Generalizable cancer detection from ultra-low-pass WGS via deep contextual modeling of cfDNA sequences
TL;DR
Fragmentia-AI™ WGS, a mutation-calling-independent framework that uses a transformer-based multiple-instance learning architecture with sequential fine-tuning across tumor fraction (TF) strata to extract latent cancer-associated signals from ULP-WGS data, enables robust cancer detection and clinically meaningful risk stratification from highly sparse cfDNA sequencing data.
Abstract
Ultra-low-pass whole-genome sequencing (ULP-WGS) of cell-free DNA (cfDNA) offers a cost-efficient strategy for cancer detection, but its clinical application is limited by extreme data sparsity and poor model generalization. We developed Fragmentia-AI™ WGS, a mutation-calling-independent framework that uses a transformer-based multiple-instance learning architecture with sequential fine-tuning across tumor fraction (TF) strata to extract latent cancer-associated signals from ULP-WGS data. Model performance was evaluated in multiple independent cohorts, including a pan-cancer test set covering 17 cancer types, an external public dataset generated on a different sequencing platform, and a technical variability cohort with heterogeneous pre-analytical and experimental conditions. Clinical relevance was assessed by correlating model predictions with progression-free survival (PFS) in patients with advanced non-small cell lung cancer receiving chemoimmunotherapy. Sequential fine-tuning across TF strata significantly improved performance in low-TF samples, achieving a 35.6% relative increase in AUC compared with high-TF-only training (0.884 vs. 0.652). In the independent test cohort, the model achieved an overall AUC of 0.930, with consistent performance across TF strata and cancer types. External validation confirmed robust cross-platform generalizability (AUC: 0.929; sensitivity: 0.78; specificity: 0.92). The model maintained stable classification performance despite score fluctuations associated with pre-analytical and technical variables. Importantly, model-negative status, defined as a prediction score below the training-derived cutoff, remained significantly associated with improved PFS compared with model-positive status (HR = 0.49, 95% CI: 0.29–0.82) after multivariable adjustment. Collectively, this framework enables robust cancer detection and clinically meaningful risk stratification from highly sparse cfDNA sequencing data.