scCAM: An explainable method to annotate scRNA-seq data and identify annotation-related genes based on residual network and layer class activation maps

Single-cell RNA sequencing (scRNA-seq) has been widely used to explore gene expression and cellular heterogeneity. Cell type annotation is a crucial step in the scRNA-seq data analysis. Recently, several deep learning methods have been developed for cell annotation. However, most existing methods lack biological explainability and fail to discover key genes related to annotation. Therefore, we propose an explainable automatic cell annotation method: scCAM. Our method combines residual networks and layer class activation maps, constructs grayscale images to represent gene expression, and utilizes backward class-specific gradients and the spatial location importance to explore the cell annotation decision-making processes and discover annotation-related genes. We performed experiments on benchmark datasets in multiple situations, the experimental results show that scCAM outperforms other state-of-the-art methods, especially on the large-scale dataset, exceeds other methods by 6.4 %, 18 %, 39.2 %, 16.5 % and 9.8 % on the accuracy, respectively. Explainable analysis on the Segerstolpe pancreas dataset successfully identifies annotation-related genes including marker genes and differentially expressed genes, provides reference and support for the discovery of new marker genes. The source code of scCAM is available at https://github.com/zhangya10956/scCAM-cell-anno.