scRecover: Discriminating true and false zeros in single-cell RNA-seq data for imputation

Zhun Miao, Jiaqi Li, Xuegong ZhangbioRxiv 2019 · DOI

TL;DR

Single-cell RNA sequencing (scRNA-seq) data is plagued by "dropout" zeros, which are technical artifacts rather than true biological zeros. This paper introduces scRecover, a novel method that accurately distinguishes true zeros from dropout zeros and selectively imputes only the latter. By doing so, scRecover improves the accuracy of downstream analyses like cell clustering and visualization, offering a more biologically realistic representation of gene expression.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) is a powerful technology for understanding cellular heterogeneity, but it suffers from a significant technical challenge: an abundance of zero values in gene expression matrices. These zeros arise from two distinct sources: true biological zeros, where a gene is genuinely not expressed in a cell, and "dropout" zeros, which are artifacts caused by technical limitations like low mRNA capture efficiency. Existing imputation methods often treat all zeros equally, imputing both true and dropout zeros. This can lead to "over-imputation," introducing noise and potentially distorting biological signals, especially when dealing with genes that are truly not expressed. The paper aims to address this by developing a method that can differentiate between these two types of zeros and only impute the dropout zeros, thereby preserving the biological information of true zeros.

Methods

The core of scRecover lies in its ability to distinguish true zeros from dropout zeros. This is achieved through a two-pronged approach:

  1. Estimating Dropout Probabilities: The method models gene expression using a zero-inflated negative binomial (ZINB) distribution. This statistical framework allows for the estimation of the probability that a zero observation for a specific gene in a specific cell is a "dropout" zero versus a "true" zero.
  2. Estimating Expressed Gene Number: To further refine the identification of dropout zeros, scRecover estimates the number of truly expressed genes in each cell. This is done by adapting concepts from species accumulation curves and employing the Good-Toulmin series and its rational function approximation (as implemented in the preseqR package). By comparing the observed number of expressed genes with the estimated total number of expressed genes, the method can infer the number of dropout zeros.

Once dropout zeros are identified, scRecover acts as a "modest" imputation layer. It can be integrated with existing imputation methods like scImpute, SAVER, and MAGIC. Instead of imputing all zeros, scRecover guides these methods to selectively impute only the identified dropout zeros, leaving the true zeros untouched.

Datasets

The paper uses the following datasets for evaluation:

  • Down-sampling data: A public scRNA-seq dataset of human preimplantation embryonic cells (E-MTAB-3929) was used. This dataset was down-sampled to simulate dropout scenarios, allowing for a controlled evaluation of scRecover's performance against the original, high-coverage data.
  • Real datasets:
    • 1k heart cells from an E18 mouse (available from 10x genomics).
  • Human pluripotent stem cells dataset (GSE75748).

Key Insights and Ideas

  • The importance of distinguishing zero types: The paper highlights that not all zeros are created equal in scRNA-seq data. Treating true biological zeros as technical artifacts can lead to significant downstream analysis errors.
  • "Modest" imputation as a key advantage: By selectively imputing only dropout zeros, scRecover avoids over-imputation and preserves the biological signal of truly unexpressed genes. This is crucial for accurate downstream analyses.
  • Improved downstream analysis: The integration of scRecover with existing imputation methods demonstrably improves the separation of cell clusters and enhances visualization, suggesting a more biologically accurate representation of the data.
  • Biological relevance: The improved clustering and separation of cell types, particularly in distinguishing subtle differences between cell lines (e.g., H1 vs. H9), indicates that scRecover's approach aligns better with biological reality.

Key Results

  • Improved zero prediction accuracy: scRecover, when combined with intermediate imputation methods (SAVER, scImpute, MAGIC), showed at least a 20% improvement in zero-prediction accuracy compared to using those methods alone, especially for SAVER and MAGIC. This improvement became more significant with increasing down-sampling rates.
  • Accurate dropout zero number prediction: Results from down-sampling experiments showed that scRecover's dropout zero number predictions were much closer to the reference results, even with increased prediction depth.
  • Enhanced cell clustering: On real datasets, scRecover processing led to better separation of cell clusters, as evidenced by increased Adjusted Rand Index (ARI) and Jaccard index values. For instance, in the human pluripotent stem cells dataset, scRecover significantly improved the separation of H1 and H9 cell clusters.
  • Reduced imputed genes: Figure 4 shows that scRecover reduces the number of imputed genes compared to methods that impute all zeros, aligning with the concept of "modest" imputation.

Limitations and Future Work

  • Computational cost: While not explicitly detailed as a limitation, the multi-step process of scRecover (ZINB modeling, preseqR estimation, and then integration with other imputers) might imply a higher computational cost compared to single-step imputation methods.
  • Parameter tuning: The performance of ZINB models and species accumulation curves can be sensitive to parameter choices, although the paper does not elaborate on potential challenges in this regard.
  • Future work: The authors express hope that scRecover will provide a "new angle for recovering single cell dropout events and make more contributions on downstream analyses," suggesting that further exploration and application of the method in various biological contexts are anticipated.

Summary updated

Abstract

Abstract High-throughput single-cell RNA-seq (scRNA-seq) data contains excess zero values, including those of genes not expressed in the cell, and those produced due to dropout events. Existing imputation methods do not distinguish these two types of zeros. We present a modest imputation method scRecover to only impute the dropout zeros. It estimates the zero dropout probability of each gene in each cell, and predicts the number of truly expressed genes in the cell. scRecover is combined with other imputation methods like scImpute, SAVER and MAGIC to fulfil the imputation. Down-sampling experiments show that it recovers dropout zeros with higher accuracy and avoids over-imputing true zero values. Experiments on real data illustrate scRecover improves downstream analysis and visualization.
BibTeX
@misc{miao2019screcover,
  author = {Zhun Miao and Jiaqi Li and Xuegong Zhang},
  title = {{scRecover: Discriminating true and false zeros in single-cell RNA-seq data for imputation}},
  year = {2019},
  doi = {10.1101/665323},
  howpublished = {bioRxiv},
}
Figure