Graph Contrastive Learning as a Versatile Foundation for Advanced scRNA-seq Data Analysis

Zhenhao Zhang, Yuxi Liu, Meichen Xiao, Kun Wang, Yu Huang, Jiang Bian, Ruolin Yang, Fuyi Li

TL;DR

This paper introduces scSimGCL, a novel framework that leverages graph contrastive learning for self-supervised pretraining of graph neural networks. This approach effectively generates high-quality cell representations from scRNA-seq data, leading to improved cell clustering and data imputation, and offers a versatile foundation for various downstream analyses.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity by providing gene expression profiles at the individual cell level. A key task in scRNA-seq analysis is cell clustering, which groups cells with similar expression patterns to identify distinct cell types or states. However, raw scRNA-seq data presents significant challenges, including high dimensionality and a high proportion of "dropout" events (where genes appear unexpressed due to technical limitations rather than biological reality). While deep learning methods have shown promise in addressing these issues, there's a need for a simple yet effective framework that can consistently learn robust, high-quality cell representations essential for accurate clustering. Existing methods often struggle to balance the complexities of graph construction, contrastive learning, and data imputation in a unified manner.

Methods

The core of scSimGCL is its graph contrastive learning (GCL) paradigm, applied in a self-supervised manner. The framework consists of several key components:

  • Cell-Cell Graph Structure Learning: A multi-head attention module is used to construct a cell-cell graph from the scRNA-seq data. This graph represents cells as nodes and their biological associations as edges, aiming to capture nuanced relationships.
  • Data Augmentation for Contrastive Learning: To create diverse "views" of the graph for contrastive learning, the framework employs two augmentation techniques:
  • Edge Dropping: Randomly removing edges from the learned cell-cell graph.
  • Gene Masking: Randomly masking gene expression values within cells.
  • Graph Neural Network (GNN) Encoder: A GNN-based encoder is used to generate initial cell representations from both the original and augmented graph views.
  • Projector: These representations are then projected into a lower-dimensional latent space.
  • Contrastive Loss: The contrastive learning objective aims to maximize the mutual information between representations from different views. It encourages similar representations for "positive pairs" (e.g., connected nodes, anchor and its counterpart) and dissimilar representations for "negative pairs."
  • Joint Learning Strategy: Crucially, scSimGCL integrates contrastive learning with data imputation using a mean absolute error loss. This joint strategy allows the model to simultaneously learn better representations for clustering and improve the quality of the gene expression data itself.

Datasets

The study evaluated scSimGCL on a combination of:

  • Ten real scRNA-seq datasets: These include datasets from various tissues and sequencing platforms, such as Shekhar mouse retina cells, Baron, 10X PBMC, Camp, Mouse bladder cells, Zeisel, and Tabula Sapiens datasets.
  • Twelve simulated scRNA-seq datasets: These were generated using the Splatter toolkit to specifically test the model's performance under varying conditions, including:
  • Different dropout rates.
  • Varying strengths of low signals for clustering.
  • High dimensionality with extreme sparsity.

Key Insights and Ideas

  • Homophily is Key: The paper emphasizes that preserving the intrinsic "homophily" of the graph (where connected nodes share similar properties) is crucial for effective contrastive learning. Directly integrating pre-processed graphs into contrastive learning can disrupt this, leading to inferior representations. scSimGCL's approach to graph augmentation and pair construction aims to maintain this homophily.
  • Unsupervised Pretraining for Robust Representations: By framing the problem as self-supervised pretraining, scSimGCL learns generalizable cell representations without requiring labeled data. This makes it a versatile foundation for various downstream tasks.
  • Synergy of Imputation and Clustering: The joint learning of contrastive loss and imputation loss is a significant contribution. It suggests that improving the data quality (imputation) and learning discriminative representations (clustering) are not independent tasks but can mutually benefit each other.
  • Versatility Beyond Clustering: The framework's ability to generate high-quality representations means it can be readily integrated with existing clustering algorithms (like K-Means and Affinity Propagation), offering flexibility to users.

Key Results

  • Superior Clustering Performance: scSimGCL consistently outperformed state-of-the-art methods in terms of clustering accuracy (CA), normalized mutual information (NMI), and adjusted rand index (ARI) across multiple real and simulated datasets. For instance, it showed significant improvement over K-Means.
  • Improved Data Imputation: The model also demonstrated competitive performance in scRNA-seq data imputation, measured by Pearson correlation coefficients (PCCs) and L1 distance, often achieving comparable or better results than specialized imputation methods.
  • Robustness and Flexibility: Ablation studies confirmed the efficacy of the GCL framework and the joint learning strategy. Hyperparameter analysis showed the model's robustness across various settings. The framework's ability to integrate with different clustering algorithms highlights its versatility.
  • Visualization Insights: UMAP visualizations showed that scSimGCL produced more distinct and visible clusters compared to several baseline methods.

Limitations and Future Work

  • Lack of Uncertainty Estimation: The authors acknowledge a limitation in the current framework: the absence of uncertainty estimation in model decisions. This "over-confidence problem" in deep learning models warrants caution regarding the generalizability of findings.
  • Batch Effects: The paper notes that batch effects in scRNA-seq data can lead to false conclusions. While scSimGCL is not explicitly designed to handle batch effects, this is identified as an area for future research.
  • Extension to Other Data Modalities: The authors suggest exploring the potential of scSimGCL on scATAC-seq datasets, where dropout effects are even more pronounced, to assess its effectiveness and robustness in such scenarios.
  • Improving Transparency: Future work will focus on improving the transparency and reliability of scSimGCL, potentially through interpretable analytical approaches.

Summary updated

Abstract

Abstract Single-cell RNA sequencing (scRNA-seq) offers unprecedented insights into transcriptome-wide gene expression at the single-cell level. Cell clustering has been long established in the analysis of scRNA-seq data to identify the groups of cells with similar expression profiles. However, cell clustering is technically challenging, as raw scRNA-seq data have various analytical issues, including high dimensionality and dropout values. Existing research has developed deep learning models, such as graph machine learning models and contrastive learning-based models, for cell clustering using scRNA-seq data and has summarized the unsupervised learning of cell clustering into a human-interpretable format. While advances in cell clustering have been profound, we are no closer to finding a simple yet effective framework for learning high-quality representations necessary for robust clustering. In this study, we propose scSimGCL, a novel framework based on the graph contrastive learning paradigm for self-supervised pretraining of graph neural networks. This framework facilitates the generation of high-quality representations crucial for cell clustering. Our scSimGCL incorporates cell-cell graph structure and contrastive learning to enhance the performance of cell clustering. Extensive experimental results on simulated and real scRNA-seq datasets suggest the superiority of the proposed scSimGCL. Moreover, clustering assignment analysis confirms the general applicability of scSimGCL, including state-of-the-art clustering algorithms. Further, ablation study and hyperparameter analysis suggest the efficacy of our network architecture with the robustness of decisions in the self-supervised learning setting. The proposed scSimGCL can serve as a robust framework for practitioners developing tools for cell clustering. The source code of scSimGCL is publicly available at https://github.com/zhangzh1328/scSimGCL .
BibTeX
@article{zhang2024graph,
  author = {Zhenhao Zhang and Yuxi Liu and Meichen Xiao and Kun Wang and Yu Huang and Jiang Bian and Ruolin Yang and Fuyi Li},
  title = {{Graph Contrastive Learning as a Versatile Foundation for Advanced scRNA-seq Data Analysis}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae558},
}
Figure