A Message Passing Framework for Precise Cell State Identification with scClassify2

In single-cell analysis, the ability to accurately annotate cells is crucial for downstream exploration. To date, a wide range of approaches have been developed for cell annotation, spanning from classic statistical models to the latest large language models. However, most of the current methods focus on annotating distinct cell types and overlook the identification of sequential cell populations such as transitioning cells. Here, we propose a message-passing-neural-network-based cell annotation method, scClassify2, to specifically focus on adjacent cell state identification. By incorporating prior biological knowledge through a novel dual-layer architecture and employing ordinal regression and conditional training to differentiate adjacent cell states, scClassify2 achieves superior performance compared to other state-of-the-art methods. In addition to single-cell RNA-sequencing data, scClassify2 is generalizable to annotation from different platforms including subcellular spatial transcriptomics data. To facilitate ease of use, we provide a web server hosting over 30 human tissues.