Chemiluminescence signature arrays coupling with machine learning for Alzheimer’s disease serum diagnosis

Tremendous efforts have been made to directly identify serum components using traditional omics approaches. However, several unmet medical needs persist, particularly for chronic diseases that lack reliable biomarkers. The subtle physicochemical abnormality of serum has been widely overlooked and currently lacks detection methods. Inspired by the bat echolocation mechanism, we proposed a chemiluminescence “echoes” approach to depict the disease-specific signatures in biofluids. Specifically, Alzheimer’s disease (AD) serums were used for proof-of-concept study. We first demonstrated the discrepancy in physicochemical properties between AD and healthy control (HC) serums. On this basis, we developed a simple, fast and versatile UNICODE (UNiversal Interaction of Chemiluminescence echOes for Disease Evaluation) array for AD diagnosis. By employing a "bat" probe (ADLumin-1), which generates chemiluminescence autonomously, and combined with a panel of “flag” molecules that enable “echo” formation, we were able to create distinct signatures for various serum components and subtle physicochemical environments. To develop an AD-specific UNICODE diagnosis, we screened a library of over 1000 small molecules, and identified 12 “flag” molecules (top-12) that optimally depict the differences between AD and HC serums. Finally, we used the top-12 array for AD diagnosis. By modeling the UNICODE signatures with seven machine learning methods, we successfully differentiated AD (n = 31) and HC (n = 37) serums, and our best model of random forest provided accuracy = 85.48%, precision = 85.00%, recall = 88.60%, f1 = 85.63%, and AUC = 90.24%. Our strategy could provide new insights into biofluid abnormality and prototype tools for developing liquid biopsy diagnoses for AD and other diseases.