Usefulness of multigene liquid biopsy of bile for identifying driver genes of biliary duct cancers

Abstract

Liquid biopsy (LB) is an essential tool for obtaining tumor-derived materials with minimum invasion. Bile has been shown to contain much higher free nucleic acid levels than blood plasma and can be collected through endoscopic procedures. Therefore, bile possesses high potential as a source of tumor derived cell-free DNA (cfDNA) for bile duct cancers. In this study, we show that a multigene panel for plasma LB can also be applied to bile cfDNA for comparing driver gene mutation detection in other sources (plasma and tumor tissues of the corresponding patients). We collected cfDNA samples from the bile of 24 biliary tract cancer cases. These included 17 cholangiocarcinomas, three ampullary carcinoma, two pancreatic cancers, one intraductal papillary mucinous carcinoma, and one insulinoma. Seventeen plasma samples were obtained from the corresponding patients before surgical resection and subjected to the LiquidPlex multigene panel LB system. We applied a machine learning approach to classify possible tumor-derived variants among the prefiltered variant calls by a LiquidPlex analytical package with high fidelity. Among the 17 cholangiocarcinomas, we could detect cancer driver mutations in the bile of 10 cases using the LiquidPlex system. Of the biliary tract cancer cases examined with this method, 13 (54%) and 4 (17%) resulted in positive cancer driver mutation detection in the bile and plasma cfDNAs, respectively. These results suggest that bile is a more reliable source for LB than plasma for multigene panel analyses of biliary tract cancers.

Keywords: bile; biliary duct cancer; liquid biopsy; machine learning; multigene panel.

FIGURE 1

Machine learning process to classify the variant calls by multigene panel liquid biopsy (LB). (A) Numbers of variants at each step are indicated near the box. The result of the classification is indicated in the table at the bottom right. (B) Plot of “somatic” variant numbers in Random Forest modeling by changing the “negative and polymorphic” variant numbers for the training data. Vertical and horizontal axes indicate the numbers of variants classified as “somatic” in the test data and numbers of “negative and polymorphic” variants in the training data, respectively. Arrow indicates the number of “negative and polymorphic” variants that we picked up for modeling. (C) Plot of mean decrease in Gini coefficient in this study. The horizontal axis indicates the mean decrease in Gini score and SDs of the scores among the 15 models. The vertical axis indicates the parameters used in the modeling. The parameters were collected by R script.

FIGURE 2

Annotation patterns of the variants classified by machine learning. (A) Ratio of variants of three functional annotations for three categories. Vertical axis shows the three categories classified by the Random Forest method (from top to bottom: “Polymorphism,” “Somatic,” and “Negative”). Horizontal axis indicates the relative ratio of the variants in the categories, the ratio of matched variants divided by that of all tested variants. Gray, orange, and blue bars indicate the ratios for COSMIC >10 cases, 54KJPN >0.33, and ClinVar Pathogenic (ClinVar P) and likely_pathogenic (LP), respectively. (B) Categorical box‐whisker plots of log of variant allele frequency (VAF) for “negative,” “positive,” and “polymorphism.” X‐axis indicates the three categories. Y‐axis indicates the log of VAF. Upper and lower ends of the boxes indicate 75% and 25% of each category, respectively. Horizontal black lines in the boxes indicate the averages, while the whiskers indicate the 1.5‐fold length of the box vertical plane size of each category. Outliers are shown as dots.

FIGURE 3

Oncoprint of the bile and plasma multigene liquid biopsies (LBs) for biliary tract cancer. Vertical axis indicates the genes in which a driver mutation was detected in this study. Horizontal axis indicates the patients. Results of both the plasma and bile LBs are indicated. Patient identification numbers and tumor types are shown on the top of the oncoprint. Black, blue, and red rectangles indicate the driver mutations as nonsynonymous in tumor suppressor genes, loss of function in tumor suppressor genes, and nonsynonymous in oncogenes, respectively. AC, ampullary carcinoma; CC, cholangiocarcinoma or gallbladder cancer; PC, pancreatic cancer.

FIGURE 4

Droplet digital PCR (ddPCR) analysis of bile cell‐free DNA (cfDNA) samples. (A) ddPCR scatter plot for SMAD4 R361C cfDNAs in BIL12. Vertical axes indicate the VIC signals (mutant), while the horizontal axes indicate the FAM/HEX signals (WT). Magenta lines in the planes indicate the positive signal thresholds. Dots indicate the FAM/HEX and VIC signals of pixels with the colors of the genotype decisions (blue, mutated; green, WT; orange, heterozygous; black, no signal). (B) Corresponding Integrative Genomics Viewer (IGV) images of the BIL12 cfDNA LiquidPlex data. (C) Corresponding IGV images of the PL12 cfDNA LiquidPlex data.