Abstract P23: A deep-learning model for quantifying circulating tumour DNA from the density distribution of DNA-fragment lengths

Abstract Background: Frequent, low-cost ctDNA profiling promises real-time therapy monitoring and economical triage of scarce volume plasma samples. Yet, current methods either require tumour-informed deep-sequencing or plateau at a ∼3% limit of detection (LoD) when tumour-naïve low-pass WGS (lpWGS) is used. A general method that detects ctDNA at the 1% level is still lacking. Methods: We developed FRAGLE, a multistage neural network that estimates the ctDNA fraction from the cfDNA fragment-length density distribution. Training was performed on 426 lpWGS samples (325 cancer, 101 healthy) from four tumour types, along with ∼2,500 in silico dilutions spanning ∼0.1–70% ctDNA. External validation comprised 506 unseen lpWGS samples from an additional six cancer types, two in vitro titration series, and a large in silico panel. Results: Cross-validation yielded mean absolute errors of 3.1-3.9% and Pearson r values ranging from 0.67 to 0.94 across tumor types. In external cohorts, FRAGLE detected ctDNA at ∼1% with an AUC of 0.93, surpassing ichorCNA (0.88). In vitro and in silico dilutions confirmed the linearity of ctDNA fraction estimations down to 1%. Accurate quantification held at 0.05x WGS and on off-target reads from targeted panels (r ≥ 0.96). Longitudinal colorectal-cancer sampling showed high concordance with radiographic response. In the MEDAL trial (162 resected lung cancers), ctDNA > 1% on day 30 predicted inferior disease-free survival (HR 2.43, 95% CI 0.98–6.03; p = 0.035). Conclusion: FRAGLE converts lpWGS (even 0.05x) or panel off-target data into 1-minute-fast, NGS <US$50, tumour-agnostic ctDNA estimates with a LoD of ∼1%. This capability supports high-frequency surveillance and risk stratification for minimal residual disease across diverse cancer patients. Citation Format: Guanhua Zhu, Chowdhury Rafeed Rahman, Victor Getty, Denis Odinokov, Probhonjon Baruah, Hanaé Carrié, Avril Joy Lim, Yu Amanda Guo, Zhong Wee Poh, Ngak Leng Sim, Ahmed Abdelmoneim, Yutong Cai, Lakshmi Narayanan Lakshmanan, Danliang Ho, Saranya Thangaraju, Polly Poon, Yi Ting Lau, Anna Gan, Sarah Ng, Si-Lin Koo, Dawn Q. Chong, Brenda Tay, Tira J. Tan, Yoon Sim Yap, Aik Yong Chok, Matthew Chau Hsien Ng, Patrick Tan, Daniel Tan, Limsoon Wong, Pui Mun Wong, Iain Beehuat Tan, Anders Jacobsen Skanderup. A deep-learning model for quantifying circulating tumour DNA from the density distribution of DNA-fragment lengths [abstract]. In: Proceedings of Frontiers in Cancer Science 2025; 2025 Nov 5-7; Singapore. Philadelphia (PA): AACR; Cancer Res 2026;86(13_Suppl):Abstract nr P23.