Abstract 5471: Genie-ADLA: A deep learning algorithm for methylation-based multiple cancer early detection (MCED).

作者：Kezhong Chen, Ziyu Li, Xiaojian Wu, Jian Huang, Guoyue Lv, Weiping Wen, Dahong Zhang, Xinyu Zhao, Danbo Wang, Zhihua Liu, Lixin Sun, Shuna Wang, Xiangnan Li, Jun Li, Jiandong Tai, Jiayin Yang, Zhentong Wei, Ming Cai, Qiang Zhang, S. He, Shuhua Yi, Shenhong Qu, Wenhui Zhao, Xiaoyan Yu, R X Guo, Jianhong Lian, Desong Yang, Huaiwu Lu, Xi Guo, Yan Zhang, Zhuowei Liu, Yingjiang Ye, C Koo Seen Lin, Jie Gao, Xuanhui Liu, Yushu Guo, Suying Ding, Guoqiang Zhao, Yanzhan Yang, Jing Li, Shiqing Chen, Hui Yu, Fang Liu, Yang Wang, Min Li, Baoliang Zhu, Yonghui Li, X. Wu, Fan Yang, Jun Wang · 发表于：Cancer Research · 年份：2026 · DOI：10.1158/1538-7445.am2026-5471 · 研究领域：Machine Learning in Bioinformatics、Machine Learning in Healthcare、Epigenetics and DNA Methylation

Abstract Background: Methylation-based analysis of cell-free DNA (cfDNA) has emerged as a key technology for MCED. However, existing approaches rely on traditional machine learning algorithms, which inherently limit detection performance. With the rapid advancement of artificial intelligence (AI), we have developed Genie-ADLA, a deep learning algorithm designed specifically for MCED. By integrating state-of-the-art deep neural network architectures with the intrinsic patterns inherent in methylation data, Genie-ADLA significantly enhanced MCED performance. Methods: Genie-ADLA was trained and evaluated on a dataset of 4,781 participants aged 40-75 years, including 2,702 pathologically confirmed cancer cases across 16 cancer types and 2,079 non-cancer controls (NCT06217900). The training set comprised 3,217 samples (1,756 cancer cases and 1,461 non-cancer controls), and the model’s performance was evaluated on an independent test set of 1,564 samples (618 non-cancer controls and 946 cancer cases). To address challenges inherent to methylation data—high dimensionality, sparsity, and noise—we applied feature dimensionality reduction and embedding strategies, reducing computational burden, mitigating overfitting, and improving learning efficiency. An ensemble learning approach further strengthened robustness and generalization. Results: Across all stages of 16 cancer types, Genie-ADLA achieved an overall sensitivity of 63.43% (600/946, 95% CI: [60.26%, 66.50%]) at 99.3% (612/618, ...