplotype block, MHL methylated haplotype load, MVI microvascular invasion, UMHL un-methylated haplotype load, SVM support vector machine

Table 1 Baseline demographic and clinical characteristics in MVI− and MVI + groupsTissue methylation markers to distinguish HCC tissues from non-tumor tissues

To screen MHB for classifiers for HCC, we quantified the DNA methylation status of all sequenced MHBs by MHL, and then performed unsupervised clustering based on those scores to visualize the degree of separation between HCC tissues and non-tumor tissues. Results showed that based on MHL scores, the HCC tissues had consistently been separated from the normal liver tissues (Fig. 2A), which demonstrated that there were profound differences in DNA methylation patterns between HCC and peritumoral normal liver tissues. We further performed principal component analyses (PCA) on the MHL scores of normal liver and HCC tissue, and the results showed that the normal liver tissues were also separated from the HCC tissues (Fig. 2B), in agreement with the results observed in unsupervised clustering.

Fig. 2

DNA methylation markers classify normal liver and HCC tissues with high degree of accuracy. A Unsupervised clustering of normal and HCC tissues based on top 100 MHBs that have the highest degree of variations in their MHL scores; B PCA analyses show clear separation between normal liver tissues and HCC tissues; C supervised analyses identified 65 MHBs as classifiers for normal liver tissues and HCC tissues; D RF-built classification models using the 65 MHB markers accurately classified normal liver tissues and HCC tissues, as was demonstrated by the AUC of their ROC curves; E top biological function categories of the identified MHB markers

After filtering MHBs and libraries, we performed Wilcoxon signed-rank test to the dynamic MHBs based on their MHL scores, and identified 65 MHBs whose methylation scores were significantly different between normal liver tissues and HCC tissues (False Discovery Rate (FDR) < 0.05) (Fig. 2C). Using these MHBs and their MHL scores as independent variables, and the liver tissues from our study cohort as training and validation sample sets, we separately employed two supervised machine learning algorithms, RF and SVM to train and cross-validate binary predictive models to classify normal or HCC liver tissues. Results showed both RF- and SVM-built models were highly accurate in classifying HCC and normal liver tissues: the AUC is no less than 0.98 (RF model: AUC = 98.0%, 95% confidence interval CI 97.3–98.8%, 10-time repeat, Fig. 2D; SVM model: AUC = 99.9%, 95% CI 99.9–99.9%, 10-time repeat, Additional file 1: Figure S1A).

Gene Ontology analysis of HCC methylation markers

To investigate the potential biological functions of these methylation markers, especially their roles in the pathology of HCC, genes associated with identified MHBs were annotated and analyzed based on their known biological and molecular functions. Results showed the biological function categories that had the highest level of enrichment were those involved in the general or specific processes of embryonic development/differentiation (Fig. 2E), such as pattern formation, embryonic organ morphogenesis and development, regionalization, sensory organ and skeletal system morphogenesis, etc. This was consistent with the finding that many embryonic genes were re-expressed in cancer cells [27], suggesting that these methylation markers might regulate these embryonic genes’ expression during HCC progression. Indeed, when those genes were re-analyzed based on their molecular function, the top categories with the highest level of enrichment were transcription factors that regulate the expression of other genes (Additional file 1: Figure S1B). This suggests that some of the identified MHB classifiers may be pivotal to the progression of HCC by regulating cascades of gene expression that are signatory to HCC.

cfDNA methylation signature to distinguish HCC patients from healthy individuals

To construct a cfDNA methylation signature for HCC, we sequenced methylation libraries of cfDNA samples from 35 HCC plasma samples of our study cohort and from 24 healthy individuals. We applied the RF-trained classification model to classify the filtered plasma DNA libraries (Fig. 3A). Results showed that it had an AUC of 96.0% (95% CI 95.1–96.9%) from the ROC curve (Fig. 3B) in identifying HCC plasma.

Fig. 3

HCC tissue markers were able to classify normal and HCC plasmas. A Heatmap of the MHL scores of the 65 MHB tissue classifiers in normal and HCC plasmas samples; B normal and HCC plasma samples were classified using the 65 MHB markers by RF method, which demonstrated high degree of accuracy in classification

DNA methylation markers for the prediction of MVI status

To identify DNA methylation markers that differentiate MVI− and MVI + HCC tissues, we performed Wilcoxon rank-sum test (FDR < 0.05) on the MVI− and MVI + HCC tissues’ DNA methylation profiles on their MHL and UMHL scores, and identified 3 MHL-quantified MHBs and 5 UMHL-quantified MHBs as classifiers for MVI− and MVI + tissues (Fig. 4A). We combined these MHBs’ and trained a Random-forest MVI classification models on our training cohort samples. When being cross-validated, this model showed an AUC of the model was 85.9% (Fig. 4B, 10-time repeats, 95% CI 83.5–88.3%), suggesting that the identified DNA methylation markers consistently and robustly differentiate MVI− and MVI + HCC tissues. Further, survival analysis revealed that RFS rate and OS rate were significantly worse in MVI + group predicted by our methylation markers than MVI- group predicted by our methylation markers, which validated high accuracy of our tissue DNA methylation markers for MVI status prediction (Fig. 4C–D).

Fig. 4

DNA methylation markers differentiate MVI- and MVI + tissues. A Heatmap of MHL and UMHL scores of the 8 MVI markers in MVI- and MVI + tissues; B RF-built models using discovered MVI markers accurately classified MVI- and MVI + tissues in cross-validation; C–D recurrence-free survival rate (C) and overall survival rate (D) of HCC patients in MVI+ and MVI− groups predicted by tissue DNA methylation markers. Log-rank test was used

The performance of DNA methylation signature to predict MVI status

To compare the prediction performance of the DNA methylation signature to the clinical characteristics, we performed ROC analysis for predicting MVI based on the diagnosis results of DNA methylation signature and clinical characteristics, respectively. We found that the AUROC of the DNA methylation signature was up to 91.5% (95% CI 82.1–100.0%) (Fig. 5A). In contrast, none of the AUROCs of clinical characteristics were more than 80.0%: AFP, 75.6% (95% CI 55.1–89.2%); TNM, 58.5% (95% CI 47.1–69.9%); BCLC, 56.5% (95% CI 41.0–72.1%); Tumor Size, 52.7% (95% CI 33.8–70.8%); and HBsAg, 52.6% (95% CI 43.2–62.0%) (Fig. 5B−F). In addition, we also integrated AFP into our tissue DNA methylation markers to examine whether AFP could improve performance of our tissue methylation makers for MVI status prediction (Additional file 2: Figure S2A). Unfortunately, the integrated methylation markers only achieved an AUC of 86.3%, which is not superior to our tissue DNA methylation markers (Additional file 2: Figure S2B).

Fig. 5

ROC analysis for diagnosing MVI based on the diagnosis results of DNA methylation signature and clinical characteristics. A ROC analysis for diagnosing MVI based on the diagnosis results of DNA methylation signature. B–F ROC analysis for diagnosing MVI based on clinical characteristics included AFP (B), TNM (C), BCLC (D), tumor size (E), and HBsAg (F)

Furthermore, we performed a univariable and multivariable logistic regression analysis to assess the associations between MVI and the DNA methylation signature or other clinical characteristics. On univariate analysis, variables associated with MVI were the DNA methylation signature (p < 0.001) and AFP ≥ 400 ng/mL (p = 0.028) (Table 2). The other parameters were not significantly correlated with MVI. At multivariate analysis, only the DNA methylation signature (p < 0.001; odds ratio, 47.51, 95% CI 5.74–393.27) was independent risk factors for MVI (Table 2).

Table 2 Univariate and multivariate analysis of risk factors for MVI of HCCDNA methylation signature and RFS

On univariate analysis, DNA methylation signature was associated with RFS (HR 7.89, 95% CI 2.16–28.88, p = 0.002) and by MVI status (HR 32.22, 95% CI 4.06–255.62, p = 0.001) (Additional file 5: Table S2). Other factors associated with RFS were AFP, ≥ 400 ng/mL (HR 2.96, 95% CI 1.04–8.39, p = 0.042) (Additional file 5: Table S2) and imaging tumor thrombus (HR 4.69, 95% CI 1.44–15.28, p = 0.010) (Additional file 5: Table S2). On multivariate analysis, the DNA methylation signature was independently associated with RFS (HR 97.85, 95% CI 3.21–2.98e + 03, p = 0.009) (Fig. 6) and by MVI status (HR 8.96e + 02, 95% CI 8.57–9.38e + 04, p = 0.004) (Additional file 3: Figure S3).

Fig. 6

Multivariate Cox analysis of clinicopathologic factors and DNA methylation signature with recurrence-free survival