Fragmentia-AI-COAD reveals cancer-associated fragmentomic signatures and enables sensitive blood-based detection of colorectal cancer

Colorectal cancer (CRC) and advanced adenomas (AA) release cell-free DNA (cfDNA) fragments that reflect tumor-associated chromatin organization and regulatory states. However, most liquid biopsy approaches rely on predefined molecular features and require substantial sequencing depth. We developed Fragmentia-AI-COAD, a transformer-based cfDNA sequence modeling framework that identifies biologically relevant fragmentomic signatures directly from low-pass whole-genome sequencing data. Fragmentia-AI-COAD integrates transformer-based sequence encoding with attention-based multiple instance learning to capture contextual sequence patterns from millions of cfDNA fragments. The model was trained and validated using plasma samples from clinically ascertained CRC and AA cases and controls confirmed negative for CRC and AA by colonoscopy, and subsequently evaluated in an independent test cohort. Model robustness was assessed through repeated inference and in silico fragment-count subsampling analyses. Explainability analyses were performed to identify fragmentomic features contributing to model predictions. In the independent test cohort, Fragmentia-AI-COAD achieved an area under the curve of 0.971 for CRC detection and 0.914 for AA detection, with an overall sensitivity of 92.9% at 93.3% specificity. Sensitivity for stage I CRC reached 90.5%, whereas AA sensitivity was 76.2%. Diagnostic performance was largely retained when fewer fragments were sampled from the existing sequencing libraries. Explainability analyses revealed cancer-associated fragmentomic signatures characterized by fragment shortening, increased sequence complexity, and enrichment of specific sequence motifs, indicating that biologically meaningful cfDNA patterns contribute to model predictions. Transformer-based cfDNA sequence modeling showed promising performance for detecting CRC and advanced adenomas from low-pass whole-genome sequencing data while identifying cancer-associated fragmentomic signatures. These findings support further prospective evaluation of this approach for blood-based colorectal cancer detection.