Ultra-low-pass whole-genome sequencing of bile cfDNA in patients with biliary strictures

Abstract

Background

Accurate discrimination of malignant versus benign biliary strictures remains clinically challenging. We previously showed that targeted bile cell-free DNA (cfDNA) mutation analysis (Bilemut assay) enables sensitive detection of malignancy, but it does not distinguish cholangiocarcinoma (CCA) from pancreatic ductal adenocarcinoma (PDAC), and its prognostic value remains unclear. In this study, we evaluated ultra-low-pass whole-genome sequencing (ULP-WGS) of bile cfDNA as a potentially cost-effective, technically simple and low-DNA input demanding genomic approach for early tumour detection, tumour-origin discrimination and prognostic stratification.

Methods

Bile samples from 95 patients with biliary strictures under radiological suspicion of malignancy, classified as malignant (n=26) or indeterminate (n=69) based on the result of cytology or biopsy at the first endoscopic retrograde cholangiopancreatography (ERCP), were analysed by ULP-WGS. Tumour fraction (TF) and copy-number alterations (CNAs) were inferred using ichorCNA. Tumour mutational burden (TMB) was assessed with the Bilemut panel. Overall survival was analysed by Kaplan-Meier analysis and log-rank test. A supervised classifier based on discriminant analysis of principal components was used to distinguish CCA from PDAC.

Results

In matched samples, bile LP-WGS showed higher sensitivity than plasma for tumour detection. Among patients with an initial malignant diagnosis, bile ULP-WGS detected circulating tumour DNA (ctDNA) in 73% of cases. In patients initially classified as indeterminate but later confirmed as malignant, tumour DNA was detected in 51%, anticipating final diagnosis by a median of 38 days; notably, in seven patients, ctDNA was detectable more than 100 days before routine diagnosis, highlighting its potential for very early tumour detection. Specificity in strictures with a final benign outcome was 96%. CNA profiling revealed recurrent chromosomal alterations consistent with tissue-based data, and a 50-region CNA classifier discriminated CCA from PDAC with an accuracy of 0.864. Although neither TF nor TMB alone was associated with overall survival, their combination identified a subgroup with poorer prognosis.

Conclusions

Our exploratory study shows that bile ULP-WGS performed at the first ERCP could represent a complementary, scalable and potentially cost-effective genomic strategy for early malignancy detection in biliary strictures, with additional value for tumour-origin classification and, combined with TMB analysis, for identifying patients with poorer prognosis.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Distinguishing benign from malignant biliary strictures remains challenging. Bile cell-free DNA mutational analysis can improve malignancy detection, but whether a lower-cost genome-wide approach could provide additional diagnostic, tumour-origin and prognostic information was unknown.

WHAT THIS STUDY ADDS

• Bile ultra-low-pass whole-genome sequencing (ULP-WGS) at the first endoscopic retrograde cholangiopancreatography (ERCP) detects tumour DNA with high specificity and can identify malignancy before routine diagnosis. It also captures copy-number alterations that help distinguish cholangiocarcinoma from pancreatic ductal adenocarcinoma, and combined with mutational burden identifies patients with poorer outcomes.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Bile genomic analysis could complement routine ERCP-based evaluation of biliary strictures, especially in indeterminate cases. As a low-cost, simple and accessible strategy, ULP-WGS could serve as a companion early diagnostic tool helping in tumour-origin assignment and risk stratification.

Visual abstract. Schematic overview of the obtained results. Bile cfDNA from patients with biliary strictures obtained at the first ERCP was analysed by ULP-WGS and NGS of 52 targeted genes (Bilemut assay). ULP-WGS enables early malignancy detection with high specificity and tumour-origin discrimination through TF and CNA profile. Integration of TF with the targeted TMB identifies a subgroup of patients with poorer prognosis. Figure created using BioRender. com, CCA, cholangiocarcinoma; cfDNA, cell-free DNA; ERCP, endoscopic retrograde cholangiopancreatography; NGS, next- generation sequencing; PDAC, pancreatic ductal adenocarcinoma; ULP-WGS, ultra-low-pass whole-genome sequencing .

Introduction

The most common malignant causes of biliary strictures include cholangiocarcinoma (CCA) and pancreatic ductal adenocarcinoma (PDAC).^1,2 However, biliary strictures under radiological suspicion of malignancy remain a frequent and clinically challenging scenario. Despite imaging findings suggestive of malignancy, definitive histological diagnosis is often difficult to establish, as conventional tools including endoscopic retrograde cholangiopancreatography (ERCP) with brush cytology or biopsy and endoscopic ultrasound (EUS)-guided sampling often yield inconclusive results or lack sufficient sensitivity, particularly in early disease or when tissue is limited.^3–6 Moreover, the performance of serum carbohydrate antigen 19-9 (CA19-9) is suboptimal, as levels are not elevated in many tumours and may also increase in benign conditions such as cholestasis and inflammation.^6,7 As a consequence, many patients experience prolonged diagnostic pathways with multiple invasive procedures, delayed therapy initiation and avoidable morbidity.^8–10

Over the past years, liquid biopsy has been increasingly explored in oncology as a complementary approach that detects tumour-derived molecules in body fluids to support diagnosis and tumour stratification.^11–14 In the context of patients with biliary strictures, ERCP is routinely performed and, without additional risk, allows bile collection. Bile is an especially attractive biofluid because it is obtained at the lesion site and, in pancreatobiliary cancers or premalignant scenarios, typically harbours higher concentrations of tumour-derived DNA than peripheral blood.^15–18 In a recent prospective proof-of-concept study including patients with indeterminate biliary strictures—many of whom were later confirmed as malignant on follow-up—we demonstrated the robust diagnostic potential of the mutational analysis of bile cell-free DNA (cfDNA) collected at the time of the first ERCP with a commercialised next-generation sequencing (NGS) panel, an assay that we named Bilemut.¹⁶ These findings are concordant with independent series demonstrating the feasibility and diagnostic value of bile cfDNA mutational profiling.^17,19–21 In parallel, we recently extended the utility of the Bilemut assay to the management of patients with primary sclerosing cholangitis (PSC), demonstrating that bile cfDNA can reveal molecular risk features in a population predisposed to CCA.²²

Beyond point mutations, somatic copy-number alterations (CNAs) are pervasive across solid tumours and carry biological and prognostic significance.^23,24 These aneuploidy patterns can also be captured in circulating tumour DNA (ctDNA), not only by deep whole-genome sequencing (WGS, >5×) but also at lower depth using low-pass WGS (LP-WGS, approximately 1–3×), with reported utility for early detection and for monitoring progression in several cancers.^25–27 In pancreatobiliary tumours, LP-WGS-based liquid biopsy studies remain limited and are based on small series, although their findings broadly point in the same direction. Specifically, an approximately 3× plasma study in biliary tract cancer (n=29) found that higher CNA burden was associated with worse prognosis²⁸; a bile LP-WGS study in CCA (n=18) was largely descriptive without outcome associations²⁹; an approximately 5× plasma study in metastatic PDAC (n=70) found that higher CNA burden was associated with shorter survival³⁰; and a recent 2−3×plasma study in PDAC (n=40) reported limited correlation of LP-WGS outcomes with clinical response.³¹ However, sequencing at approximately 1–3× is relatively costly, limiting scalability for routine clinical application.³² To mitigate costs while retaining informative CNA profiling, ultra-LP-WGS (ULP-WGS; ≤0.5×) has emerged as a cost-efficient alternative.^33–37 Moreover, pairing ULP-WGS with an open-source workflow such as ichorCNA—which estimates tumour fraction (TF) in addition to arm-level CNAs²⁶—ensures transparency and reproducibility across centres. To our knowledge, ULP-WGS (≤0.5×) has not been evaluated in liquid biopsy matrices (plasma or bile) in the context of pancreatobiliary tumours.

Therefore, using bile collected at the first ERCP from patients with biliary strictures under radiological suspicion of malignancy, classified as malignant or indeterminate based on the cytology or biopsy result obtained at that procedure, and given that the Bilemut targeted panel alone did not stratify prognosis or distinguish tumour origin (CCA vs PDAC), we performed an exploratory study to evaluate whether ULP-WGS, as a cost-effective, technically simple and low-input genomic approach, could: (1) serve as a diagnostic complement for indeterminate biliary strictures; (2) provide genome-wide information (arm-level CNAs and TF) to inform prognosis and (3) discriminate tumour origin (CCA vs PDAC).

Methods

All methods were carried out in accordance with relevant guidelines and regulations.

Study design

Bile samples were prospectively collected at the first ERCP from 95 adults with biliary strictures and radiological suspicion of malignancy at Navarra University Hospital between January 2017 and December 2024; paired plasma was available for 10 patients. Patients with clearly benign radiological aetiologies were excluded. Initial diagnosis was based on ERCP-guided brushing or biopsy, EUS-guided fine needle aspiration (FNA) or percutaneous biopsy, classifying strictures as malignant (n=26) or indeterminate (n=69). For initially indeterminate cases, final diagnosis was established after up to 12 months of follow-up. Malignant stricture was defined by histological evidence of adenocarcinoma, clinical or radiographic progression consistent with malignancy within 12 months, or death attributed to a malignancy involving the bile duct. Benign stricture was defined by absence of clinical and radiographic progression or resolution of the bile duct stricture after 12 months of follow-up, or by histopathological examination of surgically resected specimens confirming benign cholangiopathy. CA19-9 was analysed using prespecified cut-offs of 44 U/mL for diagnostic discrimination and 500 U/mL for prognostic stratification.

Bile collection, sample processing and cfDNA extraction

Bile was collected during standard ERCP from overnight-fasting patients. After bile duct cannulation and usually before contrast injection, 2–6 mL of bile was aspirated through the sphincterotome, kept at 4°C, centrifuged at 3500 ×g for 10 min at 4°C, aliquoted and stored at −80°C within 2 hours of collection. When available, plasma was obtained from 10 mL of ethylenediaminetetraacetic acid (EDTA) blood by two sequential centrifugations and stored at −80°C.

Before cfDNA isolation, bile and plasma were thawed at 4°C and centrifuged at 13 000 ×g for 10 min at 4°C. cfDNA was extracted using the QIAamp Circulating Nucleic Acid Kit (Ref 55114; QIAGEN, Hilden, Germany), quantified with the Qubit dsDNA High Sensitivity assay, and stored at −80°C until analysis.

Library preparation and sequencing

An input of 2 ng of cfDNA was used for both LP-WGS and ULP-WGS libraries, constructed with the NEBNext Ultra II DNA Library Prep Kit (Ref NEB#E7645L; New England Biolabs, Ipswich, Massachusetts, USA). Libraries were pooled and sequenced on a NextSeq 2000 platform (Illumina, San Diego, CA, USA) using 100 bp paired-end runs, targeting approximately 3× (LP-WGS) and approximately 0.2× (ULP-WGS) genome-wide coverage.

Copy-number analysis and TF estimation

Raw reads (FASTQ) were quality-filtered with Trim Galore (minimum length 50 bp) and aligned to hg19 with Bowtie2. BAM files were processed with Samtools and polymerase chain reaction (PCR) duplicates marked with Picard MarkDuplicates. Copy-number analysis and TF estimation were performed with ichorCNA,²⁶ which bins reads in non-overlapping 1 megabase (Mb) windows, corrects for guanine–cytosine (GC) content and mappability and applies a hidden Markov model to identify large-scale CNAs and estimate the tumour-derived fraction. Samples were classified as ctDNA-positive when TF ≥10%.²⁶

Bilemut targeted sequencing assay

Bile cfDNA was also analysed using the Bilemut assay, based on the Oncomine Pan-Cancer Cell-Free Assay, a targeted NGS panel covering hotspot regions in 52 cancer-related genes, including genes frequently altered in pancreatobiliary malignancies. Libraries were prepared from 50 ng of cfDNA and sequenced on an Ion S5 System. Only variants with mutant allele fraction (MAF) ≥0.1% were reported. Tumour mutational burden (TMB) was approximated as the number of mutated genes detected within the Bilemut panel.

Machine-learning models for bile cfDNA–based classification

Synthetic training data were generated using the algorithm described in a previous study by Urman et al,³⁸ modelling the joint distribution of ichorCNA-derived copy-number features from a reference cohort of 22 TF-positive samples by preserving intervariable correlations. Recurrent neural networks were trained exclusively on synthetic data using the caret (Version 6.0-86) and neuralnet R packages^39–41 for binary classification (benign vs malignant; CCA vs PDAC). Model performance was then evaluated on real patient data, including the reference cohort (n=22) used to generate the synthetic data and a separate validation cohort of 20 TF-positive samples.

For CCA versus PDAC discrimination, discriminant analysis of principal components (DAPC),⁴² implemented with the adegenet R package, was applied to select the 50 ichorCNA-derived genomic regions best separating the two tumour types.

Statistical analysis

Quantitative variables were expressed as medians and interquartile ranges (IQRs) and compared using the Mann-Whitney U test. Categorical variables were reported as frequencies and percentages and compared using Fisher’s exact test or the χ² test, as appropriate. Diagnostic sensitivity and specificity were calculated with 95% confidence intervals (CIs) using the Wilson score method. Overall survival was estimated by the Kaplan-Meier method and compared with the log-rank test. Cox proportional hazards models were used to assess associations between bile-derived genomic markers and overall survival, adjusting for relevant clinical covariates. All p values were two-sided, and p values <0.05 were considered statistically significant. Analyses were performed using R Version 4.4.0 and SPSS Version 25.0 (IBM Corp., Armonk, NY, USA).

Additional methodological details are provided in the online supplemental data.

Results

Patient cohort and baseline characteristics

A cohort of 95 patients undergoing ERCP for a biliary stricture with radiological suspicion of malignancy and with an available bile sample was accrued (figure 1). Based on the pathological result of the cytology or biopsy obtained at the first ERCP, strictures were classified as malignant (n=26; 17 CCA and 9 PDAC) or indeterminate (n=69). After a follow-up period of up to 12 months, 45 out of the 69 initially indeterminate cases were ultimately confirmed as malignant (23 CCA and 22 PDAC), whereas the other 24 were confirmed as benign, yielding a final distribution of 71 malignant cases (40 CCA and 31 PDAC) and 24 benign cases (figure 1).

Figure 1. Patient cohort. Flow chart showing the initial (at the time of ERCP) and final clinical diagnosis after up to 12 months of follow-up. CCA, cholangiocarcinoma; ERCP, endoscopic retrograde cholangiopancreatography; PDAC, pancreatic ductal adenocarcinoma.

The cohort comprised individuals with a median age of 72 years (range 42–94), 60% of whom were male. Most strictures were distal (75%), with perihilar (18%) and intrahepatic (7%) locations less frequent. CA19-9 levels were ≥44 U/mL in 63% of evaluable cases, and ≥500 U/mL in 33% (table 1). These CA19-9 cut-offs were prespecified for diagnostic^16,43 and prognostic purposes,⁴⁴ respectively.

Table 1. Baseline demographics and clinical characteristics of the study cohort (n=95). Values are median (range) or n (%), unless otherwise specified.

Characteristics	All n=95
Age, years (median; range)	72 (42–94)
Sex (male), n (%)	57 (60)
Type of stricture, (%)
Malignant	26 (27)
CCA	17 (18)
PDAC	9 (9)
Suspicion of malignancy	69 (73)
CA19-9 ≥44 U/mL
No	28 (30)
Yes	60 (63)
Missing	7 (7)
CA19-9 ≥500 U/mL
No	57 (60)
Yes	31 (33)
Missing	7 (7)
CEA
<5 ng/mL	45 (47)
≥5 ng/mL	35 (37)
Missing	15 (16)
Location of stricture, n (%)
Distal	71 (75)
Intrahepatic	7 (7)
Perihilar	17 (18)

CA19-9 and CEA levels were dichotomised according to clinically relevant cut-offs (44 U/mL and 500 U/mL for CA19-9 and 5 ng/mL for CEA, respectively).

CA19-9, carbohydrate antigen 19-9; CCA, cholangiocarcinoma; CEA, carcinoembryonic antigen; PDAC, pancreatic ductal adenocarcinoma.

Bile outperformed plasma for tumour detection by low-pass and ultra-low-pass whole-genome sequencing

The primary goal of this study was to evaluate ULP-WGS (≤0.5× depth) of bile cfDNA as a potentially cost-effective, technically simple and low-input genomic approach. However, a recent study using approximately 3× depth and a proprietary workflow detected CNAs in plasma from patients with biliary tract cancer with 89% sensitivity,²⁸ exceeding rates reported in other tumours.^25,30,35,45 Given prior evidence, including our own, that bile cfDNA outperforms plasma for pancreatobiliary cancer mutation detection,^16,17,19,20 we compared plasma and bile cfDNA using LP-WGS and ULP-WGS, with the open-source pipeline ichorCNA^30,35,46,47 to profile CNAs and estimate TF.²⁶

Following the sequencing depth reported by Wang et al,²⁸ we first applied LP-WGS at a sequencing depth of 3× to plasma samples from 10 patients with malignant biliary strictures. A TF of ≥10% was detected in only 1 of the 10 plasma samples tested (10%) (figure 2). This sensitivity is markedly lower than the 89% sensitivity reported,²⁸ which may reflect differences in patient cohorts and, importantly, methodological differences between analytical pipelines; particularly because the proprietary workflow implemented by Wang et al²⁸ is not publicly available.

Figure 2. Bile outperforms plasma at matched depth and under ultra-low-pass coverage. (A) Results of plasma LP-WGS (3×), bile LP-WGS (3×), bile ULP-WGS (0.2×) and Bilemut in 10 paired malignant cases. Each square represents one case (red indicates positive and green negative). For LP-WGS and ULP-WGS, positivity was defined as TF ≥10% estimated by ichorCNA. For Bilemut, positivity was defined as ≥1 somatic mutation with MAF ≥0.1%. Statistical comparisons were performed using two-sided Fisher’s exact tests. (B) TF (%) estimated by ichorCNA for the same 10 malignant patients shown in (A), comparing plasma approximately 3× (red), bile approximately 3× (dark green), bile approximately 0.2× (light green). LP-WGS, low-pass whole-genome sequencing; MAF, mutant allele fraction; TF, tumour fraction; ULP-WGS, ultra-LP-WGS.

We then analysed the paired bile samples from the same 10 patients by LP-WGS (3×). A TF of ≥10% was detected in six bile samples (60% sensitivity; median TF 27.32%, range 11.64–44.34%), showing that at matched coverage (3×) and with the same pipeline (ichorCNA), bile outperformed plasma for tumour detection by LP-WGS (Fisher’s exact test, two-sided p=0.05; figure 2A–B) and among the five patients with bile TF >20%, only one was positive in plasma.

Although approximately 3× LP-WGS already represents a cost reduction compared with conventional high-depth WGS (online supplemental table S1), it remains relatively costly for routine diagnostic implementation.^32,33,46 We estimated the per-sample sequencing costs for different whole-genome depths on two widely used Illumina platforms (NextSeq 2000 and NovaSeq X; online supplemental table S1). Reducing coverage from 3× LP-WGS to 0.2× ULP-WGS provided an additional approximately 13–15 fold decrease in cost, depending on platform and flow-cell configuration, resulting in an overall economic impact of approximately 130–215 fold reduction when comparing conventional 30–50× WGS with 0.2× ULP-WGS (online supplemental table S1). Beyond cost, ultra-low coverage substantially reduces workflow time and data footprint, and can be implemented on more widely available low- to-mid throughput sequencers, further supporting its potential for routine clinical implementation.

With this in mind, we next evaluated the performance of ULP-WGS (0.2×) for TF detection in bile. Of the 10 paired bile samples analysed, six were TF-positive at 3× and five were TF-positive at 0.2× (figure 2A), with the single discordant case being the sample with the lowest TF at 3× (11.64%; figure 2A–B), reflecting the expected loss in sensitivity when reducing sequencing depth. Importantly, and in line with prior reports,^26,30,46 the TF estimates obtained with ichorCNA at 0.2× in the five positive samples was nearly identical to those obtained at 3× (figure 2B). Although additional studies including larger cohorts of patients will be needed, these findings suggest that intermediate sequencing depths, such as 0.5× or 1× might recover borderline cases while retaining most of the cost advantage of ULP-WGS.

Malignant diagnostic performance of bile ULP-WGS in patients with biliary strictures

We first interrogated the performance of bile ULP-WGS in the 26 individuals with an initial malignant diagnosis, established by pathological evidence of adenocarcinoma (17 CCA and 9 PDAC). A TF of ≥10% estimated by ichorCNA was detected in 19/26 bile samples (73%), including 14/17 CCA (82%) and 5/9 PDAC (56%) (figure 3A–B). There were no significant differences in TF levels between CCA and PDAC, and among TF-positive samples, the median TF was 26.51% (range 11.03–51.57%; figure 3B). Serum CA19-9 levels at the time of bile collection were not available for two cases (figure 3A). Among the remaining 24, CA19-9 values were ≥44 U/mL in 19/24 samples, yielding a sensitivity of 79% for malignancy detection, which was comparable to that of ULP-WGS (73%, 19/26; p=0.745, Fisher’s exact test). By tumour type, CA19-9 sensitivities were 73% (11/15) in CCA and 89% (8/9) in PDAC, again not significantly different from those of ULP-WGS (82% and 56%, respectively; both p>0.05).

Figure 3. Bile ULP-WGS, serum CA19-9 and Bilemut performance at the first ERCP in individuals with biliary strictures. (A) Individuals with an initial malignant diagnosis (n=26; 17 CCA, 9 PDAC). Each column represents one individual, grouped by tumour type. Rows indicate the result of bile ULP-WGS, serum CA19-9 and Bilemut. Red squares denote a positive result, green squares denote a negative result and white squares denote missing data. The table below summarises the sensitivity with 95% CIs of each assay by tumour type and shows p values for pairwise comparisons between ULP-WGS and CA19-9 and between Bilemut and ULP-WGS (two-sided Fisher’s exact test); bold red values indicate p<0.05. †CA19-9: 1 missing value. (B) TF (%) estimated from bile ULP-WGS using ichorCNA in malignant biliary strictures, stratified by tumour type (CCA, n=17; PDAC, n=9). Boxes show means±standard deviation. (C) Individuals with an initial indeterminate diagnosis (n=69) and their final outcome after up to 12 months of follow-up (45 malignant: 23 CCA and 22 PDAC; and 24 benign). Each column represents one individual. Rows indicate the result of bile ULP-WGS, serum CA19-9 and Bilemut as in (A). The table below summarises the sensitivity of each assay in malignant strictures, and the table on the right summarises the specificity of ULP-WGS and CA19-9 in benign strictures. Sensitivity and specificity are shown with 95% CIs in brackets. P values were calculated using two-sided Fisher’s exact tests; bold red values indicate p<0.05. †CA19-9: 1 missing value among malignant cases and 5 missing values among benign cases. § Bilemut: ND. (D) Serum CA19-9 levels at the time of bile collection in individuals from (C), stratified by final diagnosis. Data are shown on a logarithmic scale (E) TF (%) estimated by ichorCNA in bile samples from individuals in (C), stratified by final diagnosis. (F) Interval between bile collection and final diagnosis of malignancy for TF-positive cases within the initially indeterminate cohort (n=23). Each dot represents one individual. CA19-9, carbohydrate antigen 19-9; CCA, cholangiocarcinoma; ERCP, endoscopic retrograde cholangiopancreatography; MAF, mutant allele fraction; ND, not determined; PDAC, pancreatic ductal adenocarcinoma; TF, tumour fraction; ULP-WGS, ultra-low-pass whole-genome sequencing.

We next evaluated serum CA19-9 and bile ULP-WGS in the 69 individuals who had an initial indeterminate diagnosis of biliary strictures at the first ERCP. CA19-9 levels were at or above 44 U/mL cut-off in 75% (33/44) of finally malignant cases (n=45), including 17 of 22 CCA (77%) and 16 of 22 PDAC (73%) (figure 3C–D). Among 44 evaluable patients who were ultimately diagnosed with malignancy, CA19-9–positive malignant cases, the median CA19-9 level was 580 U/mL (range 66–12 000 U/mL; figure 3D). Notably, and in line with previous reports,^6,7 CA19-9 levels were also elevated (≥44 U/mL) in 9 of 19 evaluable individuals whose final diagnosis was benign (n=24) (median 423 U/mL, range 61–11 939 U/mL), resulting in a specificity of only 53% (figure 3C–D). Although CA19-9 levels differ between benign and malignant strictures, substantial overlap was observed between groups (figure 3D), underscoring the limited specificity of this biomarker in the context of biliary strictures.

In this same subgroup, bile ULP-WGS classified 23 of the 24 individuals who had a final benign diagnosis as ctDNA-negative (TF <10%), resulting in a specificity of 96% in this cohort (figure 3C and figure 3E). The overall sensitivity was 51% (23/45), with 15/23 (65%) CCA and 8/22 (36%) PDAC cases testing positive (figure 3C and figure 3E). When we combined these results with CA19-9 values, specificity reached 100% and sensitivity fell to 45% (figure 3C).

Among TF-positive samples, the median TF was 24.53% (range 11.25–68.79%; figure 3E), which was similar to that observed in individuals with an initial malignant diagnosis (figure 3B). The time interval between bile collection and the final diagnosis of malignancy in these 23 TF-positive cases is shown in figure 3F, and demonstrates that bile ULP-WGS allows an earlier diagnosis of malignancy (median 38 days, range 7–499 days). Notably, in 7 of these 23 patients, ctDNA was detectable more than 100 days before the final diagnosis of malignancy was established by conventional means during the follow-up, highlighting the potential of this approach for early tumour detection.

Malignant diagnostic performance of bile ULP-WGS and Bilemut

We then benchmarked bile ULP-WGS against Bilemut, the targeted NGS assay we previously described.¹⁶ In the cohort with an initial pathological diagnosis of malignancy (n=26), Bilemut demonstrated a sensitivity of 100%, which was significantly higher than that of ULP-WGS (19/26, 73%; Fisher’s exact test, two-sided p=0.010; figure 3A).

In the clinically most challenging subgroup of initially indeterminate strictures that were ultimately malignant, Bilemut detected ctDNA in 90% of cases, clearly outperforming ULP-WGS (51%, 23/45), and was significantly superior in strictures due to PDAC (14/17, 82% vs 8/22, 36%; p=0.013) and in CCA cases (14/14, 100% vs 15/23, 65%; p=0.041; figure 3C). As observed for ULP-WGS, Bilemut was negative in all evaluable strictures that were ultimately benign, thus maintaining 100% specificity in this group (figure 3C).

Globally in this cohort, 54/57 (94.7%) biliary strictures that were eventually confirmed as malignant were Bilemut-positive. In line with our previous report,¹⁶ the mutational landscape was dominated by KRAS, GNAS and TP53, with additional, less frequent alterations in BRAF, ERBB2, ERBB3 and FGFR2 (online supplemental figure S1). TMB was estimated as the number of genes mutated within the Bilemut panel. Patients harboured a median of 2 mutated genes (range 0–8; online supplemental figure S1), and 19 patients carried mutations in three or more genes, hereafter considered as high TMB.

Genomic features captured by bile ULP-WGS and Bilemut

We next evaluated whether the proportion of ctDNA detected in bile was related to the number or genomic extent of tumour alterations. Strikingly, the TF did not correlate with the percentage of the genome affected by CNAs (figure 4A–B and online supplemental figure S2A). In fact, patients with the highest TF (>50%) had the lowest percentage of genome altered; conversely, patients with TF of 10% had more than 50% of genome altered (figure 4A–B). As expected,³⁰ the MAF measured by Bilemut positively correlated with the TF estimated by ULP-WGS (online supplemental figure S2B), supporting the interpretation that both metrics are largely driven by the amount of ctDNA present in bile cfDNA. In contrast, TF did not correlate with the number of mutated genes per sample (TMB estimated by Bilemut; online supplemental figure S2C). Together, these findings indicate that TF detected in bile mainly reflects ctDNA abundance, whereas mutational and copy-number complexity represent independent dimensions of tumour biology.

Figure 4. CNAs in patients with bile positive TF. (A) Tumour type (CCA or PDAC; shown in greyscale), TF (%) estimated by ichorCNA and percentage of the autosomal genome affected by CNAs (losses in blue, gains in orange and no change in white) in the 42 patients with detectable bile TF. Each column represents one patient, ordered from highest to lowest ichorCNA-estimated TF. (B) The heatmap shows the distribution of large CNAs across chromosome arms (rows, from 1p to 22q and X) in patients from (A). Blue indicates loss, orange indicates gain, purple indicates both loss and gain within the same arm, and grey indicates no detected alteration. For each arm, the percentage of patients altered is shown on the side. (C) The histogram shows, for each arm, the proportion of patients with losses (blue), gains (orange) or both (purple). The accompanying pie chart summarises the overall proportion of altered arms accounted for by losses, gains and combined losses and gains across the cohort. CCA, cholangiocarcinoma; CNA, copy-number alteration; PDAC, pancreatic ductal adenocarcinoma; TF, tumour fraction; ULP-WGS, ultra-low-pass whole-genome sequencing.

We next characterised the landscape of CNAs using the 42 TF-positive bile samples. ichorCNA aggregates ULP-WGS reads into non-overlapping 1 Mb bins and compares these profiles with a reference panel of healthy controls to infer broad CNAs. This analysis revealed highly heterogeneous patterns of large-scale CNAs among the patients (figure 4A-bottom and figure 4B), with a median of 746.5 Mb of the genome altered (range 124–1813 Mb), corresponding to a median of 30.2% of the autosomal genome (range 5.0–73.3%). The distribution of these altered Mb across chromosomes in each patient is depicted in the heatmap in figure 4B. As expected, samples with a higher number of altered Mb also displayed a greater number of affected chromosome arms (figure 4A-bottom and figure 4B).

Losses involved a larger fraction (62%) of CNAs than gains (46%) (figure 4C). The most recurrent losses involved 18q (detected in 64% of patients), 17p (57%), 9p (48%) and 22q (45%), whereas gains most commonly affected 19q (62%), 7p (50%), 7q (48%), 19p (45%) and 8q (43%) (figure 4B–C). This pattern of CNAs is consistent with tissue-based profiles reported for pancreatobiliary malignancies, which commonly show losses of 9p, 17p and 18q, encompassing canonical tumour suppressors such as CDKN2A (9p), TP53 (17p), SMAD4 (18q) and gains of 7p, 7q and 8q, with the 8q locus harbouring MYC.^48–50

Positive TF together with high TMB identifies patients with worse prognosis

Given previous evidence that plasma CNAs, TF and TMB correlate with clinical features and overall survival in several tumour types,^25,35 including CCA²⁸ and PDAC,³⁰ we assessed whether bile-derived genomic markers were associated with clinical and laboratory characteristics in the 57 malignant cases with both ULP-WGS and Bilemut data available (online supplemental table S2). Among them, 34 (60%) were TF-positive, predominantly CCA (23/34, 68%) rather than PDAC (11/34, 32%). CA19-9 levels tended to be higher in patients who were TF-positive (p=0.084), whereas age, sex, carcinoembryonic antigen (CEA), stricture location, first management and Eastern Cooperative Oncology Group status did not differ significantly (all p>0.05). Median follow-up was 10.6 months (range, 2.5–83.5). Median overall survival was shorter in patients who are TF-positive than TF-negative, although the difference was not significant (7.2 vs 13.6 months; log-rank p=0.737; online supplemental figure S3A).

We next evaluated CNA burden and TMB. Patients with ≥20% of the genome affected by CNAs, all of whom were TF-positive, had similar overall survival to the remaining cases (9.8 vs 10.6 months; log-rank p=0.475; online supplemental figure S3B), indicating that neither TF nor CNA burden alone was significantly associated with OS. Likewise, high TMB (≥3 mutated genes) was associated with shorter, but non-significantly different, median overall survival (5.5 vs 13.6 months; log-rank p=0.354; online supplemental figure S3C), consistent with reports linking TMB more strongly to immune checkpoint inhibitor benefit than to prognosis.^51,52

Because TF and TMB capture complementary facets of tumour biology, we evaluated a combined criterion. We defined a dual-positive group as TF-positive with high TMB (≥3 mutated genes); 15 of the 57 patients (26%) met this definition. Although the differences did not reach statistical significance (p=0.086), dual-positive patients tended to have CCA more frequently than PDAC and to present with more advanced-stage disease compared with the remaining patients (online supplemental table S3). Similarly, no significant differences were observed in CA19-9, CEA or stricture location between groups (all p>0.05; online supplemental table S3). Unlike the single-variable analyses, the combination of TF positivity and high TMB (≥3 mutated genes) was associated with significantly worse survival compared with patients who were TF-negative and/or had TMB <3 (5.1 vs 13.7 months; log-rank p=0.012; figure 5). Moreover, multivariable Cox proportional hazards regression, adjusted for first management (surgery vs systemic treatment+best supportive care), CA19-9 (≥500 U/mL) and CEA (≥ 5 ng/mL), demonstrated that the combination of TF positivity and high TMB (≥3 mutated genes) was independently associated with poorer overall survival (adjusted hazard ratio (aHR), 2.51; 95% CI 1.22 to 5.14; figure 5). HRs for first management, CA19-9 and CEA were 3.31 (95% CI 1.75 to 6.25), 0.61 (95% CI 0.30 to 1.22) and 1.23 (95% CI 0.65 to 2.35), respectively.

Figure 5. Prognostic impact of bile-derived TF together with mutational burden. Kaplan-Meier curves comparing overall survival between the dual-positive group (TF-positive and TMB ≥3 mutated genes) versus all other malignant cases. Shaded areas represent 95% CIs and the aHR from the multivariable Cox model is shown; tick marks indicate censored observations; p values are from log-rank tests. aHR, adjusted hazard ratio; TF, tumour fraction; TMB, tumour mutational burden.

Bile CNAs by ULP-WGS help to discriminate the tissue of origin of the malignancy

Clinical and routine biochemical variables failed to discriminate between CCA and PDAC beyond the expected difference in stricture location (p<0.001; online supplemental table S4). Moreover, in our cohort, the first management of patients showed significant differences between CCA and PDAC (p=0.025; online supplemental table S4). The proportion of patients receiving best supportive care was significantly higher in CCA than in PDAC (58% CCA vs 26% PDAC; p=0.009).

We asked whether CNA patterns identified by ULP-WGS could aid in the differential diagnosis between CCA- and PDAC-driven biliary strictures. We focused on the 42 patients who were TF-positive, including 29 with CCA and 13 with PDAC. TF was comparable between groups (medians 26.8% vs 26.2%; online supplemental figure S4A), as were the overall CNA burden (balance of losses vs gains; online supplemental figure S4B–C) and the number of altered chromosome arms (medians 17 vs 21.5; online supplemental figure S4D).

To maximise between-group separation, we applied a supervised DAPC.⁴² The classifier or algorithm generated using synthetic data as described in the Methods section, and selected 50 genomic regions that best separated CCA from PDAC. These regions mapped predominantly to chr12q (16 regions), chr9p (16 regions), and chr1p (8 regions), with additional regions on chr5q (6 regions), chr18q (2 regions) and one region each on chr19q and chr8p (online supplemental figure S4E and online supplemental table S5). We tested the classifier in two cohorts of patients who are TF-positive: the reference cohort of 22 patients (15 CCA and 7 PDAC), yielding an overall accuracy of 0.864 (95% CI 0.651 to 0.971; online supplemental table S6) and a validation cohort of 20 patients (14 CCA and 6 PDAC) with an accuracy of 0.850 (95% CI 0.621 to 0.968; online supplemental table S6). The consistency of accuracy across both cohorts (0.864 vs 0.850) supports the robustness of the classifier despite the limited sample size. However, these findings should be considered exploratory, and prospective validation in larger independent cohorts is required.

Because the 50 regions were selected in a supervised manner, we next asked whether any of the seven chromosomal arms harbouring these regions (12q, 9p, 1p, 5q, 18q, 19q, 8p; online supplemental figure S4E and online supplemental table S5) showed differential alteration frequencies between CCA and PDAC. We observed enrichment of 1p losses in PDAC (p=0.046), consistent with prior PDAC copy-number studies,⁵³ and significantly higher frequencies of 9p losses in CCA (p=0.007; online supplemental table S7). These arm-level differences were modest compared with the performance of the 50-region signature, suggesting that discrimination arises from the integrated CNA pattern rather than a single dominant locus.

Discussion

The usefulness of bile as a liquid biopsy matrix is still not widely appreciated.^15,54,55 In patients with biliary strictures, ERCP is routinely performed and bile is usually discarded, despite being obtained directly from the vicinity of the lesion. Our group has actively promoted systematic bile collection and analysis^15,16,22,38 and is leading the Bilebank project (https://bilebank.org/). In this context, we previously demonstrated that mutational analysis of bile cfDNA using a commercially available targeted NGS panel (Bilemut assay) is substantially more sensitive for uncovering the malignant origin of biliary strictures than the same analysis performed on plasma and than currently used technique for the differential diagnosis of malignant versus benign strictures.¹⁶ Moreover, Bilemut has shown potential utility for the management of individuals at high risk of CCA development, such as those with PSC.²²

However, beyond the limited use of bile itself, widespread implementation of NGS-based strategies in routine clinical practice is often constrained by cost. Based on current reagent and sequencing prices on widely used Illumina platforms, reducing coverage from conventional 30–50× WGS to 0.2× lowers sequencing costs by roughly 130–215-fold (online supplemental table 1), in line with our previous analyses.³² In addition to cost, ultra-low coverage substantially reduces workflow time and data footprint, and can be implemented on more widely available low- to mid- throughput sequencers, facilitating accessibility and downstream data handling and storage. Provided that ULP-WGS preserves clinically relevant information, its utility as a feasible and potentially cost-effective complementary genomic approach deserves further evaluation in prospective studies.

In this study, we evaluated 0.2× ULP-WGS of bile cfDNA as a potentially cost-efficient, low-input genomic approach to support the management of patients with biliary strictures. This study applied ULP-WGS (≤0.5×) to bile cfDNA for the diagnostic evaluation of pancreatobiliary tumours. Prior similar liquid-biopsy studies in these tumours used higher-coverage depths (approximately 3–5×).^28–31 In this context, the study by Wang et al reported high sensitivity (approximately 89%) of 3× LP-WGS using a proprietary pipeline in plasma samples from patients with biliary tract cancers.²⁸ Because plasma-based approaches could, in principle, obviate the need for bile—which, although easily available during ERCP, is inherently less generalisable than blood—we first performed a head-to-head comparison of bile versus plasma, at 3× LP-WGS using ichorCNA as a transparent, open-source, widely used workflow.²⁶ In our cohort, a TF of ≥10% was detected only in one plasma sample (10%) but in six paired bile samples (60%), supporting, as previously shown,^16,17 the matrix advantage of bile in pancreatobiliary disease and contrasting with the performance reported by Wang et al.²⁸ Several factors likely contribute to these discrepancies. First, we used the open-source ichorCNA pipeline, which facilitates reproducibility and standardisation across centres, whereas their analysis relied on a proprietary workflow that cannot be independently benchmarked. Second, the composition of the cohorts differed, as ours included both CCA and PDAC, whereas their series was enriched for CCA and gallbladder carcinoma. These differences limit direct comparison and underscore the need for harmonised, head-to-head evaluations of plasma and bile using shared analytical frameworks. Although our main goal was to evaluate a potentially cost-effective methodology based on 0.2× ULP-WGS, we compared its performance against 3× LP-WGS in a reduced number of positive tumour samples. The single discordant case, negative at 0.2× and positive at 3×, had the lowest TF at 3× (11.64%), consistent with the expected relationship between sequencing depth and sensitivity. These findings suggest that ULP-WGS may preserve informative signal in samples with sufficient tumour-derived DNA while substantially reducing sequencing demand and cost. Increasing sequencing depth would improve sensitivity, and intermediate depths representing different cost-effectiveness trade-offs could be explored in future studies.

In terms of malignant diagnostic performance, and as expected, Bilemut yielded the highest sensitivity. Bile ULP-WGS showed similar sensitivity to CA19-9 in patients with an initial malignant diagnosis (73% vs 79%), but lower sensitivity in initially indeterminate strictures later proven malignant (51% vs 75%; p=0.028). However, this apparent higher sensitivity of CA19-9 occurred at the expense of specificity: CA19-9 ≥44 U/mL was observed in 9 of 19 evaluable benign cases, yielding a specificity of only 53%, by contrast, ULP-WGS showed a specificity of 96% (p=0.008 CA19-9 vs ULP-WGS). Moreover, when positive, bile ULP-WGS detected tumour-derived DNA a median of 38 days before the final diagnosis of malignancy. Notably, in seven individuals ctDNA was detectable in bile more than 100 days before malignancy was routinely diagnosed (113, 126, 169, 186, 203, 370 and 499 days). These findings warrant further prospective studies in larger independent cohorts to confirm the potential of bile ULP-WGS for early tumour detection. However, our results suggest that in the current clinical setting, an elevated CA19-9 value accompanied by a positive ULP-WGS result would help to identify malignancy with higher specificity.

We found that TF correlates with MAF but not with the number of mutated genes or with the fraction of the genome affected by CNAs which is consistent with the notion that TF primarily reflects ctDNA shedding rather than the intrinsic genomic complexity of the tumour. Several studies have shown that ctDNA levels, measured either by TF or by the MAF of recurrent mutations, track tumour burden and disease dynamics and are influenced by factors such as tumour size, site, vascularisation, proliferation and cell death, rather than by the absolute number of somatic alterations carried by tumour cells⁵⁶ and that some entities are bona fide low-shedding tumours.^57,58 In line with this, blood TMB has been reported to depend critically on ctDNA abundance, underestimating tissue TMB in low-shedding tumours and losing prognostic value in that context.⁵⁹ Our data extend these observations to bile cfDNA, suggesting that TF is a robust proxy for the amount of tumour-derived DNA available for analysis, whereas mutational and copy-number complexity constitute largely layers of information that may capture complementary aspects of tumour biology. Importantly, what any given liquid-biopsy assay detects will depend both on the biomarker interrogated (mutations, CNAs, methylation, etc) and on the amount of ctDNA present, with low-shedding tumours often yielding non-informative or apparently negative profiles despite substantial genomic complexity in the tumour tissue.

Plasma studies using LP-WGS have shown that cfDNA CNAs closely track tissue aneuploidy and tumour burden in biliary and pancreatic cancers.^28,30,31 In line with this, bile ULP-WGS in our cohort reproduced the canonical CNA patterns in pancreatobiliary malignancies—most notably losses of 9p, 17p and 18q and gains of 7p, 7q, 8q and 19q^48–50—supporting the biological validity of bile-derived CNAs rather than matrix artefacts.

Recent meta-analyses across solid tumours support the prognostic potential of ctDNA profiling by LP-WGS or ULP-WGS in plasma, including studies with sequencing depths of ≤0.5×, with higher TF consistently associated with poorer overall survival in hepatocellular carcinoma, lung, prostate and osteosarcoma among others.³² For mutational burden, however, the evidence is mixed. In biliary tract cancer, some studies linked higher TMB to worse overall survival,^60,61 whereas others found no clear prognostic value.⁶² In PDAC, TMB is typically ultra-low and its clinical relevance remains uncertain.⁶³ Real-world data suggest that high TMB is primarily associated with improved outcomes in patients receiving immune-checkpoint inhibitors, but shows inconsistent associations otherwise, supporting a mainly predictive—rather than prognostic—role.⁵² In our cohort, although higher TF and TMB values tended to be associated with shorter survival, they did not significantly stratify overall survival, likely due to limited sample size and clinical heterogeneity. However, combining TF positivity with high TMB (≥3 mutated genes) identified a subgroup with clearly worse outcomes, in line with studies showing that integrating TMB with copy-number or aneuploidy metrics can refine prognostic stratification.⁶⁴ It should be emphasised that these prognostic analyses are exploratory in nature and that given the small sample size, these results should be interpreted with caution and not used to inform clinical decisions until prospectively validated.

Because distinguishing PDAC from CCA can be clinically challenging, and additional tumour biology might aid classification, we next asked whether CNAs identified by ULP-WGS could help separate the two entities. Leveraging the CNA signal, we applied DAPC that distilled genome-wide variability into a 50-region signature—enriched on 12q, 9p and 1p, with additional bins on 5q, 18q, 19q and 8p—capable of separating CCA from PDAC with high accuracy in the reference cohort and good performance in independent validation cohort. Despite the modest sample sizes (n=22 reference cohort; n=20 independent validation cohort), these results are encouraging for a supervised approach and suggest that bile-derived CNA patterns may capture tissue-of-origin information that merits validation in larger, independent cohorts.

We acknowledge limitations of both the ULP-WGS approach and our study. First, this is a single-centre analysis of modest size, with relatively few TF-positive cases and a limited benign cohort, which broadens the CIs around specificity estimates. Second, the DAPC classifier for CCA–PDAC discrimination was trained and tested in small datasets and will require external, multicentre validation with harmonised analytical thresholds. Third, we used 2 ng of cfDNA input and 0.2× coverage; at this depth, ULP-WGS retained most TF calls observed at 3×, but we did not explore intermediate coverages (eg, 0.5×) or higher inputs, which could further improve sensitivity and merit prospective evaluation in larger cohorts.

Conclusions

Our exploratory study shows that bile ULP-WGS at 0.2× could represent a scalable, reproducible and potentially low-cost complementary approach to support the clinical management of patients with biliary strictures. By capturing TF and large-scale CNAs at 1 Mb resolution, it may improve the diagnostic assessment of indeterminate cases and provide signals that may help distinguish tumour origin (CCA vs PDAC). Moreover, when combined with Bilemut, it may also add prognostic information, although this will require confirmation in larger, independent cohorts.

Data availability statement

Data are available upon reasonable request.