Regulation of gene expression is accomplished through binding of transcription factors (TFs) to distinct regions of DNA (TF binding sites (TFBSs)), and, after anchoring at these sites, transmission of the regulatory signal to the basal transcription complex. Indeed, regions around TFBSs can be interrogated with regards to binding and interaction with other TFs (so-called composite modules) as well as local sequence properties that favor recruitment of TFs, bending and looping of DNA and nucleosome positioning. Some of these TFs are specific for a particular tissue, a definite stage of development, or a given extracellular signal, but most TFs are involved in gene regulation under a rather wide spectrum of cellular conditions. It is clear by now that combinations of TFs rather than single factors drive gene transcription and define its specificity. Dynamic and function-specific complexes of many different TFs, so-called enhanceosomes 1, are formed at gene promoters and enhancers to drive gene expression in a specific manner. At the DNA level, the blueprints for assembling such variable TF complexes on promoter regions may be seen as specific combinations of TFBSs located in close proximity to each other. They are termed 'composite modules' (CMs) or 'composite regulatory modules' 2 or cis-regulatory modules 3. There may be several different types of CMs located in the regulatory region of one gene, which may be distant from each other (for example, liver- and muscle-specific enhancers of one gene) or overlapping. Taking this into account, it becomes more and more evident that the 'local sequence context' in the vicinity of the TFBS, as well as 'global context' of the whole promoter/enhancer where the TF site is located, influences binding and functioning of the corresponding TF. Numerous examples of so called composite regulatory elements are reported (see the TRANSCompel database 4) when TF binding and proper functioning of a site is strongly dependent on other sites located in the close vicinity (adjacent or even overlapping sites) or quite distant from each other (up to 100 and more nucleotides). For instance, for the TF family of nuclear receptors (to which hepatocyte nuclear factor (HNF)4 factors belong) there is experimental evidence showing clear dependence between functioning of HNF4 factors at their cognate sites and binding of other factors to the neighboring sites, both synergistically and antagonistically 4. There is a need to develop computational models to predict TFBSs that are functional and are involved in the control of gene transcription. Recent developments in the field of machine learning techniques allow us to apply them to build highly sensitive and specific methods for predicting functional TFBSs in human and other genomes.

Because of our continued interest in the regulation of liver-enriched TFs 56, we were particularly interested in identifying novel genes regulated by the hepatic nuclear factor (HNF)4α. Indeed, HNF4α is a versatile TF, and several investigators have reported the identification of genes targeted by HNF4α. These studies included various experimental approaches, including transient transfection of HNF4α into a human hepatoma cell line, a rat insulinoma cell line, and a human kidney cell line 789. Additionally, findings with conditional knock-outs of HNF4α 10 were recently reported. Notably, in the study of Odom et al. the genome-wide identification of binding sites for TFs HNF4α, HNF1α, and HNF6 was reported by use of the ChIP-chip assay with a 13,000 human promoter sequence containing microarray 11. Strikingly, in the case of HNF4α, the number of contacted promoters was unexpectedly high; 1,575 potential HNF4α target genes were identified. In addition, 42% of the genes occupied by RNA polymerase II were also occupied by HNF4α, suggesting that nearly 50% of all liver-expressed genes are regulated by HNF4α alone. Similarly, in another recent ChIP-chip experiment of ENCODE (Encyclopedia of DNA Elements) genomic regions (about 1% of the human genome), 663 novel HNF4α binding sites were identified in 100 genes 12, which suggests there are a large number of HNF4α targets (over 60,000 sites in the vicinity of about 10,000 genes) if extrapolated to the entire genome. This unprecedented high number of HNF4α binding sites revealed by the ChIP-chip method raises the question of the functional role of all these sites in the regulation of gene transcription.

Indeed, the ChIP-chip assay is a much wanted and a highly advanced method for the genome-wide search and identification of TFBSs. Nonetheless, it suffers from unacceptably high false positive findings. In the study of Odom et al. 11 252 (16%) false positive binding sites were predicted by the authors. Another problem with this method is that a surprisingly small fraction of identified ChIP fragments possesses the canonical binding motif for the corresponding TF 3. This limitation needs to be overcome, and it is highly desirable to identify functional binding sites relevant for the regulation of gene transcription. Furthermore, in the current studies there is often no rationale for the selection of promoters spotted on the array; for example, no bioinformatic approach is applied to identify relevant sequences for the design of the ChIP-chip assay.

Here, we report a computational approach based on a novel machine learning technique, which enabled the identification of genome-wide TFBSs. This method was applied to search for HNF4α gene targets. A genetic algorithm and an exhaustive feature selection algorithm were trained on 73 known and well characterized HNF4α target sequences in promoters and enhancers of different mammalian genes (Additional data file 1). By genome-wide scanning of all human gene promoters we identified novel genes targeted by HNF4α. Then, a subset of predicted binding sites was confirmed by electromobility shift assay. We further interrogated promoter sequences for HNF4α binding sites identified by the ChIP-chip assay. We also analyzed expression of genes targeted by HNF4α and observed a good correlation between computationally annotated HNF4α binding sites and expression of targeted genes. Notably, ChIP-chip experiments tend to report a rather high number of TFBSs in promoters of genes whose regulation by HNF4α is not observed, whereas our computational method for the prediction of HNF4 regulatory sites enabled improved specificity with the method encompassing rules for the regulation of gene expression.

Overall, we demonstrate the power of our computational approach in identifying novel genes targeted by HNF4α. Our machine learning technique significantly improved the overall recognition and, therefore, the identification of faithful HNF4α targets. This method enabled refinement of TF site predictions based on the ChIP-chip assay and identification from among them of potentially functional sites, as reported here. Furthermore, our method can easily be applied to the genome-wide identification of genes targeted by any mammalian TF and is not limited to promoter sequences alone, with an overall success of approximately 80% based on experimental confirmation.

De novo computational identification of genes targeted by various TFs is a challenging task, especially in genomes of higher eukaryotic organisms, which are characterized by extremely large gene regulatory regions. Indeed, binding of TFs to their cognate sites on DNA is a complex process that requires the presence of a specific short sequence pattern in DNA, commonly described by a PWM. Furthermore, the specific local sequence context in the vicinity of the binding site is required to provide favorable conditions for DNA confirmation and DNA flexibility 22. In addition, local structures such as short repeats and palindromes are often observed and, as discussed before, are needed to enable an optimal environment for homo- and heterodimerization of TFs 4. The particularly important role of the global context of TFBSs in determining cooperative binding of factors with other TFs to their neighboring DNA sites is broadly recognized 1. A broad collection of experimentally proven facts on cooperative binding of two and more TFs to so-called composite regulatory elements with synergistic effects on the regulation of gene expression is provided by the TRANSCompel^® database 4. Among these are several known examples of nuclear receptors that are involved in such composite elements (for example, glucocorticoid receptor, androgen receptor and others). But there are no bioinformatics tools available so far that enable a systematic analysis of the combinatorial sequence context of genomic binding sites.

In general, there is a definitive need to develop novel computational approaches to improve the description of the DNA patterns required for TF binding. Ellrott and co-workers 23 applied a Markov chain model to identify HNF4α binding sites in order to improve recognition accuracy of the DNA binding pattern. They have demonstrated that the approach performs better than PWMs alone, but this approach does not consider any local context on the flanks of sites that indeed play a crucial role in promoter activation and DNA binding in vivo.

Recently, local context in the form of short repeats has been successfully implemented to improve recognition of binding sites for nuclear receptors 2425. Extending the previously published approach 24 to the application of hidden Markov models, Sandelin and Wasserman 25 modeled various known constellations of direct, inverted and everted repeats for different sites of nuclear receptors and were able to improve the general precision of the recognition. This approach looks very promising, although it lacks any capability to classify predicted sites in order to identify which particular TF from the large family of nuclear receptors is able to bind to the predicted sites. In addition, we show here that binding sites for such nuclear receptors as HNF4α are highly enriched by various different repeat structures, which does not completely fit with the existing paradigm that the DR1 repeat comprises the canonical structure of HNF4 sites. This makes it extremely difficult to judge factor recognition based on an oversimplified model based on the repeat structures of sites.

We therefore developed a novel approach for the recognition of functional HNF4α binding sites by analyzing the local and global contexts of targeted genes. The method is based on the assumption that the sequence context surrounding TFBSs in DNA is very important for the process of TF binding to the site and, most importantly, for providing specificity of the TF in the regulation of gene expression - by either activation or repression of the gene in particular cellular situations. The sequence contexts of the TFBSs actually makes them functional: in the absence of the proper context, the possible binding of a TF to a particular site on DNA can be impaired or made functionally neutral (which means that the factors are bound to the DNA, but do not influence expression of the gene; such sites are, therefore, non-functional).

In the current work, we performed a thorough analysis of the local nucleotide context on the flanks of known functional HNF4α sites, as well as in the whole local region occupied by the sites. We improved our earlier published approach to analyzing local context 2, which is based on a SiteVideo method 14, and introduced new types of contextual features that modeled various repeated structures in the sequences on the flanks of the sites. Interestingly, the revealed short oligonucleotide features and repeats can be classified into three categories. The first category includes oligonucleotides like ANGD and MDDR and repeats like AV-VS, VS-YA and (RBNH)² that fit to different parts of the HNF4 consensus sequence and appear to be overrepresented in a rather wide area around the center of the binding sites. We can interpret such features as a signature of overrepresentation of HNF4 site-like patterns in the local area surrounding the functional HNF4 site, which may play a role in increasing the probability of HNF4α binding to this site. The second category includes oligonucleotides like CDDM and repeats like (DNCD)², which are overrepresented in a quite small area corresponding to the central positions of the sites. Such features correspond to the central HNF4 site pattern, but they reveal some contextual features of the functional HNF4 sites that cannot be described by the PWM matrix model, for example, correlation between neighboring nucleotides that can not be captured in full by the mononucleotide weight matrix. The third category includes 'negative' features that reveal oligonucleotides to be underrepresented at functional binding sites when compared with background sequences. Such negative features can be local, as in the case of the repeats (BNDK)², (NBHV)² and (NVYB)², which again describes some mutual nucleotide correlations that cannot be captured by PWMs, or distributed, such as BR-NT, which can be interpreted as an 'echo' of some physical-chemical properties of DNA that may interfere with the binding or functioning of the TFs. Notably, the length of the oligonucleotides tested by our method was restricted to four letters of the extended code, mainly because of the high computational complexity of the calculations; however, this oligoncleotide length seems quite optimal for revealing statistically significant features of DNA sequences.

We assume that, in addition to the local context, the global context of the TFBSs in the regulatory regions of genes dictates whether these sites are functional. The global context, which we model by specific combinations of binding sites of various TFs, provides some sort of 'scaffold' on DNA to enable cooperative or antagonistic interactions between TFs. These multiple and complex interactions, if correctly organized in space and time, give rise to the regulatory function of the TFBSs under investigation. It is clear by now that binding of a single TF to its cognate site on DNA alone does not guarantee the proper functional activity of the targeted gene. More interaction with other TFs in the transcription complex and in the enhanceosome are necessary to acquire the full regulatory functionality.

Specifically, known functional combinations of TFBSs were used before in a number of promoter analysis approaches, for example, for identifying muscle-specific promoters 2627, the promoters of liver-enriched genes 28, yeast genes 29, immune-specific genes 303132, and the promoters of genes regulated during the cell cycle 33 or genes involved in antibacterial defense responses 3435. A number of approaches identifying composite motifs were created, including BioProspector 36, Co-Bind 37, MITRA 38, and dyad search 39. These programs help to discover ab initio new regulatory sites for yet unknown TFs. Another set of methods has been developed to discover composite modules by utilizing information on potential binding sites for known TFs (stochastic methods such as ClusterScan 40 and TOUCAN system 41, and probabilistic methods such as reported in 42). We combined these two approaches by computing local context as an exhaustive ab initio composite motif discovery method with the global context - the powerful composite module discovery method based on application of a genetic algorithm.

Furthermore, we wish to point out that our method considers several alternative PWMs for the calculation of the global context. The use of such alternative PWMs for constructing the composite module (see Materials and methods) enables more reliable predictions. In cases with small training sets but with data derived from multiple rounds of computations, the use of different matrices is particularly meaningful. These computations are far from statistical saturation, but new sites may eventually add a certain bias and potentially drive the new PWM matrix away from the functional binding site sequence context. In the case of HNF4α we included both half-site and full-site matrices. Still, the full-length PWMs are able to capture some subtle differences in the spacing sequences between the 'repeats'. Recent reports confirm our old observation that such 'gap' sequences between two repeats of the nuclear receptor sites are often actually more conserved between species then the repeats themselves. Therefore, the combined usage of fixed-length PWMs together with the distributed oligonucleotides on the flanks provides a more robust method for site detection than each method separately.

To compare the findings from our algorithm with the best existing methods, we performed an independent run of the TOUCAN software 41 on the set of HNF4α sites. Notably, TOUCAN is similar to our method and is based on a genetic algorithm. It identified a combination of 14 PWMs for different TFs, including two matrices for HNF4α factors (V$HNF4_01 and V$HNF4_01_B) and matrices for other factors (V$USF_01, V$OCT1_02, V$SP1_Q6, V$PPARG_03 and some others). Interestingly, except for the HNF4 matrices, there were no further correlations with matrices selected by our method (Table 2). This difference can be explained by the ability of our method to identify pairs of matrices and also by the ability to optimize the cut-off values, which is not possible in TOUCAN. Another advantage of our method is the possibility to include information on the tissue specificity of the factors, through extensive use of factor expression annotation in the TRANSFAC^® database. We further compared our method with the NUBIscan algorithm 24. Our approach combines many different features of the most conserved part of the sites as well as various features of the local and global contexts, whereas the NUBIscan algorithm relies solely on the 'repeated' structure of the nuclear receptor sites, which is indeed a very profound property of these sites but not the only one. And, similar to the later published Wasserman approach 25, the NUBIscan algorithm lacks the capability to classify predicted sites in order to identify which particular TF from the large family of nuclear receptors is able to bind to a predicted site. Notably, our method was designed specifically for the recognition of HNF4 sites. Nonetheless, our strategy to define the local and global sequence contexts is a highly generalized method and can be applied to any TF.

An additional point of consideration is the 'regulatory potential score', as introduced in the work of Elnitski and co-workers 43, which refers to the five-way multi-species alignment introduced in the UCSC Genome Browser. In our study we did not restrict ourselves to conserved sites based only on multispecies homology of regulatory regions (phylogenetic footprinting). Although this concept is quite popular for the selection of evolutionarily conserved regulatory sites, the method suffers from low sensitivity because functional TFBSs are frequently missed, mainly due to the very complicated evolutionary history of mammalian regulatory sequences, which can hardly be modeled by the simple divergent concept, which is the basic concept of phylogenetic footprinting.

Furthermore, promoter regions are characterized by very specific average base composition as well as composition of di-nucleotides. It is known that whereas the overall genomic sequences are highly depleted of CG dinucleotides, promoters are often located near high concentrations of them, in or near so-called CpG islands. We therefore compared the nucleotide and dinucleotide composition of the promoter sequences of the known HNF4α sites and sequences that were used as probes in the ChIP-chip experiments of Odom et al. 11, since these were the sequences to which our method was applied. We did not find any significant difference in the nucleotide and dinucleotide compositions of these two sets of sequences, which is not a surprise since the probes of the chip were designed using known genomic promoters. Consequently, since training and test sequences are very similar in their context, the CpG bias, if any, of the HNF4-containing promoters of the training set cannot bias the results of the analysis of the sequences from the ChIP-chip experiment.

Additionally, many authors attribute a certain functional role to the CpG islands in promoters. These islands can be some sort of centers of regulated DNA methylation, which can effectively contribute to hepatocyte-specific gene regulation by providing HNF4α binding sites with necessary functional context. The absence of such a CpG context in the vicinity of HNF4α binding sites may potentially render them functionally neutral. Therefore, some CG dinucleotide-like features that were included in the local context (for example, elevated frequency of the oligonucleotide CDDM, where D is (A/T/G) and M is (A/C)) may reflect this 'CpG' bias of functionally active promoter sequences and, therefore, help to identify the functionally active HNF4α binding sites.

We further studied the influence of the distance from HNF4α sites to TSSs. As shown in Additional data file 1, the location of known HNF4 sites is variable in promoter sequences and may range from a position close to the TSS to up to +10 kb and -11 kb. Under such high variability, it is improbable that the method can 'memorize' the position of TSSs during training since sequences in the training set are not aligned in relation to TSSs.

We then applied the developed methods to analysis of the data derived from ChIP-chip experiments reported by Odom et al. 11 for HNF4α. This study is based on chromatin immunoprecipitation combined with DNA-DNA hybridization on a microarray containing 13,000 human promoter sequences. In the study of Odom et al. 11, the number of HNF4α-targeted promoters was unexpectedly high; 1,575 potential HNF4α target genes in hepatocytes were identified, corresponding to 42% of the genes occupied by RNA polymerase II. Only 48 (3%) of the 1,575 putative HNF4α targets were verified, however, in separate gene-specific ChIP-chip experiments. Additionally, HNF4α DNA binding was not distinguished from protein-protein interactions, as in vitro binding was not analyzed. We applied our algorithm to the proposed HNF4α gene targets and found merely 20% of them to obey the complex computational criteria (the presence of appropriate local and global contexts) that can predict the functional activity of these binding sites. We further stratified our approach by comparing HNF4α functional sites identified by us with independent gene expression data. This comparison shows that our computational approach is versatile and predicts expressed genes directly targeted by the HNF4α TF with similar sensitivity to chromatin immunoprecipitation (ChIP)-chip experiments. In strong contrast, the false discovery rate of the computational method is almost five times lower than that of the ChIP-chip method. This confirms our suspicion that many of the HNF4α binding sites predicted by Odom et al. 11 are functionally neutral, whereas the developed computational method is able to recognize the functionally active HNF4α binding sites based on verification of the local and global contexts of these sites.

Furthermore, in a recent paper by Gupta et al. 44, regulation of gene expression was studied in pancreatic cells of an HNF4α conditional knockout model. Expression analysis identified 133 genes as HNF4α regulated. Regulated genes could be compared with the promoter array data of Odom et al. 11. Surprisingly, the overlap between differentially expressed genes and those bound by HNF4α is rather small. In other words, of 133 genes whose expression was dependent on HNF4α, only 13 have been identified by the location analysis reported by Odom et al. 11. Likewise, of 587 promoters occupied by HNF4α in the study of Odom et al. 11, 574 showed no significant change in gene expression 44. Therefore, 86% of HNF4α-targeted genes proposed by Odom et al. 11 did not differ in gene expression in the absence of HNF4α. These estimates agree well with our computational approach where only 20% of the target genes proposed by Odom et al. 11 could be computationally confirmed.

In the most recent study 45 by the same investigators and through the application of an improved ChIP-chip assay, more then 4,000 HNF4α target genes were identified. By comparing the list of genes identified in the ChIP-chip assay with the list of genes expressed in liver, the authors determined the combinatorial co-occupancy of binding sites of different factors in promoters of HNF4α target genes. Furthermore, this feature correlated well with expression of these genes in hepatocytes. This agrees well with our findings and confirms the utility of our method in defining the local and global contexts for specific combinations of different TFBSs in the vicinity of functionally active HNF4α binding sites of promoters of genes whose expression is regulated by HNF4α. Notably, the combination of PWMs identified by the genetic algorithm (Table 2) captured two TFs, FOX and CREB. Strikingly, these factors were identified independently by Odom et al. 45 in an analysis of TFBSs that co-accrued with HNF4α sites (Table 2, matrices indicated by asterisks).

In a further study of Odom et al. 46, the authors showed that two-thirds of the binding sites identified by ChIP-chip experiments are not conserved between human and mouse. Taking into account the quite conservative liver expression patterns of genes between these two species, we can conclude that by far not all HNF4α binding sites identified by the ChIP-chip method directly contribute to the regulation of gene expression.

To experimentally validate our predictions, we selected two sets of promoters. The first set contained ten ab initio, and therefore novel, HNF4α recognition sites predicted by the computational complex recognition criteria described above. Strikingly, eight of the ten binding sites (NCOA2, TFF2, CHEK1, CD63, SH3Gl2, RND2, ESRRBL2 and DDB1) could be confirmed as HNF4α binding sites in electromobility supershift experiments (Figure 6b). In addition, we studied another set of ten promoters that were reported by Odom et al. 11 as targets for HNF4α, but our computational method rejected them because of extremely low scores for the HNF4 weight matrix as well as low scores for local and global contexts. None of these sites (NPAS2, GPHN, PPP1R3C, AKR1C3, CFL2, MDM2, CLCN3, CBX3, AZI2 and C14orf119) did in fact bind to HNF4α, as shown by electromobility supershift assays (Figure 6c). These findings suggest a high error rate concerning the proposed targets by Odom et al. 11.

Finally, another computational approach has been applied to analyze the same set of HNF4α (as well as HNF1 and HNF6) ChIP-chip data that is the focus of our current study. Indeed, Smith and colleagues 47 demonstrated that an application of combinations of motifs allowed for improvements in the prediction of the genomic location of TFBSs. In contrast to our approach, however, they performed a blind motif discovery instead of using the existing TF weight matrices. To the best of our knowledge, this makes the algorithm very complicated and increases the risk of missing important TF combinations that are characteristic of functionally active regulatory sites.

Several further improvements to our algorithm can be considered in the future. Among the most important, we should consider the possibility of taking into account sequence conservation in the non-coding regulatory regions of genes between different species, for example, human and mouse. It was demonstrated in recent studies 48495051 that sequence conservation can be a good indication of the functional importance of a region. Indeed, such regions can bear functional TFBSs. Despite being quite useful, such considerations should be taken with care since regulatory regions are characterized by a high level of convergent evolution, which can provide non-divergent means of forming the functional context of a TFBS.

Another direction for further improvements is considering known protein-protein interaction data between different TFs. Such data are partially available in databases such as TRANSFAC^®, TRANSPATH^® and BIND. Known interactions between TFs can help to find proper combinations of neighboring binding sites for these factors.

A further step in improving our method will be the use of PWMs for factors whose expression is tissue specific, as indicated in TRANSFAC^®. This will greatly improve the predictive power of the method. To achieve this, more extensive annotation of expression information of TFs is needed and will be a task of the future. One possibility for obtaining this information resides in ChIP-chip experiments in conjunction with gene expression data. This will help to identify TF activity in a given cellular environment.

Taking gene expression data into account will significantly help to determine the global and local sequence contexts and, therefore, functional TFBSs. Recently, we applied the algorithm described in this paper to analysis of promoters of differentially expressed genes 21551. Such an integrative approach is now available in the software system ExPlain™ for a mechanistic interpretation of gene expression changes in eukaryotic cells under various physiological and pathological conditions 52.

We report a new approach based on machine learning techniques for the de novo identification of novel HNF4α binding sites. The genetic algorithms developed by us significantly improved data analysis of various experimental sources. The method described here can be applied to any TF and enables computational prediction of genome-wide functional TFBSs. By applying our method, interactions between different TFs can be taken into account. This provides clues to the mechanisms responsible for promoter activation and even for antagonistic binding of TFs, for example, HNF4α and Coup-TF, which successfully compete for the same binding site but differ in activity under various biological conditions. Indeed, while both factors can bind to the same sequence, the individual local and global sequence contexts determine the actual binding activity and may, therefore, provide an estimate of TF activity in particular cellular or physiological conditions.