Evaluation of zero counts to better understand the discrepancies between bulk and single-cell RNA-Seq platforms

Joanna Żyła, Anna Papież, Jun Zhao, Rihao Qu, Xiaotong Li, Yuval Kluger, Joanna Polańska, Christos Hatzis, Lajos Pusztai, Michał MarczykComputational 2023 · DOI

TL;DR

Single-cell RNA sequencing (scRNA-Seq) offers unprecedented resolution but often suffers from "dropout" events (zero counts) and expression discrepancies compared to bulk RNA-Seq. This study identifies key technical factors beyond just sequencing depth, like RNA integrity and gene structure, that contribute to these differences. Understanding these factors is crucial for developing better methods to integrate and interpret data from both bulk and single-cell platforms.

What Problem Does This Paper Address?

Bulk RNA sequencing provides a robust average gene expression profile from a population of cells. However, single-cell RNA sequencing (scRNA-Seq) allows us to dissect heterogeneity within that population, revealing cell-type-specific expression patterns and dynamic cellular states. A major challenge with scRNA-Seq is the high proportion of "zero counts" or "dropout events," where a gene is expressed but not detected. While low sequencing depth is a known contributor, especially for lowly expressed genes, the paper investigates whether other technical and biological factors also drive these zero counts and broader expression discrepancies between bulk and scRNA-Seq. Accurately interpreting scRNA-Seq data requires understanding and potentially correcting for these platform-specific biases.

Methods

The researchers employed a multi-pronged approach:

  • Machine Learning for Dropout Prediction: They trained two machine learning models (XGBoost and Bayesian Regularized Neural Networks - BRNN) to predict the dropout rate in scRNA-Seq data. The models used various technical factors as input, including gene expression from bulk RNA-Seq, RNA integrity (TIN), gene length, transcript number, and mappability. This allowed them to quantify the relative importance of each factor.
  • Non-linear Regression for Expression Shift: A five-parameter logistic curve (5PL) model was used to describe the relationship between bulk RNA-Seq expression and scRNA-Seq dropout rate. Residuals from this model were analyzed to identify genes that were consistently over- or under-expressed between the two platforms.
  • Gene Set Enrichment Analysis (GSEA): GSEA was applied to the residuals from the 5PL model to identify biological pathways that were consistently differentially expressed between bulk and scRNA-Seq across different datasets.
  • Differential Analysis: For one dataset, they performed differential gene expression analysis comparing conditions within each platform (bulk vs. scRNA-Seq) to assess how platform differences might affect biological interpretations.

Datasets

The study utilized four publicly available paired bulk and single-cell RNA-Seq datasets:

  • ABT263: Human triple-negative breast cancer cell line (MDAMB-468) treated with cytotoxic agents.
  • RH41: Human alveolar rhabdomyosarcoma cell line.
  • BM: Human bone marrow mononuclear cells from healthy patients.
  • PR: Prostate specimens from healthy organ donors.

These datasets represent diverse tissues and experimental conditions, allowing for a robust evaluation of cross-platform discrepancies.

Key Insights and Ideas

  • Dropout is More Than Just Depth: While low sequencing depth is a significant driver of zero counts, the study strongly implicates other technical factors. This challenges the common assumption that dropout is solely a consequence of insufficient sequencing.
  • RNA Integrity Matters: The Transcript Integrity Score (TIN), a measure of RNA quality and degradation, emerged as a key factor influencing dropout rates. This suggests that degraded RNA can lead to missed transcript detection, even if the gene is expressed.
  • Gene Architecture Plays a Role: Gene length and the number of transcripts per gene were also identified as important predictors of dropout. This implies that the physical characteristics of a gene can influence its detectability in scRNA-Seq.
  • Consistent Biological Pathways Affected: The study identified specific biological pathways that are consistently discordant between bulk and single-cell platforms. This is crucial for understanding potential misinterpretations when integrating these data types.
  • Technical Factors Drive Pathway Differences: Many of the identified discordant pathways were correlated with the key technical factors (bulk expression, TIN, gene length), suggesting that technical biases, rather than purely biological differences, might be responsible for these observed pathway shifts.

Key Results

  • Technical Factors Identified: Machine learning models identified bulk gene expression and RNA integrity (TIN) as the most important factors influencing dropout rates, followed by gene length and the number of transcripts. Mappability, UTR5' length, and GC content showed less influence.
  • Discordant Genes and Pathways: At the gene level, only 25 genes (0.12%) were consistently discordant across all four datasets. However, at the pathway level, 7 out of 347 pathways (2.02%) showed shared enrichment in discordant genes, indicating a more systemic bias affecting groups of related genes.
  • Under-expressed Pathways: Pathways showing consistent under-expression in scRNA-Seq compared to bulk RNA-Seq were more prevalent and showed stronger shared enrichment across datasets. Examples include "Parathyroid hormone synthesis, secretion and action" and "Chemical carcinogenesis - reactive oxygen species."
  • Impact on Biological Interpretation: Differential analysis on a treated dataset showed that platform differences could potentially lead to misinterpretations of biological responses.

Limitations and Future Work

  • Limited scRNA-Seq Platforms: The study focused on the 10x Genomics Chromium droplet platform. Future work could explore other scRNA-Seq technologies (e.g., full-length transcript methods) to see if these technical factors are equally relevant.
  • Generalizability: While four diverse datasets were used, the authors acknowledge that the generalizability of pathway-level results to other biological contexts might be uncertain.
  • Computational Correction: The findings pave the way for developing novel computational methods to correct for these cross-platform expression shifts, improving the accuracy and reliability of scRNA-Seq data analysis.
  • Further Investigation of Biological Factors: While technical factors were the primary focus, further investigation into biological factors that might contribute to expression discrepancies is warranted.

Summary updated

Abstract

Recent advances in sample preparation and sequencing technology have made it possible to profile the transcriptomes of individual cells using single-cell RNA sequencing (scRNA-Seq). Compared to bulk RNA-Seq data, single-cell data often contain a higher percentage of zero reads, mainly due to lower sequencing depth per cell, which affects mostly measurements of low-expression genes. However, discrepancies between platforms are observed regardless of expression level. Using four paired datasets with multiple samples each, we investigated technical and biological factors that can contribute to this expression shift. Using two separate machine learning models we found that, in addition to expression level, RNA integrity, gene or UTR3 length, and the number of transcripts potentially also influence the occurrence of zeros. These findings could enable the development of novel analytical methods for cross-platform expression shift correction. We also identified genes and biological pathways in our diverse datasets that consistently showed differences when assessed at the single cell versus bulk level to assist in interpreting analysis across transcriptomic platforms. At the gene level, 25 genes (0.12%) were found in all datasets as discordant, but at the pathway level, 7 pathways (2.02%) showed shared enrichment in discordant genes.
BibTeX
@inproceedings{zya2023evaluation,
  author = {Joanna Żyła and Anna Papież and Jun Zhao and Rihao Qu and Xiaotong Li and Yuval Kluger and Joanna Polańska and Christos Hatzis and Lajos Pusztai and Michał Marczyk},
  title = {{Evaluation of zero counts to better understand the discrepancies between bulk and single-cell RNA-Seq platforms}},
  booktitle = {Computational},
  year = {2023},
  doi = {10.1016/j.csbj.2023.09.035},
}
Figure