For many years the mouse and human genomes, which diverged from a common ancestor about 65-75 million years ago, have been extensively used for comparative studies 9101112; but it is still an open question as to which species should be added to this comparative analysis to derive the most information content. Among several recent studies providing guidance for selecting additional species is the investigation of the stem cell leukemia (SCL) genomic interval, originally based on a human-mouse sequence comparison 23, and later expanded to include three additional species: chicken, pufferfish and zebrafish 12. This analysis demonstrates that mouse-human alignments show high levels of sequence similarity for all coding exons and for all eight known murine regulatory regions of the SCL locus. Human-mouse-chicken alignments identified the similarity of all coding exons and also discovered protein-binding motifs in five of the known regulatory regions. Thus, inclusion of the chicken DNA sequences allowed for superior functional annotation of a subset of the regulatory regions that had already been identified by the human-mouse comparison.

Pairwise mouse-pufferfish and mouse-zebrafish sequence alignments identified only some of the coding exons, and found similarity for only two of the eight known regulatory regions in the pufferfish comparison; and no significant similarity was found for any known regulatory region in the zebrafish comparison 12. This analysis suggests that comparative analysis of zebrafish and mammalian genomic sequences might be of limited value for the identification of functionally significant noncoding sequences in the SCL region; and these results are consistent with what is expected on the basis of the evolutionary distance of the species analyzed.

The analysis of conservation between Drosophila melanogaster and four other Drosophila species (D. erecta, D. pseudoobscura, D. willistoni, and D. littoralis) that have different divergence times (6-15, 46, 53 and 61-65 million years, respectively) 24 has generated several important conclusions to guide further functional studies of these species 25. One conclusion is that the addition of a third species could reveal functional constraints in otherwise nonsignificant pairwise exon comparisons. All D. melanogaster genes identified in divergent species show evidence of functional constraint; and including more distantly related species defines the exact position of short regulatory elements that are hard to find in the long regions of non-coding sequence conservation observed in closely related species. Non-coding conserved sequences have also been found to be spatially clustered, and these clusters can be used to predict enhancer sequences 25. This work provided a solid basis for choosing species whose genome sequences would be most useful in aiding the functional annotation of coding and cis-regulatory sequences in D. melanogaster: D. pseudoobscura, which has recently been sequenced, was recognized as the best for discovery of functional genomic features, and adding D. willistoni to the comparison allows the dissection of regions of the Drosophila genome under different levels of functional constraint.

Using multiple alignments in the S. cerevisiae and related fungal genome annotation projects 26 provided a powerful demonstration of functional analysis, yielding results that would be difficult to obtain by other computational and experimental methods. A comprehensive comparison of the genome of the yeast S. cerevisiae with those of three related Saccharomyces species (S. paradoxus, S. mikatae and S. bayanus) 26 yielded a major revision to the yeast gene catalog, reducing the total count by about 500 genes. In addition, motif analysis automatically identified a number of genome-wide elements, including most known regulatory motifs and numerous new motifs suitable for biological study.

Another approach to multiple species sequence analysis, 'phylogenetic shadowing' 21, is used for comparison of evolutionarily closely related species. It demonstrates the utility of sequence comparisons within the primate group for discovering common mammalian, as well as primate-specific, functional elements in the human genome, which could not be achieved by comparison of more evolutionarily distant species. Rubin and colleagues 21 showed that the high information content of comprehensive primate sequence comparisons could be captured with a small subset of phylogenetically close primates, such that sequence from as few as four or six primate species compared with humans might be sufficient for the identification of a large fraction of functional elements in the human genome, many of which are likely to be missed by human-mouse comparisons. While the number of multi-species comparative studies grows, it is becoming clear that reasonable selection of species for comparison of a particular genomic interval is still to a large extent an intuitive process, with some guidance from previous successful comparative studies.

Two recently developed algorithms, MLAGAN 1630 and MAVID 3132, are designed for global alignment of both evolutionarily close and distant megabase-length genomic sequences. The MLAGAN 1630 algorithm assumes that the phylogenetic tree is known, as is usually the case for large vertebrate genomes. The program is based on progressive alignment: a multiple alignment of K sequences is constructed in K-1 pairwise alignment steps, such that in each step two sequences, or intermediate multiple alignments, are aligned.

MLAGAN uses LAGAN as the global pairwise-alignment subroutine, and introduces new methods for scoring and refining a multiple alignment. It also aligns the sequences in the order of the given phylogenetic tree. For example, MLAGAN aligns sequences from human, chimpanzee, mouse, rat, and chicken, in the following order: first, human-chimpanzee; second, mouse-rat; third, human-chimpanzee to mouse-rat; fourth, human-chimpanzee-mouse-rat to chicken. Each alignment step merges two sequences or alignments into a larger alignment, effectively building a profile of all the sequences. The results obtained with MLAGAN on the cystic fibrosis (CFTR) genomic region 16, suggest that multiple alignments are better than pairwise alignments at aligning conserved exons between distant species: it was precise enough to refine mis-annotated splicing sites.

MAVID 3132 is a progressive global alignment program that works by recursively aligning the 'alignments' at ancestral nodes of the guide phylogenetic tree. At each internal node, ancestral sequences are inferred from the existing alignments using maximum likelihood, and these alignments are then aligned using the global aligner AVID 33. The multiple alignment is used to build a phylogenetic tree for the sequences, which is subsequently used as a basis for identifying conserved regions in the alignment.

MultiPipMaker 3435 uses multiple pairwise local alignments of secondary sequences against a reference sequence to create a crude multiple alignment that is subsequently refined to generate a true multiple alignment. Analysis of multiple alignments generated by MultiPipMaker 35 allowed for discovery of regulatory elements in the mammalian WNT2 genomic region, and confirmed the phylogenetic inference that horses are evolutionarily more closely related to cats than to cows 17. Alignments between the human sequence and the sequence of each of the other 12 species used in the analysis 17 showed, as expected, that the fraction of sequence that can be aligned generally decreases with increasing evolutionary distance from humans (except for mouse and rat).

Another program, Multiz, developed for large-scale comparison of multiple sequences, takes BLASTZ/axtBest 35 as the pairwise input. This program has been used for the alignment of the mouse and the rat draft assemblies to the human genome 36.

'Phylogenetic footprinting' 37 aims to discover specific protein-binding sites within regulatory regions of multiple sequences on the basis of phylogenetic relationships. It is a method that is mostly applied to promoter regions of orthologous genes. Sumiyama with coauthors 38 attained good results by using multiple sequence comparison combined with a small window size (where the window is the region analyzed in each sub-comparison). This high-resolution procedure can predict the binding sites of transcription factors and reveal polymorphisms in control elements between phylogenetic clades. Phylogenetic footprinting was applied to the Hoxc8 early enhancer region, where it successfully identified a known protein-binding cis-regulatory motif that had previously been analyzed in depth by functional methods 38. The authors demonstrated that an eight-species analysis is clearly superior to the conventional two-species methodology for this type of study.

Another group of specialized phylogenetic footprinting algorithms finds the most conserved motifs among the input sequences, as measured by a parsimony score of the underlying phylogenetic tree 3940. These algorithms have been used successfully to identify a variety of regulatory elements, some known and some novel, in sets of diverse vertebrate DNA sequences. Although phylogenetic footprinting methods show a lot of promising results, their use requires prior information about the location of orthologous regions in genomic intervals of interest. Multiple sequence alignments can help in defining these regions by finding longer conserved regions that can serve as guides to functionally important elements 10.

The most obvious but difficult question to ask in comparative studies is how to define a functionally significant level of sequence conservation between species. Although two-way comparison is effective for discovery of evolutionarily constrained elements, distinguishing them from conserved sequences that are present due to lack of sufficient divergence time is not straightforward and requires knowledge of the neutral substitution rate 1341. In the majority of comparative genomics studies the definition of a significant level of conservation between two species has been intuitive, or based on biological experience. For example, aligning sequences in divergent noncoding regions proved useful in analyzing the enhancer in the β-globin locus-control region 42. The conventional cutoff of 70-75% conservation over 100 base-pairs for the human-mouse comparison has produced discoveries of several important biologically functional elements 1043. One of the major obstacles to applying a single universal conservation criterion for potential regulatory regions is the substantial variation in the underlying mutation rates from region to region 1341. Conservation scores that incorporate the local neutral substitution rate are now available for the human and mouse genomes 13, and they can help to determine if a particular sequence is likely to be functional.

A more detailed analysis of interspecies pairwise genomic sequence alignments, aiming to distinguish regulatory regions from neutrally evolving DNA, has appeared recently 44. This study proposed scoring procedures that evaluate alignments for properties other than overall percentage identity, although highly conserved noncoding sequences have proven to be good indicators of regulatory elements; among these procedures are discrimination on the basis of frequencies of nucleotide pairs or gaps, in combination with scoring procedures that include the alignment context, using frequencies of short runs of alignment columns. This study 44 thus gave a good start for extensive testing of these measures.

Adding genomic sequences from multiple vertebrates to the analysis makes the problem of estimating conservation even less trivial. Expanding pairwise analyses of conservation to multiple sequence alignments would require calculation of the neutral substitution rate between all pairs of sequences. That would give a weighted contribution of each sequence in the multiple alignment, but would also requires much more detailed evolutionary information than is available now. A three-way comparison makes it possible to enrich a pairwise alignment, and a simplified method for calculating a level of active non-coding conservation in such a comparison 45 is based on the supposition that actively conserved human-mouse noncoding sequences are likely to be present in additional mammals, whereas noncoding regions that are similar because of an insufficient accumulation of random mutations will not be present in other mammals.