EL-PACA: ensembled low-dimensional projections for accurate cell-type annotation using deep learning for single-cell transcriptomics of human tissues and organoids

Abstract

Single-cell transcriptomics has been revolutionising the discovery of human cell types and disease mechanisms. Accurate human cell-subtype annotation in complex neural tissues and disorders remains challenging due to the scarcity of postmortem tissue availability, in vitro culture-specific influences in surrogate models, and low cell abundance or transient states that arise during the disease process. Here, we present an Ensembled Low-dimensional Projections for Accurate Cell-type Annotation (EL-PACA), a computationally efficient framework that ensembles unsupervised principal component analysis with supervised multiple discriminant analysis to increase class separability prior to training a deep classifier. EL-PACA outperforms state-of-the-art methods on most of the benchmarking datasets of blood and pancreatic cells. Importantly, it enables accurate, fine-grained cell-type identification in real-world datasets derived from human neural tissues. Applying EL-PACA to single-cell and single-nucleus RNA-sequencing datasets from postmortem human brain and 3D brain organoids, we show that combining supervised and unsupervised projections enables efficient, high-precision annotation with competitive performance across different analytical approaches. Altogether, EL-PACA facilitates the discovery of rare or transient cell types relevant to complex disease mechanisms. The framework is openly available at https://github.com/umar1196/EL-PACA.