Fragmentia AI–Lymphoma: A cfDNA language model for lymphoma detection using ultra-low-pass WGS.

Abstract

7019 Background: Despite the potential of cfDNA liquid biopsy for non-invasive cancer monitoring, its clinical utility is often limited by high sequencing costs and a reliance on detectable driver mutations. To address these barriers, we introduce Fragmentia AI – Lymphoma, a novel transformer-based cfDNA language model designed for lymphoma detection using cost-effective ultra-low-pass whole genome sequencing (ULP-WGS). Methods: Trained on a cohort of 389 samples (189 lymphoma and 200 healthy), the architecture integrates genomic language model backbone with gated attention-based multiple instance learning. Fragmentia AI – lymphoma learned to identify malignancy-associated, mutation-independent signals directly from raw cfDNA sequences. We validated performance on an independent test cohort of 190 lymphoma patients and 200 healthy controls. Additionally, to evaluate clinical scalability, we conducted a read-depth titration analysis to test the minimum input requirements for sustained model performance. Results: Fragmentia AI – lymphoma achieved an AUC of 0.943 in the training cohort and 0.944 in the testing cohort. At 95% specificity, the model demonstrated a sensitivity of 0.889 (F1 score: 0.913). Notably, diagnostic performance remained robust even with a threefold reduction in sequencing reads (AUC > 0.94), significantly lowering the required depth compared to standard somatic mutation calling. Feature attribution analysis revealed that model’s decision-making was predominantly anchored in pathognomonic fragmentomic signatures, specifically GC-content biases and aberrant fragment-size distributions characteristic of malignant cfDNA. Conclusions: Our model effectively identified mutation-independent diagnostic signals from low coverage sequencing data, providing a scalable and cost-effective approach for lymphoma screening and monitoring.