IntaRNA: efficient prediction of bacterial sRNA targets incorporating target site accessibility and seed regions

Abstract

Motivation: During the last few years, several new small regulatory RNAs (sRNAs) have been discovered in bacteria. Most of them act as post-transcriptional regulators by base pairing to a target mRNA, causing translational repression or activation, or mRNA degradation. Numerous sRNAs have already been identified, but the number of experimentally verified targets is considerably lower. Consequently, computational target prediction is in great demand. Many existing target prediction programs neglect the accessibility of target sites and the existence of a seed, while other approaches are either specialized to certain types of RNAs or too slow for genome-wide searches.

Results: We introduce INTARNA, a new general and fast approach to the prediction of RNA-RNA interactions incorporating accessibility of target sites as well as the existence of a user-definable seed. We successfully applied INTARNA to the prediction of bacterial sRNA targets and determined the exact locations of the interactions with a higher accuracy than competing programs.

Availability: http://www.bioinf.uni-freiburg.de/Software/

Fig. 1.

Interpretation of the matrix C^k,l(i,j).

Fig. 2.

Visualization of the recursion for C^k,l(i,j). The hybridized part is shown in red, while the energy required to make the mRNA and the sRNA target site accessible (ED) is given in blue and green, respectively. Since ED-values are not additive, e.g. ED(i,k)≠ED(i,p)+ED(p,k), we need to substract ED(p,k) and ED(q,l), and add ED(i,k) and ED(j,l) to get the final result of C^k,l(i,j).

Fig. 3.

C^k,l_seed(i,j) matrix and its relation to the other matrices. Note that both C^k,l(p,q) and seed(i,j,p,q;5) consistently assume that (p,q) is a pair. Here, a seed of 5 bp and one unpaired base is shown.

Fig. 4.

Performance of INTARNA. (a) Comparison of INTARNA and other leading methods in the prediction of targets on our test set of 18 sRNAs with experimentally verified targets. The sensitivity (true positive rate) is shown as a function of the false positive rate (1−specificity). For each prediction method, the target candidates for each sRNA were sorted by energy score. Each ROC curve was generated from the rate of true and false predictions, while varying the number of considered interactions per sRNA. (b) Comparison of resource requirements of INTARNA (including computation of ED-values) and RNAUP for a GcvB target search in Salmonella. Without restricting the interaction length, RNAUP uses up the available complete memory and, as a consequence, crashes.