Circulating cell-free DNA methylation-based multi-omics analysis allows early diagnosis of pancreatic ductal adenocarcinoma

Abstract

Pancreatic ductal adenocarcinoma (PDAC) is a highly aggressive cancer with a 5-year survival rate of 7.2% in China. However, effective approaches for diagnosis of PDAC are limited. Tumor-originating genomic and epigenomic aberration in circulating free DNA (cfDNA) have potential as liquid biopsy biomarkers for cancer diagnosis. Our study aims to assess the feasibility of cfDNA-based liquid biopsy assay for PDAC diagnosis. In this study, we performed parallel genomic and epigenomic profiling of plasma cfDNA from Chinese PDAC patients and healthy individuals. Diagnostic models were built to distinguish PDAC patients from healthy individuals. Cancer-specific changes in cfDNA methylation landscape were identified, and a diagnostic model based on six methylation markers achieved high sensitivity (88.7% for overall cases and 78.0% for stage I patients) and specificity (96.8%), outperforming the mutation-based model significantly. Moreover, the combination of the methylation-based model with carbohydrate antigen 19-9 (CA19-9) levels further improved the performance (sensitivity: 95.7% for overall cases and 95.5% for stage I patients; specificity: 93.3%). In conclusion, our findings suggest that both methylation-based and integrated liquid biopsy assays hold promise as non-invasive tools for detection of PDAC.

Keywords: cfDNA; liquid biopsy; machine learning; methylation; mutation; pancreatic ductal adenocarcinoma.

Fig. 1

Flowchart of the study design. Analyses marked with asterisk were conducted in samples with complete measurements. CA19‐9, carbohydrate antigen 19‐9; cfDNA, circulating cell‐free DNA; DMR, differentially methylated regions; NAT, normal tissue adjacent to tumor; PDAC, pancreatic ductal adenocarcinoma.

Fig. 2

Mutation landscape of plasma cfDNA and mutation‐based diagnostic models for PDAC. (A) Correlation of AFs for shared mutations between cfDNA and paired WBC. Mutational landscape of plasma cfDNA in healthy controls (B) and (C) PDAC patients. Each column represents a PDAC or healthy plasma sample. Upper bar chart represents the number of mutations in each sample. Lower waterfall diagram depicts the mutated genes in each sample. Top 10 mutated genes in healthy (D) and PDAC plasma cfDNA (E). (F) The upper heatmap shows the mutant hotspots, and color depicts the level of mutant AFs. The Bottom plot demonstrates AFs of variants detected by hotspot status (G) Performance of the diagnostic models in the training (left) and testing (right) dataset using different indicators of mutational status. (H) PDAC sensitivity in the testing set by stage at the specificity of 95.9%. AF, allele fractions; cfDNA, circulating cell‐free DNA; PDAC, pancreatic ductal adenocarcinoma; WBC, white blood cells.

Fig. 3

Differentially methylated regions (DMRs) discovered by targeted bisulfite sequencing of PDAC tumor and NAT tissues. (A) Circos plot showing the distribution of PDAC‐specific DMRs across the genome. Red points: hyper‐DMRs. Blue points: hypo‐DMRs. Circles from outer to inner circle were the overview of DMRs, the area statistics of hypermethylated regions, and hypomethylated regions, respectively. (B) Heatmaps showing DMR methylation levels in tissue data. (C) Locations of DMRs in genome. (D) GO term annotation of DMRs. DMR, differentially methylated regions; HMG, high Mobility Group; NAT, normal tissue adjacent to tumor; PDAC, pancreatic ductal adenocarcinoma; UTRs, untranslated regions.

Fig. 4

Methylation‐based PDAC diagnostic models. Performance of random forest model in the training set using all DMRs (A), and 200 DMRs selected by Boruta algorithm (B) as features. Performance in the training set using 13 DMRs selected by Boruta followed by RFE (C) and 6 DMRs in addition filtered by tissue‐plasma concordance (D). (E) Comparison of performance of the above 4 diagnostic models in the testing set. (F) PDAC sensitivity in the testing set by stage (specificity = 96.8%). AUC, area under the curve; RFE, recursive feature elimination; ROC, receiver operation characteristics.

Fig. 5

Multi‐omics based PDAC diagnostic models. (A) Performance of the diagnostic model in the training set based on 6 selected DMR markers. Performance of the diagnostic model in the training set based on methylation in combination with KRAS/TP53 mutation (B) or mutation in top 10 genes (C) or hotspot mutation (D). (E) Performance in methylation‐based diagnostic model in the testing set with and without mutational status. (F) Sensitivities of diagnostic models in testing set based on different combination of analytes. AUC, area under the curve; CA19‐9, carbohydrate antigen 19‐9; RFE, recursive feature elimination; ROC, receiver operation characteristics.