Blood‐based detection of early‐stage Alzheimer using multiomics and machine learning

Abstract Background The diagnosis of AD is based on cognitive symptoms, CSF assays and PET imaging. A blood test capable of detecting patients suffering from AD, at a prodromal or even pre‐symptomatic stage, will reduce the invasive and costly diagnostic tools currently used and, would allow an earlier clinical intervention. Here we assess the performance of a multiomics blood test using Machine Learning to combine metabolomic and proteomic biomarkers, pre‐identified in a brand‐new rat model. Method By sampling the plasma of a non‐transgenic animal model successfully reproducing the continuum of Alzheimer's disease progression at the brain level (Audrain et al., 2017), we identified the 105 most informative biomarkers using Artificial Intelligence. Then we analyzed the behavior of these biomarkers in 232 human plasma samples collected up to 15 years before the dementia. Three independent cohorts were used: two with the sporadic form of AD and one with Down Syndrome individuals. For each sample the 105 pre‐identified biomarkers (proteins, metabolites) were analyzed by global mass spectrometry. Using Artificial Intelligence, we identified the 25 best‐in‐class biomarkers in humans. Then we developed a neural network based on these 25 biomarkers to detect AD patients from the pre‐symptomatic phase. Result Among the 25 biomarkers, 13 are proteins and 12 are metabolites. None of these biomarkers are produced by the brain. They are produced or regulated by peripheral organs. The neural network identify Alzheimer’s patients (including pre‐symptomatic, prodromal and demented patients) from Non Alzheimer Individuals (Healthy controls and patients suffering from a neurodegenerative disease excluding Alzheimer) with 100% sensitivity and 99% specificity on a 5‐folds cross validation. Based on non‐linear and non‐monotonic biomarker progression, the neural network can identify undemented AD patients, including pre‐symptomatic and prodromal patients, against demented patient, with 99% overall accuracy. Conclusion In our retrospective, multi‐center study, we achieved high accuracy (>99%) for Alzheimer’s from pre‐symptomatic phase. Proteic and metabolomic signals are required for an effective early detection. These results strongly suggest that the use of non‐brain‐produced biomarkers improves sensitivity and specificity of early detection.