Fragle

Quantifying circulating tumor DNA (ctDNA) in blood enables non-invasive cancer detection and monitoring, but existing tumor-naive approaches based on low-pass whole-genome sequencing (lpWGS) typically plateau at a ~3% limit of detection, while more sensitive tumor-informed assays require deep sequencing and a matched tumor biopsy.

Fragle is a multistage deep-learning model that estimates the ctDNA fraction directly from the density distribution of cell-free DNA fragment lengths, requiring neither a tumor biopsy nor a matched-normal sample.

Trained on 426 lpWGS samples across four cancer types together with ~2,500 in silico dilutions and validated on 506 independent samples spanning six additional cancer types, Fragle detects ctDNA at ~1% with an AUC of 0.93, outperforming ichorCNA (0.88), the prior best tumor-naive method. Accurate quantification is retained at ultra-low sequencing coverage (0.05x WGS) and on off-target reads from targeted sequencing panels (r >= 0.96).

In clinical cohorts, longitudinal ctDNA trajectories tracked radiographic treatment response in colorectal cancer, and in the MEDAL cohort of resected early-stage lung cancer (162 patients), elevated day-30 post-surgery ctDNA (>1%) predicted poorer disease-free survival (HR 2.43), supporting the use of minimal residual disease (MRD) risk stratification.

By delivering accurate, cost-effective ctDNA quantification from routine sequencing data, Fragle has the potential to enhance high-frequency cancer surveillance and precision oncology in both research and clinical molecular diagnostics. 

Fragle is freely available for academic use. Commercial usage requires a license. https://github.com/skandlab/FRAGLE

Zhu G, Rahman CR, Getty V, et al. A deep-learning model for quantifying circulating tumour DNA from the density distribution of DNA-fragment lengths. Nat Biomed Eng. 2025;9(3):307-319. doi:10.1038/s41551-025-01370-3

https://www.nature.com/articles/s41551-025-01370-3