Zero-preserving imputation of scRNA-seq data using low-rank approximation

George C. Linderman, Jun Zhao, Yuval KlugerbioRxiv 2018 · DOI

TL;DR

This paper introduces ALRA, a novel imputation method for single-cell RNA sequencing (scRNA-seq) data that specifically targets "dropout" events. ALRA effectively imputes technically missing gene expression while crucially preserving biologically zero expression levels, leading to improved downstream analyses and significantly better scalability for large datasets.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) is a powerful tool for understanding cellular heterogeneity, but it suffers from a significant technical challenge known as "dropout." This phenomenon occurs when a gene is actually expressed in a cell but fails to be detected due to limitations in mRNA amplification and sequencing. This results in an abundance of zero values in the gene expression matrix, some of which are "technical zeros" (imputed as zero due to dropout) and others are "biological zeros" (genes that are genuinely not expressed). Existing imputation methods often struggle to distinguish between these two types of zeros, leading to the incorrect imputation of biological zeros, which can distort downstream analyses like cell type classification and gene expression interpretation. Furthermore, many existing methods are computationally expensive and do not scale well to the ever-increasing size of scRNA-seq datasets, which can involve hundreds of thousands to millions of cells.

Methods

The core of ALRA (Adaptively-thresholded Low-Rank Approximation) lies in a clever application of low-rank approximation. The authors hypothesize that the underlying true gene expression matrix is non-negative and low-rank. They compute a low-rank approximation of the observed (sparse) scRNA-seq data using Singular Value Decomposition (SVD). A key observation is that the values corresponding to true zeros in the original matrix become symmetrically distributed around zero in the low-rank approximation. This symmetry allows ALRA to estimate the error distribution of these imputed values.

ALRA then employs an adaptive thresholding strategy: for each gene, it sets to zero any imputed value whose absolute value is smaller than the most negative value in that gene's row. This effectively preserves biological zeros (which are expected to be close to zero or negative in the approximation) while imputing technical zeros (which are expected to be further from zero). The imputed values are then rescaled to match the mean and standard deviation of the original non-zero values for each gene. This approach is designed to be "conservative," prioritizing the preservation of biological zeros even if it means missing some technical zeros.

Datasets

The paper utilizes several datasets for evaluation:

  • Purified Peripheral Blood Mononuclear Cells (PBMCs): A dataset from Zheng et al. (2017) containing 10 purified populations, used to assess ALRA's ability to preserve biological zeros in specific cell types (B cells, NK cells, cytotoxic T cells, and helper T cells) and their known marker genes.
  • Simulated scRNA-seq data: Generated from bulk RNA-seq data of 9 immune populations from the ImmGen consortium. This dataset is crucial for distinguishing true biological zeros from technical zeros, allowing for quantitative evaluation of imputation accuracy.
  • Mouse Visual Cortex Cells: A large dataset from Hrvatin et al. (2018) containing 65,539 cells, used to evaluate ALRA's impact on cell type separation visualized by t-SNE and to assess its scalability.
  • Melanoma Cell Line (Drop-seq and FISH): Data from Torre et al. (2018) comparing Drop-seq measurements with RNA Fluorescence In Situ Hybridization (FISH) for the same cell line. This dataset is used to validate the recovered expression profiles by comparing gene expression variability (Gini coefficient) between technologies.

Key Insights and Ideas

  • The "Symmetric Distribution" Insight: The core idea that imputed zeros from a low-rank approximation of non-negative data exhibit a symmetric distribution around zero is a powerful observation. This allows for a principled way to distinguish between technical and biological zeros.
  • Preserving Biological Truth: ALRA's focus on preserving biological zeros is a critical advantage. Many imputation methods, by treating all zeros as missing, can erroneously "fill in" genes that should genuinely be absent, leading to misleading biological interpretations.
  • Scalability as a Feature, Not a Bug: The paper highlights ALRA's dramatic speed advantage over existing methods, making it practical for the massive datasets now common in single-cell genomics. This opens up imputation for a much wider range of studies.
  • Conservative Imputation: ALRA is described as a "conservative" method. This means it prioritizes not making false positive imputations (i.e., imputing a biological zero) over making false negative imputations (i.e., failing to impute a technical zero). This is a valuable trade-off for many downstream analyses.

Key Results

  • Biological Zero Preservation: ALRA preserves 99% of true biological zeros in simulated data, significantly outperforming MAGIC (47% positive values) and SAVER (incorrectly completes more biological zeros than ALRA).
  • Balanced Accuracy: ALRA achieves a balanced accuracy of 80% for preserving biological zeros and completing technical zeros, compared to 65% for MAGIC and 75% for SAVER.
  • Improved Cell Type Separation: ALRA significantly improves the separation of cell types in t-SNE visualizations, reducing the out-of-bag error of a random forest classifier from 6.3% to 2.6% on a large mouse visual cortex dataset.
  • Enhanced Expression Profile Correlation: ALRA increases the correlation of simulated cells to their true profiles, leading to more accurate classification, whereas MAGIC and SAVER can sometimes worsen these correlations.
  • Exceptional Scalability: ALRA is dramatically faster than MAGIC and SAVER. On a dataset of 50,000 cells, ALRA takes ~40 minutes, while MAGIC requires over 4 hours, and SAVER was not attempted on larger subsets due to its excessive runtime.

Limitations and Future Work

  • Parameter Selection: While ALRA has only one primary parameter (the rank k), the paper acknowledges that the heuristic used for choosing k is based on the "spacings between consecutive singular values" and is a heuristic, though effective. Future work could explore more robust methods for automatic rank selection.
  • Potential for Further Parallelization: The paper notes that the most time-consuming step of ALRA, randomized SVD, can be easily parallelized, suggesting that further performance gains are possible.
  • Adaptive Thresholding for Other Methods: The authors suggest that their adaptive thresholding scheme could potentially be applied as a post-processing step to other low-rank matrix completion (LRMC) methods, especially if their error distributions are also symmetric around zero. This opens up avenues for improving existing LRMC-based imputation techniques.

Summary updated

Abstract

ABSTRACT Single cell RNA-sequencing (scRNA-seq) methods have revolutionized the study of gene expression but are plagued by dropout events, a phenomenon where genes actually expressed in a given cell are incorrectly measured as unexpressed. We present a method based on low-rank approximation which successfully replaces these dropouts (zero expression levels of unobserved expressed genes) by nonzero values, while preserving biologically non-expressed genes (true biological zeros) at zero expression levels. We validate our approach and compare it to two state-of-the-art methods. We show that it recovers true expression of marker genes while preserving biological zeros, increases separation of known cell types and improves correlation of simulated cells to their true profiles. Furthermore, our method is dramatically more scalable, allowing practitioners to quickly and easily recover expression of even the largest scRNA-seq datasets.
BibTeX
@misc{linderman2018zero,
  author = {George C. Linderman and Jun Zhao and Yuval Kluger},
  title = {{Zero-preserving imputation of scRNA-seq data using low-rank approximation}},
  year = {2018},
  doi = {10.1101/397588},
  howpublished = {bioRxiv},
}
Figure