scCapsNet-mask: an updated version of scCapsNet with extended applicability in functional analysis related to scRNA-seq data

Abstract Background We recently developed the single cell Capsule Network (scCapsNet), an interpretable deep learning classifier for single cell RNA sequencing (scRNA-seq) data based on Capsule Network (CapsNet). Although scCapsNet could identify cell type related genes that determine the classification process, the random association with one-to-many and many-to-one relationships between primary capsules and type capsules adds complexity and difficulty for model interpretation. Results Here we introduce scCapsNet-mask, an updated version of scCapsNet that utilizes a mask to ease the task of model interpretation. To assess the performance of scCapsNet-mask, we conducted experiments on two scRNA-seq datasets. The results of experiments on two scRNA-seq datasets show that scCapsNet-mask could constrain the coupling coefficients, the internal parameters of the model, and make a one-to-one correspondence between the primary capsules and type capsules. Therefore, scCapsNet-mask keeps the virtue of high classification accuracy and high interpretability of the original scCapsNet, and has the advantages of automatic processing and easy interpretation. Furthermore, we show that scCapsNet-mask could extend its applicability in functional analysis. Firstly, scCapsNet-mask could estimate the lineage (fate) bias of cells with less differentiated states. After deducing the fate bias, a pseudo-temporal order of cells could be established for each lineage. Following these pseudo-temporal order, lineage specific genes exhibit a gradual increase expression pattern and HSC associated genes exhibit a gradual decrease expression pattern. Secondly, scCapsNet-mask was applied to the cell type assignment in spatial transcriptomics. Training on scRNA-seq data, the spatial map of predicted cell types generated by scCapsNet-mask model is consistent with that generated by RCTD and the anatomical structure of the mouse hippocampus, with much less time and computing resources. Conclusions scCapsNet-mask source code is freely available at https://github.com/wanglf19/scCapsNet_mask. It is an updated version of scCapsNet to identify cell type associated genes more easily, and can extend its applicability in functional analysis such as fate bias prediction in less differentiated cells and cell type assignment in spatial transcriptomics.