由新型80万位点双峰扩增子测序技术实现的液体活检多癌种早期检测

Background: Multi-cancer early detection (MCED) from plasma liquid biopsy has advanced rapidly, demonstrating that diverse cfDNA features can reveal early tumor signals. Most current assays rely on whole-genome or methylation sequencing, which are costly and require deep coverage. EarlySeek is a novel, highly multiplexed cfDNA amplicon-based alternative that targets ~800,000 SINE-enriched loci using long and short amplicons, generating a bimodal insert-size distribution and requiring as little as 0.25 ng of DNA input. This design captures complementary signals such as fragmentation, aneuploidy, genomic abundance shifts, and sequence motif patterns. We evaluated a multi-signal framework integrating these features using artificial intelligence and machine learning. Methods: Three datasets were analyzed: a training set (237 cancers, 463 normals), a calibration set (72 cancers, 140 normals), and an independent test set with tissues present in training (78 cancers, 192 normals). EarlySeek output yields six biological scores: fragment length, two aneuploidy scores, two coverage/abundance scores, and a 6-mer motif score. Scores were generated using representation and deep learning architectures, including autoencoders that compress amplicon-level and binned genomic information into informative latent features, combined with machine learning classifiers such as support vector machines and gradient-boosted trees. Each score was calibrated via quantile-to-quantile regression, and a predefined multi-signal rule was applied to generate final scores for the test cohort. Results: Across 270 independent test samples, the combined six-score framework achieved 51% sensitivity at 99% specificity (95% CI: 40%-62%). Sensitivity by stage showed meaningful early detection: 45% for stage I (CI 26%-66%), 58% for stage II ( CI 39%-74%), 57% for stage III ( CI 37%-67%), and 100% for stage IV (2/2; CI 34%-100%). Performance varied by cancer type, with strongest detection in colorectal (61%), gastric (71%), liver (80%), ovarian (75%), and pancreatic cancer (64%). Lower sensitivities in breast, and prostate cancers reflected known low cfDNA shedding. Conclusions: EarlySeek's bimodal amplicon design enables extraction of diverse cfDNA signals from a single sequencing assay. Integrating these orthogonal features through artificial intelligence and machine learning supports high-specificity MCED detection and yields meaningful early-stage performance. This approach offers a scalable, cost-effective MCED test, a desirable feature for population-level cancer screening.