Multi-level multi-view network based on structural contrastive learning for scRNA-seq data clustering

Zhenqiu Shu, Min Xia, Kaiwen Tan, Yongbing Zhang, Zhengtao YuBriefings in Bioinformatics 2024 · OpenReview

TL;DR

This paper introduces scMMN, a novel multi-level, multi-view deep learning framework for clustering single-cell RNA sequencing (scRNA-seq) data. By integrating shallow and deep feature representations with structural consistency contrastive learning, scMMN significantly improves clustering accuracy compared to existing methods, offering a more robust approach to understanding cellular heterogeneity.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) technology has revolutionized our understanding of cellular biology, but analyzing the resulting data presents significant challenges. scRNA-seq data is notoriously high-dimensional and complex, making it difficult to accurately identify and group similar cells (clustering). Traditional clustering methods often struggle with this complexity, and even existing deep learning approaches can be limited by relying on a single perspective or "view" of the data. This paper aims to overcome these limitations by proposing a method that leverages multiple views of the data at different levels of abstraction, combined with a novel contrastive learning strategy to enhance both feature and structural representations.

Methods

The core of the proposed method, scMMN, lies in its multi-level, multi-view construction and a unique contrastive learning strategy.

  1. Multi-Level Multi-View Construction:
    • Shallow Views: Two initial "shallow" views of the data are created using the k-Nearest Neighbors (kNN) algorithm and Diffusion Mapping (DM). These methods capture linear and non-linear relationships between cells, respectively, by constructing different adjacency matrices.
  2. Deep Views: To further enrich the representations, "deep" views are generated by applying Graph Laplacian Filters (GLFs) to the shallow views. This process creates four distinct views in total (two shallow, two deep, each derived from both kNN and DM initializations).
  3. Representation Learning with Contrastive Learning:
    • Graph Convolutional Networks (GCNs): Each of the four views is fed into a two-layer GCN to learn latent representations of the cells.
  4. Group Contrastive Learning: To ensure that cells within the same cluster are close and those in different clusters are far apart, a group contrastive loss is applied. This loss is computed on two groups of these learned representations, encouraging consistency across different views.
  5. Structural Consistency Contrastive Learning: A novel contrastive loss is introduced to explicitly learn and enforce structural consistency between the different views. This component focuses on aligning the relationships between cells as captured by the various views.
  6. Feature Fusion and Clustering:
    • The learned representations from the different views are fused in a two-step process: inter-group fusion followed by intra-group fusion.
  7. A Multi-Layer Perceptron (MLP) is then used to learn a common representation from these fused features.
  8. Finally, the k-means algorithm is applied to this common representation to obtain the final cell clusters.

Datasets

The authors evaluated their method on eight real-world scRNA-seq datasets: Klein, Young, Trachea, Diaphragm, Muscle, Bladder, Pollen, and Wang_Lung. These datasets vary in size and biological origin, providing a comprehensive benchmark for performance evaluation. No new datasets were introduced.

Key Insights and Ideas

  • The Power of Multi-View and Multi-Level: The paper strongly advocates for the benefit of looking at scRNA-seq data from multiple perspectives (views) and at different levels of abstraction (shallow vs. deep features). This comprehensive approach seems crucial for capturing the complex underlying cellular structures.
  • Structure Matters: Explicitly incorporating structural consistency into the contrastive learning framework is a key innovation. This suggests that not only the gene expression patterns but also the relationships between cells are vital for accurate clustering.
  • Robustness to Noise: The method's ability to achieve good results across diverse datasets, including those with high gene counts and potential noise, highlights its robustness.

Key Results

The scMMN method demonstrated superior performance across all four evaluation metrics (ACC, NMI, ARI, FMI) on six out of the eight scRNA-seq datasets. On the remaining two datasets, it achieved optimal values for three out of the four metrics. This consistent outperformance suggests the method's effectiveness and generalizability. Ablation studies also confirmed the significant contribution of each component (GCNs, feature fusion, structural contrastive learning, and multi-view strategy) to the overall performance.

Limitations and Future Work

The authors acknowledge that their current method primarily relies on the gene expression matrix. Future work could involve integrating other biological information, such as gene regulatory relationships and gene functions, to further enhance clustering performance. They also suggest exploring models that can integrate these diverse data types for a more holistic understanding of cellular states.

Summary updated

BibTeX
@article{shu2024multi,
  author = {Zhenqiu Shu and Min Xia and Kaiwen Tan and Yongbing Zhang and Zhengtao Yu},
  title = {{Multi-level multi-view network based on structural contrastive learning for scRNA-seq data clustering}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  url = {https://openreview.net/forum?id=xkJK5ZF8Yr},
}
Figure