Predicting the Operon Structure of Bacillus Subtilis Using Operon Length, Intergene Distance, and Gene Expression Information

Abstract

We predict the operon structure of the Bacillus subtilis genome using the average operon length, the distance between genes in base pairs, and the similarity in gene expression measured in time course and gene disruptant experiments. By expressing the operon prediction for each method as a Bayesian probability, we are able to combine the four prediction methods into a Bayesian classifier in a statistically rigorous manner. The discriminant value for the Bayesian classifier can be chosen by considering the associated cost of misclassifying an operon or a nonoperon gene pair. For equal costs, an overall accuracy of 88.7% was found in a leaveone-out analysis for the joint Bayesian classifier, whereas the individual information sources yielded accuracies of 58.1%, 83.1%, 77.3%, and 71.8% respectively. The predicted operon structure based on the joint Bayesian classifier is available from the DBTBS database (http://dbtbs.hgc.jp).

Operon Prediction for Sequenced Bacterial Genomes without Experimental Information

“…On the chromosome of B. anthracis, which contains 5,308 protein-coding genes, the algorithm predicted a total of 2,473 cooperonic gene pairs when we used a prediction probability cutoff of 0.5. These pairs form 1,121 multigene operons that contain between 2 and 32 genes, and the probability distribution of operon length is remarkably similar (Pearson's correlation, 0.9917) to that reported for Bacillus subtilis in a recent study (7). Although there are very few experimentally verified operons to use in testing the predictions, we note that the gene pairs within the operons that have been verified (i.e., plcCspmC, csaAB, rsbVW-sigB, asbABCDEF, and gerHABC) were all predicted successfully.…”

“…Examples of the former include methods that rely on microarray-based expression data (3,4,7,26) and others that use different forms of detailed functional annotation (5,30,35). Although these algorithms have shown great promise in terms of being able to predict operon structure with a high degree of specificity and sensitivity, the data they rely on are only available for a select subset of bacterial species, and this limits how widely they can be used.…”

“…Although these data have all proven to be valid predictors of operon structure, it is striking that these studies have also consistently demonstrated that one of the most valuable predictors is simply intergenic distance. The distances between genes within an operon tend to be considerably shorter than the distances between genes that are not cotranscribed, and in several recently developed algorithms, intergenic distance was shown to be more informative than any other data source, including even microarray-based expression data (3,7,28,34). In addition, this trend appears to be universal in bacterial genomes, making it a very attractive option for a generalized, portable prediction algorithm (7,15,33).…”

“…The distances between genes within an operon tend to be considerably shorter than the distances between genes that are not cotranscribed, and in several recently developed algorithms, intergenic distance was shown to be more informative than any other data source, including even microarray-based expression data (3,7,28,34). In addition, this trend appears to be universal in bacterial genomes, making it a very attractive option for a generalized, portable prediction algorithm (7,15,33). Unfortunately, intergenic distance alone only allows for a specificity of ϳ65 to 70% when tested on a large set of experimentally verified operons from within the Escherichia coli genome, and so other sources of information must be added to bring the total accuracy to a more acceptable level (28).…”

Prediction of Transcriptional Terminators in Bacillus subtilis and Related Species

“…Third, the operon prediction will be reliable only if the fraction of Rho-terminated operons is sufficiently small. We note that previous attempts at terminator prediction in B. subtilis [20] and Synechococcus sp. WH8102 [21,22] for the purpose of operon prediction were unsuccessful.…”