Inlernalionol Journal/or Parasilo/og.y, Vol. 26, No. 6, pp. 589417, 1996 Copyright ~0 1996 Australian Society for Parasitology. Pubhshed by Elsevia Science Ltd Pergamon Printed in Great Britain PII: SOO20-7519(96)00044-6 002&7519/96 $15.00+0.00

INVITED REVIEW Phylogenetic Tree-building

DAVID A. MORRISON

Molecular Parasitology Unit, University of Technology Sydney, Westbourne Street, Gore Hill, NS W 2065, Australia

(Received 13 October 1995; accepted 3 March 1996)

Abstract-Morrison D. A. 1996. Phylogenetic tree-building. Iniernationul Journal fir Parasitology 26: 589-617. Cladistic analysis is an approach to phylogeny reconstruction that groups taxa in such a way that those with historically more-recent ancestors form groups nested within groups of taxa with moredistant ancestors. This nested set of taxa can be represented as a branching diagram or tree (a cladogram), which is an hypothesis of the evolutionary history of the taxa. The analysis is performed by searching for nested groups of shared derived character states. These shared derived character states dehne monophyletic groups of taxa (clades), which include all of the descendants of the most recent common ancestor. If all of the characters for a set of taxa are congruent, then reconstructing the phylogenetic tree is unproblematic. However, most real data sets contain incongruent characters, and consequently a wide range of tree-building methods has been developed. These methods differ in a variety of characteristics, and they may produce topologically distinct trees for a single data set. None of the currently-available methods are simultaneously efficient, powerful, consistent and robust, and thus there is no single ideal method. However, many of them appear to perform well under a wide range of conditions, with the exception of the UPGMA method and the fnvariants method. Copyright 0 1996 Australian Society for Parasitology. Published by Elsevier Science Ltd.

Key words: Phylogeny; evolution; cladistics; cladograms.

INTRODUCTION English of Hennig (1966). These methods are now usually referred to as “cladistics”, and the evolution- If we accept the proposition that evolution exists, ary diagrams they produce “cladograms”, to distin- then meaningful comparisons among organisms must guish them from all prior phylogenetic studies, many ultimately include a phylogenetic context (Maddison of which were neither explicit nor repeatable (note & Maddison, 1992). This is because the evolutionary that I am using the word cladistics to include a wide relationships among a group of taxa constrain any class of explicit and repeatable phylogenetic analyses, other possible relationships that might exist. It is thus which may be a broader definition than would not surprising that in biology there has been a long be accepted by many phylogeneticists, who would history of attempts to deal with the reconstruction of restrict the term to the more strictly Hennigian genealogical history (Nelson & Platnick, 1981; Mayr, methods). The current interest in cladistic analyses 1982; Stevens, 1994), notwithstanding the difficulties has inevitably led to a proliferation of data analysis associated with producing testable hypotheses about techniques; and the apparent plethora of methods for historically unique events. phylogenetic analysis is typical of a young science It is, however, only in the last 30 years that still coming to terms with both its aspirations and its widespread attempts have been made by systematists constraints. to produce explicit and repeatable methods for It is therefore important for practitioners to under- the construction of phylogenetic trees (Felsenstein, stand the limitations of the available techniques as 1982), notably with the translation from German into well as to appreciate their capabilities (Felsenstein, Tel: +61-2-3304159; Fax: +61-2-3304003; 1982). Unfortunately, the current wide choice among E-mail: [email protected]. possible phylogenetic methods seems to be daunting

589 590 D. A. Morrison

for many people, and they thus acquire little knowl- PHYLOGENETIC ANALYSIS edge of the relative advantages and disadvantages of the various methods. This is unfortunate, because the Cladistics choice of data-analysis methods should be based on The study of evolutionary processes has often been their apparent appropriateness for the data at hand, considered to be unscientific because it deals with rather than on the local availability of computer historically unique events (Popper, 1957). Hypoth- programs or on historical inertia (Hillis, Allard & eses concerning these events are thus not universal (in Miyamoto, 1993). It is my purpose here to review the either space or time) and, therefore, they are consid- phylogenetic inference methods that are currently ered to be untestable in the contemporary world. The available, and to indicate what is currently known sociological development of phylogenetic analysis about their strengths and weaknesses. This will allow has consequently been based largely on the erection a more informed decision to be made when assessing of what have been called “evolutionary scenarios” which of the methods might be appropriate for a describing the presumed genealogical history of particular data set. the organisms under study. The number of such Along the way, I will attempt to make clear some scenarios that may be created is, of course, limited of the aspects of phylogenetic analysis that are obvi- solely by the imagination of the researcher, and none ously misunderstood by non-specialists, and to dispel of the scenarios are likely to be open to falsification. a few widely-held misconceptions. My discussion will Cladistic analysis can thus be seen as an attempt to focus on molecular sequence data (particularly DNA base phylogenetic analysis on a more objective foot- and RNA), since trees derived from this source are ing, where the phylogenetic hypotheses are explicitly increasingly those with which parasitologists are stated, along with the evidence supporting (and con- working (e.g., Nadler, 1990), there being a limit to tradicting) them, and are then subjected to quantita- the usefulness of phenotypic characteristics for tive testing. Its practitioners therefore claim that constructing phylogenetic trees for most parasites. cladistics is designed to make phylogenetic analysis Indeed, it is the proliferation of molecular data sets into an hypothetico-deductive science, where explicit that is introducing many people from outside of hypotheses are subjected to repeatable attempts at systematics to the science of phylogeny (Miyamoto & falsification. Cracraft, 1991; Hillis, Huelsenbeck & Cunningham, Note that the claimed advantages of cladistic 1994a); and it is for this reason that phylogenetic analysis are not intended to denigrate pre-cladistic principles need to be clearly introduced to non- biologists, nor is there any suggestion that these experts. Furthermore, many of the recent advances in biologists did not apply their minds to phylogenetic cladistics have been motivated by attempts to deal questions. However, it is clear that pre-cladistic phy- with problems that are specific to molecular data logenetic analyses were not necessarily based on (Penny et al., 1990), and the similarities and differ- repeatable methods that produced explicit hypoth- ences between traditional and molecular phylogeny eses of evolutionary relationship which could be thus need to be emphasized. subjected to falsification (Nelson & Platnick, 1981). I do not intend to describe the tree-building Furthermore, the taxonomic groups produced by techniques in any great detail, partly because most pre-cladistic biologists were not necessarily mono- of them are covered in the excellent review by phyletic (see below), and therefore did not always Swofford & Olsen (1990) and in the books by Nei reflect evolutionary history. Post-Darwinian biolo- (1987) and Li & Graur (1991), and partly because gists have had the unenviable task of producing nothing is more intimidating to most biologists than taxonomic schemes that should, in theory, reflect mathematics. I concentrate much more on those evolutionary history, without any theoretical frame- aspects of phylogenetic analysis that are likely to be work for how they should go about discovering what of most practical and theoretical interest to non- the evolutionary history actually was (Stevens, 1994). experts. I start by summarizing the principles of Cladistics is an attempt to provide this theoretical cladistic analysis, before proceeding to a review of framework. the alternative methods for constructing clado- Cladistic analysis as an approach to phylogeny grams. A number of introductory reviews covering attempts to group taxa on the basis of their ancestry. similar topics, but from different perspectives, in- In the analysis, taxa are grouped in such a way that clude those of Felsenstein (1988), Olsen (1988) those with historically more-recent ancestors form Sneath (1989), Beanland & Howe (1992), Hillis groups nested within groups of taxa with more- et al. (1993) and Stewart (1993); and there is also distant ancestors. The analytical technique is the worthwhile introductory book by Forey et al. based on a widely-held view of the mode of the (1992). evolutionary process: species are lineages undergoing Invited Review 591 divergent evolution with modification of their for Character 16 in Table 1 possession of denticles, intrinsic attributes, the attributes being transformed dermal scales, epidermal scales, feathers and hair are through time from ancestral to derived states. Thus, all considered to be homologous character states of a if a species is a group of populations, then if 1 group single character. In practice, characters and their of populations acquires a new (derived, advanced or states are postulated as homologous on the basis of apomorphic) character state while the rest do not (i.e. their structural, positional, ontogenetic, composi- they retain the ancestral, primitive or plesiomorphic tional and/or functional correspondences; and they state) then these populations constitute a new species. are postulated between different taxa so as to maxi- This new species then forms a separate historical mize the number of one-to-one correspondences of lineage that diverges from the other populations, and their parts (Stevens, 1984). The choice of characters maintains its own historical tendencies and fate. to be included in a cladistic analysis may be some- Evolution can thus be thought of as a branching what arbitrary, but can include intrinsic attributes sequence of historical lineages, and indeed the only such as morphology, anatomy, embryology, behav- diagram in Darwin (1859) represents precisely such a iour, physiology, ultrastructure, cytology, biochemis- branching sequence. try, immunology, ecology and even geography. The When a character changes from an ancestral state important points are that the characters used in the to a derived state in a lineage it is historically unique analysis are hypothesized to reflect the evolutionary (a novelty), and it will be passed on to all of the history of the taxa, and that the character states of a descendants of that lineage (even if the character is single character are hypothesized to have a unique later modified into something else). Therefore, the evolutionary origin. branching sequence of evolution can be deduced by Concepts of homology are often intuitively obvi- searching for nested groups of shared derived char- ous when dealing with, for example, morphological acter states (synapomorphies) among the taxa being data, and these concepts can be put into practice analysed. So, if a derived character state is observed through a detailed study of ontogeny (e.g., the in 2 or more taxa, then we can hypothesize that they homologies postulated in Table 1 are thus mostly share this apomorphy because they are descended unproblematic); and the alignment of molecular from a common ancestor, and that they inherited the sequences is the direct equivalent of these homology apomorphy from that ancestor. The possession of a assessments (Mindell, 1991). The concepts of hom- shared ancestral state (a symplesiomorphy) tells us ology in molecular and morphological studies are nothing about the phylogeny of the taxa, since this thus fundamentally the same (de Pinna, 1991; state was inherited from an ancestor that is also held Williams, 1993). So, alignment of molecular in common with those taxa possessing the derived sequences involves a series of hypotheses of hom- state. Thus, cladistic analysis is simply the search for ology among the taxa, with 1 hypothesis of homology nested sets (a hierarchy) of synapomorphies among for each position (nucleotide or amino acid) in the the taxa. Each synapomorphy represents an ancestral sequence. It is important to recognize this, because lineage that has diverged from its related lineages, it is clear that often very little attention is paid thus being contemporary evidence for a prior by molecular biologists to this point - to a evolutionary event (Table 1). Clearly, the only morphologist the assessment of homology is an characters that are of use for a cladistic analysis of a important (and time-consuming) component of phy- group of contemporary taxa are those features that logenetic analysis, but to a molecular biologist the reflect their evolutionary history. There is, of course, assessment of homology is apparently often an after- no simple method for determining which features thought. The correct formulation of hypotheses of these are, but at least the method forces the prac- homology is just as important for molecular data as it titioner to be explicit about the characters that have is for morphological data; and it is clear from the been chosen for the analysis. literature that many of the so-called controversies about the phylogeny of particular groups of para- sites are based as much on disagreements about Homology sequence alignment as on disagreements about actual For this analytical technique to work, homologous evolutionary events (e.g., Ellis & Morrison, 1995). rather than analogous character states must be com- Unfortunately, for molecular data there is pared across the taxa (Hall, 1994). That is, for all of little possibility of further investigations (such as the taxa we must compare like with like, particularly ontogeny) to assess homology, and so in practice with reference to the evolutionary origin of the homology assessment is very different for molecular attribuies. Characters and their states are thus hy- studies (Mindell, 1991). Positional homology can potheses of evolutionary homology. As an example, be represented by either identical character states Table I-Some phenotypic data for extant vertebrates that might be useful for reconstructing their evolutionary history

Character Lampreys Sharks Teleosts Lungfish Frogs Salamanders Turtles Lizards Snakes Crocodiles Birds Mammals

1 Internal skeleton yes yes yes yes yes yes yes yes yes yes 2 Jaws no yes yes yes yes yes yes yes yes yes yes yes 3 Ossified skeleton no no yes yes yes yes yes yes yes yes yes yes 4 Internal nostrils no no no yes yes Yes yes yes yes yes yes yes 5 Atria1 septum no no no yes yes yes yes yes yes yes yes yes 6 Four limbs no no no no yes yes yes yes yes yes yes yes 7 Teeth pedicillate no no no no yes yes no no no no no no 8 Amniotic egg IlO no no no no no yes yes yes yes yes yes 9 Temporal fenestrae none none none none none none none two two two two one IO Hemipenes no no no no II0 no no yes yes no no no 11 Suspensorium streptosylous no no no no no no no yes yes no no no I2 Antorbital fenestrae no no no no no no no no no yes yes no 13 Lateral fenestrae ossified no no no no no no no no no Yes yes no I4 Gizzard no no no no no no no no no yes yes no I5 Homeothermy no no no no no no no no yes yes Em al dermal Ermal smooth smooth epidermal epidermal epidermal epidermal 16 Body covering scale-less denticles scales scales epidermis epidermis scales scales scales feathers hair

These 16 characters each have at least 2 character states that are hypothesized to be homologous, one of which is postulated to be ancestral (the states “no”, “none”, and “scale-less”). The other (one or more) states are postulated to be derived from this ancestral state, and the taxa that share these derived states are then hypothesized to have inherited these states from a common ancestor. These shared derived character states (synapomorphies) thus group the taxa into sets that reflect their evolutionary history. Note that Character I has only I state in this data set, because the derived state has been inherited by all of these taxa, and it thus forms a synapomorphy for the group (vertebrates) as a whole. Furthermore, Characters 9 and I6 have more than 2 homologous states, thus implying that the ancestral state has been modified more than once during the evolutionary history of these taxa; each of these derived states forms its own set of taxa. A possible cladogram derived from these data is shown in Fig. I. The data are from Romer (1971) and Maddison & Maddison (1992). Note that this non-parasitological example has been chosen because the homology assessments and the resulting cladogram are not controversial, which is something that cannot be said for any parasitological example known to me. Invited Review 593

(nucleotides or amino acids) in all sequences, substi- very little is known about their theoretical or practi- tutions in 1 or more sequences (representing point cal limitations (Thorne & Kishino, 1992; McClure mutations), or insertions/deletions (indels) in 1 or et al., 1994). However, 1 important generalization more sequences. The most problematic aspect of may be made: the differences between the alignments sequence alignment is the positioning of indels, and produced by the various algorithms is often less than this problem becomes more acute for more divergent the differences produced by varying the “gap sequences. It is worthwhile in this context to distin- weights”. These weights refer to the relative cost of guish between gaps, which are introduced into inserting a new gap into a sequence or extending the sequences by the alignment procedure, and an already-existing gap, and there is no way of deter- indels, which are actual mutation events (Olsen, mining analytically what these weights should be 1988); clearly, the objective is to introduce into the (Rinsma-Melchert, 1993). Most of the computer pro- sequences only gaps that truly represent indels. grams that implement the alignment algorithms have There are 3 possible scenarios for the degree of default values for the weights that are designed to difficulty of sequence alignment. Firstly, there may be produce “biologically interesting” results, and very relatively few indels, in which case a robust sequence few molecular biologists seem to be willing to deviate alignment can usually be produced by hand. Such a from these default choices. It is, however, clear that situation is shown, for example, by protein-coding to simply report that a particular computer program parts of mtDNA (Miyamoto & Cracraft, 1991) and was used to align the sequences is meaningless (since the plant rbcL gene (Chase et al., 1993). Secondly, the the work cannot be verified) unless the weight values sequence may represent a molecule for which there is are also reported (Wheeler, 1995). an a priori biological model of secondary structure in In dealing with the problematic nature of sequence which certain active sites must be maintained; the alignment, molecular biologists often delete parts of alignment is then constrained by the base-pairing of their sequences from the cladistic analysis. The the model (Olsen, 1988). Such a situation is shown, rationale for this is that those parts of the sequences for example, by the 5s (Specht, Wolters & Erdman, that cannot be aligned reliably should be excluded 1990), 12s (Gutell et al., 1985) and 18s (Van de Peer from the estimation of the phylogeny (Olsen, 1988). et a/., 1994) rRNA genes. This is probably very sensible (Smith, 1994), but Thirdly, there may be many indels and no a priori unfortunately there is often no objective criterion structure model. Under these circumstances it is given for deciding which parts of the alignment are usual to use a mathematical algorithm to produce the ambiguous, the decisions usually being made by alignment. These algorithms all attempt to produce a “visual inspection”. Gatesy, DeSalle & Wheeler sequence alignment that optimizes some chosen cri- (1993) have suggested that those parts of the align- terion of match between the individual sequences ment that are sensitive to the gap weights (i.e. where (cost). That is, the sequences are compared using a the alignment varies significantly when the gap pattern-matching process that searches for corre- weights are changed) may constitute unreliable spondence between the elements of the sequences, hypotheses of homology and may therefore be can- introducing gaps into the sequences as required to didates for exclusion; and Ellis & Morrison (1995) maximize some criterion for optimality of the corre- have shown that for some organisms it is the double- spondence. There are many algorithms currently stranded parts of rRNA that may contain most of the available (see Waterman, 1989; Doolittle, 1990; phylogenetic information when the sequences are Chan, Wong & Chiu, 1992; and McClure, Vasi & aligned according to secondary structure. Thus, there Fitch, 1994), which optimize a variety of mathemati- is considerable room for further research into the cal functions measuring the overall alignment cost. problems of sequence alignment. When there are more than 2 sequences, most of these It is also important to recognize that there is also algorithms use exact procedures (see below for a a more general level of homology assessment for definition) to align the sequences pair-wise, but then molecular data. The sequences being compared must use heuristic procedures (see below for a definition) themselves be homologous rather than analogous to braid these alignments into a multiple alignment. (Fitch, 1970). Thus, only orthologous sequences will Thus, these procedures do not guarantee to produce reflect the historical relationships of species, while the globally optimal alignment; nor do they guaran- paralogous sequences (e.g., the 2 sequences that tee that the optimal alignment (even if they could find result from a gene duplication) will reflect only it) represents the true alignment (Thorne & Kishino, gene history, xenologous sequences (e.g., recently- 1992). incorporated sequences such as result from horizon- I am not going to discuss these various algorithms tal gene transfer) will only partly reflect gene history, here, because this is an area of active research and and plerology (e.g., the inter-mixture of exons and 594 D. A. Morrison

introns) will only reconstruct a composite gene Polarity history (Patterson, 1988; Williams, 1993). All Having determined the homology of the character molecular studies thus rely on the assumption that states, the key to cladistic analysis is the distinction the sequences from the taxa being compared are between derived character states and ancestral states orthologous, and this may be a dubious assumption (character polarity). It is important to note that this for distantly related taxa (Sneath, 1989). Any is a local concept that applies only to a particular sequence for which orthology has not been estab- set of taxa. By this I mean that a character state lished should be omitted from the analysis (Olsen, is only considered to be derived relative to a speci- 1988). fied ancestral state, and it may well be the ancestral Furthermore, for orthologous sequences there is state for a further derived state. As an example, the implicit assumption that the sequences being for Character 16 in Table 1 possession of epider- compared are actually from the organisms being ma1 scales is a derived state relative to possession studied, rather than from some other co-habiting of dermal scales, but is an ancestral state rela- organism. For example, it is now recognized for the tive to possession of either feathers or hair. ITS gene that many of the published sequences Thus, possession of epidermal scales is a synapo- purporting to be from conifers and ferns are actually morphy for Turtles+Lizards+Snakes+Crocodiles+ those of fungi, while the published Mimulus Birds+Mammals, while possession of feathers is a (monkey-flower) sequences are actually those of synapomorphy for Birds and possession of hair is a green algae (P. Weston, personal communication). synapomorphy for Mammals; this character does not As a tinal observation on molecular data, it is supply a synapomorphy for the grouping only of worth emphasizing the distinction between species Turtles+Lizards+Snakes+Crocodiles. Possession of trees and gene trees. The result of cladistic analysis of epidermal scales is thus a synapomorphy at a more molecular data is a gene tree (provided that the entire general level than is possession of either feathers gene has been sequenced), hypothesizing relation- or hair. There is thus a hierarchical relationship ships among the genes or genomes that have been between ancestral and derived character states, sampled, whereas a species tree reflects the actual and recognizing synapomorphies therefore involves evolutionary pathways (Pamilo & Nei, 1988). The determining the correct level of generality of gene tree may be fundamentally incongruent with homologies. the true species phylogeny, with the genome tree, Clearly then, the success of a cladistic analysis and with other gene trees (e.g., Cao et al., 1994; rests on the correct determination of the relative Cummings, Otto & Wakeley, 1995), due to various apomorphy of the character states, and numerous phenomena such as allelic polymorphism, intro- criteria have been proposed for doing this (Crisci & gression, lineage sorting, unequal rates of speciation Stuessy, 1980; Stevens, 1980; Bryant, 1992; Nixon & and gene mutation, lateral transfer, hybridization, or Carpenter, 1993). Most of these criteria rely on mistaken orthology (Pamilo & Nei, 1988; Penny illogical arguments or on assumptions that are either et al., 1990; Doyle, 1992; de Queiroz, Donoghue & false or untestable; and it is worthwhile recognizing Kim, 1995). Thus, for the reconstruction of phylo- only 2 objective possibilities: the direct method; genetic history, a single gene may be, in practice, no and outgroup analysis. These are complementary more useful than a single morphological character methods, and both may thus be used in any 1 data (Doyle, 1992). It may, therefore, be unwise to assume set. For the direct method, the hierarchical relation- tacitly in molecular studies that the number of ship between ancestral and derived character states is characters being used to reconstruct a phylogeny observed directly, and it does not require any pre- is equivalent to the number of positions in the existing hypotheses of character polarity. On the sequence. Furthermore, a living organism is an other hand, the outgroup analysis method does not integrated functioning whole, not just a collection of directly observe character polarity, and it relies upon unrelated genetic attributes. Thus, an organism is a an hypothesis of the relationship of the taxa under collection of interactions between genes, and between study to their near relatives. genes and their environment (i.e. a phenotypic The direct method (Weston, 1988) states that: if 1 whole), and it is the organism as a whole that takes character state is possessedby all of the taxa that also part in the evolutionary process. Consequently, there possess the alternative state, and in addition it is is no more reason for genetics to reflect phylogeny possessedby some taxa that do not possessthe other than for anything else to do so (de Queiroz et al., state, then it is postulated to be the ancestral state. 1995). In fact, morphological characters may be a For example, open gill slits are possessed by all better reflection, because they integrate many genetic chordates in at least the embryonic stage, but in and phenotypic characters. tetrapod chordates these close early in development; Invited Review 595

Fig. 1. A phylogenetic tree derived from the character data shown in Table 1. Almost all of the characters shown in Table 1 have derived states that are congruent with this cladogram; that is, the synapomorphies form a perfect series of nested sets. For example, Characters 10 and 11 form a set consisting of the Lizards+Snakes, while Characters 12-14 form a set of the Birds+Crocodiles; these characters are plotted on the branch referring to the hypothesized ancestor where the derived character state arose (and the other characters can be plotted in similar fashion). The only exception to this congruence concerns Character 15, for which 2 evolutionary origins of the derived state must be postulated (i.e. it appears twice on the cladogram): that is, there is an apparent convergence on this cladogram, where the derived state is postulated to have arisen in 2 separate ancestors. This implies a mistaken hypothesis of homology (a homoplasy) for homeothermy in Birds and Mammals. Other cladograms could be constructed in which this convergence does not occur (i.e. in which Character 15 is a synapomorphy uniting Birds+Mammals), but in all of these cases at least 2 of the other characters must then be homoplasious.

thus, all chordates possess open gill slits but only consisting of Frogs+Salamanders+Turtles+Lizards+ some possess both open gill slits (early in develop- Snakes+Crocodiles+Birds+Mammals), then the ment) and closed gill slits (later in development). relevant sister groups (the outgroup) would be Consequently, possession of closed gill slits is at least the Lunglish and the Teleosts. Thus for hypothesized to be derived relative to possession of Character 8 in Table 1, possession of an amniotic egg open gill slits. This type of argumentation can be is hypothesized to be apomorphous relative to the applied to many types of characters (Nelson, 1978; lack of such an egg because all members of the Weston, 1988, 1994), but it is probably of limited outgroup and some members of the ingroup lack it, utility for molecular data. However, Weston (1994) while only some members of the ingroup possess it. has successfully applied the direct method to gene This type of argumentation relies on the existence of duplications (paralogy) to polarize the a and b sub- a corroborated higher level phylogeny for the taxa units of ATPase based on taxa from the archaebac- being studied, because we need a priori knowledge of teria, eubacteria and eukaryotes, thus providing a the sister groups of the ingroup. Such higher level “root” for the tree of life. phylogenies in turn may rely on other outgroup The outgroup analysis (or indirect) method comparisons, and so on in a regress back to the origin (Watrous & Wheeler, 1981) states that: if a character of life. Ultimately, we must rely on the direct method state is found in both the ingroup (the group of taxa for at least some of the characters in some of the under study) and also in the outgroup (the sister analyses. group of taxa), then it is postulated to be the Outgroup analysis is the most common type of ancestral state. For this method to provide unequivo- argumentation in cladistic analysis, but it is now cal evidence of character polarity, the outgroup usually implemented in a different way from its should consist of at least 2 sequential sister groups original formulation, particularly when analysing (Maddison, Donoghue & Maddison, 1984). For molecular data (Smith, 1994). Instead of first deter- example in Fig. 1, if we were interested in the mining character state polarity and then producing phylogeny of the tetrapods (the ingroup thus the cladogram, it is now more common to produce an 596 D. A. Morrison unrooted cladogram (a network) based on simul- different hypothesis of phylogenetic history for the taneous analysis of all of the characters of both the taxa being analysed. If a data set is perfectly con- ingroup and the outgroup and then to determine the gruent (i.e. all of the characters reflect the same root of the tree using the position where the outgroup speciation and phylesis events), then all of the tech- joins the ingroup (see below). This type of method niques will produce the same cladogram, which will was first described by Farris (1972), and Nixon & be the cladogram produced by the original method of Carpenter (1993) provide a recent detailed summary. Hennig (1966). However, apparent incongruences In this method, there are no guarantees that any almost always exist in the real world, and therefore a particular number of outgroups or any particular range of methods has been developed to try to detect choice of outgroups will ensure that the cladogram the phylogenetic pattern that underlies the apparent accurately reflects the evolutionary history (Li et al., contradictions. 1987; Smith, 1994; Adachi & Hasegawa, 1995). In Cladistics is basically axiomatic, in the sense that if practice, the phylogenetic inferences will be more the assumptions (the axioms) are accepted then the robust if more outgroup taxa are chosen rather than rest (the corollaries) follows directly from them. fewer, and the more closely-related these taxa are to Thus, like all axiomatic propositions, if one or more the ingroup (Hendy & Penny, 1989; Wheeler, 1990; of the assumptions are invalid, then the rest of the Nixon & Carpenter, 1993). For molecular data, a edifice falls with them. The key assumption is that distantly-related outgroup may have surpassed the evolution is mainly a divergent process. Homoplasies point of saturation of base substitutions, and there are thus assumed to represent “errors”, rather than will thus be a loss of phylogenetic signal through evidence in favor of some form of alternative evolutionary time as a result of random sequence evolutionary process (i.e. that does not produce a noise (Smith, 1994). The outgroup may thus, in tree-like set of relationships), such as hybridization, effect, be random, and as such it will simply join the endosymbiosis, recombination, gene duplication rest of the cladogram on the longest internal branch (producing pseudogenes), or lateral transfer. You in the ingroup and will itself have a long terminal might like to keep this in mind when you next branch (Wheeler, 1990). The best strategy may be for attempt to interpret an apparently fully-bifurcating the outgroup to consist of several taxa from the sister cladogram, as such alternative evolutionary processes group to the ingroup (Smith, 1994). are increasingly being recognized as relatively common.

Conclusion Ideally, all hypotheses of homology and relative Phylogenetic trees apomorphy will be congruent with one another. That For the purposes of cladistics, it is assumed that is, the sets of synapomorphies for a group of taxa will the phylogenetic pattern of evolutionary history can form a perfect nested hierarchy, and the polarization be represented as a branching diagram like a tree (a of the character states provides a root for the tree. cladogram), with the terminal branches (or leaves) The construction of a cladogram from the data linking the taxa being analysed and the internal would then be unproblematic (Fig. 1); and it was branches (or internodes) linking hypothesized ances- indeed under these circumstances that Hennig (1966) tral taxa (nodes) (Fig. 1). The cladogram is thus introduced his phylogenetic method. In most of the usually constructed purely from a knowledge of real world, there are, however, always characters that contemporary taxa, but it implies the existence of are incongruent with each other (homoplasies). These ancestors with particular characteristics. Tradition- incongruent characters are postulated to result from ally, the descendants of an ancestor on a cladogram either reversals, where a derived character state re- are called daughters, while the siblings after a verts to the ancestral state, parallelisms, where the speciation event are called sisters (so a descendant is same derived character state arises in separate a daughter relative to its ancestor and a sister relative evolutionary lineages, or convergences, where super- to its other sibling). Note that if either of the daugh- ficially similar character states have arisen in separate ters undergoes further speciation then the sister to a lineages. Homoplasies are thus mistaken hypotheses particular contemporary taxon may actually be a of homology. group of contemporary taxa (i.e. all of the descend- In order to deal with the common existence of ants of its sister); thus the sister to the Turtles in homoplasies, a range of tree-building techniques has Fig. 1 is the group formed by the Lizards+Snakes+ been developed in cladistics. Each technique imple- Crocodiles+Birds+Mammals. ments a different stratagem for how homoplasies are Two types of evolutionary information can be to be treated, and each technique may thus produce a represented on a cladogram. Firstly, the speciation Invited Review 591

Fig. 2. The same phylogenetic tree as shown in Fig. I, but with the branch lengths drawn proportional to the number of characters being transformed from an ancestral to a derived state along that branch (the number of character state changes is also marked on each branch). This cladogram thus shows both the phylesis and the speciation implied by the data in Table I.

events in evolutionary history are represented by the by a cladogram is in terms of which taxa form branching sequence of the tree. Secondly, the phylesis monophyletic groups on the tree. A monophyletic events may be represented by the relative length of group is: a group of taxa descended from a single the branches on the tree (in which case it is sometimes ancestor and which includes all of the taxa descended referred to as a phylogram). Thus there is an explicit from this ancestor. Groups that do not include all of distinction between those evolutionary events that the descendants of their most recent common ances- modify a taxon without causing new taxa to originate tor are called paraphyletic groups. Thus in Fig. 1 (phylesis) and those evolutionary events that do lead Crocodiles+Birds form a monophyletic group, as to the origin of new taxa (speciation). Both of these do Lizards+Snakes+Crocodiles+Birds, and also events are involved in evolutionary divergence among Lizards+Snakes+CrocodiIes+Birds+Mammals, but taxa. However, only the branching sequence (the Lizards+Snakes+Crocodiles form a paraphyletic topology) is essential to a cladogram, and indeed this group (because their most recent common ancestor is the only information shown in Fig. 1; the equiva- also has Birds as a descendant); the Class Reptilia as lent tree showing both speciation and phylesis is usually defined is thus not a monophyletic group (it shown in Fig. 2. This is not to say that the phylesis would have to include all of the members of the events are not of interest, but merely that their traditional Class Aves and Class Mammalia as well). pattern is not an essential part of the interpretation of Simple though it may seem, it is my impression that the tree; and indeed only some of the phylesis can be the recognition of monophyletic groups, and thus shown on the tree (i.e. only for those characters being the description and interpretation of the evolution- used) while all of the speciation may be included. ary pattern shown by a cladogram, appears to be Interestingly, the book of Darwin (1859) is basically one of the more difficult aspects of cladistics for about phylesis (the transformation of 1 species into non-experts (see also the specific examples discussed another) rather than about speciation (the formation by Penny et al., 1990 and Brooks & McLennan, of 2 daughter species from 1 ancestral species), in 1992). spite of its title. It should also be recognized that any 1 clado- The branching sequence of a cladogram organizes gram can actually be drawn as a tree in many taxa into monophyletic groups (clades), and the best different ways (Fig. 3). This is because the cladogram way to discuss the evolutionary pattern represented only specifies the temporal order of the branching D. A. Morrison

Fig. 3. Several ways of presenting the same rooted phylogenetic tree. (A) Shows a cladogram of the hypothetical relationships among 5 taxa; (B) is the same tree as in A but with 1 of the branches (marked) rotated; (C) is the same tree as in (B), but with another one of the branches also rotated; (D) is the same tree as in (C) but with another 2 of the branches rotated. All 4 of these trees represent exactly the same evolutionary history for the 5 taxa involved - that is, they specify exactly the same cladistic relationships among the taxa (e.g. Taxonl is shown as the sister to Taxon in all of the trees, and Taxon is the sister to this pair). There are a further 11 ways of drawing exactly this same tree, as well.

sequence from the root to the tips, and which sequence data) involves producing an unrooted tree daughter is drawn on the tree to the left of the and then rooting the tree using the outgroup. This ancestor and which is drawn to the right has no rooting of the tree then indicates the direction of evolutionary meaning. Consequently, the taxa can be evolutionary change on the cladogram; and this rotated around each internal branch on the tree allows hypotheses of relative character-state polarity without in any way affecting the meaning of the (ancestral vs derived) to be produced, and also allows cladogram. The best analogy here is to think of a inferences about the composition of monophyletic mobile hanging from the ceiling, and to think about groups to made. So, unrooted trees do not indicate the way in which its parts can rotate freely. The character polarity, nor do they unequivocally show mobile does not change its basic form no matter how monophyletic groups. This is because a group that much the parts are rotated, and similarly the phylo- appears to be monophyletic on an unrooted tree will genetic information in the cladogram does not be paraphyletic if the tree is rooted within this group. change no matter which way we rotate the branches. For example, in Fig. 4 the group Taxonl+Taxon2+ It is thus very easy for non-experts to think that 2 Taxon is only indicated as monophyletic in Tree E, trees are different, when in fact the cladogram still because this is the only tree of those shown in which specifies exactly the same evolutionary relationship the root lies outside of this group. Any specified among the taxa (Fig. 3). unrooted tree can actually be rooted on any of The distinction between rooted and unrooted trees its branches (Fig. 4), which is why it is important to is very important (Fig. 4), but it is also often prob- have an objective criterion (such as an outgroup, or lematical for non-experts. As noted above, the most some characters polarized by the direct method) for common method of cladistic analysis (particularly for determining the position of the root. Invited Review 599

A

B C

Fig. 4. Several of the rooted phylogenetic trees that can be derived from a single unrooted phylogenetic tree. (A) Shows an unrooted cladogram of the hypothetical relationships among 5 taxa; (B) is the rooted cladogram that results from rooting the tree in (A) on the terminal branch leading to Taxonl; (C) is the rooted cladogram that results from rooting the tree in (A) on the internal branch connecting Taxonl+TaxonZ with the other taxa; (D) is the rooted cladogram that results from rooting the tree in (A) on the terminal branch leading to Taxon3; (E) is the rooted cladogram that results from rooting the tree in (A) on the terminal branch leading to Taxon5. The unrooted tree in (A) could also be rooted on the terminal branch leading to Taxon2, on the terminal branch leading to Taxon4, and on the internal branch connecting Taxon4+TaxonS with the other taxa. There are thus as many possible rooting points on an unrooted tree as there are branches on the tree. Note that the unrooted tree does not specify the direction of evolutionary change, while each of the rooted trees specifies a different direction of evolutionary change.

It is worthwhile pointing out here that, in spite of Furthermore, it will be possible to answer many its limitations, an unrooted tree is, in itself, still a phylogenetic questions using only the unrooted tree. valuable phylogenetic commodity. This is because it For example, none of the rooted trees that can be limits the possible cladograms (rooted trees) that are derived from the unrooted tree shown in Fig. 4 has supported by the data. For example, for 5 taxa there Taxonl+Taxon3 as a monophyletic group (since are 105 possible rooted trees, and yet the unrooted Taxon must also be included). Therefore, if this tree shown in Fig. 4 supports only 7 of these pos- grouping was one of the hypotheses of interest, then sible trees (i.e. we have rejected 98 possibilities). we can test it adequately without ever needing to 600 D. A. Morrison

Table 2-Some characteristics of commonly-used tree-building methods

Method Procedure” Datab Taxa’ Data*

1 Unweighted pair-group ak dist all morph 2 Neighbor-joining ak dist all morph 3 Neighborliness ak dist quart morph 4 Minimum-evolution opt dist all morph 5 Distance-Wagner opt dist all morph 6 Least-squares opt dist all morph 7 Maximum-parsimony opt orig all morph 8 Weighted-parsimony opt orig all morph 9 Compatibility opt orig all morph 10 Maximum-likelihood opt orig all molec 11 Invariants opt orig quart molec 12 Spectral analysis opt orig all molec

“Method chooses a tree based on an optimality criterion (opt) (i.e. a search method) or simply uses an algorithm (alg) (i.e. a constructive method). bMethod analyses the original data (orig) or data that have been converted to distances between taxa (dist). ‘Method performs calculations on all taxa (all) or only on subsets of taxa (quart). dMethod suitable for both morphological and molecular studies (morph) or only for molecular studies (molec).

determine the root of the tree. Consequently, the the data matrix directly, while others require that the production of an unrooted tree is a very big step data have been converted to distances between taxa towards the goal of producing a cladogram. before the analysis (i.e. a matrix of all possible Finally, it is worthwhile remembering that a clado- pair-wise distances between the taxa is calculated). gram represents a set of hypotheses concerning Furthermore, some of the methods that use an evolutionary history. There are hypotheses about: optimality criterion evaluate that criterion over all character homology; character polarity; and phylo- taxa, while some methods evaluate the criterion on genetic relationships of taxa. All of these hypotheses subsets of the taxa (usually 4 taxa, called quartets). are stated explicitly by the cladogram itself, irrespec- Some methods have been developed specifically for tive of whether they have actually been shown on sequence data, while others may be applied to any the diagram. The cladogram does not constitute an type of data. The more commonly-used methods experimental test of the evolutionary history of the (Huelsenbeck, 1995a) and their characteristics are taxa, but it is a test of hypotheses of possible phylo- summarized in Table 2 , and are discussed below. genetic patterns. Becausethese hypotheses of pattern For those methods that have an optimality crite- are stated explicitly, they are open to objective tests rion, it is important to recognize that they are based of their robustness. Furthermore, the hypotheses of on a double-level optimization, where first the opti- pattern are open to further objective testing through mality criterion is optimized for a given tree topol- cladistic analyses of other data sets for the same set of ogy, and then the tree that optimizes these optimal taxa. This may be as close as we will ever get to a values is selected (i.e. first you make each tree as hypothetico-deductive science of phylogeny. good as you can, and then you find the tree that is the best). Methods thus need to be known that Methods for constructing trees accomplish the optimization for a specified tree, and then methods need to be known for searching among Tree building. The tree-building methods that have the trees for the optimal one. The first class of been developed vary in many properties (Swofford & methods will be unique for each tree-building pro- Olsen, 1990; Penny, Hendy & Steel, 1992). Some cedure, but the second class of methods is more methods choose the tree from among all of those general. possible that either maximizes or minimizes some There are several strategies for finding the optimal optimality criterion (search methods), while other tree (or trees). Exact methods are those that guaran- methods merely follow an algorithm (a pre-defined tee to find the optimal solution, while heuristic sequence of operations) to produce a single tree methods are computationally-efficient strategies that (constructive methods). Some of the methods analyse should produce a solution that is at least close to the Invited Review 601 optimal one even if it does not find the optimum. The nucleotide differences between them (Gojobori, most efficient (see below) of the exact methods in- Moriyama & Kimura, 1990). However, rather than volve the use of the branch-and-bound strategy just count the number of observed differences in (Hendy & Penny, 1982). The need for heuristic nucleotides between 2 taxa, it is usual to recognize methods is simply one of practicality - the number that there are likely to have been many superimposed of possible trees to be tested by the exact methods substitution/indel events through evolutionary time increases exponentially with increasing numbers of (e.g., multiple substitutions at any one nucleotide taxa (Felsenstein, 1978a). Heuristic search strategies position), and that these superimposed events cannot first try to find a tree that is a good estimate of the be observed in the contemporary taxa. Many final solution, usually by sequentially adding the taxa methods have thus been proposed to try to correctly to the growing tree, adding each taxon in the optimal estimate sequence divergence from observed place on the tree. Most heuristic strategies then nucleotide differences (see Gojobori et al., 1990; search through those trees that are similar to this Pollock & Goldstein, 1995; Rzhetsky & Nei, 1995), initial tree in order to find a better solution, usually based on models of nucleotide evolution that vary using branch swapping (Swofford & Olsen, 1990). from very simple to quite sophisticated. These The important point to note is that heuristic searches methods work by estimating the expected amount do not necessarily find the optimal solution, nor of nucleotide divergence between the sequences, do they necessarily get particularly close to it. It assuming some specified probability that nucleotide is therefore important to accurately describe any changes will occur during an evolutionary time- heuristic procedures used in an analysis, so that the period (such as that the number of changes follows a likely success of the method can be assessed. Poisson frequency distribution), usually by diIIeren- There are many methods for calculating the tially weighting the various transitions and transver- distances for those methods that require them. sions that are possible among the 4 nucleotide Distances specify a relationship between pairs of character-states (A, C, G, T/U) to reflect their taxa, with taxa that have identical states for all of the relative probability of occurrence. The differences characters having zero distance between them, and observed between the taxa are thus “corrected” for taxa that have no character states in common having unobserved multiple substitutioniindel events, and an arbitrarily large distance. Some data that are the distances derived from these corrected data useful for cladistic purposes, such as those from should provide a better estimate of the evolutionary immunology, nucleic acid hybridization and breeding divergence among the taxa. experiments, are automatically expressed as pair-wise The studies to date indicate that the simplest of the distances (West & Faith, 1990); but most character various models of nucleotide evolution that is likely data need to be mathematically transformed into to apply to real data sets is the 2-parameter model distances. The basic assumption when analysing dis- of Kimura (1980), but that it may not be applicable tances is that the distances used provide an estimate when the sequences are distantly related and/or the of the evolutionary divergence among the taxa. total number of nucleotides examined is large The wide array of possible general measures of (Rzhetsky & Nei, 1995). Part of the problem here inter-taxon relationship is summarized by Clifford & is that the rate of nucleotide substitution varies Stephenson (1975). The most commonly used direct within evolutionary hneages (Li er al., 1987) as well distance measures are the manhattan (sum of the as among sites within a sequence (Olsen, 1987; character-state differences for all of the characters) Shoemaker & Fitch, 1989). Consequently, the nucleo- and euclidean (square-root of the sum of the squared tide evolution models may be better if the variance of differences) distances, although the former may have the estimates is reduced (Pollock & Goldstein, 1993, more biological meaning (West & Faith, 1990). For and if they assume that the rate of nucleotide data with a mixture of multi-state characters (i.e. substitution varies from site to site according to a where different characters have different numbers of gamma frequency distribution (Jin & Nei, 1990). states) it will be necessary to standardize the data Many other methods have also been proposed first, and the most popular standardized distance is specifically for calculating distances from protein- the gower (range-standardized manhattan) distance. coding DNA sequences, based on comparing codons It is also possible to convert measures of similarity rather than nucleotides. They recognize the distinc- into distances (Swofford & Olsen, 1990), and there tion between non-synonymous and synonymous are many published versions of such measures nucleotide substitutions (i.e. those that do and do not (Hubalek, 1982). result in changes in the amino acid coded for), and For DNA/RNA sequence data, the distance be- include both weighted and unweighted pathways tween taxa can be estimated by the total number of methods (Gojobori et al., 1990). Distances may also 602 D. A. Morrison be calculated directly from amino acid sequences Neighborliness is thus a quartet method, evaluating (Ota & Nei, 1994). Furthermore, a number of the 4-point metric for all possible subsets of 4 taxa methods have been proposed specifically for convert- (quartets) to decide how the 4 taxa should be clus- ing frequency data (such as allelic or genotypic tered, and then building clusters of the taxa that are frequencies) to distances, based on different genetic consistent with the largest fraction of the quartets models (Rogers, 1991), and also for calculating (i.e. the tree is the one for which the 4-point metric is distances from restriction endonuclease data (Nei & satisfied the largest number of times). Li, 1979). Other algorithmic methods that analyse distance data have been described by Farris (1977), Tree-building methods. Here, I provide only a brief Chakraborty (1977), Klotz & Blanken (1981), Li introduction to the procedures employed by the (1981), Tateno, Nei & Tajima (1982), Hasegawa, various tree-building methods. More details of most Kishino & Yano (1985), and Charleston et al. (1994); of these methods, along with worked examples, are but none of these has been widely used. provided by Swofford & Olsen (1990), and The Minimum-evolution procedure (Edwards & Charleston, Hendy & Penny (1994) provide a brief Cavalli-Sforza, 1964) analyses distance data using mathematical exposition. an optimality criterion. The tree chosen is the one The Unweighted pair-group method using that minimizes the sum of the lengths of all of the arithmetic average (UPGMA) is an algorithmic branches on the tree, these branch lengths being method that analyses distance data (Michener & estimated using linear algebra on the observed dis- Sokal, 1957; Sokal & Michener, 1958). It groups the tances (i.e. from the original distance matrix) between taxa based on increasing distances between the taxa, 3 taxa inter-connected by a common node. Rzhetsky starting by grouping the 2 taxa with the smallest & Nei (1993) provide an exact method for optimizing distance, and then progressively adding more-distant each tree; and Rzhetsky & Nei (1992) have suggested taxa to the group or to new groups. It assumes that an heuristic method for finding the optimal tree by the data are ultrametric; that is, a tree can be pointing out that the neighbor-joining algorithm is constructed so that the observed distance between guaranteed to find the Minimum-evolution tree for 4 any 2 taxa (i.e. in the original data matrix) is equal to taxa. the sum of the branch lengths joining them (i.e. the The Distance-Wagner method (Farris, 1972) also distance as measured along the branches connecting analyses distance data using an optimality criterion. the 2 taxa on the tree), and that the tree is rooted so The optimality criterion is the same as for the that all of the taxa are equidistant from the root. Minimum-evolution procedure: the desired tree is Neighbor-joining (Saitou & Nei, 1987; Studier & the one that minimizes the sum of all of the branch Keppler, 1988) is another algorithmic method that lengths on the tree. However, in this case all of the analyses distance data. It is similar to UPGMA in patristic distances (along the tree branches) connect- that it starts by grouping the 2 taxa with the smallest ing any 2 taxa are constrained to be greater than the distance, and then progressively adds more-distant observed distances between these taxa, whereas the taxa to the group or to new groups. However, it Minimum evolution procedure assumes that some of differs in modifying the distance matrix at each step, the patristic distances will be greater and some so that each pair of groups is adjusted on the basis of smaller than the observed distances. Exact methods their average divergence from all other groups. The for optimizing each tree have been described method thus relaxes the ultrametric assumption, and (Beyer et aZ., 1974; Waterman et al., 1977), and assumes only that the data are additive; that is, a tree heuristic procedures for tinding the optimal tree are can be constructed so that the observed distance usually used (Farris, 1972; Swofford, 1981; Faith, between any 2 taxa (i.e. in the original distance 1985). matrix) is equal to the sum of the branch lengths The Weighted least-squares (Fitch & connecting them on the tree (patristic distance or Margoliash, 1967) and the Unweighted least-squares pathlength). (Cavalli-Sforza & Edwards, 1967) are related The Neighborliness procedure (Sattath & Tversky, methods that analyse distance data using an opti- 1977; Fitch, 1981) is the final algorithmic method mality criterion. They both choose the tree that discussed that analyses distance data. It is based minimizes the error (i.e. the measured discrepancy) on evaluating the 4-point metric (Dobson, 1974), a between the observed distances between the taxa (i.e. relaxed definition of additivity for a set of 4 taxa. It the original distance matrix) and the patristic dis- specifies the conditions for clustering 4 taxa into 2 tances connecting the taxa on the tree (pathlength). pairs, based on the relative ranking of the patristic So, for each possible tree they compare the sum of all distances between all possible pairs of the taxa. of the squared differences between the distances in Invited Review 603

the original matrix and those on the tree, and they Sharkey, 1989). However, many a priori weighting then choose the tree that minimizes this sum. The 2 schemes have also been developed specifically for methods differ in the formula they actually use to molecular sequences, where the genetic mechanisms measure the discrepancy between the observed and responsible for nucleotide substitutions make it seem patristic distances, the unweighted method assuming unreasonable to assume equal weights for all that all of the distance estimates have the same changes. Differential weights can be applied to (see absolute error and the weighted method assuming Hillis et al., 1993): (1) different sequence positions that they are have the same percentage error. Exact (character weighting), thus emphasizing functional methods for optimizing each tree have been described inequalities along the sequences (e.g., different codon (Kidd & Sgaramella-Zonta, 1971; Beyer et aZ., 1974; positions in translated genes, stems vs loops in struc- Waterman PZ al, 1977), and heuristic procedures for tural RNAs, hypervariable positions in mtDNA); or finding the optimal tree are usually used. (2) the same sequence positions (character-state The Maximum-parsimony method (Farris, 1970; weighting), thus emphasizing mutational biases (e.g., Fitch, 1971) is the first method to be discussed that transitions vs transversions, relative substitution analyses the original data matrix using an optimality frequencies, base composition, synonymous vs criterion. It is probably the most popular method non-synonymous changes). A posteriori weighting among cladists, and indeed many cladists restrict the schemes include successive approximations (Farris, definition of cladistics to include this method alone. 1969), quadratic weighting (Fitch & Yasunobu, The optimal tree is the one that has the minimum 1975), transformation series (Mickevich, 1982), number of inferred character state changes (i.e. short- dynamic weighting (Williams & Fitch, 1989) and est tree length or fewest evolutionary steps). Exact homoplasy weighting (Goloboff, 1993). Exact methods for optimizing each tree have been described methods for optimizing each tree using weights have (Fitch, 1971; Swofford & Maddison, 1987). Exact been described (Sankoff & Cedergren, 1983). methods for finding the optimal tree include an The Compatibility method (Le Quesne, 1969, exhaustive search (i.e. testing every possible tree 1982) also analyses the original data matrix using an topology) and branch-and-bound methods, but optimality criterion. The optimal tree is the one that heuristic methods must be used for analyses with has the largest set (clique) of mutually compatible more than about 20 taxa (Swofford & Olsen, 1990). characters (i.e. characters for which there is no Special methods have been developed for frequency homoplasy on the tree). The method proceeds by data, such as allelic or genotypic frequencies testing the pair-wise compatibility of all of the (Swofford & Berlocher, 1987). characters, and then examining all possible cliques of The Weighted-parsimony method (Farris, 1969; characters. The method thus relies on the theorem Sankoff, 1975) is a generalization of the Maximum- that a group of characters which are pair-wise com- parsimony method that allows the characters and patible are jointly compatible. This is always true for character states to be weighted so that they con- both polarized and unpolarized 2-state characters tribute unequally to the calculation of optimality (Estabrook & McMorris, 1980), but for unpolarized (note that character weighting and character-state multi-state characters it is possible for a collection of weighting are separate ideas; Williams & Fitch, pair-wise compatible characters not to have any 1989). The optimal tree is the one that has the phylogeny for which they are all compatible (Fitch, minimum sum of the weights. The rationale for this is 1975). Branch-and-bound methods are available for that the unequal weights reflect unequal probabilities finding the maximal cliques. of change in the characters (Felsenstein, 1981 b), the The Maximum-likelihood procedure (Felsenstein, weighting usually being applied so that evolutionarily 1973a,b) is another method that analyses the original conservative characters or states can contribute more data matrix using an optimality criterion. So far, to the reconstruction of the phylogeny (Olsen, 1987). practical likelihood methods have only been The magnitude of the weights to be applied can be developed for analysing nucleotide sequence data decided either a priori (i.e. decided before the analy- (Felsenstein, 198la), protein sequence data (Kishino, sis) or a posteriori (i.e. calculated after an initial Miyata & Hasegawa, 1990), and restriction site data cladogram has been constructed) (Neff, 1986). (Felsenstein, 1992), with restricted maximum- Several general a priori weighting schemes have been likelihood methods also available for continuous proposed, such as Dollo-parsimony (Le Quesne, characters (Felsenstein, 198lc). The optimal tree is 1974), where no parallelisms or convergences are the one that maximizes the statistical likelihood permitted, and Camin-Sokal-parsimony (Camin & that the specified evolutionary model produced the Sokal, 1965), where no reversals are allowed, as well observed data. For example, for a given model of as compatibility weighting (Penny & Hendy, 1986; nucleotide evolution (the most commonly-used one is 604 D. A. Morrison the generalized Kimura 2-parameter model described of support for every possible subset of taxa, by by Kishino & Hasegawa, 1989), formulae are derived summing the number of the expected character-state that describe the probability that an initial nucleotide changes that are compatible with each subset. The will be transformed into a specified nucleotide during Closest-tree method (Hendy, 1991; Steel et d, 1992), an evolutionary time-period. The likelihood for each a compatibility-type method based on minimizing nucleotide position is then equal to the prior prob- the sum-of-squares, is then used to select the optimal ability of finding the initial nucleotide at that position tree as the one with the largest set of mutually multiplied by the probability of transformation. The compatible subsets of taxa. Exact methods for find- likelihood of the divergence of 2 sequences during the ing the optimal tree include branch-and-bound time-period is then the product of the likelihoods at methods (Hendy & Penny, 1989), with a limit of each position, and the overall likelihood for a tree is about 20 taxa. the product of the likelihoods along the branches. No exact methods for optimizing each tree have been Characteristics. Clearly, it is impossible to compare described, and iterative (“hill-climbing”) methods are the many tree-building methods based on their used (Fukami & Tateno, 1989). Exact methods for relative ability to detect the phylogenetic pattern finding the optimal tree involve an exhaustive search underlying the apparent incongruences in real data (since suitable branch-and-bound methods are un- sets, because the real phylogenetic pattern is almost known), and so heuristic methods (Felsenstein, never known. Instead, many other criteria have been 1981a) are usually used for analyses with more than proposed to assess the relative usefulness of the about 5 taxa. various tree-building methods, although many of The Invariants (or Evolutionary parsimony) these criteria have not yet been effectively evaluated method (Lake, 1987; Cavender & Felsenstein, 1987) for most of the methods. Those criteria about which also analyses the original data matrix using an opti- we know the most are outlined here - no single mality criterion, although it was developed specifi- method meets all of the criteria (Penny et al., 1990, cally for analysing nucleotide sequence data. It is 1992). based on evaluating the number of operator invari- The Computational Eficiency of a method refers to ants (linear for the method of Lake; quadratic for the the relative speed with which a solution is produced method of Cavender & Felsenstein) for 4 taxa, which (Penny et al., 1992). All of the algorithmic procedures specifies the conditions for clustering the 4 taxa into 2 are relatively efficient, as the computation time does pairs, based on the relative number of transversions not increase dramatically as the number of taxa is on the internal branch of the unrooted tree to those increased. On the other hand, none of the optimality on the terminal branches. The use of Invariants is methods is efficient, because the number of trees from thus a quartet method, evaluating the invariants for which the optimal one needs to be chosen (i.e. at the all possible subsets of 4 taxa (quartets) to decide how second step in the double-optimization) increases the 4 taxa should be clustered, and then building exponentially as the number of taxa increases clusters of the taxa that are consistent with the largest (Felsenstein, 1978a). Finding the optimal tree using fraction of the quartets (Lake, 1988). Exact methods exact methods is an example of the large class of are used for the evaluation of the invariants. NP-complete mathematical problems (Graham & Spectral analysis (Penny, Hendy & Henderson, Foulds, 1982), for which no efficient procedure is 1987; Hendy & Penny, 1993) is the final procedure known (or is expected ever to be known). So, for that analyses the original data matrix using an opti- large data sets (both in terms of number of taxa and mality criterion. It has been developed only for number of characters) heuristic methods must be nucleotide data, originally with 2-state characters used for all of the optimality methods if a solution is (Hendy, 1989) and more recently with 4 states to be produced within a reasonable time. This means (Hendy, Steel & Penny, 1994). This procedure that finding the optimal solution cannot be guaran- first uses the hadamard transformation (Hendy & teed for the analysis of large data sets - just how Charleston, 1993) together with a model of sequence large “large” is depends on the computing power evolution (such as the 2-character-state model (both software and hardware) available for the analy- of Cavender, 1978a,b) to adjust (or correct) the sis and the patience of the cladist. This situation is observed data for unobserved character-state compounded for data sets with a relatively low changes (e.g., multiple substitutions at a sequence phylogenetic signal (see below), which are likely to position), thus producing a matrix of expected data have multiple optimal trees that are structurally quite (i.e. expected number of character-state changes different (Maddison, 1991) -these “islands” of trees under the evolutionary model adopted). These may not easily be found by heuristic techniques corrected data are then used to estimate the degree (Page, 1993). Invited Review 605

The Maximum-likelihood method is even more quartet methods ignore that part of the information inefficient than all of the other methods that use that comes from the incompatibilities among the optimality criteria. For Maximum-likelihood, the quartets, while Compatibility analysis assumes that first step in the double-optimization (i.e. optimizing incompatible characters have no phylogenetic each tree topology) has no simple analytical solution information at all (Hill, 1975). Furthermore, the (Fukami & Tateno, 1989), unlike all of the other Parsimony and Compatibility methods do not use methods, and so computationally-intensive iterative characters that have constant states across all of the procedures must be used. This problem is exacer- taxa, nor characters where only 1 of the states bated by the fact that there may be more than 1 way occurs in more than 1 taxon (autapomorphies); and of optimizing a particular tree topology (Steel, 1994). these may form a large proportion of sequence data These problems are then compounded by the lack of (for example, I looked at an arbitrary selection of a branch-and-bound algorithm for the second step in 20 molecular studies published in 1994, and found the double-optimization; and so there is a severe limit that on average 84% of the nucleotide positions to the size of the data set that it is practical to analyse were uninformative for Parsimony/Compatibility by the Maximum-likelihood method. analysis using these 2 criteria). The Maximum- The Power of a method refers to the response of likelihood, Invariants and Spectral-analysis methods the method to increasing amounts of data (Olsen, ignore characters with missing character-state data 1988) - with an increasing number of characters in for any of the taxa, including indels for sequence the original data matrix we expect a method to data, which may also constitute a sizable pro- converge to a single tree (i.e. a limit is reached portion of the original data matrix. The Invariants beyond which further increases in the data do not method is also based entirely on transversion substi- change the tree produced by the method) (Penny & tutions, and there may not be enough in a data set Hendy, 1986). A powerful method is one in which the to effectively infer the branching pattern, or the convergence occurs with a relatively small number of frequency of transitions may confound the pattern characters (Hillis, Huelsenbeck & Swofford, 1994b). (Jin & Nei, 1990; West & Faith, 1990); it thus Power is also related to the amount of information appears to have very low power (Hillis et al., in the original data matrix that is actually used to 1994a,b). construct the tree (Penny et al., 1992, 1993) - all of The Consistency of a tree-building method refers to the tree-building methods omit some of the character its ability to converge to the correct tree (Felsenstein, information, and so the power of these methods may 1988). That is, with an increasing number of charac- be reduced. However, it is important to note that not ters in the data matrix a consistent method converges all of the character information is necessary for the to the single correct tree, and a method is inconsistent reconstruction of an evolutionary tree, and that if there are any possible situations where it converges therefore it is not loss per se but loss of relevant to the wrong tree (i.e. there are systematic or non- information that is important (Shoemaker & Fitch, random errors). Consistency is related to the assump- 1989; Penny et al., 1993). tions (e.g., the evolutionary model) made by the Compared with the character-based methods, all tree-building method - if the assumptions are not of the distance methods involve a very large loss of met for a particular data set then consistency cannot information in converting the original data into be expected. distances. This is because there will be multiple There are fundamental assumptions underlying all character data sets that produce exactly the same of the tree-building methods, such as that evolution distance matrix (Felsenstein, 1982; Penny, 1982), and proceeds as a continuous-time markov process this problem becomes worse for increasing numbers (Beanland & Howe, 1992), whether they are explicit of taxa and characters (Steel, Hendy & Penny, 1988). or not for a particular type of data (e.g Rodriguez Moreover, for sequence data it is not clear how to et aZ., 1990). The most notable assumption is that calculate distances when the aligned sequences con- character-state changes are independent and identi- tain terminal length variations, indels or ambiguous cally distributed (Cavender, 1978a; Olsen, 1987; nucleotides (since the models of nucleotide evolution Shoemaker & Fitch, 1989; Sanderson, 1995); which are based on estimating rates of transitions and means that no character-state change affects any transversions only) - including or excluding these other character-state change, and that all character- features in the calculations can produce differ- state changes are equally likely. If evolution has ent distance estimates (Swofford & Olsen, 1990; occurred then this assumption is generally not true, Beanland & Howe, 1992). because of the common history of the taxa during The character-based methods also do not use all at least some evolutionary period. The distance of the data (Penny et aZ., 1990). For example, the methods will be particularly sensitive to this problem 606 D. A. Morrison

(Felsenstein, 1988), although all of the methods suffer problems listed above may thus not necessarily pre- from it (Felsenstein, 1982). clude use of the various methods if they are robust. The distance methods will also be inconsistent if The distance methods, for example, may be quite the data are not additive (Felsenstein, 1988), as there robust to the problem of non-additivity if the is no exact solution for estimating the tree under distance estimates are corrected for multiple these circumstances (Swofford & Olsen, 1990). character-state changes and the problems of negative Additivity of taxonomic data is unlikely because branch-lengths are addressed (Felsenstein, 1988); and the distances are only estimates of evolutionary even UPGMA does not require perfectly ultrametric divergence, and are therefore expected to vary data (Colless, 1970). The consistency problems randomly around the true divergence value (i.e. the associated with homoplasy can also potentially be data will be additive only if there is no homoplasy). solved (i.e. the methods can be made robust) by any One of the consequences of this is that trees with method that provides a framework for making cor- negative branch lengths are theoretically possible rections for multiple character-state changes (Penny when using the Neighbor-joining, Minimum- er al., 1993), such as the use of transformations and evolution and Least-squares methods (Swofford & corrected distances (e.g., for unequal nucleotide Olsen, 1990). UPGMA adds the further constraint composition, use of the LogDet transformation, that the data are ultrametric; that is, that evolution- Lockhart et al., 1994; compositional statistics, Sidow ary divergence among the taxa occurs at a uniform & Wilson, 1990, 1991; or character weighting, rate through time (such as occurs with a molecular Marshall, 1992). Consistency can often be restored clock). This is also unrealistic. (i.e. the methods made robust) when there are juxta- All of the methods will also be inconsistent if the posed long and short branches by adding taxa to the amount of homoplasy is large, because homoplasy analysis that join the tree on the long branches makes unrelated taxa appear to be more similar (Hendy & Penny, 1989; Penny et al., 1990; Lento (Olsen, 1987; Swofford & Olsen, 1990). This will be a et al., 1995), or by using character weighting schemes particular problem if the frequency of character- that shorten the branches (Swofford & Olsen, 1990; states varies between taxa (Penny et al., 1990); for Hillis et al., 1993). However, the preferred method example, for sequence data similar GC-content often is to make corrections for multiple character-state unites unwanted branches on a tree (Sidow & Wilson, changes using transformations (Jin & Nei, 1990; de 1990, 1991; Lockhart et al., 1992; Marshall, 1992; Bry, 1992; Steel et al., 1993; Huelsenbeck, 1995a). Hasegawa & Hashimoto, 1993) and all tree-building Note that the attempts to improve the consistency of methods are sensitive to this (Lockhart e? al., 1992, the tree-building methods by using corrections and 1994). transformations apply only to molecular data, be- Methods may also be inconsistent when there are cause it is possible to use a consistent model of juxtaposed long and short branches on the tree (e.g., character-state changes that will apply to all of the unequal rates of phylesis or speciation), as these characters in any 1 data set. It is not obvious how to branches will tend to join together (often referred to appropriately correct or transform other types of as “long branches attracting”). This has long been data, since each character will have its own model of recognized as a particular problem for the Parsimony character-state change. and Compatibility methods (Cavender, 1978a; A popular approach to studying the robustness of Felsenstein, 1978b; Hendy & Penny, 1989; Zharkikh the various methods is by using simulated data, & Li, 1993; Takezaki & Nei, 1994), but we now where the original tree and the characteristics of the recognize that it also applies to all of the distance data are known, and the behavior of the tree-building methods as well as to the Invariants and Closest-tree methods under these conditions is evaluated (e.g., methods (Jin & Nei, 1990; Penny et al., 1990; de Bry, Tateno et al., 1982; Li et a!., 1987; Sourdis & 1992; Huelsenbeck & Hillis, 1993; Steel, Hendy & Krimbas, 1987; Kim & Burgman, 1988; Saitou & Penny, 1993; Zharkikh & Li, 1993; Hillis et al., Imanishi, 1989; Jin & Nei, 1990; Rohlf et al., 1990; 1994b). The Maximum-likelihood method appears to Nei, 1991; Sidow & Wilson, 1991; Hasegawa & suffer least from this problem (Hillis et al., 1994b). Fujiwara, 1993; Huelsenbeck & Hillis, 1993; The Robustness of a tree-building method refers to Charleston et al., 1994; Hillis et al., 1994a,b; Kuhner how much the assumptions of the method can be & Felsenstein, 1994; Tateno, Takezaki & Nei, 1994; violated before the method becomes inconsistent Gaut & Lewis, 1995; Huelsenbeck, 1995a,b; Tillier & (Penny et al., 1990). This may be the most important Collins, 1995; Yang, 1995). There is clearly a large criterion, because the idealized assumptions of the conceptual gap between simulated data and real data methods are likely to be violated by real data sets (Miyamoto & Cracraft, 1991; Hillis et al., 1994a), (Huelsenbeck, 1995a,b). Many of the consistency and most of the simulations apply only to the Invited Review 607

branching order of a limited number of taxa under important to quantitatively assess the magnitude of specific evolutionary models. The jury is still out on the phylogenetic signal (i.e. that part of the character the results (cf. Felsenstein, 1988), but these studies variation that is potentially informative about the show that many of the tree-building methods are evolutionary history) in a set of data - is the support actually quite effective (i.e. robust) over a wide range for a particular tree any better than would be of evolutionary conditions. The most robust of the expected from a random data set? Quantifying this methods appears to be the Maximum-likelihood pro- information also allows cladograms to be compared cedure, with UPGMA and the Invariants method effectively with each other - a cladogram with a being the least robust. The robustness of the other stronger phylogenetic signal is to be preferred. Four methods seems to depend very much on the ability classes of techniques have been developed: optimality to correctly utilize transformations for multiple measurements (for those methods that use an opti- character-state changes or to appropriately apply mality criterion), such as consistency and homoplasy character or character-state weighting; clearly this indices (Archie, 1989; Farris, 1989, 1991) for the requires a priori knowledge about the evolutionary Parsimony methods, likelihood values for the process in the taxa being studied that is unlikely to be Maximum-likelihood methods, and sums-of-squares available. It is unfortunate that Spectral analysis has for the Least-squares methods; skewness indices not been incorporated into most of the simulation (Huelsenbeck, 1991; Hillis & Huelsenbeck, 1992); studies to date, in order to test whether its theoretical analytical statistical tests (Felsenstein, 1988; Li & advantages are met in practice. Gouy, 1991; Li & Zharkikh, 1995); and randomiza- Finally, there are methodological and philo- tion (or permutation) statistical tests (Goloboff, 1991; sophical considerations that are often used to com- KBllersjij et aZ., 1992; Ahoy, 1994; Faith & Ballard, pare the techniques (Miyamoto & Cracraft, 1991). 1994). In particular, the algorithmic procedures are often A cladogram is interpreted in terms of the mono- criticized because they provide no method for rank- phyletic groups of taxa that it hypothesizes. It ing alternative tree topologies (i.e. there is no is therefore important to quantitatively assess the criterion to compare trees and decide how “good” robustness of all of the groups indicated by the any of them are as estimates of phylogeny) (Swofford branching sequence of a cladogram (i.e. the degree of & Olsen, 1990). This is a serious limitation if we treat support for each branch in the tree) - is the support a cladogram as being an attempt to estimate the true for a particular group any better than would be phylogeny, because we are using incomplete informa- expected from a random data set? Many classes of tion and there is thus a degree of uncertainty about techniques have been developed, including: analytical the correctness of our estimate (Felsenstein, 1988). It procedures, such as confidence limits, branch-length is thus always a good idea to assess how much better variances and likelihood-ratio tests (Li & Gouy, the optimal tree is than the next-most-optimal tree 1991; Li & Zharkikh, 1995); resampling procedures, (or set of trees) (Swofford, 1991), especially as these such as the bootstrap (Li & Zharkikh, 1994, 1995; trees may not be independent of each other (Penny Sanderson, 1995), the jacknife (Felsenstein, 1988) et al., 1995). and topology-dependent permutation (Faith, 1991); and non-statistical procedures, such as the decay After tree construction. We can never know index (Bremer, 1994), clade stability (Davis, 1993), whether our cladogram actually represents the true and spectral signals (Hendy & Penny, 1993). phylogenetic history of the taxa concerned. There- It is unlikely that different tree-building methods fore, there are 3 further issues that need to be dealt will always produce identical trees for the same data with once a phylogenetic tree has been constructed set, given their different underlying assumptions, and using 1 or more of the methods outlined above: it is also possible for multiple optimal trees to be assessing the magnitude of the phylogenetic signal in produced by a single tree-building method. Further- the cladogram; assessing the robustness of the mono- more, we may wish to compare cladograms produced phyletic groups represented on the cladogram; and using different sets of characters (since congruence comparing the structure of multiple cladograms from among data sets may constitute the strongest achiev- the same set of taxa. Although these issues are able evidence for the true phylogeny; Penny & extremely important for cladistic analysis, they are Hendy, 1986), and these may differ as a result of only briefly introduced here because a full coverage sampling error, different stochastic processes, or dif- would require a separate (although shorter) review. ferent branching histories (de Queiroz et al., 1995). It Even randomly-generated data will lead to the is therefore usually necessary to quantitatively com- construction of a cladogram if any of the tree- pare the structure of multiple trees from the same set building techniques is applied to them. It is therefore of taxa. For example, which monophyletic groups 608 D. A. Morrison are supported by (congruent among) which set of Baverstock & Johnson, 1992; Ellis et al., 1994). In trees? - monophyletic groups that are supported particular, these studies suggest a strong correlation by several trees are to be preferred. Several classesof between the parasite phylogeny and that of their techniques have been developed (most of them re- definitive hosts, with species of the Sarcocystidae viewed by Swofford, 1991): consensus trees; largest forming 2 clades based on the use of either canids or common pruned trees; consensus indices; tree felids as their definitive hosts. Thus Surcocystis(felid comparison metrics; partition tests (Farris et al., or canid hosts) is only monophyletic if both 1995); and randomization tests (Rodrigo et aZ., 1993). Toxoplasma (felid host) and Neospora (unknown host) are included within it. A parasitological example. To explore the practical Second, we need to consider the data that are consequences for phylogeny reconstruction of the available to tackle this problem (i.e. the ingroup). wide range of available tree-building methods, it There are currently complete ssrRNA sequences is worthwhile exploring a concrete example. This available for 8 species of the Sarcocystidae parasitological example illustrates and emphasizes (Sarcocystis arieticanis, S. fusiformis, S. gigantea, many of the points made in earlier sections, as well as S. muris, S. neurona, S. tenella, Neospora caninum, providing a brief step-by-step guide to a cladistic Toxoplasma gondii), which can be used to test the analysis. monophyly of Surcocystis. Several published se- quences are available for some of the species (9 for Methods. First, we need to consider the back- T. gondii, 2 for N. caninum), and for the purposes of ground to the problem that we intend to address with this example consensus sequences were derived for the cladistic analysis. The members of the genus these species (after alignment; see below) using the Sarcocystis are parasitic protozoans with heterox- MacClade 3.04 computer program (Maddison & enous life cycles, in which both hosts are mammals. Maddison, 1992). The standard IUPAC ambiguity The sexual stage is in the definitive host (often a codes were used for those few nucleotide positions predator species) and the asexual stage is in the with more than 1 possible character-state in the intermediate host (often a prey species), with the sequence (1 for S. muris, 1 for N. caninum, 26 for sporocysts being passed in the faeces of the definitive T. gondii). Thus, we have a representative sample of host and ingested by the intermediate host. Studies of the family (e.g., species with both felid and canid this genus may be of economic importance, because hosts), and we have taken inter-individual variation some of the species are important pathogens of into account as far as possible. livestock and humans. Third, we need to consider how the cladogram will The current classification of the phylum Apicom- be rooted (i.e. the outgroup), since monophyly can be plexa (Sporozoa) is based largely on ultrastructural assessed only on a rooted tree. The members of analysis, on details of the life-cycles, and on the the Eimeriida that are most closely-related to the location of developmental stages. The internal Sarcocystidae (see Levine, 1985) are the members of arrangement of the phylum is currently rather un- Eimeriu (family Eimeriidae), for which 7 species stable, at least partly because of a lack of knowledge have published ssrRNA sequences (E. acervulina, E. of the life-cycles for many taxa. Cox (1994) treats the brunetti, E. maxima, E. mitis, E. necatrix, E. praecox, coccidial parasites as a class (the Coccidea), with E. tenella). For the outgroup, a single consensus Sarcocystis in the order Eimeriida. The family sequence was derived from the sequencesfor these 7 Sarcocystidae is often subdivided into 2 subfamilies, Eimeria species(after alignment) using the MacClade the Sarcocystinae (3 genera, 100 species) and the program, with the standard IUPAC ambiguity codes Toxoplasmatinae (2-3 genera, 15 species) (Levine, being used for those 68 nucleotide positions 1985). with more than 1 character-state in the consensus However, the phylogenetic relationships within the sequence. These ambiguous character-states can be Sarcocystidae remain ambiguous, at least partly be- dealt with effectively by some analysis techniques, cause of difficulties in assessing the homology of while others treat them as missing data. For this data phenotypic characters. It is possible that molecular set this different treatment has very little effect, data might alleviate these problems, although there because in almost all casesthe Sarcocystis sequences has been little consistency among molecular studies are invariant for these characters (i.e. the variability to date. While the family as a whole is generally is only within the Eimeria phylogeny itself and it considered to be monophyletic, several analyses of thus does not influence the reconstruction of the small-subunit ribosomal RNA (ssrRNA) sequences Sarcocystis phylogeny). have questioned the monophyly of Sarcocystis Fourth, we need to consider character homologies itself (Barta, Jenkins & Danforth, 1991; Tenter, (i.e. sequence alignment). There are 9 sequences in Invited Review 609 the analysis, 8 for the ingroup and 1 for the out- matrix was used, with transitions weighted as 1 and group. The ssrRNA sequences for all of these species transversions weighted as either 1 or 2. The nucleo- were aligned using the secondary structure model tide frequencies are approximately equal for all of the of Van de Peer et al. (1994) - all of the sequences sequences, and so there is probably no need to were obtained from this computer database. For the correct for GC-bias (the GC content varies from 50 cladistic analysis, the single-stranded regions were to 60% across the sequences). omitted, as Ellis & Morrison (1995) have indicated Next, we need to consider the magnitude of the that it is the double-stranded region that contains phylogenetic signal in the data (i.e. the robustness of most of the phylogenetic information for these taxa. the cladograms). For those methods with an optimal- There are thus 694 nucleotide positions in the analy- ity criterion, permutation tests were used. Two- sis. This procedure effectively deals with both the hundred permutations of the data were produced for problems associated with positions of equivocal each model and each tree-building method, and the alignment (since there are very few indels in the percentage of these replicates that produced trees double-stranded region), as well as those associated that were at least as optimal as the observed tree was with relative character weighting based on structural/ calculated. The Phylip program implements this op- functional inequalities along the sequences. tion for its tree-building methods. For the Parsimony Fifth, we need to consider the tree-building analyses, some of the permutations were performed methods. Four tree-building methods were used (cov- using the RandomCladistics 3.0 computer program ering the full range of possibilities), in order to test (Siddall, 1994); for the remainder, the Phylip pro- the sensitivity of the analysis to this source of varia- gram was used to create the permuted replicates and bility. Two of these were distance-based methods, 1 these were then individually fed into the PAUP of them algorithmic (Neighbor-joining) and the other program. with an optimality criterion (Unweighted least- Finally, we need to decide how we are going to squares). The other 2 methods were character-based assess the magnitude of the support that the data with optimality criteria, one developed specifically have for the monophyly of Sarcocystis. Bootstrap for nucleotide sequences (Maximum-likelihood) and resampling was used, which is implemented in both one not (Weighted-parsimony). The Phylip 3.57 com- the Phylip and PAUP programs. Two-hundred boot- puter program (Felsenstein, 1995) was used for strap replicates were produced for each model and the first 3 methods, employing the heuristic search each tree-building method, and the percentage of the with global branch-swapping for those methods trees derived from these replicates that indicated with an optimality criterion. The PAUP 3.1.1 com- monophyly of Surcocystis was calculated. puter program (Swofford, 1993) was used for the final method, employing the branch-and-bound Results. Three cladograms resulted from the 8 search. Note that in order to test the monophyly analyses (Fig. 5), with each of the optimality methods of Surcocystis, the only part of the cladogram that finding a single most-optimal tree for each analysis. needs to be interpreted is the branching order (not The Neighbor-joining and Maximum-likelihood the branch lengths), since we will be independently analyses produced the same tree irrespective of assessing the degree of support for monophyly (see the model used (Table 3), this tree supporting below). monophyly of Sarcocystis. Sixth, we need to consider the evolutionary model However, the Weighted-parsimony method pro- or models that we wish to employ. Two models of duced different trees for the 2 models (Table 3), both character-state change were used, in order to test the supporting monophyly of Surcocystis, but differing in sensitivity of the analysis to this source of variability. the relative placement of S. muris and S. neurona. The first model assumed that all nucleotide changes However, for the model in which transitions and were equally probable at each sequence position, transversions were equally probable, the tree in Fig. while the second model assumed that transitions were 5A is only one step longer (i.e. one more character- twice as likely as transversions (for the trees shown in state change) than the tree in Fig. 5B, and so there is Fig. 5 the transition : transversion ratio is about 1 S, actually very little conflict between the 2 methods. as determined by the MacClade program). For the This result emphasizes the importance of considering maximum-likelihood method, the transition : trans- sub-optimal trees in a cladistic analysis. version ratio was set as part of the specified model. The Unweighted least-squares method produced For the distance-based methods, the Kimura distance the same tree irrespective of the model used {Table measure was used to estimate sequence divergence, 3), but this tree does not support monophyly of and the transition : transversion ratio was set to Surcocystis, as S. muris and S. neurona are removed either 1 or 2. For the Parsimony method, a step- from the clade. The contradiction of monophyly in 610 D. A. Morrison

, Eimeria

T. gondii gondii

N. caninum caninum

S. murk neurona

S. neurona muris

S. gigantea gigantea

S. fusiformis fusiformis

S. tenella S. tenella \/ S. arieticanis \S. arieticanis

Eimeria C / S. muris

S. neurona

T. gondii

N. caninum

S. gigantea

S. fusiformis

S. tenella

S. arieticanis Fig. 5. The 3 phylogenetic trees produced from the cladistic analysis of SarcocysfisSarcocystis ssrRNA. Table 3 shows which analysis produced which of these trees. Trees (A) and (B) support monophyly of Sarcocystis, whereas tree (C) does not,

this cladogram is supported only by a branch with produced by the original data set (Table 3). The negative length (i.e. the branch from the ancestor phylogeny of Sarcocystis can thus be usefully of the ingroup to the ancestor of T. gondii+ evaluated using this data set. N. caninum+S gigantea+S. fusiformis+S. tenella+S. The bootstrap resampling indicates relatively arieticanis). This situation is not realistic, and so strong support for the Sarcocystis clade (i.e. the tree cannot be considered seriously as a possible Sarcocystis is monophyletic in more than 85% of the representation of the real evolutionary history of replicates) for all of the methods in which it occurs these taxa. If this tree-building method is constrained (Table 3). This support is apparently stronger (i.e. the to use only positive branch lengths (which is an bootstrap percentage is higher) when transitions are option in the Phylip program), then the clado- considered to be more likely than transversions. Even gram produced does indeed support monophyly of the Unweighted least-squares method provides some Sarcocystis using both models (they produce the support for monophyly, because the negative branch- cladogram shown in Fig. 5A). length that contradicts the monophyly is actually The permutation tests indicate that there is very quite short. strong phylogenetic structure in these ssrRNA data, It is thus reasonable to conclude from this analysis as no permutations were found that produced that the ssrRNA sequence data do support mono- trees that were anywhere near as optimal as those phyly of Sarcocystis, and thus also support the way in Invited Review 611

Table 3-Results of the cladistic analysis of Surcocystis ssrRNA

Tree-building Optimality Permutation Bootstrap method” measure Treeb percentagec percentaged

Neighbor-joining ti=tv -e A -e 85 ti=2tv -e A e 91.5 Unweighted least-squares ti=tv 0.00241’ C 0 25 ti=2tv 0.00335r c 0 42.5 Weighted-parsimony ti=tv 3298 B 0 91 ti=hv 473h A 0 98 Maximum-likelihood ti=tv - 2499.7’ A 0 99 ti=2tv -2502.1’ A 0 99.5

“ti=tv, transitions and transversions equal; ti=2tv. transitions twice as likely as transversions. ‘As shown in Fig. 5. ‘Percentage of the permutation replicates that produced trees at least as optimal as the observed tree. dPercentage of the bootstrap replicates that support monophyly of Sarcocystis. ‘No optimality criterion, and so no calculations possible. ‘Sum of squares. aNumber of character-state changes. ‘Number of weighted character-state changes. ‘Log-likelihood.

which the Sarcocystidae is currently divided into 2 affect the outcome of the analysis. This is one of subfamilies (if the recently-described Neospora is the biggest sources of problems when analysing included in the Toxoplasmatinae) (note that in a molecular data. Popperian sense the data actually “fail to disprove” Include samples from more than 1 individual or the hypothesis rather than “support” it). This con- population if possible. This is the norm when assessing clusion is robust both to variations in the evolution- morphological characters, because if intra-taxon ary model and to the tree-building method used. It is variability exists then it will be confounded with inter- not, however, necessarily robust to the sequence taxon variability in the analysis. For molecular data, a alignment used (see Ellis & Morrison, 1995). consensus sequence can be used for each taxon. There does, however, seem to be some uncertainty The ingroup must be monophyletic for the clado- with respect to the placement of Smuris and S. gram to accurately reflect phylogenetic history. If neurona (which is the source of non-monophyly de- the analysis shows that the ingroup and outgroup tected by the previous studies of ssrRNA referred to taxa are inter-mixed in the cladogram, then it is above). This uncertainty is only likely to be resolved not possible to root the cladogram in such a way by including in the analysis other species that are that the ingroup is monophyletic. You should closely-related to these 2 species (i.e. that will join the thus seriously consider your original rationale for tree on the terminal branches leading to these 2 assuming monophyly. species). This is the obvious direction to take for Long branches on a cladogram (indicating future research into the question posed for the apparently extensive phylesis) should be avoided if analysis presented here. this is possible. They can often be shortened by including taxa that join the tree on these branches Some advice for phylogeny reconstruction. The fol- (i.e. the branches actually show speciation rather lowing is a list of topics that need to be considered than phylesis, but they are not doing so because the when carrying out a cladistic analysis. The topics are relevant taxa are not in the data set). more-or-less in the order in which they will need to be The outgroup should consist of more than 1 taxon considered in practice. closely-related to the ingroup. The more distantly- Include in the analysis all of the taxa for which related is the outgroup then the more arbitrary is the there are data. Adding or deleting any taxon may apparent position of the root on the cladogram. 612 D. A. Morrison

For molecular data, use entire sequencesif they are by statistics, in which a series of mathematical tests available. There is no reason to expect part of a gene have been developed (and are still being actively sequence (effectively a sample) to reflect accurately developed) that make different assumptions about the total gene sequence, unless all of the nucleotide the nature of the data to be analysed, so that after the positions agree on the phylogenetic branching order. data have been evaluated an informed decision can Homology assessments should be taken very be made about which test is the most appropriate for seriously. For molecular data, this means that se- the particular task at hand. Thus, out of the current quence alignment should be given careful considera- plethora of phylogenetic techniques should emerge a tion, using models of secondary structure if they are subset of mathematical methods that are known to be available. If computer programs are used, then vary robust and consistent under certain circumstances, the weighting parameters available in order to detect and it will be from amongst these methods that an those parts of the sequence alignment that are not objective choice (based on efficiency and power) will robust, and delete these regions from the analysis. be made by the cladist. Assess the characteristics of the data before the The current vogue for simply presenting the results cladogram is constructed (e.g., GC-content for of analyses using several tree-building methods is nucleotide data), in order to choose rationally among thus both creditable and naive. It is creditable the available tree-construction algorithms. because it emphasizes our current uncertainty about Use several cladogram-construction algorithms if cladistic analysis; and it is naive because there this seemsto be appropriate for the data. If heuristic appears to be no rationale being used for the choice tree-search methods are used then describe them from among the available methods. Clearly, what precisely, so that the likely successof having found must be done for each phylogenetic data set is for the the optimal tree can be assessed. For nucleotide nature of the data to be evaluated, and then for a data, correct for multiple character-state changes if reasoned choice to be made about which methods possible. might be appropriate for subsequent analysis, along Assessthe magnitude of the phylogenetic signal in with an assessment of the support for the phylo- the cladogram. genetic hypotheses generated. Two recent studies Assess the robustness of the monophyletic groups that incorporate a carefully-reasoned evaluation of represented on the cladogram. the data, followed by a series of phylogenetic Quantitatively compare the cladograms if there is analyses that are themselves then subjected to re- more than one (e.g., multiple optimal trees, trees evaluation, are those of Lento et al. (1995) and from different cladogram-construction algorithms, or Penny et al. (1995), and these could serve as useful trees from different data sets). models. Remember that the cladogram represents a series of hypotheses that are amenable to further test. It is not a representation of reality, nor is it the final word Acknowledgements-Thanks to Steve Barker, Peter Baverstock, Dave Booth, Mike Crisp, John Ellis, Dan on the phylogeny of the taxa analysed. Faith, Alan Johnson and John Trueman for commenting on an earlier version of the manuscript; and to John Ellis for Conclusion. Many biologists treat data analysis in providing the ssrRNA data file. a manner that is completely antithetical to their apparent attitude to the other, more biological aspects of their scientific pursuits (cf. Felsenstein, REFERENCES 1988). In particular, there is a widespread view Adachi J. & Hasegawa M. 1995. Phylogeny of whales: among non-experts that there should be only 1 dependence of the inference on species sampling. Molecular Biology and Evolution 12: 177-I 79. correct method for phylogenetic data analysis, and Alroy J. 1994. Four permutation tests for the presence of that it is about time that the experts told us what it is. phylogenetic structure. Systematic Biology 43: 430-437. This view is clearly rather naive, because it assumes Archie J. W. 1989. Homoplasy excess ratios: new indices for that all phylogenetic data sets have identical charac- measuring levels of homoplasy in phylogenetic teristics. This is patently untrue, as it will be for UN systematics and a critique of the consistency index. Systematic Zoology 38: 2533269. data sets (whether biological or not) that incorporate Barta J. R., Jenkins M. C. & Danforth H. D. 1991. an historical component. Furthermore, currently no Evolutionary relationships of avian Eimeria species methods are simultaneously efficient, powerful, con- among other apicomplexan protozoa: monophyly of sistent and robust (Penny et al., 1990, 1992), and so the Apicomplexa is supported. Molecular Biology and Evolution 8: 345-355. there is no single ideal method. Beanland T., J. & Howe C. J. 1992. The inference of The future development of cladistic techniques will evolutionary trees from molecular data. Comparative presumably follow the well-documented path taken Biochemistry and Physiology 102B: 643-659. Invited Review 613

Beyer W. A., Stein M. L., Smith T. F. & Ulam S. M. 1974. Darwin C. 1859. On the Origin of Species by Means of A molecular-sequence metric and evolutionary trees. Natural Selection. John Murray, London. Mathematical Biosciences 19: 9-25. Davis J. I. 1993. Character removal as a means for assessing Bremer K. 1994. Branch support and tree stability. the stability of clades. Cladistics 9: 201-210. Cladistics 10: 295-304. deBry R. W. 1992. The consistency of several phylogeny- Brooks D. R. & McLennan D. A. 1992. The evolutionary inference methods under varying evolutionary rates. origin of Plasmodium falciparum. Journal of Parasitology Molecular Biology and Evolution 91 537-551. 78: 564566. de Pinna M. C. C. 1991. Concepts and tests of homology -. in Bryant H. N. 1992. The polarization of character transfor- the cladistic paradigm. Clad&tics 7: 367-394. mations in phylogenetic systematics: role of axiomatic and de Queiroz A., Donoghue M. J. & Kim J. 1995. Separate auxiliary assumptions. Systematic Zoology 40: 433-5. versus combined analysis of phylogenetic evidence. Camin J. H. & Sokal R. R. 1965. A method for deduc- Annual Review of Ecology and Systematics 26: 657-68 1. ing branching sequences in phylogeny. Evolution 19: Dobson A. J. 1974. Unrooted trees for numerical 31 l-326. taxonomy. Journal of Applied Probability 11: 32-42. Cao Y., Adachi J., Janke A., Plabo S. & Hasegawa M. Doolittle R. F. (Editor) 1990. Molecular evolution: 1994. Phylogenetic relationships among eutherian orders computer analysis of protein and nucleic acid sequences. estimated from inferred sequences of mitochondrial pro- Methods in Enzymology 183: 303-502. teins: instability of a tree based on a single gene. Journal Doyle J. J. 1992. Gene trees and species trees: molecular of Molecular Evolution 39: 519-527. systematics as one-character taxonomy. Systematic Cavalli-Sforza L. L. & Edwards A. W. F. 1967. Phylo- Botany 17: 144163. genetic analysis: models and estimation procedures. Edwards A. W. F. & Cavalli-Sforza L. L. 1964. Reconstruc- American Journal of Human Genetics 19: 233-257 (also tion of evolutionary trees. In: Phenetic and Phylogenetic published in Evolution 32: 550-570). Classification (Edited by Heywood V. H. & McNeil J.). Cavender J. A. 1978. Taxonomy with confidence. pp. 67-76. Systematics Association, London. Mathematical Biosciences 40: 271-280. Ellis J. & Morrison D. 1995. Effects of sequence alignment Cavender J. A. 1978. Erratum. Mathematical Biosciences 44: on the phylogeny of Sarcocyscis deduced from 18s 308. rDNA seauences. Parasitologv Research 81: 696699. Cavender J. A. & Felsenstein J. 1987. Invariants of phylo- Ellis .I., Lutoh K., Baverstock PTR., Brindley P. J., Nimmo genies in a simple case with discrete states. Journal of K. A. & Johnson A. M. 1994. The phylogeny of Classification 4: 57-71. Neospora car&urn. Molecular Biochemistry and Chakraborty R. 1977. Estimation of time of divergence Parasitology 64: 303-3 11. from phylogenetic studies. Canadian Journal of Genetics Estabrook G. F. & McMorris F. R. 1980. When is one and Cytology 19: 217-223. estimate of evolutionary relationships a refinement Chan S. C., Wong A. K. C. & Chiu D. K. Y. 1992. A survey of another? Journal of Mathematical Biology 10: of multiple sequence comparison methods. Bulletin of 367-373. Mathematical Biology 54: 563-598. Faith D. P. 1985. Distance methods and the approximation Charleston M. A., Hendy M. D. & Penny D. 1994. The of most-parsimonious trees. Syscernaric Zoology 34: effects of sequence length, tree topology, and number 312-325. of taxa on the performance of phylogenetic methods. Faith D. P. 1991. Cladistic permutation tests for monophyly Journal of Computational Biology 1: 133-I 5 1. and nonmonophyly. Systematic Zoology 40: 366375. Chase M. W., Soltis D. E., Olmstead R. G., Morgan D., Les Faith D. P. & Ballard J. W. 0. 1994. Length differences and D. H., Mishler B. D., Duvall M. R., Price R. A., Hills topology-dependent tests: a response-to KIllersjo et al. H. G., Qiu Y.-L., Kron K. A., Rettig J. H.. Conti E., Cladistics 10: 57&t. Palmer J. D., Manhart J. R., Sytsma K. J., Michaels Farris J. S. 1969. A successive approximations approach to H. J., Kress W. J., Karol K. G., Clark W. D., Hedrin character weighting. Systematic Zoology 18: 374385. M., Gaut B. S., Jansen R. K., Kim K.-J., Wimpee C. F., Farris J. S. 1970. Methods for computing Wagner trees. Smith J. F., Furnier G. R., Strauss S. H., Xiang Q.. Systematic Zoology 19: 83-92. Plunkett, G. M.. Soltis P. S., Swensen S. M., Williams Farris J. S. 1972. Estimating phylogenetic trees from S. E., Gadek P. A., Quinn C. J., Eguiarte L. E., distance matrices. American Naturalist 106: 645-668. Golenberg E.. Learn G. H., Graham S. W., Barrett Farris J. S. 1977. On the phenetic approach__ to vertebrate S. C. H.,bayanandan S. & Albert V. A. 1993. Phylo- classitication. In: Major Patterns in Vertebrate Evolution genetics of seed plants: an analysis of nucleotide (Edited bv Hecht M. D.. Goodv P. C. & Hecht B. M.). sequences from the plastid gene rbcL. Annals of the pp. 8233850. Plenum Press, New York. Missouri Botanical Garden 80: 5288580. Farris J. S. 1989. The retention index and the resealed Clifford H. T. & Stephenson W. 1975. An Introduction to consistency index. Cladistics 5: 417419. Numerical Classification. Academic Press, New York. Farris J. S. 1995. Excess homoplasy ratios. Ciadistics 7: Colless D. H. 1970. The phenogram as an estimate of 81-91. ohvlogenv. Svstematic Zoologv 19: 352-362. Farris J. S., KHllersjG M., Kluge A. G. & Bult C. 1995. Co; F. E. G. -1994. The evolutionary expansion of the Testing significance of incongruence. Cladistics 10: 315- Sporozoa. International Journal for Parasitology 24: 319. 1301-1316. Felsenstein J. 1973. Maximum-likelihood estimation of evo- Crisci J. V. & Stuessy T. F. 1980. Determining primitive lutionary trees from continuous characters. American character states for phylogenetic reconstruction. System- Journal of Human Genetics 25: 471492. atic Botany 5: 112-135. Felsenstein J. 1973. Maximum-likelihood and minimum- Cummings M. P., Otto S. P. & Wakeley J. 1995. Sampling steps methods for estimating evolutionary trees from properties of DNA sequence data in phylogenetic analy- data on discrete characters. Systematic Zoology 22: sis. Molecular Biology and Evolution 12: 814-822. 24&249. 614 D. A. Morrison

Felsenstein J. 1978. The number of evolutionary trees. Graham R. L. & Foulds L. R. 1982. Unlikelihood that Systematic Zoology 21: 27-33. minimal phylogenies for a realistic biological study Felsenstein J. 1978. Cases in which parsimony and compat- can be constructed in reasonable computational time. ibility methods will be positively misleading. Systematic Mathematical Biosciences 60: 133-142. Zoology 27: 401410. Gutell R. R., Weiser B., Woese C. R. & Noller H. F. 1985. Felsenstein J. 1981. Evolutionary trees from DNA Comparative anatomy of 16-S-like ribosomal RNA. sequences: a maximum likelihood approach. Journal of Progress in Nucleic Acids Research and Molecular Molecular Evolution 17: 368-316. Biology 32: 155-216. Felsenstein J. 1981. A likelihood approach to character Hall B. (Editor) 1994. Homology: The Hierarchical Basis of weighting and what it tells us about parsimony and Comparative Biology. Academic Press, San Diego. compatibility. Biological Journal of the Linnean Society Hasegawa M., Kishino H. & Yano T. 1985. Dating of the 16: 183-196. human-ape splitting by a molecular clock of mito- Felsenstein J. 1981. Evolutionary trees from gene fre- chondrial DNA. Journal of Molecular Evolution 22: quencies and quantitative characters: finding maximum 160-174. likelihood estimates. Evolution 35: 122991242. Hasegawa M. & Fujiwara M. 1993. Relative efficiencies Felsenstein J. 1982. Numerical methods for inferring of the maximum likelihood, maximum parsimony, evolutionary trees. Quartet+ Review of Biology 51: and neighbor-joining methods for estimating protein 379404. phylogeny. Molecular Phylogenetics and Evolution 2: Felsenstein J. 1988. Phylogenies from molecular sequences: l-5. inference and reliability. Annual Review of Genetics 22: Hasegawa M. & Hashimoto T. 1993. Ribosomal RNA trees 521-565. misleading? Nature 361: 23. Hendy M. D. 1989. The relationship between simple Felsenstein J. 1992. Phylogenies from restriction sites: a maximum-likelihood approach. Evolution 46: 159-173. evolutionary tree models and observable sequence data. Systematic Zoology 38: 310-321. Felsenstein J. 1995. PHYLIP (Phylogeny Inference Hendy M. D. 1991. A combinatorial description of the Package). University of Washington, Seattle. closest tree algorithm for finding evolutionary trees. Fitch W. M. 1970. Distinguishing homologous from analo- Discrete Mathematics 96: 5 l-58. gous proteins. Systematic Zoology 19: 99-l 13. Hendy M. D. & Charleston M. A. 1993. Hadamard con- Fitch W. M. 1971. Toward defining the course of evolution: jugation: a versatile tool for modelling nucleotide minimum change for a specific tree topology. Systematic sequence evolution. New Zealand Journal of Botany 31: Zoology 20: 406416. 23 l-237. Fitch W. J. 1975. Toward finding the tree of maximum Hendy M. D. & Penny D. 1982. Branch and bound parsimony. In: Proceedings of the Eighth International algorithms to determine minimal evolutionary trees. Conference on Numerical Taxonomy (Edited by Mathematical Biosciences 59: 277-290. Estabrook G. F.), pp. 189-230. Freeman, San Hendy M. D. & Penny D. 1989. A framework for the Francisco. quantitative study of evolutionary trees. Systematic Fitch W. M. 1981. A non-sequential method for construct- Zoology 38: 297-309. ing trees and hierarchical classifications. Journal of Hendy M. D. & Penny D. 1993. Spectral analysis of Molecular Evolution 18: 3&37. phylogenetic data. Journal of Classification 10: 5-24. Fitch W. M. & Margoliash E. 1967. Construction of Hendy M. D., Steel M. A. & Penny D. 1994. A discrete phylogenetic trees. Science 155: 279-284. Fourier analysis for evolutionary ttees. Proceedings of Fitch W. M. & Yasunobu K. T. 1975. Phylogenies from the National Academy of Sciences of the U.S.A. 91: amino acid sequences with gaps. Journal of Molecular 3339-3343. Evolution 5: l-24. Hennig W. 1966. Phylogenetic Systematics. University of Forey P. L., Humphries C. J., Kitching I. J., Scotland Illinois Press, Urbana IL. (Translated by Davis D. D. & R. W., Siebert D. J. & Williams D. M. 1992. Cladistics. Zangerl R. from Hennig W. 1950. Grund:iige einer A Practical Course in Systematics. Clarendon Press, Theorie der Phylogenetischen Systematik. Deutscher Oxford. Zentralverlag. Berlin.). Hill L. R. 1975. Problems arising from some tests of Le Fukami K. & Tateno Y. 1989. On the uniqueness of the Quesne’s concept of uniquely derived characters. In: maximum likelihood method for estimating molecular Proceedings af the Eighth International Conference on trees: uniqueness of the likelihood point.- Journal of Numerical Taxonomy (Edited by Estabrook G. F.), Molecular Evolution 31: 5 1 I-523. pp. 375-398. Freeman, San Fmncisco. Gatesy J., DeSalle R. & Wheeler W. 1993. Alignment- Hillis D. M. & Huelsenbeck J. P. 1992. Signal, noise. and ambiguous nucleotide sites and the exclusion of system- reliability in molecular phylogenetic analyses. Journal of atic data. Molecular Phvlogenetics and Evolution 2: Heredity 83: 189-195. 152-157. Hillis D. M., Allard M. W. & Miyamoto M. M. 1993. Gaut B. S. & Lewis P. 0. 1995. Success of maximum Analysis of DNA sequence data: phylogenetic inference. likelihood phylogeny inference in the four-taxon case. Methods in Enzymology 224: 456487. Molecular Biology and Evolution 12: 152162. Hillis D. M.. Huelsenbeck J. P. & Cunningham C. W. 1994. Gojobori T., Moriyama E. N. & Kimura M. 1990. Statisti- Application and accuracy of molecular phylogenies. cal methods for estimating sequence divergence. Methods Science 264: 671-677. in Enzymology 183: 431-550. Hillis D. M., Huelsenbeck J. P. & SwotTord D. L. 1994. Goloboff P. A. 1991. Homoplasy and the choice among Hobgoblin of phylogenetics? Nature 369: 363-364. cladograms. Cladistics 7: 215-232. Hubalek Z. 1982. Coefficients of association and similarity Golobolf P. A. 1993. Estimating character weights during based on binary (presence-absence) data: an evaluation. tree search. Cladistics 9: 83-91. Biological Reviews 51: 669689. Invited Review 615

Huelsenbeck J. P. 1991. Tree-length distribution skewness: Li W.-H. 198 1. Simple method for constructing phylo- an indicator of phylogenetic information. Systematic genetic trees from distance matrices. Proceedings of zoology 40: 257-270. the National Academy of Sciences of the U.S.A. 78: Huelsenbeck J. P. 1995. Performance of phylogenetic 1085-1089. methods in simulation. Systematic Biology 44: 1748. Li W.-H. & Gouy M. 1991. Statistical methods for testing Huelsenbeck J. P. 1995. The robustness of two phylo- molecular phylogenies. In: Phylogenetic Analysis of DNA genetic methods: four-taxon simulations reveal a slight Sequences (Edited by Miyamoto M. M. & Cracraft J.), superiority of maximum likelihood over neighbor pp. 2499277. Oxford University Press, New York. joining. Molecular Biology and Evolution 12: 843-849. Li W.-H. & Graur D. 1991. Fundamentals of Molecular Huelsenbeck J. P. & Hillis D. M. 1993. Success of phylo- Evolution. Sinauer Associates, Sunderland. genetic methods in the four-taxon case. Sysrematic Li W.-H. & Zharkikh A. 1994. What is the bootstrap Biology 4L: 247-264. technique? Systematic Biology 43: 424430. Jin L. & Nei M. 1990. Limitations of the evolutionary Li W.-H. & Zharkikh A. 199.5. Statistical tests of DNA parsimony method of phylogenetic analysis. Molecular phylogenies. Systematic Biology 44: 49-63. Biology & Evolution 7: 82-102. Li W.-H., Wolfe K. H., Sourdis J. & Sharp P. M. 1987. Kallersjii M., Farris J. S., Kluge A. G. & Bult C. 1992. Reconstruction of phylogenetic trees and estimation of Skewness and vermutation. Cladistics 8: 275-287. divergence times under nonconstant rates of evolution. Kidd K. K. & Sgaramella-Zonta L. A. 1971. Phylogenetic Cold Spring Harbor Symposium on Quantitative Biology analysis: concepts and methods. American Journal of 52: 847-856. Human Genetics 23: 235-252. Lockhart P. J., Penny D., Hendy M. D., Howe C. J., Kim J. & Burgman M. A. 1988. Accuracy of phylogenetic- Beanland T. J. & Larkum A. W. D. 1992. Contro- estimation methods under unequal evolutionary rates. versy on chloroplast origins. Federation of European Evolution 42: 596602. Biochemical Societies Letters 301: 127-131. Kimura M. 1980. A simple method for estimating evolu- Lockhart P. J., Steel M. A., Hendy M. D. & Penny D. 1994. tionary rates of base substitutions through comparative Recovering evolutionary trees under a more realistic studies of nucleotide sequences. Journal of Molecular model of sequence evolution. Molecular Bio1og.v and Evolution 16: 11 I-120. Evolution 11: 605-612. Kishino H. & Hasegawa M. 1989. Evaluation of the maxi- Maddison D. R. 1991. The discovery and importance of mum likelihood estimate of the evolutionary tree topolo- multiple islands of most-parsimonious trees. Systematic gies from DNA sequence data, and the branching order Zoology 40: 315-328. in Hominoidea. Journal of Molecular Evolution 29: Maddison W. P. & Maddison D. R. 1992. MacClade: 170-179. Analysis of Phylogeny and Character Evolution. Sinauer Kishino H., Miyata T. & Hasegawa M. 1990. Maximum Associates, Sunderland. likelihood inference of protein phylogeny, and the origin Maddison W. P., Donoghue M. J. & Maddison D. R. 1984. of chlorovlasts. Journal of Molecular Evolution 30: Outgroup analysis and parsimony. Systematic Zoology 151-160. - 33: 83-103. Klotz L., C. & Blanken R. L. 1981. A practical method Marshall C. R. 1992. Substitution bias, weighted parsimony, for calculating evolutionary trees from sequence data. and amniote phylogeny as inferred from 18s rRNA Journal of Theoretical Biology 91: 261-272. sequences. Molecular Biology and Evolution 9: 37c373. Kuhner M. K. & Felsenstein J. 1994. A simulation compari- Mayr E. 1982. The Growth of Biological Thought. Harvard son of phylogeny algorithms under equal and unequal University Press, Harvard. evolutionary rates. Molecular Biology and Evolution 11: McClure M. A., Vasi T. K. & Fitch W. M. 1994. Compara- 459468. tive analysis of multiple protein-sequence alignment Lake J. A. 1987. A rate-independent technique for analysis methods. Molecular Biologv and Evolution 11: 571-592. of nucleic acid sequences: evolutionary parsimony. Michener C. D. & Sokal R. R. 1957. A quantitative Molecular Biology and Evolution 4: 167-191. ;;:;;;I to a problem in classification. Evolution 11: Lake J. A. 1988. Origin of the eukaryotic nucleus as determined by rate-invariant analysis of rRNA. Nature Mickevich M. F. 1982. Transformation series analysis. 331: 184188. Systematic Zoology 31: 461478. Le Quesne W. J. 1969. A method of selection of charac- Mindell D. P. 1991. Aligning DNA sequences: homology ters in numerical taxonomy. Sysfematic Zoology 18: and phylogenetic weighting. In: Phylogenetic Analysis 201-205. of DNA Sequences (Edited by Miyamoto M. M. & Le Quesne W. J. 1974. The uniquely evolved character Cracraft J.), pp. 73-89. Oxford University Press, New concept and its cladistic application. Sysfemaric Zoology York. 23: 513-517. Miyamoto M. M. & Cracraft J. 1991. Phylogenetic Le Quesne W. J. 1982. Compatibility analysis and its inference, DNA sequence analysis, and the future of applications. Zoological Journal of the Linnean Society molecular systematics. In: Phylogenetic Analysis of DNA 74: 267-275. Sequences (Edited by Miyamoto M. M. & Cracraft J.), Lento G. M., Hickson R. E., Chambers G. K. & Penny D. pp. 3317. Oxford University Press, New York. 1995. Use of spectral analysis to test hypotheses on the Nadler S. A. 1990. Molecular approaches to studying origin of pinnipeds. Molecular Biology and Evolution 12: helminth population genetics and phylogeny. inter- 28852. national Journai for Parasitology 20: 11-29. Levine N. D. 1985. Phylum II. Apicomplexa Levine 1970. Neff N. A. 1986. A rational basis for a priori character In: An Illustrated Guide to the Protozoa (Edited by Lee weighting. Systematic Zoology 35: 110-123. J. J., Hutner S. H. & Bovee E. C.), pp. 322-374. Society Nei M. 1987. Molecular Evolutionary Genetics. Columbia of Protozoologists, Lawrence. University Press. New York. 616 D. A. Morrison

Nei M. 1991. Relative efficiencies of different tree-making Rinsma-Melchert I. 1993. The expected number of matches methods for molecular data. In: Phylogenetic Analysis of in optimal global sequence alignments. New Zealand DNA Sequences (Edited by Miyamoto M. M. & Cracraft Journal of Botany 31: 219-230. J.), pp. 90-128. Oxford University Press, New York. Rodrigo A. G., Kelly-Borges M., Bergquist P. R. & Nei M. & Li W.-H. 1979. Mathematical model for studying Bergquist P. L. 1993. A randomisation test of the null genetic variation in terms of restriction endonucleases. hypothesis that two cladograms are sample estimates of a Proceedings of the National Academy of Sciences of the parametric phylogenetic tree. New Zealand Journal of U.S.A. 16: 526995213. Botany 31: 257-268. Nelson G. J. 1978. Ontogeny, phylogeny, palaeontology, Rodriguez F., Oliver J. L., Marin A. & Medina J. R. 1990. and the biogenetic law. Systematic Zoology 27: The general stochastic model of nucleotide substitution. 324345. Journal of Theoretical Biology 142: 485-501. Nelson G. J. & Platnick N. I. 1981. Systematics and Rogers J. S. 1991. A comparison of the suitability of the Biogeography. Cladistics and Vicariance. Columbia Rogers, Modified Rogers, Manhattan, and Cavalli- University Press, New York. Sforza and Edwards distances for inferring phylogenetic Nixon K. C. & Carpenter J. M. 1993. On outgroups. trees from allele frequencies. Systematic Zoology 40: Cladistics 9: 4133426. 63-73. Olsen G. J. 1987. Earliest phylogenetic branchings: compar- Rohlf F. J., Chang W. S., Sokal R. R. & Kim J. 1990. Accuracy of estimated phylogenies: effects of tree topol- ing rRNA-based evolutionary trees inferred with ogy and evolutionary model. Evolution 44: 1671-1684. various techniques. Cold Spring Harbor Symposium on Romer A. S. 1971. The Vertebrate Body, 4th edition. Quantitative Biology 52: 825-837. Saunders, Philadelphia. Olsen G. J. 1988. Phylogenetic analysis using ribosomal Rzhetsky A. & Nei M. 1992. A simple method for esti- RNA. Methods in Enzymology 164: 793-812. mating and testing minimum-evolution trees. Molecular Ota T. & Nei M. 1994. Estimation of the number of amino Biology & Evolution 9: 945-967. acid substitutions per site when the substitution rate Rzhetsky A. & Nei M. 1993. Theoretical foundation of the varies among sites. Journal of Molecular Evolution 38: minimum-evolution method of phylogenetic inference. 642-643. Molecular Biology and Evolution IO: 1073-1095. Page R. D. M. 1993. On the islands of trees and the Rzhetsky A. & Nei M. 1995. Tests of applicability of several efficiency of different methods of branch swapping in substitution models for DNA sequence data. Molecular finding most-parsimonious trees. Systematic Biology 42: Biology and Evolution 12: 131-151. 200-210. Saitou N. & Imanishi T. 1989. Relative efficiences of Pamilo P. & Nei M. 1988. Relationship between gene trees the Fitch-Margoliash, maximum-parsimony, maximum- and species trees. Molecular Biology and Evolution 5: likelihood, minimum-evolution, and neighbor-joining 568-583. methods of phylogenetic tree construction in obtaining Patterson C. 1988. Homology in classical and molecular the correct tree. Molecular Biology and Evolution 6: biology. Molecular Biology and Evolution 5: 603-625. 514-525. Penny D. 1982. Towards a basis for classification: the Saitou N. & Nei M. 1987. The neighbor-joining method: incompleteness of distance measures, incompatibility a new method for reconstructing phylogenetic trees. analysis and phenetic classification. Journal of Molecular Biology and Evolution 4: 406425. Theoretical Biology 96: 129-142. Sanderson M. J. 1995. Objections to bootstrapping phylo- Penny D. & Hendy M. D. 1986. Estimating the reliability of genies: a critique. Systematic Biology 44: 299-320. evolutionary trees. Molecular Biology and Evolution 3: Sankoff D. 1975. Minimal mutation trees of sequences. 403417. SIAM Journal of Applied Mathematics 28: 3542. Penny D., Hendy M. D. & Henderson I. M. 1987. Reliabil- Sankoff D. & Cedergren R. J. 1983. Simultaneous compari- ity of evolutionary trees. Cold Spring Harbor Symposium son of three or more sequences related by a tree. In: Time on Quantitative Biology 52: 8577862. Warps, String Edits, and Macromolecules: The Theory and Practice of Sequence Comparisons (Edited by Sankotf Penny D., Hendy M. D., Zimmer E. A. & Hamby R. K. 1990. Trees from sequences: panacea or Pandora’s box? D. & Kruskal J. B.), pp. 253-263. Addison-Wesley, London. Australian Systematic Botany 3: 21-38. Sattath S. & Tversky A. 1977. Additive similarity trees. Penny D., Hendy M. D. & Steel M. A. 1992. Progress with Psychometrika 42: 319-345. methods for constructing evolutionary trees. Trends in Sharkey M. J. 1989. A hypothesis-independent method of Ecology and Evolution 7: 73-19. character weighting for cladistic analysis. Cladistics 5: Penny D., Watson E. E., Hickson R. E. & Lockhart P. J. 63-86. 1993. Some recent progress with methods for evolution- Shoemaker J. S. & Fitch W. M. 1989. Evidence from ary trees. New Zealand Journal of Botany 31: 275-288. nuclear sequences that invariable sites should be consid- Penny D., Steel M., Waddell P. J. & Hendy M. D. 1995. ered when sequence divergence is calculated. Molecular Improved analyses of human mtDNA sequences support Biologv and Evolution 6: 210-289. a recent African origin of Homo sapiens. Molecular Siddall %. E. 1994. Randomcladistics. University of Biology and Evolution 12: 863-882. Toronto, Toronto. Pollock D. D. & Goldstein D. B. 1995. A comparison of Sidow A. & Wilson A. C. 1990. Compositional statistics: an two methods for constructing evolutionary distances improvement of evolutionary parsimony and its applica- from a weighted contribution of transition and transver- tion to deep branches in the tree of life. Journal of sion differences. Molecular Biology and Evolution 12: Molecular Evolution 31: 51-68. 713-717. Sidow A. & Wilson A. C. 1991. Compositional statistics Popper K. R. 1957. The Poverty of Historicism. Routledge evaluated by computer simulations. In: Phylogenetic & Kegan Paul, London. Analysis of DNA Sequences (Edited by Miyamoto M. M. Invited Review 617

& Cracraft J.), pp. 129-146. Oxford University Press, Tateno Y., Nei M. & Tajima F. 1982. Accuracy of estimated New York. phylogenetic trees from molecular data. I. Distantly Smith A. B. 1994. Rooting molecular trees: problems and related trees. Journal of Molecular Evolution 18: strategies. Biological Journal of the Linnean Society 51: 387404. 279-292. Tateno Y., Takezaki N. & Nei M. 1994. Relative efficiencies Sneath P. H. A. 1989. Analysis and interpretation of of the maximum-likelihood, neighbor-joining, and sequence data for bacterial systematics: the view of a maximum-parsimony methods when substitution rates numerical taxonomist. Systematic and Applied Microbi- vary with site. Molecular Biology and Evolution 11: ology 12: 15-31. 261-277. Sokal R. R. & Michener C. D. 1958. A statistical method Tenter A. M., Baverstock P. R. & Johnson A. M. 1992. for evaluating systematic relationships. University of Phylogenetic relationships of Sarcocysfis species from Kansas Science Bulletin 28: 1409-1438. sheep, goats, cattle and mice based on ribosomal RNA Sourdis J. & Krimbas C. 1987. Accuracy of phylogenetic sequences. International Journal for Parasitology 22: trees estimated from DNA sequence data. Molecular 503-513. Biology and Evolution 4: 159-l 66. Thorne J. L. & Kishino H. 1992. Freeing phylogenies from Specht T., Wolters J. & Erdman V. A. 1990. Compilation of artifacts of alignment. Molecular Biology & Evolution 9: 5s rRNA and 5s rRNA gene sequences. Nucleic Acids 1148-1162. Research Suppl. 18: 2215-2235. Tillier E. R. M. & Collins R. A. 1995. Neighbor joining and Steel M. A. 1994. The maximum likelihood point for a maximum likelihood with RNA sequences: addressing phylogenetic tree is not unique. Systematic Biology 43: the interdependence of sites. Molecular Biology and 560-564. Evolution 12: 7-15. Steel M. A., Hendy M. D. & Penny D. 1988. Loss of Van de Peer Y., Van den Broeck I., De Rijk P. & De information in genetic distances. Nature 336: 118. Wachter R. 1994. Database on the structure of small Steel M. A., Hendy M. D. & Penny D. 1993. Parsimony can ribosomal subunit RNA. Nucleic Acids Research 21 be consistent! Systematic Biology 42: 581-587. 3488-3494. Steel M. A., Hendy M. D., Szbkely L. A. & Erd(is P. L. 1992. Spectral analysis and a closest tree method for Waterman M. S., Smith T. F., Singh M. & Beyer W. A. genetic sequences. Applied Mathematics Letters 5: 63-67. 1977. Additive evolutionary trees. Journal of Theoretical Stevens P. F. 1980. Evolutionary polarity of character Biology 64: 199-213. * states. Annuaf Review of Ecology and Systemafics 11: Waterman M. S. 1989. Mathematical Methods for DNA 333-358. Sequences. CRC Press, Boca Raton. Stevens P. F. 1984. Homology and phylogeny: morphology Watrous L. E. & Wheeler Q. D. 1981. The outgroup and systematics. Systema& Bot&y 9: 395409. comparison method of character analysis. Systematic Stevens P. F. 1994. The Development of Biological System- zoology 30: l-l 1. atics: Anroine-Laurent de Jusiieu, Nature, and the kafural West J. G. & Faith D. P. 1990. Data, methods and System. Columbia University Press, New York. assumptions in phylogenetic inference. Australian Stewart C.-B. 1993. The powers and pitfalls of parsimony. Systematic Botany 3: 9-20. Nature 361: 603-607. Weston P. H. 1988. Indirect and direct methods in Studier J. A. & Keppler K. J. 1988. A note on the systematics. In: Ontogeny and Sysfematics (Edited by neighbor-joining algorithm of Saitou and Nei. Molecular Humphries C. J.), pp. 27-56. Columbia University Press, Biology and Evolution 5: 729-731. New York. Swofford D. L. 1981. On the utility of the distance Wagner Weston P. H. 1994. Methods for rooting cladistic trees. In: procedure. In: Advances in Cladistics (Edited by Funk Models in Phvlopenv Reconstruction (Edited bv Scotland V. A. & Brooks D. R.), pp. 2543. New York Botanical R. W., Siebeit b. j. & Williams D: M.), pp- 125-155. Garden, New York. Clarendon Press, Oxford. Swofford D. L. 1991. When are phylogeny estimates from Wheeler W. C. 1990. Nucleic acid sequence phylogeny and molecular and morphological data incongruent? In: random outgroups. Cladisfics 6: 363-367. Phylogenetic Analysis of DNA Sequences (Edited by Wheeler W. C. 1995. Sequence alignment, parameter Miyamoto M. M. & Cracraft J.), pp. 295-333. Oxford sensitivity, and phylogenetic analysis of molecular data. University Press, New York. Systematic Biology 44: 321-331. Swofford D. L. 1993. PA UP: Phylogenetic Analysis Using Williams D. M. 1993. A note on molecular homology: Parsimonv. Smithsonian Institution, Washington. multiple patterns from single datasets. Cladistics 9: Swofford D. L. & Berlocher S. H. 1987. Inferring evolution- 233-245. ary trees from gene frequency data under the-principle of maximum oarsimonv. Svstematic Zoology 36: 293-325. Williams P. L. & Fitch W. M. 1989. Finding the minimal Swofford D. i. & Mahdiion W. P. 1987ykeconstructing change in a given tree. In: The Hierarchy of Life: ancestral character states under Wagner parsimony. Mole&es and Morphology in Phylogenetk knal&is Mathematical Biosciences 87: 199-229. (Edited bv Fernholm B.. Bremer K. & Jiirnvall H.), Swofford D. L. & Olsen G. J. 1990. Phylogeny reconstruc- pp. 453470. Elsevier, Amsterdam. tion. In: Molecular Sysfematics (Edited by Hillis D. M. Yang 2. 1995. Evaluation of several methods for estimating & Moritz C.), pp. 411-501. Sinauer Associates, phylogenetic trees when substitution rates differ over Sunderland. nucleotide sites. Journal of Molecular Evolution 40: Takezaki N. & Nei M. 1994. Inconsistency of the maximum 689-697. parsimony method when the rate of nucleotide substitu- Zharkikh A. & Li W.-H. 1993. Inconsistency of the tion is constant. Journal oj’ Molecular Evolution 39: maximum-parsimony method: the case of five taxa with a 210-218. molecular clock. Systematic Biology 42: 113-125.