Single-Cell Capsule Attention : an interpretable method of cell type classification for single-cell RNA-sequencing data

Tianxu Wang, Xiuli Ma

Keywords: Single-Cell RNA-sequencing, cell type classification, capsule network, attention, interpretable model

Abstract: Single-cell RNA-sequencing technique can obtain genes’ expression level of every cell. Cell type classification (also known as cell type annotation) on single-cell RNA-seq data helps to explore cellular heterogeneity and diversity. Previous methods for cell type classification are either based on statistical hypotheses of gene expression or deep neural networks. However, the hypotheses may not reflect the true expression level. Deep neural networks lack interpretation for the result. Here we present an interpretable neural-network based method single-cell capsule attention(scCA) which assigns cells to different cell types based on their different feature patterns. In our model, we first generate capsules which extract different features of the cells. Then we obtain compound features which combine a set of features’ information through a LSTM model. In the end, we train attention weights and apply them to the compound features. scCA provides a strong interpretation for cell type classification result. Cells from the same cell type share a similar pattern of capsules’ relationship and similar distribution of attention weights for compound features. Compared with previous methods for cell type classification on nine datasets, scCA shows high accuracy on all datasets with robustness and reliable interpretation.

One-sentence Summary: We present a interpretable cell type classification method single-cell capsule attention(scCA) which assigns cells to different cell types based on their different feature patterns for single-cell sequencing data.