Footer: a quantitative comparative genomics method for efficient recognition of cis-regulatory elements

Abstract

The search for mammalian DNA regulatory regions poses a challenging problem in computational biology. The short length of the DNA patterns compared with the size of the promoter regions and the degeneracy of the patterns makes their identification difficult. One way to overcome this problem is to use evolutionary information to reduce the number of false-positive predictions. We developed a novel method for pattern identification that compares a pair of putative binding sites in two species (e.g., human and mouse) and assigns two probability scores based on the relative position of the sites in the promoter and their agreement with a known model of binding preferences. We tested the algorithm's ability to predict known binding sites on various promoters. Overall, it exhibited 83% sensitivity and the specificity was 72%, which is a clear improvement over existing methods. Our algorithm also successfully predicted two novel NF-kappaB binding sites in the promoter region of the mouse autotaxin gene (ATX, ENPP2), which we were able to verify by using chromatin immunoprecipitation assay coupled with quantitative real-time PCR.

Figure 1.

AP-1 binding site preferences in different species clusters. The LOGOs for the AP-1 binding preferences from (A) TRANSFAC weight matrix M00199 (includes sites from various organisms like human, mouse, rat, chicken and frog) and (B) human, mouse, and rat. Most of current algorithms are using one of the two variants of weight matrices to scan the promoter regions. (C, D) Further partition of the mammalian sequences into human and rodent, revealing differences in the suboptimal patterns. LOGOs are created by using the program enoLOGOS (Workman et al. 2005) available on the Web http://biodev.hgen.pitt.edu/cgi-bin/enologos/enologos.cgi.

Figure 2.

Performance of Footer in relation to the WAP threshold. This graph presents the sensitivity (black line) and specificity (gray line) over all promoter regions analyzed (see text). According to this graph, Footer performs best on a WAP threshold of 0.05%.

Figure 3.

Sensitivity and specificity performance of Footer in relation to the analyzed promoter length. This graph presents the sensitivity (black line) and specificity (gray line) over all promoter regions analyzed (see text). According to this graph, Footer performance increases with examined promoter length up to 3 kb, while specificity remains essentially constant.

Figure 4.

Real-time PCR data on NF-κB association to one known and three predicted binding sites. The real-time PCR data correspond to the in vivo NF-κB association to known and predicted sites in the promoters of genes iNOS and autotaxin (two sites in each promoter). This graph confirms the direct association of NF-κB with iNOS114 (known site at position -114 from the TSS) and reveals that NF-κB binds on both predicted sites in autotaxin gene (at positions -375 and -596, respectively). The peaks show the maximum fold association compared to non-induced cells (control sample) in the time interval between 30-min and 2-h post-induction. This analysis does not confirm a second predicted iNOS site (at position -2760). All predictions were made by Footer, using the default parameters described in the text.