Abstract A035: Denoising Models Enhance Detection of Tumor-Derived cfDNA fragments and Cancer Tissue signal in Liquid Biopsy

Abstract Cell-free DNA (cfDNA) methylation profiling enables non-invasive early cancer detection and tissue-of-origin (TOO) localization. In early-stage disease, circulating tumor DNA (ctDNA) comprises ∼0.05% of the total cfDNA, with the remainder originating from hematopoietic and other non-malignant tissues, significantly limiting detection sensitivity. Filtering non-tumor fragments can improve tumor-specific signal, but distinguishing ctDNA from non-tumor methylation patterns remains challenging due to biological overlap and technical noise. Generative models, particularly denoising diffusion models (DDMs), are well-suited for cfDNA analysis due to their capacity to model complex data distributions perturbed by structured and stochastic noise. Building on this, we developed a novel approach that adapts DDMs to learn non-cancer cfDNA methylation patterns and filter fragment-level data, thereby enhancing tumor-derived signal and TOO inference. Peripheral blood samples (CORE-HH study; NCT05435066) were collected in 9 cancer types from treatment-naïve cancer patients (N=401) and individuals with no reported cancer, evenly distributed across ages (N=100). cfDNA was extracted from plasma and profiled using a custom 18.6 Mb targeted bisulfite sequencing hybrid capture assay. cfDNA fragments were pre-processed to retain only CpG sites, with methylation states encoded as binary sequences. U-Net based Denoising Diffusion Implicit models (DDIMs) were trained on these sequences to learn the underlying methylation patterns in non-cancer cfDNA (N=80). Mean Squared Error was used as reconstruction error to identify and retain tumor derived cfDNA fragments. Tumor-derived fragments, being out-of-distribution (OOD) relative to the training data, exhibit higher errors than non-cancer fragments, allowing for their effective identification. OOD thresholds per genomic start location were established from cfDNA fragments in held-out non-cancer samples (N=20). To predict TOO across all cancer samples, we trained a four-layer feed-forward neural network using five distinct metrics that quantify informative methylation signals at each genomic region of interest (Farashahi et al. (2024), Clin Cancer Res, 30). Performance was evaluated using 10-fold cross validation and reported for samples with tumor fraction greater than 0.1% (N=227). Filtering the samples using 90th percentile OOD thresholds revealed a significant correlation (Kendall’s τ=0.5) between the estimated tumor fraction of samples and the fraction of retained reads. Keeping only these reads increased the estimated tumor fraction ten-fold. Our TOO multi-class classification on filtered cfDNA achieved 87% balanced accuracy (BA) versus 78% BA on unfiltered cfDNA. These findings demonstrate that denoising-based OOD models can effectively learn non-tumor cfDNA methylation patterns despite inherent noise in cfDNA and filter fragment-level data to enhance tumor-derived signal. This novel approach improves TOO resolution and can increase the clinical utility of multi-cancer early detection by guiding diagnostic workup. Citation Format: Shiva Farashahi, Yifan Wu, Elie Massaad, Feras Hantash, Hutan Ashrafian, Dorna Kashef, Kieran Chacko. Denoising Models Enhance Detection of Tumor-Derived cfDNA fragments and Cancer Tissue signal in Liquid Biopsy [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A035.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.