A scalable deep-learning framework for cancer detection using cell-free DNA shallow whole-genome sequencing
Menée à partir de 2 063 échantillons plasmatiques prélevés sur 631 témoins sains et 1 432 patients atteints d'un cancer (26 localisations), cette étude compare la sensibilité et la spécificité, pour détecter la maladie, de deux approches utilisant des données de séquençage à faible profondeur de l'ADN libre circulant : l'une basée sur XGBoost, un algorithme d'apprentissage automatique classique et l'autre sur les réseaux de neurones convolutifs
Cell-free DNA (cfDNA) in body fluids enables noninvasive cancer detection. Multifeature artificial intelligence (AI) can improve sensitivity by integrating diverse biomarkers when cancer signals are sparse. Tumor-informed assays that rely on mutations have limited practicality for early cancer detection. Emerging fragmentomic and epigenetic features underpin tumor-naive approaches to screening for individuals with low tumor burden. Here, we designed UNITE—a universal cfDNA feature ensemble framework that provides scalable cancer detection methods based on “genomic bin–fragment length” matrices derived from shallow whole-genome sequencing (sWGS) data at 0.1× depth. Using sWGS data from 2063 plasma samples (631 controls and 1432 cases from 26 cancer types), we systematically evaluated both XGBoost (UNITE-XGB) and convolutional neural networks (UNITE-CNN) across multiple feature spaces and cancer stages. In stage I-II cancer, UNITE-XGB and UNITE-CNN achieved 31 and 21% sensitivity, respectively, at 95% specificity. These findings provide roadmaps for developing multifeature AI beyond plasma biopsies. Multifeature AI detects cancer across all stages with 47% sensitivity at 95% specificity using sWGS data.
Science Advances , article en libre accès, 2026