Machine Learning Reveals the Unique Biomarkers of Clonal Hematopoiesis in Patient With Early-Stage Colorectal Neoplasia: A Case Control Study

Abstract BACKGROUND: Blood test has a better uptake for colorectal cancer screening than stool test and colonoscopy but suboptimal detection of early-stage colorectal neoplasia (CRN), including advanced adenoma and stage I cancer, limits its application. The present study aimed to evaluate whether clonal hematopoiesis (CH) from peripheral blood can be used as a biomarker for early-stage CRN screening and improve the detection of blood tests by machine-learning approach.METHODS: The CH profile was evaluated in 63 early-stage CRNs and 32 controls by error-corrected sequencing and classified by machine-learning method. Diagnostic performance was measured by receiver operator characteristic analysis. Additional 20 early-stage CRNs and 10 controls were used to validate the machine-learning model. We simultaneously used mutational signature analysis to study predictors based on CH.RESULTS: We identified 1,446 variants and clarified the uniqueness of variants from the peripheral bloods of early-stage CRNs. The machine learning model identified early-stage CRNs from controls and its AUC, sensitivity and specificity were 0.988, 94.2% and 99.3%, respectively. The CH-based CRN detection model was further verified. The accuracy, sensitivity, and specificity were 0.933 (p=0.00065), 95.0%, and 90.0%, respectively. Furthermore, the mutational signature analysis of those unique variants in CRNs revealed the influence of genetic architecture on DNA damages.CONCLUSIONS: Our results reveal the potential of CH to a mark produced by the carcinogenesis in early-stage CRN. We developed a CH-based blood test with machine learning approach, which not only increase screening uptake but also improve the detection rate of early-stage CRN.