Discovering cell types and states from reference atlases with heterogeneous single-cell ATAC-seq features

Abstract

Despite substantial recent advances in query mapping and cell type or cell state discovery tools, their application to single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) data remains challenging. The heterogeneous nature of peak feature spaces across samples hinders the effectiveness of existing methods, while the absence of dedicated tools for detecting perturbed cell types and states in scATAC-seq data further limits the depth of downstream analyses. To address these limitations, we present EpiPack, an integrative computational toolkit that leverages heterogeneous transfer learning and graph-based modeling strategies to advance scATAC-seq analysis. At its core, the Peak Embedding Informed Variational Inference (PEIVI) framework within EpiPack enhances mappable reference construction, query mapping, and label transfer, demonstrating that leveraging heterogeneous features in scATAC-seq data outperforms methods relying solely on conventional homogeneous features. In addition, EpiPacks global-local out-of-reference (OOR) detection framework achieves robust and efficient detection of perturbed cell types and states, extending the utility of scATAC-seq to disease and perturbation contexts. With its modular design and transferable pre-trained references, EpiPack can be readily applied to diverse analytical tasks and is available as a Python package at https://github.com/ZhangLabGT/EpiPack.