Machine Learning to Detect Alzheimer's Disease from Circulating Non-coding RNAs

Blood-borne small non-coding (sncRNAs) are among the prominent candidates for blood-based diagnostic tests. Often, high-throughput approaches are applied to discover biomarker signatures. These have to be validated in larger cohorts and evaluated by adequate statistical learning approaches. Previously, we published high-throughput sequencing based microRNA (miRNA) signatures in Alzheimer's disease (AD) patients in the United States (US) and Germany. Here, we determined abundance levels of 21 known circulating miRNAs in 465 individuals encompassing AD patients and controls by RT-qPCR. We computed models to assess the relation between miRNA expression and phenotypes, gender, age, or disease severity (Mini-Mental State Examination; MMSE). Of the 21 miRNAs, expression levels of 20 miRNAs were consistently de-regulated in the US and German cohorts. 18 miRNAs were significantly correlated with neurodegeneration (Benjamini-Hochberg adjusted P < 0.05) with highest significance for miR-532-5p (Benjamini-Hochberg adjusted P = 4.8 × 10^-30). Machine learning models reached an area under the curve (AUC) value of 87.6% in differentiating AD patients from controls. Further, ten miRNAs were significantly correlated with MMSE, in particular miR-26a/26b-5p (adjusted P = 0.0002). Interestingly, the miRNAs with lower abundance in AD were enriched in monocytes and T-helper cells, while those up-regulated in AD were enriched in serum, exosomes, cytotoxic t-cells, and B-cells. Our study represents the next important step in translational research for a miRNA-based AD test.