The Bipad server can be accessed in either guest or registration mode. Guests are required to enter a valid email address in order to receive results, since individual information theory is patented. Graphical output (sequence logo and gap histogram) is produced only for registered users. A project title is required, which is also used to label the sequence logo of the binding site alignment. Unless explicitly specified, text output is sent by electronic mail.

A search pattern is defined in the most general sense. The search pattern can be either a single-block or bipartite motif which is specified by the lengths of the conserved sequence elements and lengths of the sequences, if any, which separate them. The sequence models may be constrained by specifying that the search identify either one site per input sequence (OOPS) or zero or one sites per sequence (ZOOPS). The respective half-site motifs may be homogeneous (consisting of perfect or imperfect repeats of the same sequence) or heterogeneous (containing different motif widths and/or sequence patterns). The configuration of a bipartite pattern may have four potential half-site orientations: direct repeat (DR), reversed DR (RDR), inverted repeat (IR), and everted repeat (ER).

The Bipad program can search for all possible orientations on either a single strand (DR) or on both strands (DR, RDR, IR, ER). The alignment mode should be specified based on biological evidence or hypothesis, for example, structural or experimental evidence that a protein contacts sequences on a single or both strands. Searches for the best alignments of sequences on both strands are slower than those involving a single strand.

The motif width for a single-block search is specified in the first box of the site width field. It is recommended that a variety of motif widths be explored based either on functional binding site laboratory data or empirical methods (see discussion below;9). If the optimal bipartite search pattern is requested, the left- and right-half motifs are specified in the first and second fill-in boxes, respectively. The minimum and maximum gap lengths are also respectively entered in adjacent fill-in boxes.

Typical searches for bipartite patterns require about twice as long as single-block motifs having the same total width. Note that a single-block is equivalent to zero-gap bipartite pattern, in which two motifs are merged together. Lengthy gaps or broad gap length ranges can increase the time required to determine bipartite motifs, but in most instances, program completes within a minute (Figure 1). The default parameters for half-site and gap length widths have been restricted in order to ensure reasonable multiuser server performance. Half-site widths are permitted to range between 4 to 500 nucleotides in length. Gap lengths between half sites can range from 0 to 200 nucleotides.

In a single Monte Carlo cycle, it is possible for the bipartite local alignment to converge to a local optimum. However, large datasets, searches for subtle motifs characterized by low average information, and bipartite alignment on both strands each require additional cycles to ensure that the preferred alignment will be produced. Increasing the number of cycles increases the confidence that results obtained are the best alignment achievable with this algorithm, at the expense of only a modest drop in server performance (Figure 1). A 500 cycle limit is imposed on model building due to finite server capacity.

The relationships between cycle number and run time, and between cycle number and total information content are illustrated by the bipartite alignment of chromatin scaffold/matrix attachment region binding sites (Figure 1). Increases in cycle number are correlated linearly with the time needed to find the optimal alignment. In this instance, two cycles were sufficient to determine the best bipartite alignment, and further cycling was unnecessary. This default parameter is set to 10 cycles, however it is recommended that a variety of cycling criteria be tested to ensure that a stable solution is obtained.

FASTA-formatted DNA sequences are entered in the sequence field text box or files can be uploaded. Only unambiguous lower or upper case nucleotide symbols are permissible {A, C, G, and T}. Each sequence may be up to 5 kb in length and the number of input sequences is limited to a maximum of 2,000.

Model refinement is not performed unless this option is specifically requested. The refinement procedure batch executes a series of runs over the specified range of site and gap length widths. The procedure outputs a unit information incremental (UII) plot, which facilitates comparison of information contents gained among a series of potential binding site models. The best model achievable with the Bipad algorithm is the one exhibiting the maximum UII9.

The specified range of core motif widths should be guided by experimental binding site evidence, if available. Generally speaking, a range of half-site widths averaging 5-mer length half-site would be a reasonable starting point to perform multiple trials of many DNA-protein interactions. Input site lengths shorter than 4 nucleotides are more likely to generate false positive motifs and generally do not represent biologically meaningful binding sites 13. Very high sequence conservation across a putative binding site warrants further exploration of a wider range of binding site motifs in order to mitigate against the possibility that the model may be truncated at either end. False positive motifs with high information contents may be mitigated by eliminating or pre-filtering, recurrent, low-complexity tandem repeated sequences, if possible, from input sequences 2. This is particularly important for sets of eukaryotic intergenic non-coding sequences which frequently contain runs of often imperfect homopolymeric sequences that produce minimum entropies that may not be relevant to protein binding.

The lengths of sequences flanking a motif, and the output delivery method may be specified. Flanking nucleotides are displayed in lower case and motifs are shown as upper case. If the specified flanking sequences extend beyond the boundaries of the input sequence, a dash is indicated at those positions. By default, flanking nucleotides are not displayed.

Constraints on the number of aligned DNA sequences, the maximum number of Monte Carlo cycles, maximum half-site widths and gap lengths can be relaxed upon request.

Given either a single frequency matrix or two half site matrices (termed "first" and "second"), the plotter will draw the corresponding sequence logos. The plotter is capable of performing several operations on the original matrices including transforming the "first" matrix by reverse complementation, transforming the "second" matrix through the same operation, transformation of "both" or "none" of the matrices. Only untransformed ("none") and "first" matrix transformation operations can be carried out on single-block matrices. The central gap length of the sequence logo may be specified for bipartite matrices, however the default size is 4 bp. Other options, ie. logo name and size, can also be defined by the user. Detailed input instructions and a working example can be found at 16. The bipartite output file also includes a separate frequency matrix specifically formatted for use with the bipartite logo plotter.

Figure 3 shows the progressive refinement of the bipartite CAR/RXRα binding motif models. To enable this program function, the 'refine' option is selected and the initial or basis search pattern is defined. The program generates and evaluated models with left and right motifs of increasing lengths. Treating 5<[0,8]>5 as the basis motif, we calculate the UII value for each motif. The UII plot orders these models along the X-axis, with model 1 corresponding to the motif pattern: 5<[0,8]>5, Model 2 to 5<[0,8]>6, through Model 30, which has the pattern, 10<[0,8]>9. The 6<[0,8]>6 motif (Figure 3) displays the highest information increment (see bipartite logo in Figure 2B), ie. the highest level of information density (bits per unit length) over all of the motifs analyzed, making it arguably the optimal binding site model.

To examine the performance of Bipad for detection of true binding sites in sequences of varying lengths, we embedded the MRS binding sites in background sequences having either a uniform equiprobable nucleotide distribution or an average human genomic composition (described above). We embedded exactly one MRS site in each background sequence (23 sequences in total) and varied the lengths of each of the background sequences from 250 bp to 2000 bp (repeated three times for each such simulation). The average performance (based on detection of the embedded sequence) in each group is shown in Figure 4. For sequences less than 1 kb in length, binding sites were detected with accuracy of over 80% regardless of background composition; however, as the sequence length increases, the performance decreases monotonically. It is interesting to note, however, that MRS sites embedded in a background having a composition similar to that of the human genome were more easily detected in longer sequences (e.g. 2 kb) compared to sites embedded in uniformly-distributed background.