Comprehensive ctDNA Measurements Improve Prediction of Clinical Outcomes and Enable Dynamic Tracking of Disease Progression in Advanced Pancreatic Cancer

Pancreatic cancer is a leading cause of cancer-related mortality, and most patients present with advanced disease with median overall survival under one year. CA19-9, the current blood biomarker for monitoring, suffers from poor sensitivity, false positives with obstructive jaundice, and false negatives in Lewis-negative patients, necessitating reliance on imaging for progression assessment. ctDNA has emerged as a tool for prognosis and response monitoring, but prior pancreatic cancer studies often focused only on KRAS mutations, limiting sensitivity, and some tumors lack KRAS mutations. This study combined highly sensitive detection of mutations in eight frequently altered genes (KRAS, TP53, SMAD4, CDKN2A, ARID1A, TGFBR2, RNF43, GNAS) using a hybridization capture NGS method (HYTEC-seq) with genome-wide copy-number analysis to improve ctDNA detection. The objectives were to assess ctDNA as a prognostic marker and as a dynamic marker for treatment monitoring in advanced pancreatic cancer, and to compare performance with CA19-9 and radiological imaging.

Prior work established ctDNA as prognostic and useful for monitoring across multiple cancers. In pancreatic cancer, studies have been limited by small cohorts and often focused solely on KRAS mutations, reducing sensitivity. Some reports showed ctDNA can precede radiologic progression with lead times ranging from several weeks to months, though findings compared with CA19-9 are mixed. Additional methods such as cfDNA fragment analysis and methylation profiling can improve detection over mutation-only approaches. RECIST-based imaging has known limitations, including interobserver variability and challenges in fibrotic pancreatic tumors, potentially contributing to discordance with liquid biopsy findings.

Design: Prospective cohort of 56 patients with locally advanced (n=8) or metastatic (n=48) pancreatic cancer treated at Stavanger University Hospital (Sept 2012–Oct 2020). First-line regimens: gemcitabine (n=7), gemcitabine+nab-paclitaxel (n=32), or FOLFIRINOX (n=17). Blood sampling: baseline before chemotherapy (n=56) and monthly during treatment (n=268). Healthy controls: 60; 20 used to build CNA reference. CA19-9 measured routinely. Radiologic response by RECIST 1.1. Sample processing: Plasma separated by density centrifugation; cfDNA isolated from approximately 1–2 to 4 mL plasma using QIAamp Circulating Nucleic Acid kit. cfDNA concentration and mononucleosomal/dinucleosomal fractions measured by Bioanalyzer. Sequencing: HYTEC-seq libraries prepared with UMI-tagged adapters (Kapa HyperPrep), target capture (Agilent SureSelect) covering eight genes (KRAS, TP53, SMAD4, CDKN2A, ARID1A, TGFBR2, RNF43, GNAS), sequenced deeply on Ion Proton (median depth 66,457; median 2,402x per ng input). UMI families constructed (median unique templates 3,584; 22.3% recovery). Bioinformatics: Ion Torrent Suite for initial processing; TagXtractor for adapter removal/UMI extraction; BWA-MEM realignment; SSCScreator to build single-strand consensus sequences; variant calling with PlasmaMutationDetector2 using background error profile from healthy controls; filters excluded known SNPs (ExAC), short indels (<3bp unless called by VarScan2 with somatic P<0.001), strand bias, and required bidirectional SSCS support. Visual IGV review to exclude artifacts. In follow-up, previously detected variants could be tracked with relaxed thresholds if present in both strands above control levels. Annotation via ANNOVAR. CHIP assessment with ddPCR on leukocyte DNA; mutation validation by ddPCR where applicable. CNA analysis: CNVkit on both on- and off-target reads from HYTEC-seq; off-target bins 500 kb; problematic regions and acrocentric short arms excluded. Reference from 20 healthy controls with normalization and bias correction; segmentation by CBS with drop-low-coverage. Samples with >5% missing bins or failed normalization excluded. Global aneuploidy z-score computed by summing squared arm-level z-scores; z-score ≥4 (~5% tumor fraction) defined CNA-positive. Statistics: SPSS v26; two-sided tests with P<0.05. Group comparisons by chi-square/Fisher’s exact and t-test/Mann–Whitney; Wilcoxon for paired samples. Survival by Kaplan–Meier/log-rank and Cox regression. PFS from inclusion to RECIST progression or death; OS from inclusion to death. For analyses at 1–2 months post-treatment start, immortal time excluded. Multivariable Cox models evaluated ctDNA mutation status or CNA status separately (due to correlation) along with CA19-9, tumor location, stage, ECOG, and treatment; stepwise selection used.

ctDNA detection at baseline and during therapy: - After validation, somatic point mutations were detected in 34/56 (61%; median 2 mutations) baseline samples and in 90/268 (33.6%) on-treatment samples. Median baseline VAF 7.7%. Most frequent genes: KRAS (n=33), TP53 (n=29), CDKN2A (n=6). KRAS variants included G12D (n=18), G12V (n=10), G12R (n=3), G12C and Q61K (two patients), and a rare Q22K. - CNA-positive by global z-score ≥4: 19/55 (34.5%) at baseline (1 failed QC), including one sample without point mutations; 30/243 (12.3%) during treatment (25 failed QC). Frequent CNA patterns: gains in chr12p (KRAS), chr8q (MYC), chr20q (GNAS); loss in chr3p (TGFBR2). Nearly all chr12p gains co-occurred with KRAS mutations. Prognostic value: - PFS: mutation-negative vs mutation-positive median 6.9 months (95% CI, 2.8–11) vs 2.6 months (95% CI, 1.4–3.7); P=2.4E-04. CNA-negative vs CNA-positive also significantly different (data consistent with strong separation; OS shown below). - OS: mutation-negative vs mutation-positive median 8.8 months (95% CI, 5.9–11.7) vs 4.7 months (95% CI, 2.9–6.6); P=0.001. CNA-negative vs CNA-positive median 8.1 months (95% CI, 7.3–9) vs 2.8 months (95% CI, 1.6–4); P=5.2E-07. - Multivariable Cox regression: ctDNA mutation status (Model A) independently predicted PFS (HR 2.847; P=0.003) and OS (HR 2.600; P=0.004). CNA status (Model B) independently predicted PFS (HR 3.539; P=3E-04) and OS (HR 3.862; P=6.6E-05). ECOG and first-line treatment were also independent prognostic factors. Dynamics under therapy: - Significant VAF decrease at 1 month post-chemotherapy (P=2E-04), further decrease by 2 months (P=0.060). VAF increased at progression compared with 2-month samples (P=0.003); no significant increase vs baseline (P=0.434). - ctDNA persistence (<10-fold VAF reduction): at 1 month, no PFS difference (P=0.822); at 2 months, markedly shorter median PFS with persistence (0 months; 95% CI, 0–0) vs without (3.4 months; 95% CI, 1.7–5.1); P=0.002. OS worse with persistence at 1 month (3.3 vs 6.9 months; P=0.016) and at 2 months (1 vs 4.2 months; P=0.025). Monitoring and lead time: - In 27 monitored patients (≥2 follow-up draws and ≥1 post-treatment imaging), 25/27 (93%) had radiologically confirmed progression (median time to progression 126 days); 2 died without radiologic progression at 152 days. - >25% ctDNA increase detected progression/death in 17/27 (63%) including both patients who died without confirmed progression; incorporating ctDNA persistence increased detection to 19/27 (70%). - CA19-9 (>50% increase) detected progression/death in 17/27 (63%). - Median lead time over radiology/death: ctDNA increase 22 days (P=0.002); ctDNA increase or persistence 19 days (P=0.002); CA19-9 6 days (P=0.007). Lead time differences vs ctDNA were not statistically significant (P=0.161 and P=0.123). Baseline associations: - ctDNA-positive vs ctDNA-negative differed by sex (P=0.025), median cfDNA level (P=1.3E-04), cfDNA mode fragment size (P=3.3E-04), primary tumor location (P=0.011), and M-stage (P=0.042).

This study confirms that ctDNA is a strong prognostic biomarker in advanced pancreatic cancer and that dynamic ctDNA changes reflect treatment effects and progression. Early decreases in ctDNA after chemotherapy initiation indicate response, while ctDNA persistence suggests lack of benefit and is associated with inferior PFS and OS. Longitudinal ctDNA monitoring detected progression with meaningful lead time relative to imaging and comparable or better timeliness than CA19-9 in many cases. Instances of discordance with imaging may reflect ctDNA’s biological sensitivity or limitations and subjectivity in RECIST-based assessments, particularly challenging in fibrotic pancreatic tumors. Although CNA analysis was less sensitive (reflecting higher tumor fractions), it provided corroborative genomic context (e.g., chr12p gains with KRAS mutations). The generally low ctDNA shedding in pancreatic cancer constrains sensitivity; increasing plasma input and leveraging complementary assays such as fragmentomics and methylation may enhance detection. Clinically, serial ctDNA could help identify early non-responders, support earlier therapeutic adjustments, and augment monitoring where CA19-9 is unreliable.

ctDNA measured by a comprehensive mutation panel and CNA analysis provides independent prognostic information in advanced pancreatic cancer. ctDNA levels typically fall after chemotherapy and rise at progression; persistence of ctDNA after treatment initiation indicates treatment failure and poorer outcomes. Longitudinal monitoring can detect progression with a median lead time of about 19–22 days over radiology/time of death and shows high specificity for true progression. Prospective, interventional trials are needed to determine if ctDNA-guided management improves outcomes. Future work should increase assay sensitivity (e.g., larger plasma volumes, tumor-informed panels, fragmentomics, methylation) and validate integration with standard care.

Key limitations include lack of tumor tissue, restricting detection to panel-covered genes and precluding tumor-informed assays; limited and irregular serial sampling in some patients due to clinical condition; modest sample size for dynamic and monitoring analyses; low plasma input (<2 mL equivalent) limiting sensitivity to rare variants; and low sensitivity of CNA analysis (~≥5% tumor fraction), which added few ctDNA-positive cases beyond mutation detection.