An interpretable deep learning framework identifies proteomic drivers of Alzheimer's disease

Alzheimer's disease (AD) is the leading neurodegenerative pathology in aged individuals, but many questions remain on its pathogenesis, and a cure is still not available. Recent research efforts have generated measurements of multiple omics in individuals that were healthy or diagnosed with AD. Although machine learning approaches are well-suited to handle the complexity of omics data, the models typically lack interpretability. Additionally, while the genetic landscape of AD is somewhat more established, the proteomic landscape of the diseased brain is less well-understood. Here, we establish a deep learning method that takes advantage of an ensemble of autoencoders (AEs) - EnsembleOmicsAE-to reduce the complexity of proteomics data into a reduced space containing a small number of latent features. We combine brain proteomic data from 559 individuals across three AD cohorts and demonstrate that the ensemble autoencoder models generate stable latent features which are well-suited for downstream biological interpretation. We present an algorithm to calculate feature importance scores based on the iterative scrambling of individual input features (i.e., proteins) and show that the algorithm identifies signaling modules (AE signaling modules) that are significantly enriched in protein-protein interactions. The molecular drivers of AD identified within the AE signaling modules derived with EnsembleOmicsAE were missed by linear methods, including integrin signaling and cell adhesion. Finally, we characterize the relationship between the AE signaling modules and the age of death of the patients and identify a differential regulation of vimentin and MAPK signaling in younger compared with older AD patients.