Genomes of multicellular eukaryotes mostly consist of DNA segments which do not encode proteins. Still, a sizeable fraction of such non-coding DNA is subject to selective constraint and, thus, is conserved between species. Typically, a long intergenic region consists of alternating segments with high and low rates of evolution 1. A variety of terms have been used to refer to slowly-evolving segments 23, here we will call them CNSs (conserved non-coding sequences).

A majority of mutations in segments which evolve at high rates are presumably selectively neutral or nearly-neutral. In contrast, a large fraction of mutations within CNSs must be deleterious enough to be removed by negative selection. Indeed, data on within-population genetic variability indicate that slow evolution of CNSs is due to negative selection, and not to locally reduced mutation rate 4. In multicellular eukaryotes with compact genomes, such as Drosophila melanogaster, a majority of mutations affecting non-coding sequences may be removed by selection 56. For large-genome organisms, such as mammals, the fraction of selectively constrained non-coding sequences is probably between 3% 7 and ~10% 8.

Obviously, CNSs must perform important biological functions, but the whole range and nature of these functions remains unknown 9. Still, many CNSs are certainly involved in regulation of transcription, and harbor binding sites of a variety of transcription factors 10. Thus, we can expect some short sequence motifs to be overrepresented in at least some kinds of CNSs, as this is the case for proximal promoters 11. Indeed, analyses of samples from human CNSs demonstrated overrepresentation of some short sequence motifs 1213.

New, powerful methods of detecting overrepresented motifs [e. g., 1415], make it possible to undertake the analysis of small-scale composition of mammalian CNSs at the genomic level. Such analysis has a potential to reveal short sequence-specific function(s) common for all human CNSs. Here, we report the results of application of discriminating matrix enumerator (DME) 14 to all strong human CNSs.

We studied representation of short sequence motifs in all human CNSs against three backgrounds: unconserved or only weakly conserved segments of intergenic regions (non-CNSs), near-promoter non-coding sequences, and randomized sequences with the same nucleotide composition as that of CNSs. CNSs are relatively AT-rich 9: frequencies of nucleotides A, T, G, and C are 30.7%, 30.7%, 19.3%, and 19.3% in CNSs, 26.3%, 26.4%, 23.6%, and 23.7% in non-CNSs, and 23.7%, 23.7%, 26.3%, and 26.3% in near-promoter sequences. Dinucleotide compositions of sequences of different classes were also substantially different (Fig. 1).

CNSs from human chromosomes with odd and even numbers were analyzed separately, to check the results for consistency. The overall lengths of CNSs were 27,112,333 on odd chromosomes and 24,962,379 on even chromosomes. Tables 1, 2, and 3 list top 30 motifs, overrepresented within CNSs over these three backgrounds. Overrepresentation was calculated as the ratio of the number of occurrences of a motif within CNSs, normalized to their overall length, over normalized number of occurrences of the motif within the background sequences.

In order to study a possible similarity of the overrepresented CNS motifs with known binding sites for transcription factors (TF), we applied our recently developed method m2transfac 16, and compared all the motifs found at the previous step with the TRANSFAC library of positional weight matrices (PWMs). Relatively few matches between the motifs and the TF matrices were found. Out of 12000 motifs reported at the previous step as being overrepresented in CNS versus the three different backgrounds, we have identified just 20 motifs that match TF matrices with E-values lower than 0.001 and satisfy factor class-specific cut-offs (Table 4). The majority of these matches involved matrices for the factors of "Forkhead DNA-binding domain", especially of the FOX family, which were repeatedly found over two rather different backgrounds: of non-CNSs and randomized sequences. Among the motifs found over the background of near-promoter sequences, there was only one that matched a PWM.

We treated all human CNSs as a single class of sequences. Comparison of this class against three different backgrounds demonstrates that many short sequence motifs are substantially overrepresented within CNSs (Tables 1, 2, 3). CNSs from odd- and from even-numbered human chromosomes show very similar patterns, which is consistent with the lack of any large-scale heterogeneity within CNSs. At a first glance, these results may seem to suggest that CNSs as a whole possess some complex sequence pattern(s), with possible implications for their functioning. However, this is probably not the case. Instead, the results can be explained by simple, generic properties of CNSs.

Indeed, when CNSs are analyzed against a background of non-CNSs (Table 1) or of near-promoter sequences (Table 2), almost all overrepresented motifs possess two common features: (i) they are AT-rich (consist of 75% or more of A and/or T) and (ii) they contain runs of A's and/or T's. Feature (i) simply reflects a well-known, although poorly understood, fact that CNSs are more AT-rich than the genome as a whole 917 or that these two classes of background sequences. Feature (ii) appears to be due to general excess of AA and TT dinucleotides in CNSs, relatively to corresponding random sequences. This tendency of A's and T'e to clump is probably due to patterns in mutation, and not to any functional constraint. Indeed, context-dependence of spontaneous mutation in mammals tends to produce runs of A's and T's, because at a site preceded and followed by A's (T's) T>A (A>T) transversions are ~2 times more common than A>T transversions 1819; Table 2.

Obviously, it is neccessary to consider CNSs against a background of the same nucleotide composition, as otherwise the impact of different compositions is the leading factor causing overrepresentation of some motifs. When CNSs are analyzed against a background of random sequences of the same, AT-rich, nucleotide composition, the results are very different (Table 3), and overrepresented motifs can be naturally subdivided into two classes. The first, larger class contains a variety of GC-rich motifs which, however, are devoid of CpG dinucleotides and are correspondingly enriched with CpA and CpT dinucleotides and with CWG short motif. The second, smaller class contains several motifs which are either purine- or pyrimidine-rich. Overrepresentation of motifs from the first class appear to be due to two simple factors: i) the presence, within CNSs, of short GC-rich segments and ii) hypermutability of CpG dinucleotides 18. Indeed, CNSs are depleted of CpG's more than the other two classes of genomic sequences (Fig. 1), which might reflect strong methilation of CNSs. Overrepresentation of motifs of the second class simply reflects a well-known 20, although poorly understood, abundance of short segments with strong purine/pyrimidine imbalance between the two DNA stands within the human genome.

The analysis of all human CNSs does not reveal clear "global" patterns consistent with overrepresentation of specific, functional motifs. A small number of the observed overrepresented motifs are similar to Position Weight Matrices (PWMs) from TRANSFAC database 21 (Table 4). Among them, the strongest similarity was to the PWMs of FOX and POU families of factors which are characterized by a specific AT-rich pattern. In order to test if the identification of FOX-domain matrices is merely an effect of the general AT-richness of the CNS regions we check carefully results of alignments of all other "AT-rich" matrices in TRANSFAC. There are approximately 64% of matrices in TRANSFAC with overall AT composition higher then 50%. 16 of them are characterised by the same and even higher AT-composition then any of the FOX and POU-domain matrices (e.g. matrices for such factors as TBP, Lhx3, Evi-1, Nkx3-1 and others). Nevertheless, non of them gave statistically significant results of the alignments with the motifs under study. This confirms the similarity of some motifs from the list specifically to the FOX- and POU-domain matrices. The FOX factors are involved in many cellular processes and often control very first steps of organism development as well as cell cycle and differentiation; e. g. FOXF1 is highly expressed in mouse embryonic extraembryonic and lateral mesoderm 22 and control murine gut development 23; FOXD1 is predominantly expressed in embryonic forebrain neuroepithelium, head mesenchyme and adrenal cortex 24 and controls normal brain and kidney morphogenesis and cellularity in the renal capsule 25; FOXO1 governs cell growth in the heart 26. Factors of other families, such as POU and bZIP are often involved in regulation of basic cell cycle machinery; e.g. POU2F1 is an ubiquitous factor involved in stimulation of replication 27 and also participates in early mouse embryogenesis 28. In summary, it might be tempting to speculate that at least some motifs overrepresented in all CNSs may play crucial role in organizing the process of development of the vertebrate organisms. However, the number of such motifs is not high., More specific classes of CNSs, such as those adjacent to genes with a particular pattern in expression 1112 should be considered in order to find a larger number of functional motifs.

In contrast, small-scale composition of human CNSs, considered as a whole, is strongly affected by patterns in mutation – hypermutability of CpG's and the tendency for A's and T's to form runs. This is unexpected because CNSs must be under negative selection which can overcome any impact of mutation 4. Apparently, selective constraint on the evolution of individual nucleoitide site can be quite weak even within strongly conserved CNSs.

Abundance of short sequence motifs in all human CNSs is mostly dictated by their general features: overall AT-richness of CNSs, runs of A's and T's, GC-rich regions, avoidance of CpG's, and local purine/pyrimidine imbalance of the DNA strands. Apparently, CNSs as a whole are too broad a class to display strong overrepresentation of specific motifs. Instead, such motifs must be sought within subclasses of CNSs. In particular, tissue-specificity of expression of the genes adjacent to a CNS must be taken into account.

We used the VISTA pipeline infrastructure 29 with Shuffle-LAGAN glocal chaining algorithm 30 applied to local alignments produced by translated BLAT 31 for the construction of genome-wide pairwise human/mouse alignment. The level of conservation in the alignment was evaluated with the computational algorithm Gumby 32 that makes minimal assumptions about the statistical features of conserved noncoding regions and treating the sequence alignment as its own training set. Gumby 32 proceeds through five steps:

1. Noncoding regions in the input alignment are used to estimate the neutral mismatch frequency pN between each pair of aligned sequences. This is done simply by counting the number of mismatches in nonexonic positions and dividing by the number of aligned nonexonic positions.

2. A log-odds scoring scheme for constrained versus neutral evolution is then independently initialized for the pair of sequences, based on the assumption that the mismatch frequency pC in constrained regions equals pN/R, where the ratio R is an arbitrary parameter. For example, if R = 3/2 (default value), constrained regions are expected to evolve at 2/3 times the neutral rate, until sequence divergence begins to saturate.

The log-odds mismatch score for the sequence pair is then given by S0 = log((pN/R)/pN) = -log(R), and the match score is S1 = log((1 - pN/R)/(1 - pN)). The default R-ratio (1.5) was selected to optimize the sensitivity-specificity tradeoff in detecting empirically defined regulatory elements in the SCL locus. Gap characters in the alignment are assigned a weighted average of mismatch and match scores: SG = pNS0 + (1 - pN)S1.

3. Each alignment column is scored as a sum of pairwise log-odds scores. The resulting conservation score fulfills the requirements of Karlin-Altschul statistics, in that positive column scores are possible, though the average column score is negative 33.

4. Conserved regions appear as stretches of alignment columns with a high aggregate score.

5. The aggregate score of the alignment columns in each conserved region is translated into a P-value using Karlin-Altschul statistics. As is the case with the BLAST algorithm 34, the P-value of a given conserved element varies with the size of the search space, since one is more likely to find a given degree of conservation by random chance in a long alignment than in a short alignment. To make the P-values comparable across alignments of different lengths, Gumby normalizes them to refer to a fictitious fixed-length alignment with the same statistical properties as the true alignment. The 10-kb P-value is related to the expected number of false positives in a 10-kb region (i.e. the 10-kb E-value) as follows: P = 1-exp(-E). When P << 1, P ≈ E. Thus, the P-value also doubles as an estimate of the false-positive rate.

Intervals with P-value threshold of 0.01 produced a set of 144,165 highly conserved sequences that totaled 49 Mb in length. We eliminated all conserved regions that coincide with the coding evidence provided by the UCSC data sets of mRNA, human spliced EST and human EST. We excluded CNSs located within (-1000, +1000) from the start and end of transcription.

Non-CNSs were defined as regions that have human/mouse alignment, conserved below 50% in a 100 bp window and not containing repeats and coding evidences. Random sequences were generated using standard C library pseudo-random generator. Overrepresentation of motifs in different random sequences was calculated using DME 14 (see Additional File 1). DME identifies motifs, represented as position weight matrices that are overrepresented in one set of sequences relative to another set. The ability to directly optimize relative overrepresentation is a unique feature of DME, making DME an ideal tool for comparing two sets. In all of studies we compared 8-mers (parameter w = 8) and bits/column bound was set to 1.6 (parameter i = 1.6).

DME motifs were compared to the TRANSFAC^® database with the m2transfac program 16. The program retrieves all non-overlapping pairwise ungapped alignments of a query matrix and a TRANSFAC matrix satisfying a given threshold. The primary similarity measure is an alignment score which combines Kullback-Leibler divergence with a scoring system that was previously applied successfully to comparison of Hidden Markov Models 35

S(p, q) = C(p, q) - D(p, q)

with

C ( p , q ) = log ⁡ 2 ∑ i = 1 4 p i q i r i MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGdbWqcqGGOaakcqWGWbaCcqGGSaalcqWGXbqCcqGGPaqkcqGH9aqpcyGGSbaBcqGGVbWBcqGGNbWzdaWgaaWcbaGaeGOmaidabeaakmaaqahabaWaaSaaaeaacqWGWbaCdaWgaaWcbaGaemyAaKgabeaakiabdghaXnaaBaaaleaacqWGPbqAaeqaaaGcbaGaemOCai3aaSbaaSqaaiabdMgaPbqabaaaaaqaaiabdMgaPjabg2da9iabigdaXaqaaiabisda0aqdcqGHris5aaaa@48E5@

D ( p , q ) = 1 2 ( ∑ i = 1 4 p i log ⁡ p i q i + ∑ i = 1 4 q i log ⁡ q i p i ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@5FCB@

In equation (2), r is background model which is set to the uniform distribution. Equation (2) is based on the column score derived in 33. The term assigns a positive score to similar distributions and tends towards zero for less conserved positions. Equation (3) is a symmetrized relative entropy or Kullback-Leibler (KL) divergence. Relative entropy was used previously in applications for classification of protein as well as nucleotide patterns 3637. The m2transfac scoring system combines the advantages of both measures. The KL divergence directly assesses the difference of two distributions and therefore increases specificity for similar distributions, but makes no distinction on the basis of their conservation, which is however a property of the column score.

The m2transfac output provides E-values, the number of alignments with greater or equal score expected from searching a database with 1000 matrices. These are derived for each TRANSFAC PWM from score distribution estimates based on large-scale searches of a random matrix library. Furthermore, we apply the transcription factor classification that was developed in our group 3839 to gather matrices according to DNA-binding domain classes of their binding factors and derive factor class-specific score thresholds. We define 57 matrix groups, 15 of which comprise matrices which cannot be associated with a particular factor class, e.g. the barbiturate-inducible element, or whose binding factors are so far not assigned to a protein-structural class. Some matrices occur in more than one class if TFs of different classes are annotated as binding factors, or binding factors possess multiple DNA-binding domain types. For each PWM, score thresholds are defined at three levels of stringency above the score of the first observed false positive in a search of the TRANSFAC database.

SM designed and carried out the computational experiments; PS developed the program and analyzed TRANSFAC PWMs, A Kel provided biological insight and actively participated in discussion of the project and writing the paper, A Kondrashov and ID designed and led the project. All authors have read and approved the final version of the manuscript.