Motif Discovery in Tissue-Specific Regulatory Sequences Using Directed Information

Arvind Rao, Alfred O. Hero Iii, David J. States, James Douglas Engel

TL;DR

This paper introduces a novel method for identifying regulatory DNA sequences that are specific to certain tissues. By leveraging "directed information" and Support Vector Machines, the authors can pinpoint short DNA motifs that act as signatures for tissue-specific gene expression, outperforming traditional methods.

What Problem Does This Paper Address?

Understanding how genes are turned on or off in specific cell types is fundamental to biology, especially for processes like development and disease. While all cells have the same genome, they only express the genes needed for their function. This tissue-specific gene expression is controlled by regulatory elements in the DNA, including promoters and enhancers. A major challenge is to find these regulatory elements genome-wide and understand which ones are responsible for specific tissue expression. Current methods often rely on known transcription factor binding sites, but many regulatory elements might not fit these known patterns, or their function might be more complex. This paper aims to discover novel, potentially unknown, regulatory motifs that can reliably distinguish tissue-specific sequences from general ones.

Methods

The core of this paper lies in a two-step approach:

  1. Feature Selection using Directed Information (DI): Instead of just looking at whether a DNA motif is present or absent (like mutual information), DI quantifies the direction of information flow. The authors hypothesize that certain DNA motifs (specifically, 6-nucleotide sequences called hexamers) induce or cause tissue-specific gene expression. They use DI to measure how much information a hexamer provides about whether a sequence is tissue-specific or not. This is a novel application of DI in bioinformatics for motif discovery.
  2. Classification using Support Vector Machines (SVM): Once the most informative (discriminative) hexamers are identified using DI, an SVM classifier is trained. The SVM uses these selected hexamers as features to learn how to distinguish between tissue-specific and non-tissue-specific sequences. The authors compare their DI-based feature selection to a more traditional mutual information approach, showing DI's superiority.

The process involves:

  • Parsing DNA sequences (promoters and enhancers) to count hexamer occurrences.
  • Normalizing these counts to account for sequence length variations.
  • Using a t-test to initially identify hexamers that show significant differences between tissue-specific and non-tissue-specific groups.
  • Applying DI to rank the top 1000 hexamers based on their directed information content.
  • Using bootstrapping to estimate confidence intervals for the DI values and applying the Benjamini-Hochberg procedure for multiple-testing correction.
  • Training an SVM classifier with the top 'd' ranked hexamers.

Datasets

The study utilizes publicly available datasets:

  • Promoter Data: Microarray gene expression data from the Novartis Foundation (GNF) SymAtlas database was used to identify tissue-specific genes. Promoters for brain-specific and heart-specific genes were extracted.
  • Enhancer Data: The Mouse Enhancer Browser (http://enhancer.lbl.gov) was used to obtain experimentally validated ultraconserved elements. These were categorized into those showing expression in the developing brain and a neutral/background set.

Key Insights and Ideas

  • Directed Information for Causality: The paper highlights that "directed information" is crucial for understanding regulatory relationships. It's not just about correlation (mutual information) but about which sequence features drive or cause tissue-specific expression.
  • Novel Motifs Beyond Known TFBS: The method successfully identifies motifs that are not necessarily known transcription factor binding sites (TFBS). This suggests a broader role for sequence elements in gene regulation, potentially including epigenetic mechanisms or novel binding site configurations.
  • Efficiency of DI: The DI-based feature selection requires fewer motifs (hexamers) to achieve high classification accuracy compared to mutual information, making the resulting models more parsimonious and potentially more robust.
  • Sequence Variability and Specificity: Promoters show higher sequence variability (acting in concert with other elements), while enhancers/LREs are more specific, leading to motifs that are more indicative of tissue-specific expression.

Key Results

  • High Classification Accuracy: The SVM classifier, using DI-selected features, achieved high accuracy in distinguishing tissue-specific from non-tissue-specific sequences. For brain promoters, it reached approximately 93% accuracy. For enhancers, it also showed strong performance, with DI outperforming MI in achieving a given misclassification rate with fewer features.
  • Identification of Novel Motifs: The study identified several novel motifs that discriminate tissue-specific sequences. Some of these matched known transcription factor binding sites (e.g., GATA, MEF2, HNF-4, Pax2), providing experimental validation for the method's ability to find biologically relevant features.
  • DI vs. MI Performance: Figures and tables consistently show that the DI-based feature selection method requires a smaller number of top-ranking features to achieve comparable or better classification accuracy than mutual information.

Limitations and Future Work

  • Enhancer Dataset Specificity: The enhancer dataset used was based on sequences cloned upstream of a heat shock promoter (hsp68-lacz). This means the results might not fully capture promoter-specific enhancer roles, as the experimental setup introduces a non-native promoter.
  • Motif Ranking Variability: The authors note that the ranking of motifs can be highly variable across different data draws, suggesting a need for robust methods to select features. They propose using median DI/MI values across multiple draws for more stable feature selection.
  • Future Directions:
    • Cross-species Alignment: Incorporating multiple sequence alignment of regulatory regions across species could improve feature selection by reducing false positives.
  • Alternative Metrics: Exploring other metrics for distributional divergence between classes.
  • High-Dimensional Entropy Estimation: Developing more consistent high-dimensional entropy estimators, especially for small sample sizes.
  • Stepwise Selection: Formulating a stepwise hexamer selection algorithm that combines DI for relevance and MI for redundancy reduction.
  • Broader Applications: Extending the framework to multiclass problems and integrating it with other data sources (literature, expression data) for a more comprehensive understanding of TF roles.

Summary updated

Abstract

EURASIP Journal on Bioinformatics and Systems Biology is a peer-reviewed open access journal published under the brand SpringerOpen. The overall aim of ...
BibTeX
@article{rao2007motif,
  author = {Arvind Rao and Alfred O. Hero III and David J. States and James Douglas Engel},
  title = {{Motif Discovery in Tissue-Specific Regulatory Sequences Using Directed Information}},
  journal = {EURASIP J. Bioinform. Syst. Biol.},
  year = {2007},
  doi = {10.1155/2007/13853},
}
Figure