The characteristic direction: a geometrical approach to identify differentially expressed genes

Neil R. Clark, Kevin S. Hu, Axel S. Feldmann, Yan Kou, Edward Y. Chen, Qiaonan Duan, Avi Ma'ayanBMC Bioinformatics 2014 · DOI

TL;DR

This paper introduces a novel geometrical approach called the "Characteristic Direction" to identify differentially expressed genes (DEGs). This method offers improved sensitivity and a more intuitive understanding of gene expression changes compared to traditional univariate methods, particularly in high-dimensional datasets.

What Problem Does This Paper Address?

Identifying differentially expressed genes (DEGs) is a cornerstone of understanding biological differences between conditions, like healthy versus diseased tissues or treated versus untreated cells. Traditional methods often analyze genes one by one (univariate), ignoring the complex interdependencies between gene expression levels. This can lead to missed signals, especially when individual genes don't show strong changes but their collective behavior does. Furthermore, modern genomic studies generate vast amounts of data (high-dimensional), where the number of genes far exceeds the number of samples, posing statistical challenges for many existing methods. The authors aim to provide a more robust and sensitive multivariate approach that can handle this complexity and better capture the subtle, coordinated changes in gene expression.

Methods

The core of this paper is the "Characteristic Direction" method, which repurposes linear classification techniques, specifically Linear Discriminant Analysis (LDA). Instead of directly classifying samples, the authors focus on the orientation of the separating hyperplane found by LDA. This orientation, represented by a vector (the "characteristic direction"), is interpreted as the direction in gene expression space that best characterizes the difference between the two conditions.

Key aspects of the method include:

  • Geometrical Interpretation: The characteristic direction provides an intuitive visual representation of differential expression. The components of this direction vector indicate the significance and direction of change for each gene.
  • Regularization: To address the challenges of high dimensionality and small sample sizes, the method incorporates a regularization scheme (covariance shrinkage) to stabilize estimates and prevent singularity issues.
  • Computational Efficiency: A singular-value decomposition trick is used to make the computation efficient, even with large datasets.
  • Gene Set Enrichment: The geometrical framework is extended to perform gene set enrichment analysis by calculating the "principal angle" between the characteristic direction and the subspace defined by a gene set.

Datasets

The authors validated their method using several datasets:

  • Microarray Data: A large collection of microarray experiments from the Gene Expression Omnibus (GEO) involving transcription factor (TF) perturbations (73 experiments) and drug perturbations (130 experiments).
  • RNA-Seq Data: A study profiling gene expression and STAT3 DNA binding in diffuse large B-cell lymphoma subtypes.
  • Synthetic Data: Generated datasets with controlled parameters to simulate the properties of real biological expression data (high dimensionality, inter-gene dependencies, and known differentially expressed genes).

The paper also leverages existing databases for validation:

  • ChIP-Seq Data: From databases like ChEA and ENCODE, used to identify putative target genes of TFs.
  • Known Drug Targets: Information from databases like DrugBank.
  • Protein-Protein Interaction Data: To identify genes whose protein products interact with TFs or drug targets.

Key Insights and Ideas

  • Multivariate is Better: The paper strongly suggests that considering genes collectively (multivariate) is crucial for accurately identifying differential expression, especially when individual gene changes are subtle.
  • Geometry as a Guide: The geometrical interpretation of differential expression as a single "direction" is a powerful conceptual tool that simplifies understanding and leads to novel applications like improved gene set enrichment.
  • Silver Standard Validation: Using ChIP-Seq and known drug targets as "silver standards" provides a more reliable way to benchmark DEG calling methods than purely simulated data.
  • Consistency Across Technologies: The Characteristic Direction method shows promise for both microarray and RNA-Seq data, adapting to the different data distributions.

Key Results

  • Superior Sensitivity: The Characteristic Direction method consistently outperformed traditional methods (fold change, SAM, LIMMA, t-test, DESeq) in identifying DEGs across various validation datasets, including TF and drug perturbations, and RNA-Seq data.
  • Prioritization of Relevant Genes: The method effectively prioritized genes known to be associated with TF binding sites or drug targets, demonstrating its ability to capture biologically relevant signals.
  • Improved Gene Set Enrichment: The novel gene set enrichment analysis based on the principal angle showed potential for identifying more relevant biological processes compared to GSEA and hypergeometric tests.
  • Robustness: The method demonstrated good performance even with limited sample sizes and high dimensionality, as shown by ROC curves on synthetic data.

Limitations and Future Work

  • Sample Size Requirement: While the method is designed for high-dimensional data, it still benefits from a sufficient number of replicates (at least three are recommended for practical applications) for robust estimation.
  • Further Exploration: The authors suggest that the improved clarity of differential expression provided by their method warrants further exploration to make more categorical statements about biological mechanisms.
  • Accessibility: The authors have made the method freely available through open-source implementations in R, Python, MATLAB, and Mathematica, encouraging wider adoption and application.

Summary updated

BibTeX
@article{clark2014characteristic,
  author = {Neil R. Clark and Kevin S. Hu and Axel S. Feldmann and Yan Kou and Edward Y. Chen and Qiaonan Duan and Avi Ma'ayan},
  title = {{The characteristic direction: a geometrical approach to identify differentially expressed genes}},
  journal = {BMC Bioinformatics},
  year = {2014},
  doi = {10.1186/1471-2105-15-79},
}
Figure