Abstract 3852: Ultrasensitive AI-driven framework for MRD detection based on multidimensional cfDNA features.

Abstract Minimal residual disease (MRD) detection is essential for postoperative risk stratification and recurrence prediction in cancer, yet current fixed-panel assays exhibit limited sensitivity and specificity, particularly in tumor-naïve settings. We developed Shielding Ultra, a pan-cancer MRD assay targeting hotspot mutations across 2,365 cancer-related genes. Leveraging ultra-deep unique molecular identifier-based sequencing and AI-driven bioinformatics, the assay integrates somatic mutations, copy number variations (CNVs), and fragmentomic (Frag) features to enable multidimensional MRD assessment within a unified workflow. Analytical validation established a detection limit of 0.0048% and demonstrated 94% sensitivity in late-stage preoperative plasma, with approximately 99% specificity in healthy controls. Tumor-naïve analysis achieved 98.9% concordance with tumor-informed workflows following multimodal integration, supporting applicability when tumor tissue is unavailable. Clinical validation across colorectal, cholangiocarcinoma, and lung cancer cohorts confirmed strong prognostic performance at early postsurgical timepoints while maintaining high specificity. In colorectal cancer, MRD positivity showed a robust association with relapse, yielding hazard ratios up to 32.47 and achieving longitudinal sensitivity of up to 90.9% during postoperative surveillance. In lung cancer, recurrence detection sensitivity was 51.5% at one week after surgery and increased to 72.7% with the addition of one-month postsurgical sampling. Across cancer types, multimodal cfDNA integration strengthened detection performance and enabled reliable identification of MRD across diverse clinical contexts. These findings demonstrate that Shielding Ultra enables sensitive and specific MRD detection through the integration of multidimensional cfDNA features and AI-based algorithms. Its strong prognostic performance across multiple malignancies and compatibility with tumor-naïve workflows support its utility for postoperative risk stratification and personalized disease management. Citation Format: Yunfei Shi, Hao Zhang, Zexiao Lin, Ningyou Li, Maolong Wang, Hua Bao, Jinfeng Zhang, Zhili Chang, Yong Ge, Peng Li, Pan Wang, Liang Huang, Xiangming Liu, Lu Han, Wangming Ji, Teng Sun, Dujun Hua, Xunbiao Liu, Mingya Wang, Baihan Zhu, Dongqin Zhu, Xue Wu, Haimeng Tang, Hao Zhang, Yang Shao. Ultrasensitive AI-driven framework for MRD detection based on multidimensional cfDNA features abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 3852.