G2S3: A gene graph-based imputation method for single-cell RNA sequencing data

Weimiao Wu, Yunqing Liu, Qile Dai, Xiting Yan, Zuoheng WangPLOS Computational Biology 2021 · DOI

TL;DR

Single-cell RNA sequencing (scRNA-seq) data is notoriously sparse due to "dropout" events, obscuring biological insights. This paper introduces G2S3, a novel gene graph-based imputation method that effectively recovers missing gene expression data by leveraging relationships between genes. G2S3 outperforms existing methods across various downstream analyses and is computationally efficient for large datasets.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) allows us to study gene expression at an unprecedented resolution, revealing cell heterogeneity, developmental trajectories, and disease mechanisms. However, a major technical hurdle is the high rate of "dropout" events, where a gene is detected as having zero expression even if it is present. This leads to sparse and noisy data, making downstream analyses like cell clustering, differential gene expression, and trajectory inference challenging and potentially inaccurate. Existing imputation methods often struggle to balance noise reduction with the preservation of biological variability or are computationally expensive for the massive datasets generated by modern scRNA-seq technologies.

Methods

G2S3 tackles the dropout problem by constructing a "gene graph" that represents relationships between genes. This graph is built using graph signal processing, assuming that genes with similar expression patterns across cells are "connected." The core idea is to "borrow" information from these connected genes to impute missing values. Specifically, G2S3 learns a sparse adjacency matrix (W) for the gene graph by optimizing an objective function that balances smoothness (genes with similar expression should be connected) and sparsity (avoiding overly dense connections). Once the graph is established, a lazy random walk is performed on it to impute gene expression values. The method is designed to be computationally efficient, especially for large numbers of genes, by focusing on gene-gene relationships rather than cell-cell similarities, which tend to scale poorly with the number of cells.

Datasets

The authors evaluated G2S3 and ten other imputation methods on eight publicly available scRNA-seq datasets. These datasets were chosen to cover various aspects of scRNA-seq analysis:

  • Expression Data Recovery: Reyfman (lung tissue), PBMC (peripheral blood mononuclear cells), and Zeisel (mouse cortex and hippocampus) datasets were down-sampled to assess how well methods recover true expression levels.
  • Cell Subtype Separation: Chu (human embryonic stem cells) dataset was used to evaluate the ability of imputation to improve the distinction between different cell types.
  • Cell Trajectory Inference: Petropoulos (human preimplantation embryos) dataset was used to assess how imputation affects the reconstruction of developmental trajectories.
  • Differential Gene Identification: Chu and Trapnell (differentiating human myoblasts) datasets were used to evaluate performance in identifying differentially expressed genes.
  • Gene-Gene Relationship Recovery: Paul (myeloid progenitors) and Buettner (mouse ES cells) datasets were used to assess the ability to restore gene regulatory and correlation relationships.

Key Insights and Ideas

  • Gene-centric imputation is key: By focusing on relationships between genes rather than between cells, G2S3 can better preserve biological variability and avoid over-smoothing that can occur with cell-similarity-based methods.
  • Graph signal processing for biological networks: Applying graph signal processing to gene expression data allows for a principled way to learn gene relationships that are robust to noise and sparsity.
  • Sparsity is crucial for gene networks: The assumption that gene networks are sparse is biologically relevant and helps G2S3 build a more interpretable and effective gene graph.
  • Computational efficiency for large-scale data: G2S3's gene-centric approach makes it significantly faster than cell-similarity-based methods when dealing with datasets containing tens or hundreds of thousands of cells.

Key Results

  • Superior performance across tasks: G2S3 consistently outperformed or matched other methods in expression data recovery, cell subtype separation, cell trajectory inference, differential gene identification, and gene-gene relationship recovery.
  • Expression data recovery: G2S3 achieved the highest gene-wise and cell-wise correlations with reference data in down-sampled experiments, significantly outperforming most other methods.
  • Cell subtype separation: G2S3 and an ensemble method (EnImpute) showed the best improvement in separating cell subtypes, with G2S3 being the only single method to achieve comparable performance to the ensemble.
  • Differential expression: G2S3 achieved the highest Area Under the Curve (AUC) for detecting differentially expressed genes in two datasets, indicating improved accuracy.
  • Gene-gene relationships: G2S3 demonstrated strong performance in recovering both gene regulatory networks and pairwise gene correlations, particularly for negatively correlated gene pairs, which are often challenging to impute.
  • Computational efficiency: G2S3 was on average twenty times faster than SAVER (another gene network-based method) and significantly faster than cell-similarity-based methods, making it suitable for large datasets.

Limitations and Future Work

  • Hyperparameter tuning: While G2S3 uses an MSE-based criterion for hyperparameter tuning, the authors suggest that a theoretical approach might yield even better results.
  • Subject effect: G2S3 currently does not explicitly account for "subject effects" (e.g., batch effects from different individuals or experimental batches), which can be important in multi-subject datasets. Future work could incorporate subject effects as a "batch" effect to be removed.
  • Dropout rate consideration: G2S3 imputes all values at once. Future improvements could involve calculating the probability of a dropout event for each observed count and selectively imputing only those with high dropout probability.
  • Weighting schemes: The current imputation weights depend only on gene similarity. Future work could explore incorporating read count reliability, cell library size, and gene expression dispersion into these weights, similar to SAVER.
  • Subject and disease effects: For datasets with both subject and disease effects, G2S3 could be extended to consider both simultaneously.

Summary updated

BibTeX
@article{wu2021g,
  author = {Weimiao Wu and Yunqing Liu and Qile Dai and Xiting Yan and Zuoheng Wang},
  title = {{G2S3: A gene graph-based imputation method for single-cell RNA sequencing data}},
  journal = {PLOS Computational Biology},
  year = {2021},
  doi = {10.1371/journal.pcbi.1009029},
}
Figure