Large-scale mapping and validation of Escherichia coli transcriptional regulation from a compendium of expression profiles.

Faith JJ, Hayete B, Thaden JT, Mogno I, Wierzbowski J, Cottarel G, Kasif S, Collins JJ, Gardner TS. · 2007

Read Large-scale mapping and validation of Escherichia coli transcriptional regulation from a compendium of expression profiles. by Faith JJ, Hayete B, Thaden JT, Mogno I, Wierzbowski J, Cottarel G, Kasif S, Collins JJ, Gardner TS. (2007) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc1764438.txt.

Introduction

Algorithms can learn gene regulation from expression data, but genome-scale proof was missing. Faith and co-authors supplied it in 2007 for E. coli, using 445 expression arrays and 3,216 known regulatory links. They tested four algorithm classes and added their own, CLR.

CLR beat the next best method by 36 percent in average precision. At a setting catching 60 percent of true links, it proposed 1,079 interactions, 741 of them novel. Lab tests confirmed 21 new links for three regulators. One found link ties iron transport to central metabolism, confirmed by gene measurements. Faith offers the data set as a benchmark for the field.

Catalog record

CREATORFaith JJ, Hayete B, Thaden JT, Mogno I, Wierzbowski J, Cottarel G, Kasif S, Collins JJ, Gardner TS.
YEAR2007
COLLECTIONPAPERS
FORMATTXT
FILE SIZE66 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC1764438

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Large-Scale Mapping and Validation of Escherichia coli Transcriptional Regulation from a Compendium of Expression Profiles By Jeremiah J Faith, Boris Hayete, Joshua T Thaden, Ilaria Mogno, Jamey Wierzbowski, Guillaume Cottarel, Simon Kasif, James J Collins, Timothy S Gardner, Andre Levchenko Published in: PLoS Biology DOI: 10.1371/journal.pbio.0050008 ABSTRACT Abstract Machine learning approaches offer the potential to systematically identify transcriptional regulatory interactions from a compendium of microarray expression profiles. However, experimental validation of the performance of these methods at the genome scale has remained elusive. Here we assess the global performance of four existing classes of inference algorithms using 445 Escherichia coli Affymetrix arrays and 3,216 known E. coli regulatory interactions from RegulonDB. We also developed and applied the context likelihood of relatedness (CLR) algorithm, a novel extension of the relevance networks class of algorithms. CLR demonstrates an average precision gain of 36% relative to the next-best performing algorithm. At a 60% true positive rate, CLR identifies 1,079 regulatory interactions, of which 338 were in the previously known network and 741 were novel predictions. We tested the predicted interactions for three transcription factors with chromatin immunoprecipitation, confirming 21 novel interactions and verifying our RegulonDB-based performance estimates. CLR also identified a regulatory link providing central metabolic control of iron transport, which we confirmed with real-time quantitative PCR. The compendium of expression data compiled in this study, coupled with RegulonDB, provides a valuable model system for further improvement of network inference algorithms using experimental data. INTRODUCTION High-throughput genome sequencing and bioinformatics technologies have dramatically eased the task of genomic annotation, producing parts lists of living organisms as simple as Mycoplasmas and as complex as mammals. Further progress in the understanding of an organism's biology requires development and refinement of techniques to determine the dynamic interactions among an organism's molecular parts [1]. A major difficulty of this task is the context-specific nature of gene regulation. The total space of possible transcriptional regulatory interactions for an organism is the number of transcription factors multiplied by the number of genes multiplied by the number of environmental contexts in which the cell might find itself. Methods to identify regulatory interactions must efficiently determine the thousands of true regulatory interactions out of the billions of possible ones. Pioneering efforts to identify regulatory interactions on a genome scale have used machine-learning algorithms to identify cis-regulatory motifs or transcription factor target genes using a large set of expression arrays [2–18], genome-wide location analysis chromatin immunoprecipitation (ChIP-Chip) [19,20], or a combination of these and other high-throughput methods [21–26]. In general, the precision of these methods has been evaluated by testing for functional enrichment of co-regulated genes, experimental confirmation of a few selected regulatory relationships, or cross-validation within the training dataset. However, experimental validation of the precision of these methods at the genome scale has remained elusive due to the lack of a model organism with both a known regulatory structure and compatible experimental data. Therefore, the relative merits and broader utility of these approaches remain difficult to judge. Here we demonstrate an unsupervised network inference method, context likelihood of relatedness (CLR), which uses transcriptional profiles of an organism across a diverse set of conditions to systematically determine transcriptional regulatory interactions. We take advantage of the extensive knowledge of transcriptional regulation in Escherichia coli to assess the performance of the CLR algorithm and several other algorithms on a genome scale. In E. coli, a set of 3,216 experimentally confirmed regulatory interactions among 1,211 genes have been curated in the RegulonDB database [27], which can be used for performance assessment. We assembled a compendium of 445 new and previously published E. coli Affymetrix Antisense2 microarray expression profiles collected under various conditions including pH changes, growth phases, antibiotics, heat shock, different media, varying oxygen concentrations, and numerous genetic perturbations (Figure 1 and Table 1). This compendium, combined with the knowledge in RegulonDB, allowed us to assess the genome-scale performance of the CLR algorithm and multiple other unsupervised network inference algorithms applied to experimental data. The compendium and algorithms are available on the Many Microbe Microarrays database (M3D) Web site (http://m3d.bu.edu/). We included several versions of relevance networks (the foundation for the CLR algorithm) [28,29], ARACNe [30], Bayesian networks [7], and regression networks in this algorithm comparison. CLR, the top-performing inference method, predicted 1,079 regulatory interactions at a 60% true positive rate. Sequence analysis of the promoters of the inferred gene targets yielded many known and novel promoter motifs. We also tested, via ChIP, more than 250 of the interactions inferred for three transcription factors at all confidence levels to verify the algorithm precision estimates based on RegulonDB and to confirm 21 of the novel predictions. In addition, an analysis of the CLR inferred network led to the discovery of an unexpected regulatory link between central metabolism and the regulation of iron import into the cell, which we confirmed with real-time quantitative PCR. [Figure: Overview of Our Approach for Mapping the E. coli Transcriptional Regulatory NetworkMicroarray expression profiles were obtained from several investigators. Our laboratory profiled additional conditions, focusing on DNA damage, stress responses, and persistence. These two data sources were combined into one uniformly normalized E. coli microarray compendium that was analyzed with the CLR network inference algorithm. The predicted regulatory network was validated using RegulonDB, sequence analysis, and ChIP. The validated network was then examined for cases of combinatorial regulation, one of which was explored with follow-up real-time quantitative PCR experiments.] [Table: Data Sources for the Escherichia coli Microarray Compendium] RESULTS/DISCUSSION The CLR Algorithm The CLR algorithm is an extension of the relevance networks approach [28,31] for identifying transcriptional regulatory interactions. Although the relevance networks algorithm was written primarily for clustering, its authors also suggest its utility for identifying regulatory networks [29]. The original relevance networks method used mutual information for scoring the similarity between the expression levels of two genes in a set of microarrays. A gene and a transcription factor are predicted to interact if the mutual information between the expression levels of the gene and its potential regulator is above some set threshold. Like correlation, mutual information is a metric that detects statistical dependence between two variables. But unlike correlation, it does not assume linearity, continuity, or other specific properties of the dependence [32,33]. As such, mutual information possesses the flexibility to detect regulatory interactions that might be missed by linear measures such as the correlation coefficient. In the relevance networks algorithm, there are tradeoffs between true positive and false positive rates in choosing a threshold for the identification of significant regulatory interactions. A high threshold results in a smaller network with fewer false positives, but it also eliminates potential novel interactions. Conversely, a low threshold will often capture false positive interactions due to a number of factors, including background correlation and misinterpretation of indirect dependence as direct interaction. The CLR algorithm builds upon the relevance network but applies an adaptive background correction step to eliminate false correlations and indirect influences (Figure 2A). After computing the mutual information between regulators and their potential target genes, CLR calculates the statistical likelihood of each mutual information value within its network context. The algorithm compares the mutual information between a transcription factor/gene pair to the “background” distribution of mutual information scores for all possible

Read the complete text