3066 Background: The demand for alternative, non-invasive methods for colorectal cancer (CRC) screening is substantial. Cell-free DNA (cfDNA) whole genome sequencing (WGS) offers a promising avenue, utilizing diverse fragmentomic data. We aimed to develop a new approach: integrating fragment end motif by size (FEMS) with genomic coverage (COV) of cfDNA to enhance CRC screening. Methods: Participants were comprised of 1,506 colonoscopy verified normal samples, 130 advanced adenoma (AA) patients, 302 CRC patients (stage I: 28.5%, stage II: 25.5%, stage III: 31.1%, stage IV: 14.2%, unknown: 0.7%). We generated low depth cfDNA-WGS data (a minimum 40 million reads per sample) using the Novaseq 6000 sequencer. We minimized batch bias by normalizing with the same samples for each experimental batch. The development set comprised 1,332 samples (1,023 normal, 103 AA, and 206 CRC), while the validation set consisted of 606 samples (483 normal, 27 AA, and 96 CRC). Results: Our new algorithm achieved 84.0% sensitivity (95% CI: 81%-86%) for CRC and 69.9% (95% CI: 66%-81%) for AA in the development set, with a specificity of 90.0%. These findings were consistent in the external set with 84.9% specificity (95% CI: 82%-88%) and sensitivities of 80.2% (95% CI: 72%-88%) and 63.0% (95% CI: 44%-82%) for CRC and AA, respectively. The sensitivities across stages (Stage I: 80%, Stage II: 86.8%, Stage III: 78.8%, Stage IV: 96.7%) were similarly reflected in the validation sets (Stage I: 74%, Stage II: 83%, Stage III: 79%, Stage IV: 92%). Conclusions: Our multimodal deep learning model shows promise for accurate non-invasive colon cancer and AA detection using low-coverage cfDNA WGS. Its strong performance in detecting not only CRC but also AA lesions suggests its potential for early intervention and improved patient outcomes. The use of colonoscopy-verified normal samples strengthens the study's credibility and paves the way for future clinical translation.
Read full abstract