scRMD: Imputation for single cell RNA-seq data via robust matrix decomposition

Chong Chen, Changjing Wu, Linjie Wu, Yishu Wang, Minghua Deng, Ruibin XibioRxiv 2018 · DOI

TL;DR

Single-cell RNA sequencing (scRNA-seq) data is plagued by "dropouts" – missing gene expression values that hinder downstream analysis. This paper introduces scRMD, a novel imputation method that leverages robust matrix decomposition to accurately recover these missing values. scRMD significantly improves the performance of crucial analyses like differential gene expression and cell clustering, making scRNA-seq data more reliable.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) is a powerful technology for understanding cellular heterogeneity and biological processes at an unprecedented resolution. However, a major technical challenge in scRNA-seq is the "dropout" phenomenon. This occurs when a gene is actually expressed in a cell, but the low amount of RNA and inefficiencies in the experimental process lead to it not being detected. This results in a large number of zero values in the gene expression matrix, some of which are "true zeros" (the gene is genuinely not expressed) and others are "dropout zeros" (the gene is expressed but missed). These dropout zeros introduce noise, reduce the statistical power of analyses, and can obscure important biological relationships between genes and cells. Therefore, accurately imputing these missing values is crucial for reliable downstream analysis.

Methods

The core of scRMD lies in modeling the dropout imputation problem as robust matrix decomposition. The authors propose that the true underlying gene expression matrix ($X$) can be decomposed into a low-rank matrix ($L$) representing the shared expression patterns across cell types, and a sparse matrix ($S$) representing the dropout events, plus some noise ($E_1$). This can be expressed as $Y = L - S + E_1$, where $Y$ is the observed (noisy and incomplete) expression matrix.

The method then aims to estimate $L$ and $S$ by minimizing an objective function that includes penalties for both low-rankness (using the nuclear norm for $L$) and sparsity (using the L1 norm for $S$). A key innovation is the incorporation of prior knowledge: if a gene expression value in the observed data ($Y_{ij}$) is above a certain threshold ($c$), it's highly unlikely to be a dropout, so the corresponding entry in $S$ ($S_{ij}$) is constrained to be zero. This leads to the optimization problem:

$$ \min_{L,S} ||Y - L + S||_F^2 + \lambda ||L||_* + \gamma ||S||_1 $$

subject to constraints related to the dropout events and non-negativity of expression. An efficient Alternating Direction Method of Multipliers (ADMM) algorithm is developed to solve this optimization problem. The tuning parameters $\lambda$ and $\gamma$ are selected using a random-matrix-theory-based approach, aiming to avoid computationally expensive grid searches. The imputed data is then generated as $Y + P_{\Omega>0}(L)$, where $P_{\Omega>0}$ selectively imputes values in the candidate dropout set.

Datasets

The paper evaluates scRMD using several real-world scRNA-seq datasets:

  • Ziegenhain scRNA-seq data: Used to assess imputation accuracy using spike-in synthetic RNA transcripts (ERCC).
  • Chu data (human embryonic stem cells): Used for differential expression (DE) analysis, with bulk RNA-seq data serving as a "golden standard."
  • Kowalczyk data (hematopoietic stem cells): Also used for DE analysis, employing down-sampling strategies to assess reproducibility.
  • Usoskin data, Pollen data, Ting data, Deng data: Used for clustering analysis to evaluate the impact of imputation on cell type identification.
  • Alles data, Karaiskos data, Hrvatin data: Ultra-high-throughput scRNA-seq datasets used to assess performance on large-scale data with potentially higher zero rates.

The paper also describes simulation studies with varying parameters (e.g., dropout rates, noise levels) to systematically compare scRMD with other methods.

Key Insights and Ideas

  • Dropout as a Sparse Corruption: The paper frames the dropout problem not just as missing data, but as a form of sparse corruption on an underlying low-rank structure, which is a powerful mathematical perspective.
  • Robust Matrix Decomposition is Key: By combining low-rankness and sparsity penalties, scRMD effectively disentangles true biological signal from technical noise and dropout events.
  • Leveraging Prior Information: The constraint that observed non-zero values are unlikely to be dropouts is a crucial piece of domain knowledge that significantly improves imputation accuracy.
  • Efficient Parameter Selection: Using random matrix theory for parameter tuning is an elegant way to avoid computationally intensive search procedures, making the method more practical.
  • Imputation Enhances Downstream Analysis: The most significant insight is that accurate imputation isn't just an end in itself; it demonstrably improves the reliability and performance of critical downstream analyses like DE and clustering.

Key Results

  • Simulation Studies:
    • scRMD achieved comparable or superior performance to the "oracle" (perfect knowledge) in DE analysis and outperformed scImpute and MAGIC.
  • It demonstrated the most accurate imputation results in terms of mean imputation error.
  • Clustering analysis showed scRMD consistently outperformed scImpute and MAGIC, and was comparable to the oracle.
  • scRMD was significantly faster than scImpute (about 10x) while being only slightly slower than MAGIC.
  • Real Data Analysis:
    • Imputation Accuracy: scRMD showed improved Pearson correlations with known ERCC gene concentrations compared to scImpute in several datasets.
  • Differential Expression Analysis: scRMD consistently improved DE analysis results, showing higher overlap with gold-standard gene sets and more concordant gene rankings with bulk data. It was particularly effective for genes with high dropout rates.
  • Clustering Analysis: scRMD improved clustering performance, especially on perturbed datasets (simulating higher dropout rates), leading to better separation and tighter clustering of cell types. It often outperformed other imputation methods when dropouts were more prevalent.
  • Ultra-high-throughput Data: scRMD provided better visualization and clustering results on large-scale datasets, leading to tighter clusters and better separation of cell types.

Limitations and Future Work

  • Threshold Selection: The paper acknowledges that the choice of the threshold $c$ for identifying candidate dropout events could be further optimized, potentially through bootstrapping to measure imputation stability.
  • Computational Cost: While faster than scImpute, scRMD is not as fast as MAGIC, which is a simple matrix multiplication. Future work could explore further optimizations for speed.
  • Model Assumptions: The method relies on the assumptions of low-rankness and sparsity. While these are generally reasonable for scRNA-seq data, their validity might vary across different biological contexts.

The authors express confidence that scRMD will find broad applications in various scRNA-seq studies due to its robust performance and ability to improve downstream analyses.

Summary updated

Abstract

Abstract Motivation Single cell RNA-sequencing (scRNA-seq) technology enables whole transcriptome profiling at single cell resolution and holds great promises in many biological and medical applications. Nevertheless, scRNA-seq often fails to capture expressed genes, leading to the prominent dropout problem. These dropouts cause many problems in down-stream analysis, such as significant noise increase, power loss in differential expression analysis and obscuring of gene-to-gene or cell-to-cell relationship. Imputation of these dropout values thus becomes an essential step in scRNA-seq data analysis. Results In this paper, we model the dropout imputation problem as robust matrix decomposition. This model has minimal assumptions and allows us to develop a computational efficient imputation method scRMD. Extensive data analysis shows that scRMD can accurately recover the dropout values and help to improve downstream analysis such as differential expression analysis and clustering analysis. Contact [email protected]
BibTeX
@misc{chen2018scrmd,
  author = {Chong Chen and Changjing Wu and Linjie Wu and Yishu Wang and Minghua Deng and Ruibin Xi},
  title = {{scRMD: Imputation for single cell RNA-seq data via robust matrix decomposition}},
  year = {2018},
  doi = {10.1101/459404},
  howpublished = {bioRxiv},
}
Figure