Opinions on Multiple Sequence Alignment, and an Empirical Comparison of Repeatability and Accuracy between POY and Structural Alignment

The concept of homology is pivotal to Darwin's paradigm of descent with modification. However, in molecular phylogenetics, the process of alignment is often overlooked as a critical step. The data in molecular phylogenetic studies are not individual sequences, but rather, columns of putatively homologous nucleotides, or arguably, reconstructed presumed homology pathways in direct optimization (see Wheeler, 1996). Simplified, alignment is the assignment of homology. Although it would be unthinkable for morphologists to ignore issues of homology, many investigators do not carefully consider alignment of molecular data. Some simply use default parameters with their data in Clustal (Thompson et al., 1994) or some other automated alignment program, perhaps manually deleting “unalignable” regions. This is a mistake, because there are many examples where alignments, and the assumptions that go into them, will produce different trees (e.g., Mindell, 1991; Wheeler, 1995; Morrison and Ellis, 1997; Kjer, 1995, Hickson et al., 2000; Wheeler et al., 2001; Xia et al., 2003; Ogden and Whiting, 2003; Kjer, 2004). Approaches to alignment for phylogenetic studies can be divided into two broad categories; manual alignment and computer-based alignment. A diagram of the types of alignment and their relationship to one another is provided in Figure 1. In a survey of phylogenetic papers in Systematic Biology, Molecular Biology and Evolution, and Cladistics from the last 3 years, we find that 76% of the papers that utilized rRNA were manually aligned (Table 1). Automated methods can be performed in a variety of programs (e.g., Clustal [Thompson et al., 1994], Divide and Conquer [Stoye, 1998], and T-Coffee [Notredame et al., 2000]), and these and other programs have been discussed by others (i.e., Hickson et al., 2000; Gardner et al., 2005; Katoh et al., 2005). Automated alignments that are adjusted by hand should ultimately be considered manual, because the final homology decisions are in the eye of the investigator. Manual alignment can be further categorized as “by eye,” or structural. “By eye” alignments are typically performed without a consistent criterion, and hence are considered by us to be not only subjective, but also phenetic (with nucleotide similarity being the criterion, albeit one without easily demonstrated consistency). Ribosomal RNA structure–based alignments involve the use of published structural models (available from rRNA databases) combined with a search for compensatory base changes (e.g., Kjer, 1995), which provide evidence for conserved helical interactions that are to be aligned together (for a recent historical review see Noller, 2005). In short, computer alignments minimize change among nucleotides, whereas rRNA structural alignments minimize change among rRNA secondary structures. An overview of alignment approaches and their interconnections. Survey of alignment methods from 2003 to 2005 for three high-ranking (based on ISI impact factor) phylogenetics journals. The “Journal” column, on the far left indicated the journals, Systematic Biology, Molecular Biology and Evolution, and Cladistics, each with their ISI impact factors in parentheses. “Papers” refers to the number of molecular phylogenetics papers, and “rRNA” refers to how many of these papers used rRNA for at least one of their markers. “Struct.” refers to manual alignment with reference to rRNA secondary structure, whereas “Adjust” refers to manually adjusted computer alignments. “Manual” refers to alignments only specified as manual. “Comp.” refers to unadjusted computer alignment, whereas “POY” refers to direct optimization with POY. Survey of alignment methods from 2003 to 2005 for three high-ranking (based on ISI impact factor) phylogenetics journals. The “Journal” column, on the far left indicated the journals, Systematic Biology, Molecular Biology and Evolution, and Cladistics, each with their ISI impact factors in parentheses. “Papers” refers to the number of molecular phylogenetics papers, and “rRNA” refers to how many of these papers used rRNA for at least one of their markers. “Struct.” refers to manual alignment with reference to rRNA secondary structure, whereas “Adjust” refers to manually adjusted computer alignments. “Manual” refers to alignments only specified as manual. “Comp.” refers to unadjusted computer alignment, whereas “POY” refers to direct optimization with POY. All organisms have the same basic rRNA structure, yet the nucleotides involved in base pairing can vary a great deal among lineages. Thus, maximizing nucleotide identity as a means of establishing homology ignores a higher order of conservation. The nucleotide states themselves can make very little difference in ribosomal function, but the conservation of higher order structure is imperative. Computer programs like Clustal (Thompson et al., 1994) and POY (Gladstein and Wheeler, 1997) assume that the lowest evolutionary costs involve the retention of nucleotides. In other words, aligning nucleotides together based on state is both parsimonious and algorithmically less costly, and that shorter trees are better trees. However, we argue that this is not justified in structurally conserved molecules in which the states of the nucleotides are less important than the conserved structures with which they are intimately associated. Previous studies have shown that elucidation of a secondary structure model can aid in more accurate phylogenetic inferences (e.g., Kjer, 1995; Hickson et al., 1996, 2000; Titus and Frost, 1996; Morrison and Ellis, 1997; Xia et al., 2001; Gowri-Shankar and Rattray, 2006). Another reason to consider secondary structure is that models that incorporate it are improving. MrBayes 3.0 (Ronquist and Huelsenbeck, 2003) includes a doublet model (Schöniger and von Haeseler, 1994), and the program PHASE (Jow et al., 2002; Hudelot et al., 2003; Jow et al., 2005) contains multiple models for pairing regions, including a seven-state model for RNA base pairs that offers a promising alternative to six-and sixteen-state models (e.g., Savill et al., 2001; Gibson et al., 2005; Gillespie, 2005). Similar reasoning for aligning by structure may be applied to protein sequences, but the codon organization of protein genes make the alignment of protien coding genes a very different process from the alignment of rRNA and is not discussed in this paper. When using default parameters in a phylogenetic study, it is important to consider whether these defaults were designed for protein-coding data or RNAs because the optima between these data sources should be very different (Katoh et al., 2005). All computer alignments require input parameters, and the objective selection of these parameters has been an active area of investigation. Among the most important of input parameters is the cost for inserting a gap in the shorter of two sequences relative to the cost of a substitution (the gap cost-to-substitution cost ratio, or gap cost). Under Needleman-Wunsch algorithms (Needleman and Wunsch, 1970), the computer assigns a numerical score to different alignments, providing points for nucleotides that are the same and subtracting points for inserting gaps (two extremes are shown in Fig. 2). In both panels, the two sequences are identical, except for the addition of a single nucleotide at the arrow. In Figure 2A, if the gap cost is too low, the program will produce a trivial alignment in which the nucleotides on the bottom line up with an identical nucleotide in the reference sequence. If the gap cost is too high, the program will not insert a gap (Fig. 2B). This exaggerated example is meant to introduce the idea that there may be some ideal gap cost to substitution ratio between these two obviously inappropriate extremes. Parameter selection will determine phylogenetic conclusions, and the selection of parameters may be accomplished through examining the sensitivity of phylogenetic conclusions to a variety of alignment parameters (Wheeler, 1995). Hypothetical computer alignment when gap costs are trivially low (A) and prohibitively high (B). Arrow (^) indicates the position of a single indel in otherwise identical sequences. An appealing modification of computer-based alignment would be to simultaneously evaluate alignment and tree building, bypassing fixed hypotheses of homology by reconstructing ancestral sequences on multiple trees in a process now called “direct optimization” (Sankoff et al., 1973; Sankoff, 1975; Sankoff and Cedergren, 1983; Kruskal, 1983). Wheeler (1996) discusses direct optimization (or optimization alignment) and provides an implementation in the program POY (Gladstein and Wheeler, 1997). The optimal tree in POY is that which depicts the minimal number of character transformations (substitutions, insertions, and deletions) between all taxa based on reconstructed ancestral sequences at each node of the cladogram. Hence, the program is rooted in the parsimony optimality criterion and offers a more precise implementation of the Sankoff-Morel-Cedergren method (Felsenstein, 2004). Similar methods of simultaneous alignment and tree estimation exist under Bayesian sampling (Mitchison, 1999) and distance methods (Hein, 1989, 1990). Redelings and Suchard (2005) provide a Bayesian method to simultaneously explore the joint posterior distribution of alignment and phylogeny, with software available at http://www.biomath.ucla.edu/msuchard/. Holmes (2004) discusses a probabalistic model of structural RNA evolution, following the model of Thorne et al. (1991; TKF91 model). Lunter et al. (2005a) also used the TKF91 model to provide a joint analysis of alignment and phylogeny under a Bayesian statistical framework, and Lunter et al. (2005b) discuss likelihood approaches to direct optimization. POY also offers a likelihood implementation of direct optimization (not considered here) that considers indels as independent 5th states (Wheeler, 2006). Gorodkin et al. (1997), Mathews et al. (2002), and Perriquet et al. (2003) provide programs for predicting and aligning structural motifs. Although these approaches may hold we to the parsimony implementation of POY program because it by the most used method of direct and many of the other methods are by a model of and on the number of taxa that can be will on the of indel models (i.e., et al., et al., 2005). has been the relative of structural alignments and automated alignments. as and Wheeler provide for but do not the most approaches one another two recent studies POY to methods and do not the relative and of both approaches et al., 2001; and A recent of and by et al. (2005b) some of POY when secondary structure is not used to the a of both methods not The in alignment approaches from both and alignments, both manual and require and Computer alignment analysis with to among a of input Manual alignment can of at the computer An ideal would be to a number of data each with taxa to of independent investigators for alignment by both manual and automated these would to which is more in of of phylogenetic trees. The of at least some among the taxa in these data would provide an that be as as is to evaluate alignment we make an ribosomal RNA sequences for were from taxa were because we consider the among to be at node (for review see Hudelot et al., and the for the tree (Fig. are in a combined analysis of The taxa were and their order with only the were aligned by each of the following the structural of et al. and Hickson et al. using published structural models from the RNA et al., as a aligned nucleotides were as and with the program et al., also each a phylogeny from POY with the following and an example to Kjer, which for the between their The from to and were and are available on the Systematic Biology and The also contains for the alignment process from analysis of the with with a The from a parsimony analysis of the structural alignments are shown in Figure All three alignments identical The from the POY are shown in Figure All three of structural alignments at some (see or as can be in the different and tree However, all three structural alignments the same tree only at a single node with and a (Fig. for most from a manual alignment of the rRNA data. with and by Kjer, and Gillespie, is a of two trees. the are from POY of the data. and by different in the left of each tree and by the gap cost and the tree The POY were of us on the same parameters or the same tree (Fig. performed different gap costs to change not to a higher gap cost with an gap trees and both gap costs and gap and to the variety of trees from the different input parameters, not to one tree it important to one tree and a of between the and The area of that the gap ratio, the cost (the cost of the of an at that of the gap The as the gap cost The to explore the cost between and the gap ratio at which the from the of of us different costs or of the other input available in POY. different investigators the same tree using at least not in this If selection is a of phylogenetic conclusions from and if the selection of these parameters is it would be that phylogenetic conclusions are less among investigators they are these parameters in is a similarity between alignments and input parameters in a computer alignment. The to a computer alignment from data when the parameters is more than an analysis from a available of a manual alignment. that this is not is meant when the is that manual alignments are not If the alignment is the analysis is that manual alignments are not from the likelihood that different investigators produce alignments that are not manual alignments were not identical, but were the POY parameters like the gap cost in their assignment is in this is to a sensitivity analysis on a variety of parameters and each tree some a tree based on for or between have been et al., 1995; Wheeler and Wheeler, Wheeler et al., but et al. (2005) that of the they were analysis Wheeler, 1995; et al., 1997) designed to a variety of parameters for phylogenetic including gap are with sensitivity analysis as it is most Wheeler and et al. only the gap cost the are many more parameters that a as the cost to a gap or the for gaps at the and of a or how an RNA is into All these parameters be by the and a search the simultaneous optimization of all parameters that phylogenetic If sensitivity analysis to to an of parameters, whether or not investigators would up with parameters has been and on some on an we provide an example of how three investigators to different conclusions based on different et al. (2005) to conclusions sensitivity but with different (2004) discusses the with using among in selection among alignment and (2003) argue for the of we with them, and the of sensitivity to to of However, when the same parameters were POY not produce the same performed the analysis three with all parameters the same and a tree of and a different tree of (Fig. these it is that the search with that performed not However, performed in a search for optimal parameters, and with only it a of performed and the trees. The number of performed in POY is another important to be Figure the of a sensitivity gap costs and than a with an that indicates the for these parameters, we see a with a single at gap the for on the with the other This that with this data it may not be to be to an one to it which is to do when there are an number of we do not if the at is but it the among the we sensitivity (e.g., Wheeler, 1995; et al., 1997; and Whiting, 2005) have also optima in and (2005) a sensitivity analysis with POY to less optima than a sensitivity analysis with using POY. of gap cost the of the from sensitivity with tree from POY of the and combined and rRNA data. Wheeler and and Wheeler make the that manual alignment a assumptions gap The idea that there is a for gap for a is at the very of the sensitivity analysis states that if a vary in their for fixed gap if it were would be inappropriate for some regions. this gap to that were aligned in one in a in other regions. is most in of rRNA or at the of (see et al., Hickson et al., 1996; Gillespie, but not in conserved regions. Figure provides a example of how a of fixed gap the of parameters from one of to another without the of the all of nucleotide into and all of gaps into we used a model to on the tree in aligned with The that the in in Figure the of a that the among for and the of gap This points a with the idea of as it is most applied in search of optimal parameters that be by a of fixed gap In using an that parameters some it is that at least some (or of the are This is not the with that use a variety of fixed gap If all are an of many inappropriate methods not to a This can be in Figure The taxa are whereas the bottom taxa and of the rRNA et al., can be among with of is all et al., the of is gap costs this of rRNA should vary from one to Manual alignment and gap these gap costs are This of rRNA can be from for of taxa to see the same of which is of Hypothetical how the gap cost may which individual are without the number of alignment of the et al., 1999) of the rRNA and rRNA from the of and and is at the that the of is all of these are if gap are with parameters, yet vary The of gap cost to change is not as as it An would be the implementation of a in which gap costs at each were to However, in in reasoning that is in that we as the between conserved and are and structurally gap cost to change are and Kjer, 1995; Wheeler, 1996; and et al., 2000; Hickson et al., one would that two investigators an identical data with POY will up with different parameters and Ellis, 1997; Hickson et al., in aligning with an a to use POY in that may be less consistent among investigators and more than published manual alignments. In if rRNA is into a conserved secondary structure, different investigators are more to to phylogenetic conclusions because they are all using a homology criterion that is not of the critical with POY is that it on the of ancestral states using et al. provide evidence from both and data that parsimony a of reconstructing ancestral states in the of nucleotide substitution provided a of the of character states on ancestral is POY there be of nucleotide and more to the of tree in parsimony However, we are not computer in The of and ancestral state are for POY in to how it the data rRNA regions, because these different and from the of the in rRNA are to by substitution et al., This is further in the analysis of from to an nucleotide in the (i.e., and and in the of the of some taxa (i.e., et al., 1995, 1996; et al., nucleotide provides another to automated alignments. If homology be automated of data will organisms together to nucleotide a phenetic in Figure nucleotide a in the of the sequences, the of regions. This tree should not be used as a of alignment In order to whether automated are organisms to nucleotide of et al. should be to see if there is a between nucleotide similarity and can be with structural alignments, in that can be and or with analysis (Wheeler, et al., are two to the of gaps in phylogenetic are often as or as a are not data. they are not character but of the alignment, they are the of that can be homologous and often is for and and Wheeler, However, as 5th state or as independent in a POY analysis is a very Figure offers an example of this is Wheeler that we gaps as individual independent and in a parsimony analysis the same as the ratio between gap costs and change that the alignment in Figure by using a gap ratio of to we should consider independent to the between and with gaps as a 5th there is to how many indel were for the difference between and is more parsimonious to assume a of not a make by these by the gap cost to change ratio, we find the of independent and all from have been the of a single by or two the these should is to the of the The is in POY with the of the cost for gaps that are with gaps (the but the cost is the of of multiple nucleotides will the tree and from regions. et al. (2005) that the of gaps with the gap costs in POY in all other have with this (e.g., et al., 1995; et al., 1995; Kjer, 1995; and 1996; et al., 1996; 1997). and discuss gap coding at and provide for A of fixed state not discussed is in POY (Wheeler, papers et al., the of aligned through a called alignment 2001; see Fig. 1). are used to regions, which are as Wheeler, in POY or as states in alignments et al., these are an optimal alignment is use of structural can these further into that should these methods 2004). The implementation of gap costs in POY may be a promising et al., but it would be to the of of an alignment of an In a a of may be but it is important to both minimize and to it where it we find it to structural alignment because it is data into combined POY because we we that two morphologists are to up with identical data to the implementation of methods that are justified on of and assignment of parameters is from the that gap costs are not and and the of parameters will vary among The that alignments may is to a variety of published of structural alignment, including Titus and et al. Hickson et al. et al. et al. and Morrison Morrison and (1997), Xia et al. and et al. Previous of alignment methods involve one automated method like Clustal et al., 2001; and Whiting, 2005) or (Wheeler, POY. et al. manually adjusted Clustal alignments with POY methods without data and the POY trees to be as by likelihood score or tree alignments to be a of these In their of the of computer alignment Hickson et al. demonstrated that alignments that can phylogenetic trees can be using programs and different alignment from a that are for data and for of the same data Hickson et al. the use of secondary structure alignment and is a aid to determine rRNA alignments that are alignment of rRNA a in to the a from the or to homology. optimization will when we have a model of how and as as their and et al., POY not a model of indel are other programs to with structural alignment. algorithms (e.g., Mathews et al., et al., 2003) can be utilized to the search for structures that be with However, of these algorithms is (see et al., as their often with the of structure methods et al., 1996; et al., that have been by recent structures of the (e.g., et al., 2000; et al., 2000; et al., Some alignment programs incorporate structural into the alignment (e.g., et al., 1997; Gorodkin et al., 2001; et al., most programs conserved base or interactions 2004). to use these programs as that are to manual in and of rRNA alignments. find the approaches of et al. Holmes et al. and Redelings and Suchard (2005) to be do not manual alignments simply for their but because they us that in both and to POY. have not yet the or the of other that manual alignments will be by computer methods that them, it is that this is a in the data are in and on the and are we that data should be with the available methods at the they are that they can be with methods in the if and when these methods The use of structural and POY are not a structural alignment for input into with each conserved and aligned one structural from the manual alignment, and POY to evaluate the aligned (see Fig. 1). In this is some studies have and 2000; by rRNA into and we that be by manual alignment to structural as applied et al., may be that some or all of us not use POY a that a analysis with to a manual alignment. This is of the with a of with too to for However, we are not of published on how a POY analysis should be Although we the of one we can that the or that computer alignment is objective and manual alignment is is require Manual alignments are and may be if structure is not or the sequences are in and substitution if the alignment is hypotheses of homology can be and methods would from on how to for structural alignment can be in and and now a has been that provides a that and will a in and are or with alignment will produce better are for the critical and by and phylogenetics and and by and an this from and