Abstract 2390: Noninvasive early detection of colorectal cancer by a deep read-level model for accurate identification of methylation-specific circulating tumor DNA

Le résumé fourni par la source

Abstract Background: Liquid biopsy enables noninvasive cancer screening through the detection of plasma circulating tumor DNA (ctDNA), yet faces challenges due to the low ctDNA abundance in plasma, especially in early-stage cancer. Precise discrimination of ctDNA fragment can significantly enrich tumor signals and enhance detection accuracy, while having difficulties owing to the ambiguous labelling and insufficient characterization of tumor-derived reads. Here, we propose DREAM (a Deep Read-level model for accurate Identification of Methylation-specific ctDNA) for early colorectal cancer (CRC) detection, which can robustly identify ctDNA through simultaneous capture of DNA sequence and methylation information of tumor-derived reads. Methods: The read-level model DREAM was trained on a well-labelled dataset derived from whole-genome bisulfite sequencing (WGBS) reads of 55 CRC tissues and 200 healthy plasma samples, which located in CRC-specific hyper- and hypo-methylated regions. The CRC tissue reads were firstly denoised based on the methylation continuity of healthy samples in these regions to filter non-tumor signal noise. The sequence and methylation information of the processed CRC reads and healthy plasma reads were then encoded as training set for a convolutional neural network (CNN)-based model to predict the tumor likelihood of each read. To validate the classification capability and subsequent CRC detection performance of DREAM, a plasma cohort including 208 CRC and 686 healthy plasmas with targeted bisulfite sequencing (TBS) was collected for model training. With DREAM, the regional ctDNA fraction of each sample was estimated and then utilized in an XGBoost classifier to predict sample-level cancer probabilities. The classifier was subsequently validated in the test set including 54 CRC, 170 healthy and 11 colon benign individuals. Results: DREAM reached a remarkable read-level performance enhancement by the refined denoising procedure, with an accuracy increase from 0.80 to 0.96. For the test plasma cohort, DREAM achieves 85.2% sensitivity (95% CI: 84.1%-88.5%) at 95.6% specificity (95% CI: 95.1%-96.9%) for CRC detection. Specifically, the sensitivities for stage I to IV were 60.0% (3 / 5), 81.3% (13 / 16), 88.9% (24 / 27), and 100% (6 / 6). Notably, among the 11 benign samples in the non-cancer group, DREAM successfully differentiated them from malignant CRC with an AUC of 0.966, showing the effectiveness and accuracy of capturing tumor-derived signals. Conclusions: In conclusion, DREAM showcases a significant advancement in early CRC detection through its precision in enriching tumor signals, which may contribute to more effective cancer screening and be further broadened in a wider range of cancer types. Citation Format: Ping Lan, Xiaosheng He, Hong Xu, Zhijian Xu, Nan Lin, Jingjing Hou, Xiaowen He, Yanzhen Cao, Jiandong Tai, Liyuan Zhao, Yiran Cai, Yonghui Li, Yanzhan Yang, Min Li, Yang Wang, Fang Liu, Hui Yu, Baoliang Zhu, Xiaohui Wu. Noninvasive early detection of colorectal cancer by a deep read-level model for accurate identification of methylation-specific circulating tumor DNA [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2390.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.