Cross-species comparison significantly improves genome-wide prediction of cis-regulatory modules in Drosophila
© Sinha et al; licensee BioMed Central Ltd. 2004
Received: 08 July 2004
Accepted: 09 September 2004
Published: 09 September 2004
The discovery of cis-regulatory modules in metazoan genomes is crucial for understanding the connection between genes and organism diversity. It is important to quantify how comparative genomics can improve computational detection of such modules.
We run the Stubb software on the entire D. melanogaster genome, to obtain predictions of modules involved in segmentation of the embryo. Stubb uses a probabilistic model to score sequences for clustering of transcription factor binding sites, and can exploit multiple species data within the same probabilistic framework. The predictions are evaluated using publicly available gene expression data for thousands of genes, after careful manual annotation. We demonstrate that the use of a second genome (D. pseudoobscura) for cross-species comparison significantly improves the prediction accuracy of Stubb, and is a more sensitive approach than intersecting the results of separate runs over the two genomes. The entire list of predictions is made available online.
Evolutionary conservation of modules serves as a filter to improve their detection in silico. The future availability of additional fruitfly genomes therefore carries the prospect of highly specific genome-wide predictions using Stubb.
Several computational approaches to the problem of predicting cis-regulatory modules ('CRM's) have been reported recently. Berman et al. , Markstein et al.  and Halfon et al.  predicted CRM's involved in body patterning in the fly, and experimentally verified their predictions. The underlying principle in these algorithms was to detect dense clusters of binding sites, as determined by matches (above some threshold) to catalogued transcription factor weight matrices. The algorithm of Rajewsky et al. , called Ahab, avoided the use of thresholds on weight matrix matches by a probabilistic modeling of CRM's. Ahab predictions within the segmentation gene network were subjected to extensive experimental validation, with excellent overall success (Schroeder et al. ). Most predicted CRM's, when placed upstream of a reporter gene, faithfully reproduce one or more aspects of the endogenous gene expression pattern. Moreover, an analysis of binding site composition over the entire set of validated modules reveals that Ahab's prediction of binding sites correlates well with expression patterns produced by the modules and suggests basic rules governing module composition.
The Stubb algorithm (Sinha et al. ) extended Ahab's approach by incorporating the use of two-species sequence information. Stubb also allows the option of scoring positional correlations between binding sites, but this option was not exercised in this study. For each sequence window analyzed, Stubb first computes the homologous sequence in the second species and aligns them using LAGAN (Brudno et al. ). The sequence is then partitioned into "blocks" (contiguous ungapped aligned regions of high percent identity) and non-blocks (sequence fragments between consecutive blocks, in either species). Putative binding sites in blocks are scored under an assumption of common evolutionary descent, using a probabilistic model of binding site evolution. Thus a "weak" site that is well conserved will score higher, while a "strong" site that is poorly conserved will have its score down-weighted. The score of the sequence window includes contributions from binding sites in blocks as well as in non-blocks. Stubb is implemented so that it can be run either on single species or two species data. In the single species mode, it is practically identical to the Ahab program. The Stubb software is available for download from http://edsc.rockefeller.edu/cgi-bin/stubb/download.pl
In this paper, we present evidence that the exploitation of cross-species comparison (between D. melanogaster and D. pseudoobscura) using Stubb can lead to a significant improvement in the accuracy of genome-wide CRM prediction. To our knowledge, this is the first direct evaluation of the effect of cross-species comparison on CRM prediction on a genome-wide scale. Another important contribution of this paper is to present a benchmark for evaluating genome-wide CRM prediction tools, collected from the BDGP database and the literature, and curated by manual inspection of several hundred expression patterns. Using the same benchmark, we evaluate the effect of varying how background sequence information is incorporated in the algorithm, since this is the only tunable parameter in the Stubb program, other than the module length. We are thus able to suggest the optimal parameter settings for genome-wide CRM prediction using Stubb. Finally, we report all genome-wide predictions for cis-regulatory modules involved in anterior-posterior patterning in the early fly embryo, using both single-species and two-species Stubb, many of which make a strong case for experimental validation.
Segmentation gene network
The transcription control paradigm we use as our test system is the segmentation of the anterior-posterior (ap) axis during early Drosophila embryogenesis, which has long been one of the preferred arenas for studying transcription control in vivo. The segmentation genes form a hierarchical network that, in a process of stepwise refinement, translates broad, overlapping expression gradients into periodic patterns of 14 discrete stripes, which prefigure the 14 segments of the larva (for reviews see St Johnston & Nusslein-Volhard ; Rivera-Pomar & Jackle ; Furriols & Casanova ). The maternal factors form gradients stretching along the entire ap axis of the embryo, the zygotic "gap" factors are expressed in one or more broad slightly overlapping domains; together they generate the 7-stripe patterns of the pair-rule genes; finally, the segment-polarity genes are expressed in 14 stripes. The regulation within the segmentation gene hierarchy is almost entirely transcriptional, and most of the participating genes are transcription factors themselves, activating (in the case of the maternal factors) or repressing (most gap factors) the transcription of genes at the same level or below. In most cases, the relevant binding sites are clustered within a small interval of 0.5–1 kb; these CRM's typically contain binding sites for multiple transcription factors and multiple binding sites for each factor. The clustering and the combinatorial and redundant nature of the input facilitate the computational search for segmentation control elements. Since the expression patterns of the segmentation genes are typically complex, their control regions often contain multiple separate CRM's controlling different aspects of the pattern.
The segmentation paradigm has been used as a test system for the computational detection of CRMs by us and others (Rajewsky et al. , Schroeder et al. , Berman et al. , Grad et al. ). Here, as before (Schroeder et al. ), we use the maternal and zygotic gap factors Bicoid, Hunchback, Caudal, Knirps, Krüppel, Giant, Tailless, Dstat, and the TorRE binding factor as input to Stubb. The binding site specificity of each factor is characterized by a position weight matrix that is based on a collection of experimentally verified binding sites.
The complete genomes of two fruitflies, D. melanogaster and D. pseudobscura have been sequenced, and Stubb was used to predict CRM's in the D. melanogaster genome. This was done in two modes – (i) STUBBSS, where Stubb is run on D. melanogaster genomic sequence alone, and (ii) STUBBMS, where Stubb uses orthologous sequence data from D. pseudobscura to help predict CRM's in D. melanogaster. For each mode of execution, we obtain a separate list of predicted CRM's, sorted in order of confidence in the prediction. The ideal test for our purpose would be to compare the accuracy of these two sorted lists. However, the set of experimentally verified CRM's involved in this system is sparse compared to the size of the system – roughly 50 CRM's are known (including the 15 new modules from Schroeder et al. ), while the number of target genes is several hundreds, by our estimate. Hence, direct evaluation of the success-rate of predictions is not feasible, and we use an alternative source of information to evaluate predictions, as described next.
A functional CRM directs the expression of a gene, by definition, and typically this gene is located in close proximity to the CRM. Hence, we may map the list of predicted CRM's to a list of predicted blastoderm-patterned genes – for each CRM predicted by Stubb, the nearest gene is identified, and if this gene is less than a threshold distance of 20 Kbp away, it is predicted to be a blastoderm-patterned gene. The resulting list of predicted "patterned genes" may now be evaluated for accuracy. (Any duplicates in the list are removed before evaluation.) The Berkeley Drosophila Genome Project (BDGP) has catalogued the expression patterns of a large number of genes in D. melanogaster, at various stages of development. We considered such a catalogue of 2167 genes, obtained from BDGP and from the literature. (See Test Genes [Additional File 1].) Visual inspection of the expression patterns of these genes revealed that 286 of them can be classified as having patterned expression along the anterior-posterior axis. (See Materials and Methods; also Patterned Genes [Additional File 2].) Hence, our benchmark is the entire set of 2167 genes, the "positive" set is the 286 ap-patterned genes, the remaining 1881 forming the "negative" set. This enables us to evaluate the accuracy of lists of patterned genes predicted by STUBBSS and STUBBMS, and compare their performance.
We note that some accuracy is lost in the translation of a list of predicted CRM's to the predicted genes it is mapped to, as per the mapping defined above. For instance, it is known that CRM's may control a gene located at large distances, i.e., further than the distance threshold of 20 Kb used in the mapping procedure. Also, it is possible that a CRM is located close to two genes, and directs the expression of both genes, or only of the farther gene, being somehow insulated from the nearer one. To address these concerns, we repeat our evaluation with a slightly different mapping from the one described above. A caveat that remains is that there may be genomic sequences that are functional, in the sense that they are capable of directing a specific blastoderm pattern in reporter gene constructs, but whose activity is 'silenced' in native genomic context and does not translate to patterning of any gene. Also, the CRM may direct expression of the gene only at post-blastodermal stages, so that the gene is not included in the "positive" test set of blastoderm patterned genes. Conversely, it may also happen that a predicted CRM lies close to a patterned gene, thereby being counted as a true positive, but the predicted CRM is not the sequence responsible for the gene's regulation. We assume that such effects are not biased against either algorithm.
STUBBMS performs significantly better than STUBBSS
In order to further scrutinize the difference in predictions made by the two modes of Stubb, we focused on the points where their difference is most pronounced. Thus, in the top 102 unique gene predictions (for which we have information), STUBBSS reports 39 positives, while STUBBMS scores 61 hits, an improvement of over 56%. In comparison, the random expectation is ~13.5 hits. Thus the predictions of both STUBBSS and STUBBMS are significantly enriched in patterned genes (P < 10-12 and <10-37 respectively, Binomial Proportions test). Further examination of the top 102 gene predictions made by each algorithm revealed that 24 true positives are common to both lists. STUBBMS reports 37 true positives not discovered by STUBBSS, while the latter reports 15 true positives not found by the former. Similar results are seen for the top 311 predictions (another peak in Figure 1b): 70 correct predictions were common to both algorithms, 42 were predicted by STUBBMS only, and 21 by STUBBSS only. Thus there is substantial exclusivity in the sets of true positives of each algorithm.
Expression patterns of predicted genes. Top 311 genes predicted as being patterned, by STUBBSS and STUBBMS. "INTERSECTION": Genes correctly predicted by both methods. "MS-ONLY": Genes correctly predicted by STUBBMS and not by STUBBSS. "SS-ONLY": Genes correctly predicted by STUBBSS and not by STUBBMS.
WEAK + INTERMEDIATE
One possible strategy that uses two-species sequence is to make predictions using STUBBSS on each of the two genomes separately and then intersect the respective lists. We found this strategy to be very restrictive – for instance, with a particular score threshold, STUBBSS predicts 205 unique genes in D. melanogaster, but intersecting these predictions with a similar number of top predictions in D. pseudoobscura gives only 68 unique genes, 33 of which are patterned. Of the top 68 predictions made in D. melanogaster alone, 29 are patterned. Thus the "intersection" strategy yields only a modest improvement over the single-species search, and does so at the price of significantly reducing the total number of predictions. Similar results were obtained when intersecting modules instead of gene predictions.
The default mapping from CRM's to genes used in our evaluations predicts a gene to be patterned only if its proximal end is less than 20 Kb from the CRM. Schroeder et al.  studied the range of locations of experimentally verified CRM's relative to the gene. They found that while there is a clustering of CRM's within the proximal 5 Kb region upstream, downstream or intronic of a gene, it is not unusual to have CRM's more than 10 Kb away from the regulated gene. Nelson et al.  observe that for D. melanogaster, the intergenic space on either side of a gene has a mean of 2 Kb – 10 Kb, depending on the complexity of the gene's function. We repeated our evaluation with different values of the distance threshold, and found that lower thresholds (5 Kb, 10 Kb) decrease the recovery rate, while higher thresholds (50 Kb) do not affect performance. (Data not shown.)
Characteristics of genes predicted by STUBBMS
Genes with dorsal-ventral aspects to their blastoderm pattern are more frequent at lower ranks of prediction; i.e., the top predictions are enriched in genes with anterior-posterior patterns only.
Core genes are predominantly found in the top predictions.
Genes found at higher ranks are somewhat more likely to be strongly expressed.
The first two observations imply that the genes more directly involved in the ap axis formation are recovered at better ranks, and that the lower rank genome-wide predictions are richer in derivative patterns characteristic of genes with more complex regulatory inputs (pair-rule factors, dv factors etc.). The same trends were found for the correct predictions made by STUBBSS. (Data not shown.)
Optimal parameter settings for Stubb
We next evaluate the effect of varying how background sequence information is incorporated in the Stubb algorithm. This is the only configurable aspect of the program, other than the module length. (In a separate test, we ran Stubb with a module length of 700 instead of the default value of 500, and found no significant difference in the prediction specificity curve.) One important parameter is the "Markov order" of background. A value of k for this parameter means that local correlations are assumed to be present at the level of (k+1)-mers, i.e., the random probability of seeing a particular base at a position depends on the bases seen at the previous k positions. (For readers familiar with the studies of Rajewsky et al.  and Schroeder et al. , "background k" in those studies is the same as a (k-1)th order background in the terminology of this paper.) We vary this parameter to take the values k = 1 and k = 2, in different runs. The other parameter is the actual sequence used by Stubb to measure background nucleotide frequencies. Here the two options are (i) to use the current sequence window as background, or (ii) to use a pre-specified sequence (or collection of sequences) as background. We call these two the "local" and "global" background models respectively. For the "global" model, we input into Stubb 150 Kb of sequence from non-coding regions of the D. melanogaster genome, collected from the five chromosome arms 2L, 2R, 3L, 3R, and X.
All the above runs were on genomic sequence with tandem repeats masked by the Tandem Repeats Finder program of Benson . We have found that this heuristic improves genome-wide CRM prediction by Stubb. To substantiate this claim, we ran STUBBSS and STUBBMS on raw (unmasked) genomic sequence. Figure 7b plots the results. We find that both STUBBSS and STUBBMS perform better on masked data than on unmasked data. However, when Stubb is used to analyze shorter sequences (such as the upstream and downstream regions of a gene of interest), we have found unmasked sequence to be more useful, since false positives are less of a concern.
The Stubb program is an extension of Ahab, with the important feature that it can handle two-species data within its probabilistic framework. The two programs differ in their underlying optimization method, with Stubb using an Expectation-Maximization approach in contrast to Ahab's conjugate gradient method. Performance evaluation of the two programs shows little difference between them, implying that the algorithm is robust to the actual optimization method used. Another technical difference between Ahab and Stubb is in the manner that orientation of binding sites is treated. While Stubb assumes a uniform prior on the orientation of a binding site, Ahab picks the best orientation for each site, with the caveat that probabilities are not strictly normalized.
An important component of Stubb is the alignment step where the two species are aligned (using LAGAN) and blocks of high sequence similarity are extracted. (See Methods.) The parameters used in LAGAN runs were obtained from Emberly et al. , who derived the alignment parameters that maximize the overlap between experimentally verified binding sites and blocks of sequence conservation. They also studied the effect of changing the alignment algorithm (LAGAN from Brudno et al. versus SMASH from Zavolan et al. ) for CRM's in the two fly species, and found no significant difference. Finally, the similarity thresholds we use for defining conserved blocks (10 bp or longer, with >70% identity) were obtained by trying a broad range of values, and choosing those that produced the best results, as per our genome-wide evaluation.
Tandem repeat masking is a common pre-processing step for many sequence analysis applications involving binding sites. These repeats are short locally duplicated sequences, that may or may not be related to binding sites. It is not clear a priori how tandem repeats should affect module detection – repeats similar to binding sites of the system may improve sensitivity when they occur in CRM's; but if repeats resembling binding sites occur by chance in non-functional regions, prediction specificity may suffer. The occurrence of tandem repeats marks statistical deviation from Stubb's probabilistic model of sequence generation. In our tests, we found that repeats distract the algorithm more than they help, as manifested in better performance on repeat-masked sequence. (See Figure 7b.) This may be because two of the weight matrices in our collection (Hunchback and Caudal) resemble a poly-T stretch. Therefore, the poly-A or poly-T tandem repeats that occur promiscuously in the genome may be confused with sites of these two weight matrices.
A recently published tool for genome-wide CRM prediction, called PFR-Searcher (Grad et al. ), first identifies "phylogenetically footprinted regions" or "PFR"s, that are sequences conserved between the two fly species, and then searches for a subset of these that are most similar in content to an input set of promoters. Their approach differs from Stubb in the nature of prior information input to the algorithm. While Stubb uses an input set of weight matrices, the training data for PFR-Searcher is a set of CRM's which, in their approach, is itself provided by a similarity search among PFR's of co-regulated genes. PFR-Searcher therefore has the advantage of not requiring knowledge of the transcription factor weight matrices relevant to the system. However, its ability to predict the binding site composition of potential CRM's is therefore more limited as compared to Stubb. (The Stubb program computes an average "parse" of the predicted module into its constituent binding sites for various transcription factors.) Grad et al. report an evaluation of their algorithm on a test system very similar to ours, but with enough minor differences to make a direct comparison of performance impossible. For instance, the entire list of CRM's predicted in their evaluation corresponds, as per our CRM → gene mapping, to a set of only 46 unique genes, of which 31 are patterned. Twenty of these 31 correct predictions are also found in the top 46 gene predictions of STUBBMS, indicating a good degree of overlap between the two methods, at least in their highest ranked predictions. A fair and comprehensive comparison of the predictive power of these two algorithms is an interesting topic for future work, and it will be even more interesting to run STUBBMS only on PFR's detected by their criteria.
Regarding the recovery of patterned genes by Stubb, several observations can be made. Of the 286 genes with ap patterns, we recover roughly half at a score cut-off of 10, using STUBBMS. Why is the other half not found? While it is obvious that lowering the cut-off will detect more patterned genes, there are other reasons why a patterned gene may be missed by Stubb. Some genes are likely to be lost due to the distance filter we have imposed (CRM to nearest gene <20 kb), since the regulatory regions of some genes (e.g., homeotic genes) are likely to be larger than that. More importantly, most of the patterned genes that are not part of the core transcriptional machinery have derivative patterns that reflect a more complex input (binding sites for pair rule factors, d-v factors etc.) and thus will only be recovered to the extent their input has a solid maternal/gap component. Conversely, there are at least two reasons for reporting false positives (roughly two thirds at a score cutoff of 10). The presence of an insulator could prevent the interaction between a CRM and its nearest basal promoter. More likely is a scenario where the predicted CRM's do drive expression but at post-blastoderm stages. All gap factors are active in multiple tissues in later development and therefore CRM's with dominant or exclusive gap input may well be active in these later contexts. These caveats affect all current CRM detection algorithms, and accounting for such additional axes of information as genomic context and module composition rules will be a difficult but important challenge for the future.
A very interesting observation comes from the analysis in Table 1: Genes predicted by STUBBSS only, and not by STUBBMS, have weak or intermediate expression pattern more often than strong expression. This means that the CRM's that are not well-conserved between the two species (and hence not picked up by STUBBMS) typically correspond to weakly expressed genes. This ties in with previous studies (e.g., Domazet-Loso & Tautz ) that found fast evolving genes in Drosophila to be expressed relatively weakly.
The Stubb program not only predicts cis-regulatory modules genome-wide, it additionally outputs the binding site profile of each predicted CRM, i.e., the locations and probabilities of binding sites in the CRM. Schroeder et al  use the corresponding feature in Ahab for a systematic analysis of the composition of all known or validated segmentation CRMs. The use of STUBBMS improves such binding site predictions. It is easy to adapt the program to take as input orthologous CRM's from the two species, and highlight the changes in terms of their binding site compositions. This leads to a powerful bioinformatic tool to predict regulatory changes between the two fly species. We can thus obtain hypotheses about changes in expression patterns, which can be verified experimentally. We have examined a representative collection of CRM's, and experimentally verified several of the changes predicted by Stubb, thereby building a catalogue of the different modes of cis-regulatory evolution. The results of this study will be reported in the near future.
We have seen that the use of a second fly genome significantly improves genome-wide module prediction. Since STUBBMS uses a natural "two-species" extension of the algorithm of STUBBSS, this finding is largely a statement about the inherent potential of cross-species comparison as a paradigm for improving functional genomics. The STUBBMS program also has a natural extension to incorporate more than two genomes, and it will be very interesting to see how much of a difference a third genome makes. The genome of D. yakuba is expected to be sequenced soon, and since this species is closer to D. melanogaster, it may help better discriminate conserved regulatory modules.
Alignment of D. melanogaster and D. pseudobscura
D. melanogaster sequences were obtained from Flybase Release 3. The analysis was limited to the five chromosome arms 2L, 2R, 3L, 3R, and X. D. pseudobscura contigs were obtained from http://www.hgsc.bcm.tmc.edu/projects/drosophila/ (February 2003 Release). Based on Blast results, we created a mapping, called "CONTIGMAP", between regions of the D. melanogaster genome and D. pseudobscura contigs, each region typically being tens of Kb long. This mapping is many to many, i.e., different regions of D. melanogaster may map to the same contig, and the same (or overlapping) region in D. melanogaster may map to two or more D. pseudobscura contigs. For each entry (M, P) in CONTIGMAP, where M is the D. melanogaster region and P is the D. pseudobscura contig, the LAGAN alignment program (Brudno et al. ) was run, with parameters gap start = -6, gap extension = 0, match = 1, and mismatch = -2, and all contiguous ungapped blocks of alignment, with length 10 bp or more and 70% identity or more, were extracted. In cases where the same region in D. melanogaster was mapped to multiple contigs, the density of LAGAN blocks was then used to choose exactly one mapping contig.
Tandem repeats in the input sequences were masked with the Tandem Repeat Finder program of Benson , with parameter settings: (match = 2, mismatch = 5, indel = 5, match probability = 0.75, indel probability = 0.2, minimum score = 20, maximum period = 500). STUBBSS was run on the D. melanogaster genome with a sliding window of length 500 bp, in shifts of 50 bp. The input weight matrices for the maternal and gap transcription factors Bcd, Hb, Cad, Kni, Kr, Tll, Dstat and the torRE binding factor were obtained from Rajewsky et al.  and Schroeder et al. . A weight matrix for the transcription factor Gt was constructed from known functional sites collected from the literature. STUBBMS was run on each entry (M, P) in CONTIGMAP, using a sliding window of length 500 bp on the D. melanogaster sequence M, in shifts of 50 bp. Thus, STUBBMS was not run on regions of D. melanogaster that are not aligned with some D. pseudoobscura contig. The weight matrices used were the same as in STUBBSS runs. The locations of the blocks computed in the alignment step (above) were input to STUBBMS, and the input value of the neutral mutation rate was 0.5, the value being chosen due to its better performance over alternatives tested.
Each genome-wide run of Stubb produces, for each starting position of the sliding window, a score that measures the likelihood of the sequence having a cluster of binding sites. The next step is to extract the coordinates of each window that scores better than all other windows overlapping it. Such windows correspond to local "peaks" in the score profile along the genome. All such "peak" windows with scores above a certain threshold are sorted in decreasing order of their score, to produce a sorted list of predicted CRM's. Each window in this list is annotated with useful information including the identity and relative location of its neighboring genes. The list is then filtered to retain only those predicted CRM's where Stubb predicts occurrences of at least two weight matrices. This is a heuristic that incorporates the combinatorial nature of CRM's, i.e., their tendency to have sites for multiple transcription factors (activators as well as repressors.) Finally, any predicted CRM that overlaps with an exon is removed from the list before evaluation. The predictions made by STUBBMS and STUBBSS are listed in the files "Predicted CRM's – two species" (Additional File 4) and "Predicted CRM's – single species" (Additional File 5), respectively.
Annotation of gene expression database
The 792 genes which the BDGP expression database http://www.fruitfly.org/cgi-bin/ex/insitu.pl lists as showing expression during blastoderm (embryonic stages 4–6) were visually inspected. From this list, we removed genes with ubiquitous expression (426; this also removes the presumably very small number of genes whose ubiquitous expression is controlled by separate "regional" modules), extremely faint or irreproducible expression (31), or expression in pole cells or yolk nuclei only (64), as well as genes whose expression is modulated along the dv axis only (13). The remaining 258 genes show patterned expression in the somatic portion along the ap axis of the blastoderm embryo; 28 known segmentation genes not captured in the BDGP expression database were added to the list, for a total of 286 genes showing ap patterned blastoderm expression. These genes were further categorized by expression level (strong, intermediate, weak) and type of pattern (ap, ap+dv, dv+ap). ap includes gap, pair rule and segment polarity-like patterns (e.g., Kr, fkh, eve); ap+dv denotes ap pattern with some dv modulation (e.g., kni, so, en); dv+ap denotes dv pattern with some ap modulation (e.g., neur).
Note added in proof
A recently published paper (Berman et al: Genome Biol 2004, 5:R61, published 20 August 2004.) also evaluates the effect of cross-species comparison on CRM prediction in Drosophila.
Support was provided by the NSF under grant DMR0129848, the NIH under grant GM066434-02, and the Keck Foundation (to SS).
- Berman BP, Nibu Y, Pfeiffer BD, Tomancak P, Celniker SE, Levine M, Rubin GM, Eisen MB: Exploiting transcription factor binding site clustering to identify cis-regulatory modules involved in pattern formation in the Drosophila genome. Proc Natl Acad Sci U S A 2002, 99(2):757–62. 10.1073/pnas.231608898PubMed CentralView ArticlePubMedGoogle Scholar
- Markstein M, Markstein P, Markstein V, Levine MS: Genome-wide analysis of clustered Dorsal binding sites identifies putative target genes in the Drosophila embryo. Proc Natl Acad Sci U S A 2002, 99(2):763–8. 10.1073/pnas.012591199PubMed CentralView ArticlePubMedGoogle Scholar
- Halfon MS, Grad Y, Church GM, Michelson AM: Computation-based discovery of related transcriptional regulatory modules and motifs using an experimentally validated combinatorial model. Genome Res 2002, 12(7):1019–28.PubMed CentralPubMedGoogle Scholar
- Rajewsky N, Vergassola M, Gaul U, Siggia ED: Computational detection of genomic cis-regulatory modules applied to body patterning in the early Drosophila embryo. BMC Bioinformatics 2002, 3(1):30. 10.1186/1471-2105-3-30PubMed CentralView ArticlePubMedGoogle Scholar
- Schroeder MD, Pearce M, Fak J, Fan H, Unnerstall U, Emberly E, Rajewsky N, Siggia ED, Gaul U: Transcriptional Control in the Segmentation Gene Network of Drosophila. PLoS Biology 2004., 2(9):Google Scholar
- Sinha S, van Nimwegen E, Siggia ED: A probabilistic method to detect regulatory modules. Bioinformatics 2003, 19(Suppl 1):i292–301. 10.1093/bioinformatics/btg1040View ArticlePubMedGoogle Scholar
- Brudno M, Do CB, Cooper GM, Kim MF, Davydov E, Green ED, Sidow A, Batzoglou S, NISC Comparative Sequencing Program: LAGAN and Multi-LAGAN: efficient tools for large-scale multiple alignment of genomic DNA. Genome Res 2003, 13: 721–31. 10.1101/gr.926603PubMed CentralView ArticlePubMedGoogle Scholar
- St Johnston D, Nusslein-Volhard C: The origin of pattern and polarity in the Drosophila embryo. Cell 1992, 68(2):201–219. 10.1016/0092-8674(92)90466-PView ArticlePubMedGoogle Scholar
- Rivera-Pomar R, Jackle H: From gradients to stripes in Drosophila embryogenesis: filling in the gaps. Trends Genet 1996, 12(11):478–483. 10.1016/0168-9525(96)10044-5View ArticlePubMedGoogle Scholar
- Furriols M, Casanova J: In and out of Torso RTK signalling. EMBO J 2003, 22(9):1947–1952. 10.1093/emboj/cdg224PubMed CentralView ArticlePubMedGoogle Scholar
- Grad YH, Roth FP, Halfon MS, Church GM: Prediction of similarly-acting cis-regulatory modules by subsequence profiling and comparative genomics in D. melanogaster and D. pseudoobscura. Bioinformatics, in press.Google Scholar
- Nelson CE, Hersh BM, Carroll SB: The regulatory content of intergenic DNA shapes genome architecture. Genome Biol 2004, 5(4):R25. 10.1186/gb-2004-5-4-r25PubMed CentralView ArticlePubMedGoogle Scholar
- Benson G: Tandem repeats finder: a program to analyze DNA sequences. Nucleic Acids Res 1999, 27(2):573–80. 10.1093/nar/27.2.573PubMed CentralView ArticlePubMedGoogle Scholar
- Emberly E, Rajewsky N, Siggia ED: Conservation of regulatory elements between two species of Drosophila. BMC Bioinformatics 2003, 4(1):57. 10.1186/1471-2105-4-57PubMed CentralView ArticlePubMedGoogle Scholar
- Zavolan M, Rajewsky N, Socci ND, Gaasterland T: SMASHing regulatory sites in DNA by human-mouse sequence comparisons. In Proceedings of the 2003 IEEE Bioinformatics Conference (CSB2003) 277–286.Google Scholar
- Domazet-Loso T, Tautz D: An evolutionary analysis of orphan genes in Drosophila. Genome Res 2003, 13(10):2213–9. 10.1101/gr.1311003PubMed CentralView ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.