Abstract
BACKGROUND: Circulating tumor DNA (ctDNA) is sometimes undetectable in liquid comprehensive genomic profiling (CGP) of advanced-stage pancreatic cancer, resulting in false-negative findings that may mislead treatment decisions and waste health care resources. Predicting ctDNA detectability before testing may help optimize the timing of liquid CGP and guide decisions to consider alternative approaches, such as tissue-based CGP. PATIENTS AND METHODS: We analyzed data from 7498 patients with advanced pancreatic adenocarcinoma using a nationwide real-world clinical genomic database. The tissue CGP cohort (cohort 1) included 4110 patients tested with FoundationOne CDx between June 2019 and December 2023. A prediction model for ctDNA detectability based on routinely available clinical variables before blood collection was trained on 2220 patients who underwent FoundationOne Liquid CDx between August 2021 and December 2023 (cohort 2). The model was deployed as a web application (https://pancreasliquidcgp.streamlit.app) and tested on two independent cohorts: 629 patients (cohort 3; FoundationOne Liquid CDx, January to December 2024) and 539 patients (cohort 4; Guardant360, July 2023 to December 2024). Model performance was assessed using Brier scores and calibration plots. Feature importance was evaluated using SHapley Additive exPlanations (SHAP). RESULTS: Among the 2220 patients in cohort 2, ctDNA was detected in 1130 (50.9%). The model achieved a Brier score of 0.210 and showed good calibration. SHAP analysis identified liver and bone metastases, progressive disease, and poor performance status as positive predictors of ctDNA detectability. Conversely, peritoneal and lung metastases negatively contributed to prediction. In the independent test cohorts, the model maintained robust performance with Brier scores of 0.203 (cohort 3) and 0.204 (cohort 4). CONCLUSION: This prediction model enables accurate pre-test estimation of ctDNA detectability in advanced pancreatic cancer and may enhance the clinical utility of liquid CGP by informing optimal test selection and timing.