Model-based deep embedding for constrained clustering analysis of single cell RNA-seq data

Abstract

Clustering is a critical step in single cell-based studies. Most existing methods support unsupervised clustering without the a priori exploitation of any domain knowledge. When confronted by the high dimensionality and pervasive dropout events of scRNA-Seq data, purely unsupervised clustering methods may not produce biologically interpretable clusters, which complicates cell type assignment. In such cases, the only recourse is for the user to manually and repeatedly tweak clustering parameters until acceptable clusters are found. Consequently, the path to obtaining biologically meaningful clusters can be ad hoc and laborious. Here we report a principled clustering method named scDCC, that integrates domain knowledge into the clustering step. Experiments on various scRNA-seq datasets from thousands to tens of thousands of cells show that scDCC can significantly improve clustering performance, facilitating the interpretability of clusters and downstream analyses, such as cell type assignment.

Self-supervised contrastive learning for integrative single cell RNA-seq data analysis

“…The existing commonly used methods to perform the above processing are based on different backbone algorithms and assumptions. The earliest methods utilize the traditional dimension reduction algorithms, such as Principal Component Analysis (PCA), followed by k-means or hierarchical clustering to group cells 5,[10][11][12][13][14][15] . Although these methods are widely used, their assumption, that is, the complex single-cell transcriptomics can be accurately mapped onto a low-dimensional space by a generalized linear model, may not be necessarily justified 8 .…”

“…However, the time and space complexity of such methods impede the broad applications of the methods 5 .…”

“…The choice of k for the widely used k-nearest-neighbors graph affects the size and number of final clusters 7 . Because of the model capacity and scalability of deep learning methods, almost all the recently developed methods are based on antoencoder 5,9,[20][21][22][23][24][25] (AE) or variational autoencoder 8,26,27 (VAE), which can also incorporate the biostatistical models 28,29 seamlessly. However, as AE and VAE methods are unsupervised learning methods, it is very difficult to control and decide what the deep learning models will learn, although some very recent studies try to impose constraints and our prior knowledge about the problem onto the low-dimensional space 5,27 .…”

“…Because of the model capacity and scalability of deep learning methods, almost all the recently developed methods are based on antoencoder 5,9,[20][21][22][23][24][25] (AE) or variational autoencoder 8,26,27 (VAE), which can also incorporate the biostatistical models 28,29 seamlessly. However, as AE and VAE methods are unsupervised learning methods, it is very difficult to control and decide what the deep learning models will learn, although some very recent studies try to impose constraints and our prior knowledge about the problem onto the low-dimensional space 5,27 . Researchers have also tried to utilize manual labeling as supervision for training the models, accompanied by transfer learning 22 or meta-learning 30 , but such methods encounter scalability issues and have strong assumptions on the homogeneity of different datasets, making them less popular than the above methods.…”

“…It has been facilitating researchers to investigate several critical biomedical topics, such as cancer 3 and autoimmunity 4 . Despite its promises, the unique properties of the scRNA-seq data, such as extreme sparsity and high variability 5 , have posed a number of computational challenges to researchers 6, 7 . To analyze the data, among all the steps 7 , the key processing is to obtain a reliable low-dimension representation for each cell, which can preserve the biological signature of the cell for downstream analysis while eliminating technical noise 8, 9 .…”