Generalizable cancer detection from ultra-low-pass whole-genome sequencing of cell-free DNA using a sequentially fine-tuned transformer framework.

10533 Background: Ultra-low-pass whole-genome sequencing (ULP-WGS) of cell-free DNA (cfDNA) provides a cost-effective approach for cancer screening, yet its clinical translation is constrained by extreme data sparsity and poor model generalizability, particularly in low-tumor fraction (TF) settings. Methods: We developed Fragmentia-AI WGS, a mutation-independent framework that leverages a transformer-based multiple-instance learning architecture with sequential fine-tuning across TF strata to extract latent cancer-associated signals from ULP-WGS data (~0.02× coverage, ~128k reads). Model performance and generalizability were evaluated across multiple independent cohorts, including a pan-cancer test cohort across 17 cancer types, a public online cohort sequenced on a different platform, and an analytical variability assessment cohort comprising samples processed under heterogeneous technical and pre-analytical conditions. Clinical relevance was further examined by associating prediction scores with progression-free survival (PFS) in patients with advanced non-small cell lung cancer receiving immunotherapy. Results: Sequential fine-tuning across all TF strata substantially improved performance in low-TF samples compared with high-TF-only training, yielding a 35.6% relative increase in AUC in the low-TF subgroup (from 0.652 to 0.884 in the validation cohort). In the independent test cohort, the model achieved an overall AUC of 0.930 (95% CI: 0.924 – 0.936), with consistently strong performance across TF levels (high TF: 0.984; medium TF: 0.958; low TF: 0.910) and across cancer types. External validation in the public cohort demonstrated robust cross-platform and cross-population generalizability (AUC: 0.929; sensitivity: 0.78; specificity: 0.92). Prediction scores remained stable across heterogeneous experimental conditions, supporting robustness to technical and pre-analytical variability. Notably, lower model prediction scores were independently associated with prolonged PFS (hazard ratio: 0.49, 95% CI: 0.31 – 0.77; p = 0.002) after adjusting for sex, age, clinical stage, and histologic subtype. Conclusions: Fragmentia-AI WGS enables robust and generalizable cancer detection as well as clinically meaningful risk stratification from ultra-sparse cfDNA sequencing, underscoring its potential for scalable and cost-effective liquid biopsy applications.