Abstract
Identifying key cell types and genes in Alzheimer's Disease (AD) is crucial for understanding its pathogenesis and discovering therapeutic targets. Single-cell RNA sequencing technology (scRNAseq) has provided unprecedented opportunities to study the molecular mechanisms that underlie AD at the cellular level. In this study, we address the problem of sample-level classification of AD using scRNAseq data, where we predict the disease status of entire samples from the gene expression profiles of their cells, which are not necessarily all affected by the disease. We introduce scAGG (single-cell AGGregation), a sample-level classification model that uses a sample-level pooling mechanism to aggregate single-cell embeddings, and show that it can accurately classify AD individuals and healthy controls. We then investigate the latent space learnt by the model and find that the model learns an ordering of the cells corresponding to disease severity. Genes associated with this ordering are enriched in AD-linked pathways, including cytokine signalling, apoptosis, and metal ion response. We also evaluate two attention-based models that perform on par with scAGG, but entropy analysis of their attention scores reveals limited interpretability value. As scRNAseq is increasingly applied to large cohorts and cell-level disease association annotations do not exist, our approach provides a way to classify phenotypes from single-cell measurements. The yielded cell- and sample-level severity scores may enable identification of AD-associated cell subtypes, paving the way for targeted drug development and personalized treatment strategies in AD. Code is available at: https://github.com/timoverlaan/scAGG.