Abstract:
|
Single-Cell RNA-Sequencing (scRNAseq) is a transformative technology that enables researchers to investigate heterogeneous cell population with high resolution. While many early studies focused on analyzing one biological sample, recent years have witnessed an increasing number of studies with multiple biological replicates across multiple conditions. Therefore, differential gene expression analysis for multiple scRNAseq samples is in urgent need. Existing methods either ignore the cell-to-cell variation by taking the average gene expression levels within a sample or introduce an imbalanced representation of sample-to-sample variability by considering each cell as an observation. We propose a novel method, in which we treat each biological sample as a "bag" and each cell within the sample as an "instance", then learn the attention weights for each cell via a deep learning model. We further use the attention weights to construct the weighted average of gene expression data for our sample-level representation. We apply our method on a publicly available Alzheimer disease dataset. Our method achieved high prediction performance as well as indicating important biological processes.
|