Single-Cell RNA Sequencing Data Clustering Based On Deep Transfer Learning

Single-cell transcriptome sequencing(scRNA-seq)technology,which measures gene expression at the resolution of a single cell,provides an opportunity to analyze the heterogeneity within tissues.The study of heterogeneity within tissues can lead to the discovery of new cell types,the study of the complex trajectories of cell differentiation and development,and the improvement of understanding of diseases such as human tumors.In order to more accurately analyze the heterogeneity within tissues,we need to identify the cell types contained within tissues.Faced with the sequencing data of thousands or even millions of cells generated in an experiment,the most mainstream method at present is to identify cell types by clustering the given scRNA-seq data.Although many scholars have proposed clustering algorithms for scRNA-seq data,on the one hand,facing the clustering problem of large-sample,high-dimensional,high-sparse and unsupervised scRNA-seq data,traditional machines Learning algorithms cannot accurately describe the complex non-linear mapping relationship between gene expression and cell type;on the other hand,existing algorithms focus on identifying cell types based on given scRNA-seq data,and they do not utilize the currently published reference maps with reliable cell type annotations to help cluster the target datasets.Therefore,based on deep learning and transfer learning methods,we proposed a single cell clustering algorithm based on deep transfer learning.On the one hand,the algorithm uses a deep coding model to mine the complex non-linear relationship between cell type and gene expression.On the other hand,through the idea of supervised pre-training in transfer learning,the accuracy of cell type recognition in the target dataset can be improved by using the reference atlas with cell type annotations.In this paper,a set of experiments of mutual transfer learning between five islet tissue data sets that are cross-sequencing protocols,cross-laboratory,cross-species,and the data set size and cell type complexity are different,covered with different levels of batch differences.The prediction effect of the algorithm in this paper proves that the method in this paper can effectively prevent negative migration even when there is a large batch difference between the target data set and the reference atlas.At the same time,the algorithm in this paper is compared with three traditional clustering algorithms and three deep clustering algorithms in visualization effects and clustering evaluation indicators,which verifies that this algorithm is stronger than existing algorithms.