Fragmentia-AI lymphoma: a cfDNA language model for lymphoma detection using ultra-low-pass WGS

Lymphomas are biologically and clinically heterogeneous malignancies. While liquid biopsies using circulating cell-free DNA (cfDNA) offer a non-invasive alternative, current methods typically require expensive deep sequencing. We developed Fragmentia-AI™ Lymphoma, a transformer-based genomic language model that enables lymphoma detection using ultra-low-pass whole-genome sequencing (ULP-WGS). The model integrates a genomic language model backbone with attention-based multiple instance learning to identify mutation-independent, malignancy-associated signals directly from raw cfDNA sequences. The model was trained on 389 samples (189 lymphoma, 200 healthy) and evaluated in an independent validation cohort of 190 lymphoma patients and 200 healthy controls. The model demonstrated robust diagnostic performance in the validation cohort, achieving an AUC of 0.944, with a sensitivity of 0.842 at 98% specificity and an F1 score of 0.904. Read-depth titration analysis showed sustained performance at sequencing depths as low as 0.01x, representing a substantial reduction compared with standard cfDNA mutation-based approaches. Attention-based fragment prioritization indicated that predictions were driven by fragmentomic signatures, including GC-content bias and altered fragment-size distributions characteristic of malignant cfDNA. These findings support the feasibility of a mutation-independent approach that may inform future lymphoma detection and monitoring strategies.