Multi-Task Encoder Using Peripheral Blood DNA Methylation Data for Alzheimer’s Disease Prediction

This study introduces a multi-task prediction model, MT-MBLAE, designed to use DNA methylation data from blood to predict the advancement of Alzheimer’s disease. By integrating various modules, including bi-directional long short-term memory (BiLSTM), long short-term memory (LSTM), and RepeatVector, among others, the model encodes DNA methylation profile data, capturing temporal and spatial information from instantaneous DNA methylation spectra data. Leveraging the network properties of BiLSTM and LSTM enables the consideration of both preceding and subsequent information in sequences, facilitating the extraction of richer features and enhancing the model’s comprehension of sequential data. Moreover, the model employs LSTM, time distributed to reconstruct time series DNA methylation profiles. The time-distributed layer applies identical layers at each time step of the sequence, sharing weights and biases uniformly across all time steps. This approach achieves parameter sharing, reduces the model’s parameter count, and ensures consistency in handling time series data. Experimental findings show the excellent performance of the MT-MBLAE model in predicting cognitively normal (CN) to mild cognitive impairment (MCI), and mild cognitive impairment (MCI) to Alzheimer’s disease (AD).