scSemiAAE: a semi-supervised clustering model for single-cell RNA-seq data

Abstract

Background: Single-cell RNA sequencing (scRNA-seq) strives to capture cellular diversity with higher resolution than bulk RNA sequencing. Clustering analysis is critical to transcriptome research as it allows for further identification and discovery of new cell types. Unsupervised clustering cannot integrate prior knowledge where relevant information is widely available. Purely unsupervised clustering algorithms may not yield biologically interpretable clusters when confronted with the high dimensionality of scRNA-seq data and frequent dropout events, which makes identification of cell types more challenging.

Results: We propose scSemiAAE, a semi-supervised clustering model for scRNA sequence analysis using deep generative neural networks. Specifically, scSemiAAE carefully designs a ZINB adversarial autoencoder-based architecture that inherently integrates adversarial training and semi-supervised modules in the latent space. In a series of experiments on scRNA-seq datasets spanning thousands to tens of thousands of cells, scSemiAAE can significantly improve clustering performance compared to dozens of unsupervised and semi-supervised algorithms, promoting clustering and interpretability of downstream analyses.

Conclusion: scSemiAAE is a Python-based algorithm implemented on the VSCode platform that provides efficient visualization, clustering, and cell type assignment for scRNA-seq data. The tool is available from https://github.com/WHang98/scSemiAAE .

Keywords: Adversarial autoencoder; Clustering; Deep learning; Semi-supervised; scRNA-seq.

Fig. 1

The illustration of scSemiAAE model. A The scRNA-seq count matrix X is preprocessed through gene filtering, screening of highly variable genes, and normalization. Next, it is divided into m_ and m depending on whether it contains true labels. B The encoder receives m_ and m to generate the corresponding latent variables z_ and z, respectively. C The SoftMax layer transforms the latent vector z_ into the pseudo-label c, which is then combined with the partial true label y_ to create a cross-entropy loss. D The decoder reconstructs the potential representation z with a zero-inflated negative binomial loss constraint. E Simultaneously, the latent feature z is fed to the discriminator for adversarial training, comprising the discriminator loss. F After completing training process, all the latent z and labels c are concatenated, and the final clustering results are given by a Gaussian mixture model

Fig. 2

Latent representation visualization. The images base on embedded representations of the 10X_PBMC, Human kidney cells, Worm neuron cells, Human liver and CITE_PBMC datasets. Each dot indicates a cell, and the different colors of the dots point to the predicted labels

Fig. 3

Benchmarking results on real scRNA-seq datasets. Clustering performances of scDeepCluster, scDSC, scDEC, scDHA, SC3, scGAE, scDCC, scAL, Itclust, scSemiAE and scSemiAAE, measured by ACC, NMI and ARI. The first six ones are unsupervised methods, and the remaining ones are semi-supervised clustering algorithms. A Comparison with semi-supervised clustering approaches on three datasets with the top 2000 highly scattered genes. B The results of unsupervised clustering algorithms. C scSemiAAE uses different proportions of labels on seven real datasets, measured by NMI

Fig. 4

Model performance analysis of scSemiAAE. A Comparing the scalability of different algorithms on the real datasets by ARI and NMI metrics. B Clustering effects based on large-scale datasets. C Differential expression analysis bases on Baron (human) data