AI-based predictive tool for detection of ctDNA in pancreatic adenocarcinoma using nationwide comprehensive genomic profiling data.

Abstract

4163 Background: Comprehensive genomic profiling (CGP) has become a cornerstone of precision oncology, with liquid biopsy expanding its applicability. However, in some cases, circulating tumor DNA (ctDNA) is undetectable in liquid CGP, limiting the ability to assess genetic alterations. Identifying the optimal timing for liquid CGP remains a challenge. This study focuses on pancreatic adenocarcinoma, using nationwide CGP data and explainable AI methods to identify clinical factors associated with ctDNA detection. Based on these factors, we develop an easy-to-use AI tool to predict the probability of ctDNA detection in real-world settings. Methods: We conducted a retrospective analysis of nationwide CGP data collected from Jun 2019 to Dec 2023, covering 99.7% of CGP performed in Japan. Cohort 1 included 4,110 pancreatic adenocarcinoma cases analyzed by FoundationOne CDx, while Cohort 2 comprised 2,220 cases analyzed by FoundationOne Liquid CDx (F1L). Using clinical information available prior to CGP, we developed an eXtreme Gradient Boosting (XGBoost)-based predictive model to estimate ctDNA detection and employed SHapley Additive exPlanations (SHAP) analysis to elucidate contributing clinical factors. A smartphone application was deployed using the refined model. The app's performance was tested with Cohort 3, consisting of 629 pancreatic adenocarcinoma cases tested by F1L between Jan 2024 and Dec 2024. Results: In Cohort 1 (tissue), 98.5% of cases harbored mutations in either KRAS, TP53, CDKN2A, or SMAD4, confirming their role as surrogate markers for tumor-derived DNA detection. The predictive AI model for ctDNA detection, trained on Cohort 2 (liquid) data, achieved an AUROC of 0.754. SHAP analysis identified key predictors, including liver metastasis, the number of metastatic organs, performance status, response to recent therapy, interval from diagnosis to blood collection, and treatment line. Notably, patients with liver metastases exhibited a significantly higher rate of ctDNA detection (p < 0.001) compared to those without, whereas patients with peritoneal metastases demonstrated a lower rate of ctDNA detection (p < 0.01). A refined model incorporating representative predictors was deployed as a smartphone application. When tested on Cohort 3 (liquid), the application demonstrated predictive accuracy with an AUROC of 0.769 (sensitivity: 0.707, specificity: 0.769) and a Brier score of 0.194. Conclusions: This study identified clinical factors predictive of ctDNA detection in liquid CGP for pancreatic adenocarcinoma using explainable AI methods and nationwide CGP data. Based on these findings, a smartphone application was developed to predict the probability of ctDNA detection. By facilitating optimal timing of liquid CGP, this app has the potential to enhance patient access to effective therapies, contributing to improved clinical outcomes.