A deep-learning model for quantifying circulating tumour DNA from the density distribution of DNA-fragment lengths
Menée à partir de données de séquençage à faible couverture du génome entier de patients atteints d'un cancer et de témoins en bonne santé, cette étude met en évidence la performance d'un algorithme d'apprentissage automatique, utilisant la fréquence des longueurs des fragments d'ADN détectés, pour quantifier rapidement avec précision l'ADN tumoral circulant
The quantification of circulating tumour DNA (ctDNA) in blood enables non-invasive surveillance of cancer progression. Here we show that a deep-learning model can accurately quantify ctDNA from the density distribution of cell-free DNA-fragment lengths. We validated the model, which we named ‘Fragle’, by using low-pass whole-genome-sequencing data from multiple cancer types and healthy control cohorts. In independent cohorts, Fragle outperformed tumour-naive methods, achieving higher accuracy and lower detection limits. We also show that Fragle is compatible with targeted sequencing data. In plasma samples from patients with colorectal cancer, longitudinal analysis with Fragle revealed strong concordance between ctDNA dynamics and treatment responses. In patients with resected lung cancer, Fragle outperformed a tumour-naive gene panel in the prediction of minimal residual disease for risk stratification. The method’s versatility, speed and accuracy for ctDNA quantification suggest that it may have broad clinical utility.
Nature Biomedical Engineering , résumé, 2025