scCoRR: A Data-Driven Self-correction Framework for Labeled scRNA-Seq Data

Single-cell RNA sequencing (scRNA-seq) data serves as the foundation for many studies investigating cellular heterogeneity. Numerous methodologies and evaluation metrics within single-cell research are intertwined with cell labels. While annotating cell labels often requires prior biological knowledge for clustering, this is frequently approached from a clustering perspective rather than considering the heterogeneity of individual cells. Building upon this, we introduce a data-driven self-correction framework for labeled scRNA-seq data, termed scCoRR. This framework utilizes a supervised approach trained from partially reliable anchor cells, eliminating the need for additional prior reference datasets or marker genes. Subsequently, a supervised deep neural network is trained with cross-entropy loss and a contrastive regularization term to predict the types of the remaining cells. During this process, the labels of some cells are corrected from one cell type to another, a phenomenon that can also be elucidated from various biological perspectives.