The first step in building Delila-Genome was to port the Delila individual information programs to the Linux platform. The Delila software library is distributed by the National Cancer Institute as binaries for the Sun Sparc system. Source code written in Pascal was translated to C using p2c and debugged.

The main components of the server are the scan (from Delila), promotsite, scandiff and genvis programs. The server module generally runs directly on top of the Delila system however it can be run using a reduced set of Delila binaries. Besides the scan program, the only Delila programs required by Delila-Genome are lister, mkdb, and dbbk (for displaying sequence walkers). The Delila-Genome server programs are described below.

Scan evaluates the strength (in bits) of each binding site and reports those sites whose strength (R_i) lies within a user defined range 5. The parameters for scan are defined in the front-end Java program. The minimum threshold R_i value (R_{i, minimum}) is set at or above zero bits. Genome scans with an R_i(b, l) matrix derived from a limited number of binding sites, n = 50 can significantly contribute to Type 1 errors (false positive detection of weak binding sites). To decrease the source of this error, R_{i, minimum} is generally set to the R_i value of the weakest binding site used to compute the weight matrix. Alternatively, sites whose Z scores or probabilities of the binding strengths fall within a user-defined range may be selected. The user also specifies which portion of the individual information matrix is scanned and which strand to evaluate (positive, negative or both). Scan can output data (locations and strengths of sites), scanfeatures (features for display with lister) and scaninst (instructions for extracting sites as Delila book files) files for each chromosome, however, only the data file is required as input to the promotsite program. Each record in the data file contains the R_i values of all predicted binding sites in the genome, their respective coordinates, the Z scores of these R_i values and their corresponding probabilities. Scan has numerous other features, the details of which are presented in 18. The Z score for user-defined matrices is based upon the mean of the distribution of scores derived from these matrices. The mean is determined first by simulating a set of binding sites based upon this weight matrix (with the ridi program 18) and then computing R_sequence from a book of sequences containing these sites with the encode, dalvec and rseq programs (eg. 19).

Promotsite was developed to filter the output produced by scan, since these results may potentially contain large numbers of potential binding sites (>>10⁶), many of which are distant from expressed sequences. Promotsite prunes the data file produced by scan and reports only relevant sites which are within or adjacent to expressed genomic templates. The user defines a search window either upstream or downstream (or both) relative to the beginning genomic coordinate (often the transcriptional initiation site) of each gene. The upstream and downstream window lengths may be specified independently. Promotsite modifies the data file format produced by scan so that the associated GenBank accession number is appended to the record containing the binding site (psdataop file). Typical analyses of splice sites within human coding regions selected sites up to 1 Mb downstream of the transcription initiation site in order to ensure that even the longest genes would be encompassed by these searches. We have limited the analyses of promoters to a 10 kb interval upstream (in some cases, downstream) of the transcription initiation site. However, these parameters should be set (and subsequently optimized) based upon previous experimental or published binding site studies for specific factors. For example, to comprehensively detect insulator elements bound by the protein CTCF, this window has been specified bi-directionally and increased in length (to 50 kb; not shown).

Since a site may, in some instances, fall within the search window of multiple mRNAs, the mRNA whose start position is closest to the binding site coordinate is assigned to be the associated mRNA for that site. The list of reported binding sites may also be pruned based on a range of chromosomal coordinates and by specifying particular chromosomes. Promotsite also defines a parameter known as the paralog distance. Since the same mRNA sequence may be mapped based upon its similarity to multiple genomic locations, paralogous genes on the same chromosome designated with the same mRNA accession number were distinguished from large genes containing multiple widely-dispersed exons by defining a parameter for the minimum distance between paralogous loci. Binding sites separated by less than the paralog distance are labeled with the same GenBank accession number and are considered part of the same gene, whereas sites exceeding this distance were assumed to be derived from different genes that were similar to the same GenBank accession. Typically, we set the paralog distance to 10⁵ or 10⁶ bp, depending upon the lengths and density of genes or gene families thought to contain relevant binding sites. Using the associated mRNA for each site, promotsite creates a BED-formatted file that can be uploaded as a custom track on the UCSC human genome browser http://genome.ucsc.edu.

The execution time of scan depends on the length of the chromosome and the nucleotide length, l, of the R_i(b, l) weight matrix that defines the binding site. For hardware platforms with multiple computational nodes, the server can distribute scan and promotsite runs for each chromosome between these nodes so that the execution time over the whole genome is minimized. As l is constant over the whole genome, this load-balancing is based upon the length of each chromosome. Since execution times are generally several hours, the server informs the user of job completion by email.

Relevant binding sites identified with promotsite or scandiff (see below) can be viewed with the genvis program. Like these programs, genvis also uses Javascript to generate HTML pages that display the binding site list extracted from the BED files. Since, in some instances, too many sites may be produced by promotsite and scandiff for browser uploading, genvis offers several options to select subsets of binding sites from a chromosome or genome scan. Groups of sites may be extracted by writing subsets of the BED files specified either by genomic strand, the chromosomal coordinates, or a list of accession numbers corresponding to mRNAs mapped onto the genome sequence.

Inaccuracies in the genome draft coordinates of splice junction recognition sites motivated the development of an automated strategy to select correctly localized splice sites. Information weight matrices were iteratively recomputed from the set of sites with positive R_i values 6. More recently, we have built models of transcription factor binding sites by cyclical refinement of weight matrices based on published data from established regulated gene targets, supplemented with binding sites in these genes predicted by information theory and experimentally validated 78. With Delila-Genome, potential novel binding sites identified can be verified in the laboratory and included in subsequent refinements of the weight matrix.

Previous approaches for comparing information weight matrices have involved determining the Euclidean or positional distances between related R_i(b, l) matrices 1020. Comparisons of the results of successive genome scans offer an alternative approach for monitoring the progress of weight matrix refinement. The scandiff program computes model-to-model changes in information at experimentally-proven and predicted binding sites by scanning the same genome sequence with two different information weight matrices. This enables the user to monitor genome-wide sensitivity and specificity of binding site prediction. The psdataop output file generated by promotsite is the input to the scandiff program. The output files generated by scandiff categorize binding sites based upon their identification of unique sets of sites by each of the matrices (models A and B; columns A-B and B-A; Table 1), and sites detected with both weight matrices that show differences in information content (columns A ∩ B; in Table 1). Scandiff can display differences in binding strength at the same coordinate based upon either exceeding thresholds of absolute changes in R_i (ΔR_i), changes in their respective Z scores (ΔZ) or distinct confidence intervals computed from each of the R_i(b, l) matrices 11.

The criteria of measuring changes in binding site strength is dictated by the stage of model refinement (see below). Absolute comparisons of R_i values are not as meaningful at early stages of refinement, since addition of experimentally-defined binding sites to an information model can substantially alter the distribution of R_i values of the binding sites that underlie these weight matrices. At early stages of refinement, the information models are based on fewer binding sites, resulting in larger confidence intervals for individual R_i values. Comparisons of R_i values based upon the sizes of confidence intervals are therefore not as reliable measure of significant change in information as changes in their respective Z scores.

Upon model convergence, the proportion of sites in successive models with significant differences in information content should be quite small (S/ [S+I] (S = significant, I = insignificant) for confidence intervals of = 3 S.D. The proportion of sites common to both models relative to discordant sites found in only one model ([S+I] / [A-B] + [B-A]), should stabilize as successive versions of the information weight matrix are refined.

Scandiff generates BED-formatted files and data files similar in format to that produced by promotsite from the identified and categorized binding sites. We used the following color shading convention for the different types of binding sites. The sites with significant changes in R_i are shaded gray; sites identified only by scanning the first matrix are shaded brown; and sites found only with the second matrix are shaded blue. An example of this output is shown in Figure 3, which indicates the results for PXR/RXRα models 1 and 2 in the vicinity of the CYP3A4 gene.

Table 2 indicates the execution times of complete genome scans for various types of binding sites on two different Linux hardware platforms: a Beowulf cluster of three dual 1.1 Ghz CPU nodes running the Scyld operating system and a Mosix cluster of 24 single processor 500 Mhz nodes. Due to limitations in disk storage, Scyld Beowulf cluster was used to genome scans with PXR/RXRα matrix only. The execution times given in Table 2 represent combined results of running of both scan on the genome sequence and promotsite on the results of the scan program. The execution time for both programs depends upon the length of the binding site, R_sequence of the weight matrix, and R_i,minimum (specified by the user). The length of the site contributes to the CPU time, and the last two factors contribute to the I/O access time. From the table, we can see that for successive models of PXR/RXRα, R_sequence decreases, and consequentially, the number of sites predicted, increases. Additional novel sites that are predicted by information analysis and validated by laboratory testing are introduced with each successive model. The additional sites in the model account for the decrease in R_sequence, and the increase in the number of predicted sites in the genome. R_sequence decreases from 17 bits to 14.9 bits from models 2 to 3, and there is a steep rise (more than a 2 fold increase) in the number of sites. With the addition of (somewhat weaker) binding sites to model 3, this resultant matrix is less biased towards the consensus sequence, resulting in a large genome-wide increase in predicted sites. The median execution times in the Mosix cluster were approximately 6.5 hrs and 3.5 hrs for the Scyld cluster for all PXR/RXRα models, despite an increase of 3.5 fold in the number of sites from models 1 to 4. The effect of increased I/O access time on the total execution time is evident in the case of the SR protein SC35 site (which has a low R_sequence value of 3.64 bits), where the run time is 19 hours due to 76-fold increase in the quantity of sites predicted compared with the scan of PXR/RXRα Model 4.

Analysis of the splice acceptor and donor runs required a modification of the published genome sequence. In the original genome drafts, a very large number of binding sites (>>10⁸) were initially found. Many of these sites were composed of long runs of undefined polynucleotides (ie. = N₍₁₀₎) in heterochromatin and in gaps in the draft sequence. The Delila program defaults to adenine in these cases, and in the case of splice acceptor sites, these substitutions generated sites comprised of polyadenine, which itself has an R_i value exceeding the user-defined threshold (2.4 bits; R_{i, minimum}). These runs exceeded our available disk storage, and to reduce the quantity of false positive sites, we generated and substituted random nucleotides for every sequence of undefined polynucleotides = 10 bp in length. Our previous studies have shown that sequence randomization produces fewer than 2% of binding sites with R_i values above zero bits 6, and none above the minimum R_i threshold value 11. The genome scans of the substituted genome sequences with splice donor and acceptor R_i(b, l) weight matrices were completed in 10.5 and 14.5 hours, respectively.

We have found that uploads of large BED files of binding sites to the remote UCSC genome browser can be time-consuming and sometimes fail. The BED file for all binding sites found with PXR/RXRα Model 4, for example, is ~ 30 MB and required 5–10 minutes to upload. Furthermore, the large numbers of sites found with some information weight matrices (eg. splice donor and acceptor sites; 254 Mb for acceptor sites on chromosome 1 alone) produce BED file sizes exceeding browser/server limits. We therefore created and viewed subsets of binding sites for genomic regions of specific interest with the genvis tool.

Figure 2 depicts the HTML page generated by genvis, containing a partial list of binding sites on chromosome 1 for the PXR/RXRα model 4 R_i(b, l) weight matrix. The websites linked to this page are also shown (but have been resized or truncated) to reflect only the important details of each. When the HTML page is initially loaded, a window for the UCSC browser pops up. The BED file is uploaded using a button in this window upon selecting the appropriate version of the genome draft at the UCSC website. When the genome browser target links (entries in the R_i, Seq and UCSC Browser columns) are activated, the genome browser displays the information based on this uploaded file.

The second row of the HTML table in Figure 2 corresponds to the binding site associated with the GenBank Accession L13278. This is a strong binding site (R_i value of ~ 20.1 bits) which is hyper-linked to the custom track detail in the genome browser. This track detail page indicates the size of the site and the orientation of the recognition sequence on the draft genome sequence. The user can obtain the DNA sequence of the site either from from the Seq cell in the HTML table or from the corresponding custom track detail. The pop up sequence walker indicates the relative contributions of each nucleotide in the site 5.

The linked GenBank and SOURCE database entries indicate that accession L13278 encodes the zeta-crystallin/quinone reductase gene. We selected this example to illustrate that Delila-Genome can be used to potentially discover novel transcriptional regulatory targets, since this gene has not been previously demonstrated to be regulated by PXR/RXRα. The SOURCE entry is based on a dynamic collection and compilation of gene data from many scientific databases associated with the GenBank accession, whereas the GenBank entry, in some instances, is not curated and guaranteed only to contain the corresponding sequence. The SOURCE entry also indicates other information such as the aliases for the gene name, the locus link designation, expression profile, etc.

The UCSC genome browser entry displays the binding site custom track and sequences in the proximity of the associated GenBank accession. The coordinates delineate a display window concordant with the search window defined in promotsite for generating the list of binding sites given in the HTML page. In Figure 2, the predicted site is 1112 bp upstream of L13278 and ~ 7.2 kb upstream of an as yet uncharacterized gene corresponding to both AK098237 and BC009514. Although we cannot exclude the possibility that this site regulates the gene encoded by AK098237/BC009514, its closer proximity to the zeta-crystallin gene and the common orientation of both the site and gene on the antisense strand suggests that this site may function as a potential transcriptional enhancer element. There are no other predicted binding sites in the vicinity of this gene.

The results of genome scans with successive refinements of PXR/RXRα information weight matrices were compared using scandiff. The refinement procedure was validated by detecting binding sites in well-established PXR/RXRα target genes. Initial models based on published sites were used to scan target genes that were known to be induced by PXR/RXRα binding, but where additional sites had not been previously identified. Sites detected in these scans were assayed for binding to PXR/RXRα and those found to bind were incorporated in subsequent rounds of refinement.

The genvis program was used to display scandiff results for CYP3A4, which is a single gene known to be regulated by PXR/RXRα (Figure 3). BED-custom tracks of this gene for scans of the initial and second PXR/RXRα models (1 and 2) are indicated. Both information models recognize experimentally-verified binding sites 2122: a strong, potential proximal enhancer binding site (custom track M18907_pxr_R17) 204 bp upstream of the transcription initiation site and a cluster of distal enhancer elements 7.2–7.8 kb upstream. Model 1 identified a 7 bit site (AF182273_pxr_R7) in the first intron, which is absent in the scan of model 2. However, model 2 also identifies an additional site (M18907_pxr_R7) within the distal enhancer cluster, which is consistent with the possibility that Model 2 more specifically recognizes promoter binding sites. Similar results were obtained confirming detection of experimentally-defined binding sites in the promoters of other PXR/RXRα regulated genes (CYP3A7, CYP2B6; results not shown) induced by this transcription factor.

Scandiff also produces a summary statistics file which can be used to monitor the progress of information theory-based model refinement. The following example indicates how the results of complete genome scans with four successive PXR/RXRα R_i(b, l) matrices can be interpreted from these summaries (each successive model is based on increasing numbers of experimentally validated binding sites; Table 1). The tables indicate the differences in the number of predicted binding sites in each category of these models. By selecting high thresholds for either ΔR_i values, Δ Z scores or confidence intervals, it is possible to identify binding sites with the most significant model-to-model changes. The following analysis is based on changes in information content of at least 3 bits (ΔR_i), Z score differences of = 1, and confidence intervals = 3 standard deviations, ie. 95%.

Newly identified sites (B-A) predicted with model 2 are 3.8 fold more abundant than those found only with model 1. Scanning the genome with model 3 (vs. model 2) resulted in an even greater disproportionate distribution of unique sites (9.2 fold). This trend continues in model 4, but the fraction of novel binding sites is decreased (2.4 fold). The findings indicate that increasing the diversity of the sequences underlying the matrix affects which binding sites are found in the genome scan. It is apparent that the PXR/RXRα weight matrix has not converged, since large numbers of novel sites continue to be found with successive information models.

Only a modest fraction of sites (S/ [S+I]; S = significant, I = insignificant) exhibit the largest significant changes in binding site strength (ΔR_i = 3 bits; ranging from 3–11%), regardless of which pair of scans are analyzed. Most changes in information content are = 2 bits. As ΔR_i values give no indication of the strengths of the sites that have changed (only the magnitude of those changes), we also cataloged significant changes by comparing the Z scores of the same binding sites found by successive models. The most stringent test (Δ Z = 1) revealed that the transition from model 2 to model 3 produced the largest proportion of significant changes (48% of sites; n = 48,657), in comparison with more modest changes in Z score from models 1 to 2 (0.8%) and models 3 to 4 (2.5%). We interpret these results to indicate that model 3 may have altered the strengths of binding sites at outlying R_i values to a greater extent than the transitions either from models 1 to 2 or from models 3 to 4.

Binding sites that are added to the models in subsequent rounds of experimental refinement have increasingly diverse sequences, resulting in lower measures of R_sequence and therefore detect additional predicted sites. Shorter binding sites, such as those recognized by AHR, with lower R_sequence values, are predicted to be even more abundant. The vast majority of the newly detected binding sites are considered "weak" (R_i <<R_sequence; Table 2). The lower threshold R_i value of binding sites reported by scan is typically set to the strength of the weakest binding site used to define the information weight matrix. The confidence intervals on binding sites with low R_i values are still fairly large [see Appendix to reference 11], and some of these sites may turn out to have R_i < 0 bits. In any case, the affinities for sites with low R_i values, especially those ~ R_{i, minimum} are likely to be negligible and may not be detectable experimentally 6. Nevertheless, the increased sequence diversity introduced by these refinement procedures augments the dynamic range of site binding strengths found with later versions of refined models. The increased sequence diversity affects the frequencies of the nucleotides underlying the weight matrix and can significantly alter the information contents of predicted "strong" sites 9.

Additional gene promoters are found with successive PXR/RXRα models (Table 1). In each pairwise comparison of information models, novel binding sites detected by the later model substantially outnumbered unique sites found only by the earlier model (by 4 to 11.2 fold). Nevertheless, it is encouraging that the increased number of genes containing these binding sites does not proportionally increase with the numbers of binding sites, which suggests that the subsequent models are predicting additional sites in the same genes. This is not surprising, since multiple PXR/RXRα enhancer binding elements with "moderate-to-strong" R_i values have been documented in known targets of this transcription factor, including several CYP3A gene family members. We examined the distributions of such sites in genome scans of promoters with the different PXR/RXRα weight matrices.

The "moderate-to-strong" binding sites in the genome-wide promoter scans (R_i >R_sequence; Table 2) are a small percentage of all sites detected (0.06 % in Model 1, increasing to 0.5 % in Model 4). The refinement procedure may improve the sensitivity of detecting such sites. PXR/RXRα models 1 and 2 actually detect fewer of these sites in gene promoters (and genes) than the numbers of genes that exhibit changes in expression by microarray studies 2122, suggesting that these models predict fewer binding sites, and consequently fewer target genes than expected. In subsequent models, increasingly higher frequencies of multiplex sites are found in the same promoters (8% in Model 1 versus 16% in Model 4). This degree of redundancy (in Model 4) substantially exceeds the expected frequency of promoters with multiple binding sites, and the information required to find these sites in the genome (R_frequency~ 4 bits). We also find that multiplex binding sites within promoters recognized by transcription factors with smaller footprints are considerably more frequent (NF-κB p50/p65 and AHR), as expected from their lower R_sequence values.