Multiview Transformer-Based Hierarchical Fusion Model for Cell Type Identification

Identifying cellular identities is a key initial process in the analysis of single-cell RNA sequencing (scRNA-seq) data. Although a number of methods have been developed for this purpose, such tools struggle with a limited number of curated marker gene lists, improper processing of batch effects, and struggle to maintain harmony between accuracy and interpretability. To overcome these challenges, we develop MTHCell, which introduces multi-view biological knowledge encoding and multi-scale feature learning to transformers. At the view level, the supervised learning of modal features and the imposing of distance constraints between different views allow the network to achieve a good balance between learning common information and discrepancy information across diverse views. At the instance level, the mechanism for dynamically discovering the 'most similar' class in each epoch/batch allows the network to focus on separating the samples from the most similar non-self-class samples, resulting in a more uniform distribution of the representation space. We apply MTHCell to human and mouse scRNA seq datasets from various tissues. Comprehensive and exacting benchmark studies substantiate the exceptional capabilities of MTHCell in cell type annotation, discovery of rare and new celltypes, robustness totraining sample sizes and batch effects, and interpretability of models. Unlike prior single-view pathway-informed Transformers, MTHCell integrates multi-view knowledge through view-level diversity regularization and instance-level dynamic contrastive learning, establishing a new paradigm for interpretable cell-type annotation.