Deep learning pathway-based metabolic code for screening of Alzheimer’s disease in postmortem cerebrospinal fluid

抄録

Typical metabolic markers of Alzheimer’s disease (AD) provide new insights leading to early diagnostic strategies and laboratory medicine. The metabolomics approach can further advance this goal by discovering valuable metabolic pathways. Metabolomics generates a huge amount of data for validating the fine characteristics in biological samples. Fine characteristics in many fields are detected by artificial intelligence, especially, by deep learning (DL) techniques. However, when DL is applied to huge amounts of raw data, its performance is degraded by over-training and a large number of incorrect predictions. Here we propose an accurate low-capacity pathway-based metabolic code (PBMC) for DL of widely targeted metabolomics (W-TMet) data obtained by liquid chromatography with tandem mass spectrometry (LC-MS/MS). The method distinguishes the postmortem cerebrospinal fluid (p-CSF) of AD patients from that of control subjects. The W-TMet data were detected via a derivatized LC-MS/MS assay, which evaluated the diversity of 91 detectable amine metabolites in p-CSF. After averaging and density-contrasting the exhaustive information of each metabolite, we created PBMC images for a convolutional neural network. The AD vs. controls were distinguished with >80% accuracy. Positive AD diagnoses were misclassified when their pathological evaluation matched that of the control sample. For a visual clarification, the PBMC images were analyzed by gradient-weighted class activation mapping, which highlighted the significant metabolic pathways of AD. Many of the AD metabolic-marker candidates were associated with tryptophan metabolism.