Deep feature extraction of single-cell transcriptomes by generative adversarial network

Abstract

Motivation Single-cell RNA-sequencing (scRNA-seq) offers the opportunity to dissect heterogeneous cellular compositions and interrogate the cell-type-specific gene expression patterns across diverse conditions. However, batch effects such as laboratory conditions and individual-variability hinder their usage in cross-condition designs. Results Here, we present a single-cell Generative Adversarial Network (scGAN) to simultaneously acquire patterns from raw data while minimizing the confounding effect driven by technical artifacts or other factors inherent to the data. Specifically, scGAN models the data likelihood of the raw scRNA-seq counts by projecting each cell onto a latent embedding. Meanwhile, scGAN attempts to minimize the correlation between the latent embeddings and the batch labels across all cells. We demonstrate scGAN on three public scRNA-seq datasets and show that our method confers superior performance over the state-of-the-art methods in forming clusters of known cell types and identifying known psychiatric genes that are associated with major depressive disorder. Availability The scGAN code and the information for the public scRNA-seq datasets are available at https://github.com/li-lab-mcgill/singlecell-deepfeature Supplementary information Supplementary data are available at Bioinformatics online.

Feature selection revisited in the single-cell era

“…These deep learning feature selection models share a common concept of ‘saliency’ which was initially designed for interpreting black-box deep neural networks by highlighting input features that are relevant for the prediction of the model [ 47 ]. Some examples in bioinformatics applications include a deep feature selection model that uses a neural network with a weighted layer to select key input features for the identification and understanding regulatory events [ 48 ]; and a generative adversarial network approach for identifying genes that are associated with major depressive disorders using gradient-based methods [ 49 ]. While feature selection methods that are based on deep learning generally require significantly more computational resources (e.g.…”

“…The application of ensemble and deep learning-based feature selection methods is even sparser in the field. One ensemble feature selection method is EDGE which uses a set of weak learners to vote for important genes from scRNA-seq data [ 121 ], and the current literature on deep learning-based feature selection in single cells are a study for identifying regulatory modules from scRNA-seq data through autoencoder deconvolution [ 122 ]; and another for identifying disease-associated gene from scRNA-seq data using gradient-based methods [ 49 ]. Owing to the non-linear nature of the deep learning models, feature selection methods that are based on deep learning are well-suited to learn complex non-linear relationships among features.…”

deepMNN: Deep Learning-Based Single-Cell RNA Sequencing Data Batch Correction Using Mutual Nearest Neighbors

“…However, scGen was a supervised method that required cell types in advance. scGAN (Bahrami et al, 2020) labeled multiple batches of the input cells that were represented in latent embedding space using a generative adversarial network model.…”

sciCAN: Single-cell chromatin accessibility and gene expression data integration via Cycle-consistent Adversarial Network

“…data 9,[17][18][19][20] . The core component of VAE is the use of reconstruction loss, which encodes a sample in a representation that is drawn from a certain distribution, for example, a Gaussian distribution.…”

“…The use of reconstruction loss also has an advantage of mapping noisy data to high-quality data, which further extends the ability of generative model to de-noise data or impute gene expression. Instead of using VAE to learn representation for single-cell RNA-seq data, two research groups simultaneously modified VAE to address batch effects using an adversarial approach 19,20 . Two methods, named scGAN and AD-AE, respectively, used generative adversarial network (GAN) as the main framework for learning the latent space that is not entangled with batch effects.…”