Prediction of regulatory networks: genome-wide identification of transcription factor targets from gene expression data

Jiang Qian, C. Jimmy Lin, Nicholas M. Luscombe, Haiyuan Yu, Mark GersteinBioinformatics 2003 · DOI · OpenReview

Introduction

In the post-genomic era, mapping the regulatory network—the "wiring diagram" of a cell—remains a fundamental challenge. Understanding which transcription factors (TFs) regulate specific genes is crucial for deciphering cellular responses. While high-throughput gene expression data is abundant, extracting functional regulatory relationships from it is non-trivial due to the complexity of biological control mechanisms.

Methodology: The SVM Approach

Researchers have developed a supervised machine-learning framework using Support Vector Machines (SVMs) to identify TF-target pairs. Unlike traditional clustering methods, SVMs can detect non-linear and subtle relationships, such as time lags or repression.

The model utilizes a 158-element feature vector, combining the expression levels of a TF and a potential target gene across 79 experimental conditions in Saccharomyces cerevisiae. To address the inherent class imbalance (where most TFs do not regulate most genes), the authors trained the model using a 1:10 ratio of known positive examples to negative controls. A third-order polynomial kernel was employed to map these features into a higher-dimensional space, allowing for the separation of regulatory signals from noise.

The Challenge of Simple Correlation

A common assumption is that a TF and its target should exhibit highly correlated expression. However, biological reality involves time lags, repression (inverse correlation), and complex combinatorial logic that simple statistics often miss.

Figure 4: Distribution of correlation coefficients for known positive (solid line) and negative (dashed line) regulatory pairs. The significant overlap demonstrates why simple correlation is insufficient for predicting regulation.

As shown in the analysis of correlation coefficients, the distributions for positive and negative regulatory pairs overlap significantly. This overlap confirms that "matching wiggles" in expression graphs are not definitive proof of a regulatory link, necessitating more sophisticated pattern recognition tools like SVMs that can account for shifted or inverted responses.

Key Results and Network Insights

The SVM-based approach successfully predicted 3,419 TF-target pairings for 36 different transcription factors in yeast. The model achieved a precision of 63% in cross-validation, with performance comparable to experimental ChIP-chip data.

Figure 6: The predicted yeast regulatory network. The diagram illustrates the high connectivity and delocalized nature of genomic control, with an average of 93 targets per transcription factor.

The resulting network reveals a highly delocalized control structure. On average, each TF regulates approximately 93 targets, and each target gene is influenced by roughly 1.8 TFs. Notably, these targets are spread uniformly across the genome, rather than being clustered near the TF or each other, highlighting a complex, distributed architecture in eukaryotic regulation.

Conclusion

By moving beyond simple correlation, this SVM-based method provides a robust tool for genome-wide regulatory mapping. The ability to identify subtle patterns allows for a more accurate reconstruction of the cellular "wiring diagram." Future integration of these computational predictions with physical binding data (e.g., ChIP-chip) promises to yield even higher-fidelity consensus maps of life's regulatory logic.

Summary updated

BibTeX
@article{qian2003prediction,
  author = {Jiang Qian and C. Jimmy Lin and Nicholas M. Luscombe and Haiyuan Yu and Mark Gerstein},
  title = {{Prediction of regulatory networks: genome-wide identification of transcription factor targets from gene expression data}},
  journal = {Bioinformatics},
  year = {2003},
  doi = {10.1093/bioinformatics/btg347},
}
Figure