scLEGA: an attention-based deep clustering method with a tendency for low expression of genes on single-cell RNA-seq data

Zhenze Liu, Yingjian Liang, Guohua Wang, Tianjiao ZhangBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces scLEGA, a novel deep learning method for clustering single-cell RNA sequencing (scRNA-seq) data. It addresses the common issue of ignoring lowly expressed genes, which can be crucial for cell type identification, by incorporating a specialized loss function and an attention mechanism that fuses denoising and topological information. scLEGA demonstrates superior performance in accuracy, scalability, and stability compared to existing methods across multiple datasets.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study cellular heterogeneity. A fundamental step in analyzing scRNA-seq data is identifying different cell types, which is crucial for understanding biological processes, disease mechanisms, and developmental trajectories. Many existing computational methods for cell type inference rely on clustering algorithms. However, a significant challenge is that these methods often focus on highly variable genes (HVGs) with high expression levels, potentially overlooking the importance of HVGs with lower expression. These lowly expressed genes can act as critical biomarkers for rare cell types or subtle cellular states, and their exclusion can lead to incomplete or inaccurate cell type classifications. This paper aims to bridge this gap by developing a method that effectively leverages both high and low expression genes for more robust cell type identification.

Methods

The core of scLEGA is a deep learning architecture that combines two key strategies:

  • Low-expression optimized Denoising Autoencoder (DAE): This component is designed to handle the sparsity and "dropout" events common in scRNA-seq data. It uses a novel Zero-Inflated Negative Binomial (ZINB) loss function. This modified ZINB loss introduces two hyperparameters, "tailor rate" and "loss rate," to specifically adjust the weighting of lowly expressed genes, ensuring they contribute more significantly to the learned representations.
  • Graph Convolutional Network (GCN)-based Graph Autoencoder (GAE): This part captures the topological structure of the data by constructing a cell-cell graph. It uses neighbor information to guide dimensionality reduction, preserving the relationships between cells.

These two components are integrated using a multi-head attention mechanism. This mechanism allows the model to learn from both the denoised features from the DAE and the topological embeddings from the GAE simultaneously, creating a more comprehensive and cluster-friendly representation of the cells. The final clustering is performed using an unsupervised algorithm that adaptively selects cluster centers, notably employing the Leiden algorithm, which does not require prior knowledge of the number of cell types.

Datasets

The authors evaluated scLEGA on 15 scRNA-seq datasets from various organs and platforms. They also used a subset of these datasets for hyperparameter tuning. The paper does not introduce new benchmark datasets but rather uses existing ones to compare scLEGA against 12 other state-of-the-art methods.

Key Insights and Ideas

  • Lowly expressed genes matter: The paper strongly advocates for the importance of including lowly expressed genes in cell type inference. Their findings suggest that these genes can capture subtle but crucial biological differences that high-expression genes might miss.
  • Attention is key for fusion: The multi-head attention mechanism is crucial for effectively merging the information from the denoising autoencoder and the graph autoencoder, leading to richer cell representations.
  • Adaptive clustering is powerful: Using the Leiden algorithm for adaptive cluster center selection removes the need to pre-specify the number of cell types, making the method more flexible and applicable to diverse datasets.
  • Robustness to noise: The DAE component, especially with the modified ZINB loss, demonstrates a strong ability to handle dropout events and noise inherent in scRNA-seq data.

Key Results

  • Superior Clustering Accuracy: scLEGA achieved superior performance in clustering accuracy, as measured by Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI), across 11 out of 15 datasets compared to 12 other methods.
  • Improved Average Scores: The average ARI score improved by 9.5%-147.9%, and the average NMI score increased by 7.2%-42.1% compared to other methods.
  • Scalability and Stability: The method demonstrated good scalability and stability, performing well even on larger datasets and showing robustness to manual dropout simulations.
  • Effective Visualization: UMAP visualizations of latent features showed that scLEGA better distinguishes cell clusters with clearer boundaries compared to other methods.
  • Ablation Study Success: Removing individual components of scLEGA (attention, GAE, DAE, or modified ZINB loss) led to a performance decline, validating the contribution of each part.

Limitations and Future Work

  • Computational Cost: While scLEGA shows improved performance, the multi-head attention mechanism and GAE components increase computational time and resource consumption.
  • Imputation for Noise Reduction: The authors plan to embed an imputation mechanism within scLEGA in the future to further clean raw data and reduce noise, especially for datasets with high sparsity.
  • Further Research: The paper suggests that deep learning frameworks like scLEGA will continue to be researched to enhance performance, reduce resource consumption, and improve time efficiency for scRNA-seq data analysis.

Summary updated

BibTeX
@article{liu2024sclega,
  author = {Zhenze Liu and Yingjian Liang and Guohua Wang and Tianjiao Zhang},
  title = {{scLEGA: an attention-based deep clustering method with a tendency for low expression of genes on single-cell RNA-seq data}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae371},
}
Figure