Computational Identi cation and Analysis of Eukaryotic Promoters: New Algorithms on the Traces of Gene Regulation

With the information of the complete DNA sequence of several higher eukaryotes as well as expression patterns of thousands of genes under a variety of conditions in our hands, we now have the possibility to computationally identify and analyze the parts of a genome believed to be largely responsible for transcription control { the promoters. This article gives a short overview of the state-of-the-art techniques for promoter localization and analysis, and comments on the most recent advances in the eld. Understanding gene regulation is one of the most exciting topics within molecular genetics. To learn how the interplay among thousands of genes leads to the existance of a complex eukaryotic organism is one of the great challenges, and the availability of large amounts of information gained in the sequencing and gene expression projects both demands and enables us to use computers to solve this task. A key role in gene regulation is played by promoter sequences. We de ne this here as the region proximal to the transcription start site (TSS) of protein encoding genes, those transcribed by RNA polymerase II, and leave aside distal regions such as enhancers. We want to outline the recent developments within two areas of bioinformatics that deal with promoters: The general recognition of eukaryotic promoters, and the analysis of these regions to identify the regulatory elements hidden in them. This is the rst step on the way to complex models of regulatory networks. We focus on the computational point of view, pinpointing out some classic and many recent publications, and leave a more elaborate description, especially of the underlying biology, to the the cited papers and reviews. 2 Analyzing Promoters to Find Unknown Regulatory Elements The interest in promoter analysis received a great boost with the arrival of microarray gene expression data: Once you have a group of genes with a similar expression pro le (e.g., that are activated at the same time in the cell cycle1), a natural assumption is that the similar pro le is (partly) caused by and re ected in a similar structure of the regulatory regions involved in transcription. The ultimate goal here is the automated construction of speci c promoter models containing a combination of several regulatory elements. Research so far has focused on the detection of single motifs (representing transcription factor binding sites) common to the promoter sequences of putatively coregulated genes. Although this problem might seem simple at rst, it is very complex and requires that we nd a motif of unknown size that might not be well conserved between promoters in a set of sequences that do not necessarily represent the complete promoters, and that was in many cases grouped together by a clustering algorithm that itself can be error-prone and include genes that are not co-expressed in vivo. Therefore, studies have mainly concentrated on the rather "simple" genome of the budding yeast S. cerevisiae| it was the rst fully sequenced eukaryotic organism, and the rst one for which a comprehensive amount of expression data became publicly available. Statistics on mapped transcription start sites2 show that its 5' UTR sequences are rather short (a mean of 89 bp), and most of the known regulatory elements are close to the 3 translated part of the genes, the majority being found between 10 and 700 bp upstream from the translation start codon. This means that for yeast, the region upstream of the start codon can be used as a good approximation of a promoter region. Most algorithms searching for conserved patterns in yeast promoters thus take 500{1000 bp upstream of the start codons of supposedly co-regulated genes as data set. There are two fundamentally di erent approaches to tackle the problem: Alignment methods aim at the identi cation of unknown signals by a signi cant local multiple alignment of all sequences. As a direct multiple alignment would be computationally very expensive, the methods go a di erent way. For example, the CONSENSUS algorithm approximates a multiple alignment by aligning sequences one by one3 and optimizing the information content of the weight matrix constructed from the alignment. Other algorithms use a probabilistic approach; they consider the start positions of the motifs in the sequences to be unknown and perform a local optimization over the sequence to determine which positions deliver the most conserved motif. Two important methods are Gibbs sampling4 and Expectation Maximization in the case of the MEME system5. Enumerative or exhaustive methods examine all oligomers of a certain length and report those that occur far more often than expected from the overall promoter sequence composition6;7. This approach has gained in popularity since the arrival of complete genomes and is trickier than one might believe | for example, how to count patterns that overlap themselves? From a practical point of view, the most eye-catching di erence between those methods is maybe the shape of the result: The alignment approaches deliver a model of the motifs 4 (usually a weight matrix) built from the alignment, the enumerative methods a list of overrepresented oligomers, possibly already grouped to form consensus sequences. Figure 1 shows an exempli ed owchart to illustrate this. Di erences of Motif Identi cation Approaches One important di erence among the approaches concerns the background model. For example, a simple background is to account for a di erent overall G/C content. Without such a model, you will most likely nd the obvious, e. g. mainly GC-rich motifs in organisms whose promoters have a high GC content. A better model is constructed from the set of all promoters and takes their speci c sequence composition into account. With such a model, you avoid nding motifs that are common to all promoters such as TATA boxes. But this also means that a speci c model, at least for each organism, has to be trained, and this information is not always available. Enumerative methods have to use such an elaborate background model because they need it to judge the importance of frequent patterns. In contrast, alignment methods usually incorporate only the G/C content, which makes them more prone to fail if the motif is not very well conserved among the sequences and the sequences to be examined become too large8. Most of the enumerative algorithms need to have the size of the motif speci ed in advance. Because of the xed size, they often deliver a number of similar motifs simply shifted by one base or having mismatches. Some methods provide an automatic postprocessing to group motifs to consensus strings and thus come up with a small number of putative regulatory elements that can be examined by experts more easily. A potential problem here is that parts of the consensus might come from di erent sequences. 5 The alignment approach requires di erent statistics depending on how often a pattern should be present in the sequences. For instance, MEME can be run in three modes assuming that a motif occurs exactly once, at most once, or an arbitrary number of times per promoter sequence. The Gibbs sampler implementation listed in table 1 also allows for zero or multiple ocurrences. In principle, alignment methods yield one pattern per run, but they can be run several times to detect more than one motif, masking out previously found sites. Gibbs sampling is a non-deterministic approach, meaning that even without masking out sites, it might deliver di erent motifs. New directions Some limitations of enumerative methods have been eliminated by a number of recent publications: The motif identi cation problem was maybe the most outstanding topic at the ISMB 2000 conference on intelligent systems for molecular biology1 and has also been prominent in several other publications. It is now possible to detect homoor heterodimer motifs separated by a xed9 or variable spacer length10, or a variable motif length in general11. To allow for mismatches, ambiguous nucleotide letters (such as R for the purines) are now included in the oligomer alphabet. Thus it seems as if the enumerative approach is the method of choice: It exhaustiveley searches over all possible oligomers and provides more signi cant results because of the background modeling. In practice, though, alignment methods are more exible: Because of the simple background, they are not restricted to one speci c organism. They can also nd long motifs the detection of which is simply not feasible by an exhaustive approach. 1http://ismb2000.sdsc.edu 6 Also, they deliver a weight matrix as a comprehensive model for a motif which can be used more exible than a consensus sequence for searching purposes. We therefore propose a two-step approach: First apply an enumerative approach, and use the results to initialize a weight matrix for an alignment method. Unfortunately, no such combined approach has been published yet, but the Gibbs sampler given in table 1, for example, lets you specify a weight matrix to start with. Of course, this only works if both methods are available, which so far is the case for only yeast and some microbial organisms. The described methods are often applied on a set of promoters that were rst grouped together using gene expression level measurements. Bussemaker et al. recently analyzed the whole set of yeast regulatory sequences without using any information on gene expression levels11, constructing a dictionary of oligomers of increasing length and using the previous dictionary of shorter oligos as background. A new way to look at the data is to cluster genes based on both expression levels and common motifs at the same time12. This can help to separate gene groups that are active under the same conditions but belong to separate regulatory pathways. An orthologous approach is to identify elements not by analyzing di erent promoters from the same organism, but promoters of the same gene from about 10 di erent related species13. An optimal alignment of a small region of speci ed size is constructed that takes the phylogenetic distance into account. The question remains how we can use all these methods when we move on to the analysis of higher eukaryotes with their highly complex genomes. The euchromatin of D. melanogaster has a gene density of roughly one gene every 9 kilobases and an average predicted transcript size of 3058 bp14, leaving a huge portion of the genome as potential 7 locations of regulatory elements. In this case, the alignment of noncoding sequences from two related species, also known as phylogenetic footprinting, can help to narrow the search region and reveal conserved and potentially regulatory regions15;16. A recent publication closes the gap between this approach and motif identi cation: 28 orthologous co-regulated gene pairs from human and rat were automatically aligned to identify conserved ungapped sequence blocks, and the subsequent analysis of the conserved parts with a Gibbs sampling approach reveiled the known motifs that were missed otherwise17. The main assumption for phylogenetic approaches is that the regulatory pathway itself has not diverged which would result in di erent motifs with the same function. If we do not have information from related species, we can concentrate on the analysis of proximal promoter regions close to TSSs, but the length of UTRs of higher eukaryotes prevents us from assuming that the TSSs can be found immediately upstream of the coding part of a gene. In our recent genome annotation assessment18, the 92 Drosophila genes from the set for which full-length cDNA information was available had an average UTR length of about 1,900 bp (17 transcripts had UTRs longer than 1,000 bp). This means we have to nd the start sites rst. Finding the Promoters in Genomic DNA For a long time, bioinformaticians have tried to come up with algorithms able to identify the TSSs in eukaryotes. This is not easy because promoters are very diverse, and even well-known signals such as the TATA box can be weakly conserved or missing altogether. Algorithms for general promoter recognition so far can be classi ed into two groups: 8 Search-by-signal algorithms make predictions based on the detection of core promoter elements such as the TATA box or the initiator, and/or transcription factor binding sites outside the core19. Search-by-content algorithms identify regulatory regions by using measures based on the sequence composition of promoter and non-promoter (typically coding and intron sequences) examples20. There are also methods that combine both ideas { looking for signals and for regions of speci c composition21;22. For an exact localization, promoter prediction should also mean identi cation of TSSs. But search-by-content methods do not provide good TSS predictions because they do not look for positionally conserved signals. To enable the comparison of di erent algorithms, predictions are thus counted as correct if they are made within a window around an experimentally veri ed start site. Using this scoring, an evaluation in 1997 found that many algorithms identi ed a third up to half of the start sites within genomic DNA sequences23. The programs were run ab initio, i. e. without any additional information but the sequence itself. The problem, though, was the vast number of false positives: Even the best algorithms had one false TSS prediction within 500{1,000 base pairs. New Features, New Algorithms, New Hope As a response to these rather disencouraging results, di erent approaches for promoter nding were pursued. One idea is to provide an accurate prediction of the TSS, but only for small regions known to contain a promoter24. An "opposite" algorithm provides speci c 9 predictions of regulatory regions (of a size of roughly 1,000 bp) via search-by-content, but gives no information whether the a ected gene is on the leading or lagging strand, or where within the region the TSS itself is located25. A fundamentally di erent approach is to construct speci c instead of general promoter models for groups of genes such as muscle active genes known by experiment to contain speci c combinations of regulatory elements26 (reviewed in detail by Werner27) | this is where promoter nding and analysis meet. With the complete genome of many organisms at our ngertips, interest in general promoter prediction has awoke again. First, there is the fact that even though the ab initio performance of the algorithms is not as good as desired, this is not the way the annotation of genomes is done. Many algorithms are used together, and limiting the analyzed sequence to upstream from the start of a gene prediction or a cDNA alignment reduces the number of false predictions immensely. Second, slow but constant advance has been made | in a more recent assessment of both ab initio and gene nder coupled promoter predictions for Drosophila, the ab initio methods had less false positives than before18, and coupling them with a gene nder proved to be quite successful | although it looks like it will be hard to achieve a sensitivity of more than 50%. Third, other features are known that can be derived from the DNA sequence28 and may be suitable for promoter recognition: Many vertebrate promoter regions coincide with CpG islands. These are regions where the C+G content is high and the CG dinucleotide occurs more frequently than expected, a consequence of the fact that the DNA of many promoters is unmethylated so as to be accessible to regulatory proteins. A method to discriminate 10 between CpG islands in promoters and in other parts of the genome has just beenpublished29. This method can be seen as a search-by-content approach and does notdeliver a TSS prediction. The use of CpG island feature in the latest version of ourMcPromoter predictor (see table 1) has also lead to a considerable reduction of falsepositives by roughly one third. Unfortunately, CpG islands only exist in vertebrateorganisms.Features common to promoters of all organisms are structural properties of DNA,such as bendability or conformation (a compilation was carried out by Liao et al.30).For these properties, scoring tables based on dior trinucleotides were determinedexperimentally and can be used to calculate pro les over the DNA sequence. Stud-ies have shown that in general, eukaryotic promoters indeed do have a distinct pro-le when compared to coding or non-regulatory sequences28 . Whether using thesefeatures will improve recognition remains to be seen: The pro les of individual se-quences can be very noisy and thus not easy to use, and it is not clear yet if theyprovide new information that is not accurately re ected in the sequence itself.The most recently published promoter nding and analysis tools are listed in table 1.More links can be found in comprehensive reviews23;27.Will we be able to nd the regulatory regions of eukaryotes with high accuracy, andif so, will we be able to derive complex models for transcription regulation from theirsequence? The question is open, but we certainly are on the way.11 References[1] P. Spellman, G. Sherlock, M. Zhang, V. Iyer, K. Anders, M. Eisen, P. Brown, D. Bot-stein, and B. Futcher. Comprehensive identi cation of cell-cyle regulated genes of theyeast saccharomyces cerevisiae by microarray hybridization. Mol Biol Cell, 9:3273{3297, 1998.[2] J. Zhu and M. Q. Zhang. SCPD: a promoter database of the yeast Saccharomycescerevisiae. Bioinformatics, 15(7/8):607{611, 1999.[3] G. Z. Hertz and G. D. Stormo. Identifying DNA and protein patterns with sta-tistically signi cant alignments of multiple sequences. Bioinformatics, 15:563{577,1999.[4] C. E. Lawrence, S. F. Altschul, M. S. Boguski, J. S. Liu, A. F. Neuwald, and J. C.Wootton. Detecting subtle sequence signals: a Gibbs sampling strategy for multiplealignment. Science, 262(5131):208{214, 1993.[5] T. L. Baily and C. Elkan. Unsupervised learning of multiple motifs in biolpolymersusing expectation maximization. Machine Learning, 21:51{83, 1995.[6] J. van Helden, B. Andre, and J. Collado-Vides. Extracting regulatory sites from theupstream region of yeast by computational analysis of oligonucleotide frequencies. JMol Biol., 281(5):827{842, Sep 1998.[7] A. Brazma, I. Jonassen, J. Vilo, and E. Ukkonen. Predicting gene regulatory elementsin silico on a genomic scale. Genome Res., 8(11):1202{1215, 1998.12 [8] P. Pevzner and S.-H. Sze. Combinatorial approaches to nding subtle signals in DNAsequences. In Proc. ISMB, volume 8, pages 269{278, San Diego, 2000. AAAI Press.[9] J. van Helden, A. F. Rios, and J. Collado-Vides. Discovering regulatory elements innon-coding sequences by analysis of spaced dyads. Nucleic Acids Res., 28(8):1808{1818, 2000.[10] S. Sinha and M. Tompa. A statistical method for nding transcription factor bindingsites. In Proc. ISMB, volume 8, pages 344{354, San Diego, 2000. AAAI Press.[11] H. Bussemaker, H. Li, and E. D. Siggia. Building a dictionary for genomes: Identi-cation of presumptive regulatory sites by statistical analysis. Proc. Natl Acad. Sci.U.S.A., 97:10096{10100, 2000.[12] I. Holmes and W. J. Bruno. Finding regulatory elements using joint likelihoods forsequence and expression pro le data. In Proc. ISMB, volume 8, pages 202{210, SanDiego, 2000. AAAI Press.[13] M. Blanchette, B. Schwikowski, and M. Tompa. An exact algorithm to identifymotifs in orthologous sequences from multiple species. In Proc. ISMB, volume 8,pages 37{45, San Diego, 2000. AAAI Press.[14] M. D. Adams, S. E. Celniker, R. A. Holt, C. A. Evans, J. D. Gocayne, P. G. Ama-natides, S. E. Scherer, P. W. Li, R. A. Hoskins, R. F. Galle, R. A. George, S. E.Lewis, S. Richards, M. Ashburner, S. N. Henderson, G. G. Sutton, J. R. Wortman,M. D. Yandell, Q. Zhang, L. X. Chen, R. C. Brandon, Y. H. Rogers, R. G. Blazej,M. Champe, B. D. Pfei er, K. H. Wan, C. Doyle, E. G. Baxter, G. Helt, C. R. Nelson,13 G. L. Gabor Miklos, H. O. Smith, E. W. Myers, G. M. Rubin, J. C. Venter, et al.The genome sequence of Drosophila melanogaster. Science, 287:2185{2195, 2000.[15] L. Duret and P. Bucher. Searching for regulatory elements in human noncodingsequences. Current Opinion in Structural Biology, 7:399{406, 1997.[16] R. C. Hardison. Conserved noncoding sequences are reliable guides to regulatoryelements. Trends Genet., 16(9):369{372, 2000.[17] W. W. Wasserman, M. Palumbo, W. Thompson, J. W. Fickett, and C. E. Lawrence.Human-mouse genome comparisons to locate regulatory sites. Nature Gen., 26:225{228, 2000.[18] M. G. Reese, G. Hartzell, N. L. Harris, U. Ohler, J. F. Abril, and S. E. Lewis. Genomeannotation assessment in Drosophila melanogaster. Genome Res., 10(4):483{501,2000.[19] D. S. Prestridge. Predicting Pol II promoter sequences using transcription factorbinding sites. J Mol Biol., 249:923{932, 1995.[20] G. B. Hutchinson. The prediction of vertebrate promoter regions using di erentialhexamer frequency analysis. Comp Appl Biosc., 12(5):391{398, 1996.[21] V. Solovyev and A. Salamov. The Gene-Finder computer tools for analysis of humanand model organisms genome sequences. In Proc. ISMB, volume 5, pages 294{302,Menlo Park, 1997. AAAI Press.[22] U. Ohler. Promoter prediction on a genomic scale | the Adh experience. GenomeRes., 10(4):539{542, 2000.14 [23] J. W. Fickett and A. G. Hatzigeorgiou. Eukaryotic promoter recognition. GenomeRes., 7:861{878, 1997.[24] M. Q. Zhang. Identifcation of human gene core promoters in silico. Genome Res.,8:319{326, 1998.[25] M. Scherf, A. Klingenho , and T. Werner. Highly speci c localization of promoterregions in large genomic sequences by PromoterInspector: a novel context analysisapproach. J Mol Biol., 297(3):599{606, 2000.[26] W. W. Wasserman and J. W. Fickett. Identi cation of regulatory regions whichconfer muscle-speci c gene expression. J Mol Biol., 278:167{181, 1998.[27] T. Werner. Models for prediction and recognition of eukaryotic promoters. Mam-malian Genome, 10:168{175, 1999.[28] A. G. Pedersen, P. Baldi, Y. Chauvin, and S. Brunak. The biology of eukaryoticpromoter prediction | a review. Comput Chem., 23(3/4):191{207, 1999.[29] I. P. Ioshikhes and M. Q. Zhang. Large-scale human promoter mapping using CpGislands. Nat. Genet., 26(1):61{63, 2000.[30] G.-C. Liao, E. J. Rehm, and G. M. Rubin. Insertion site preferences of the P transpos-able element in Drosophila melanogaster. Proc. Natl Acad. Sci. U.S.A., 97(7):3347{3351, 2000.15 Program Name descriptionHTTP AddressGeneral Promoter FindingPromoter2.0search-by-signal, arti cial neural network www.cbs.dtu.dk/services/PromoterNNPPsearch-by-signal, time delay neural network www.fruit y.org/seq tools/promoter.htmlPromoterInspector search-by-content, class-speci c oligomers www.gsf.de/biodvMcPromoter V3signal/content,www.mustererkennung.de/HTML/English/stochastic segment model/neural network Research/PromoterCorePromotersignal/content, discriminant analysisargon.cshl.orgPromoter Analysis ToolsRSA Toolsyeast and microbial exhaustive searchwww.ucmb.ulb.ac.be/bioinformatics/rsa-toolsGibbs sampleralignment methodbayesweb.wadsworth.org/gibbs/gibbs.htmlMEMEalignment via Expectation Maximizationmeme.sdsc.eduBBAphylogenetic footprinting bybayesweb.wadsworth.org/Bayes alignmentcgi-bin/bayes align12.plPipMakerphylogenetic footprinting by identity plots bio.cse.psu.eduTable 1: A selection of recently published WWW accessible promoter nding and analysistools.16 ...CGTCGGGC...TTCA ...TTTCCAG...CACTCACACGTGG