ROSeq: Modeling expression ranks for noise-tolerant differential expression analysis of scRNA-Seq data

Krishan Gupta, Manan Lalit, Aditya Biswas, Ujjwal Maulik, Sanghamitra Bandyopadhyay, Gaurav Ahuja, Abhik Ghosh, Debarka SenguptabioRxiv 2018 · DOI

TL;DR

Single-cell RNA sequencing (scRNA-Seq) generates noisy data with many "dropout" events, making it hard to accurately identify genes that are differentially expressed between cell types. This paper introduces ROSeq, a new method that models gene expression ranks instead of raw counts, making it more robust to noise and dropouts. ROSeq demonstrates strong performance in identifying true differential expression while controlling false positives, even with limited cell numbers.

What Problem Does This Paper Address?

Understanding the intricate workings of biological systems often requires analyzing gene expression at the single-cell level. scRNA-Seq technologies have revolutionized this by allowing us to probe the unique molecular profiles of individual cells within a tissue. However, scRNA-Seq data is notoriously "noisy." This noise arises from various technical challenges like limited starting RNA material, leading to a high proportion of "dropout" events (where a gene is detected at very low or zero levels, even if it's expressed). Accurately identifying genes that are truly different in expression between cell populations (differential expression analysis) is crucial for understanding cell function, disease mechanisms, and developmental processes. Existing methods often struggle with this noise and dropout issue, leading to unreliable results.

Methods

The core innovation of ROSeq lies in its approach to modeling gene expression. Instead of directly using the raw or normalized read counts (which are susceptible to noise and dropouts), ROSeq models the ranks of gene expression. Here's a breakdown:

  • Discretization and Binning: For each gene, expression levels across all cells are first discretized into bins. These bins are then ranked based on the frequency of cells falling into them. The bin with the highest cellular frequency gets the lowest rank.
  • Discrete Generalized Beta Distribution (DGBD): The distribution of these ranks is then modeled using a DGBD. This distribution is chosen for its flexibility in fitting various data shapes.
  • Wald-Type Test: Based on the DGBD model, ROSeq develops a Wald-type test to statistically compare the gene expression ranks between two cell groups. This test is designed to be robust to the inherent noise in scRNA-Seq data.
  • R Package: The authors have implemented ROSeq as an R package, making it accessible to the research community.

Datasets

The authors evaluated ROSeq using three publicly available scRNA-Seq datasets:

  • Trapnell data: Myoblast cells sampled before and after differentiation.
  • Tung data: Induced pluripotent stem cells (iPSCs) from three different individuals.
  • Zheng data: Jurkat cells, used for noise tolerance experiments.

They also created "null datasets" by artificially introducing noise to these datasets to specifically evaluate the Type I error rates (false positives) of different DEG callers.

Key Insights and Ideas

  • Ranks are Robust: The central idea that gene expression ranks are more stable and less affected by noise and dropouts than raw counts is a key insight. This is analogous to how rank-based statistical tests (like Wilcoxon's) are often more robust than parametric tests that rely on specific distribution assumptions.
  • DGBD Fits Well: The DGBD proved to be a good fit for modeling the rank-ordered expression bins across a wide spectrum of genes, suggesting its suitability for this novel approach.
  • Noise Tolerance: ROSeq demonstrates remarkable resilience to various types of noise introduced into the data, outperforming many other methods in controlling false positives.
  • Scalability: The method scales well with increasing sample sizes, making it suitable for the large datasets generated by modern scRNA-Seq technologies.

Key Results

  • Performance Comparison: ROSeq was compared against several established differential expression analysis tools (MAST, BPSC, SCDE, DESeq2, Wilcoxon).
  • In terms of Type I error control (false positives) on null datasets, ROSeq performed competitively, especially with larger cell group sizes (>= 100 cells). SCDE was noted as being very conservative.
  • When benchmarking against bulk RNA-Seq data (considered a more robust ground truth), ROSeq and SCDE achieved the highest accuracy (ROC-AUC values) on most datasets.
  • ROSeq showed superior noise tolerance, particularly when dealing with simulated Gaussian noise, outperforming other methods significantly.
  • Runtime efficiency was good, with ROSeq being among the faster methods, comparable to MAST and Wilcoxon.

Limitations and Future Work

  • Cell Paucity: The authors acknowledge that ROSeq might produce sub-optimal DEG calls if a cell cluster contains a very small number of cells. This is attributed to its reliance on asymptotic distributions. They suggest that a higher number of cells (at least 100 per group) is generally required for optimal performance.
  • Rare Scenarios: In rare cases where cell-population wise expression distributions of a gene are identically shaped but distantly located, ROSeq might miss a DEG status. This is because it analyzes the distribution of rank-ordered bins rather than the marginal distribution of expression.
  • Future Work: The paper highlights ROSeq's promise as a robust and scalable alternative for scRNA-Seq differential expression analysis, especially in noisy datasets. Further validation and application to diverse biological questions are implied. The availability of the R package on Bioconductor and GitHub facilitates its adoption and further development.

Summary updated

Abstract

1 Abstract Systematic delineation of complex biological systems is an ever-challenging and resource-intensive process. Single cell transcriptomics allows us to study cell-to-cell variability in complex tissues at an unprecedented resolution. Accurate modeling of gene expression plays a critical role in the statistical determination of tissue-specific gene expression patterns. In the past few years, considerable efforts have been made to identify appropriate parametric models for single cell expression data. The zero-inflated version of Poisson/Negative Binomial and Log-Normal distributions have emerged as the most popular alternatives due to their ability to accommodate high dropout rates, as commonly observed in single cell data. While the majority of the parametric approaches directly model expression estimates, we explore the potential of modeling expression-ranks, as robust surrogates for transcript abundance. Here we examined the performance of the Discrete Generalized Beta Distribution (DGBD) on real data and devised a Wald-type test for comparing gene expression across two phenotypically divergent groups of single cells. We performed a comprehensive assessment of the proposed method, to understand its advantages as compared to some of the existing best practice approaches. Besides striking a reasonable balance between Type 1 and Type 2 errors, we concluded that ROSeq, the proposed differential expression test is exceptionally robust to expression noise and scales rapidly with increasing sample size. For wider dissemination and adoption of the method, we created an R package called ROSeq, and made it available on the Bioconductor platform.
BibTeX
@misc{gupta2018roseq,
  author = {Krishan Gupta and Manan Lalit and Aditya Biswas and Ujjwal Maulik and Sanghamitra Bandyopadhyay and Gaurav Ahuja and Abhik Ghosh and Debarka Sengupta},
  title = {{ROSeq: Modeling expression ranks for noise-tolerant differential expression analysis of scRNA-Seq data}},
  year = {2018},
  doi = {10.1101/374025},
  howpublished = {bioRxiv},
}
Figure