- Top of page
- Phylogenetic assumptions and evolutionary principles
- Problems with incorporating phylogeny-neutral principles into systematics
Many have argued strongly that incorporation of evolutionary theory into systematics is dangerously circular, while others have maintained that such an integrated approach increases the accuracy of phylogenetic inference. Here, it is demonstrated that such blanket statements regarding exclusion or inclusion of evolutionary principles in systematics fail to distinguish between two very different types of principles. ‘Phylogeny-neutral’ evolutionary principles are those inferred without any recourse to specific phylogenetic hypotheses (e.g. via developmental genetics, biomechanics). In contrast, ‘phylogeny-dependent’ principles are those which can only be inferred on the basis of specific phylogenetic hypotheses (e.g. character associations detected via ‘comparative methods’). Inclusion of phylogeny-neutral principles in systematic studies as a priori assumptions can be justified, since these principles have (often strong) external empirical support from other spheres of investigation. However, inclusion of phylogeny-dependent principles in systematic studies is circular, since such principles have no external empirical support but are themselves derived from systematic studies. Advocating inclusion or exclusion of all (or as many as possible) evolutionary principles from phylogenetic analysis is therefore misguided. Rather, phylogeny-neutral principles are independently supported and can be included, while phylogeny-dependent principles are unjustified assumptions and should be excluded to avoid circularity. However, integration of complex phylogeny-neutral principles in systematics can create operational problems, even though there are no methodological reasons against their inclusion.
- Top of page
- Phylogenetic assumptions and evolutionary principles
- Problems with incorporating phylogeny-neutral principles into systematics
Incorporation of evolutionary assumptions into systematics – in particular cladistic analysis – is much debated (e.g. Hull, 1988; Scott-Ram, 1990; Panchen, 1992). Since the publication of the English version of Hennig’s pivotal book more than three decades ago ( Hennig, 1966), there has been a movement towards minimizing assumptions of evolutionary process in cladistic analyses. This culminated in the pattern cladist view that systematics (and in particular cladistic analysis) can and should be undertaken completely independently of assumptions of evolutionary process (e.g. Platnick, 1979; Patterson, 1980, 1988, 1994; Nelson & Platnick, 1981; Rosen, 1984; Brady, 1985, 1994; Schoch, 1986; Rieppel, 1988).
Certain aspects of this view have been shown to be untenable. According to Platnick’s (1979) influential paper, there are three basic axioms of pattern cladism.
1‘Nature is ordered in a single specifiable pattern which can be represented by a branching diagram’. Cladistic analysis seeks to order organisms into a (mostly dichotomous) nonoverlapping hierarchy. The attempt to force organisms into this branching order (rather than, for instance, overlapping nonhierarchical sets or something analogous to Mendeleev’s periodic table) assumes that this branching order is the ‘true’ pattern of nature. This view can only be justified by recourse to evolutionary theory: nature will show this pattern if the history of life consists of lineages of organisms which have transformed and branched (e.g. Ball, 1983; Kemp, 1985; Ridley, 1986; Sober, 1989; de Queiroz & Donoghue, 1990; Scott-Ram, 1990; Bryant, 1991; Panchen, 1992; Kluge, 1997). Even if one accepts this evolutionary justification, one must acknowledge the possibility of reticulate (i.e. nondivergent) evolution, such as hybrid species of plants. Divergence is the major pattern, but exceptions exist and might not be uncommon ( Rieseberg, 1997).
Brady (1994) has recently attempted to counter this criticism by arguing that cladistic analysis does not necessarily assume a branching order at the outset. Rather, all possible patterns of relationship are considered, and ‘the choice between alternative classificatory schemes must … be settled on descriptive grounds’ (p.19). In other words, one accepts a nested hierarchy as the natural order because one’s observations (congruent synapomorphies) fit this scheme better than they fit other schemes (e.g. a ‘periodic table’ of organisms).
This argument is questionable. Why should one choose to observe and use genetically determined traits to classify organisms, in preference to other attributes such as ecological, geographical or stratigraphic relationships? Ecological, geographical, stratigraphic and even alphabetical classifications can be constructed, and observations of the relevant attributes of organisms will also ‘fit’ such classifications tightly (but will probably not fit a nested hierarchy at all). For instance, one could argue that an alphabetical system is the ‘natural’ system since the scientific names of all species always begin with one of the 26 letters of the English alphabet! Thus, emphasis on a nested hierarchy based on genetically determined traits, in preference to other (equally tidy) schemes based on other criteria, must rest on stronger grounds than the assertion that ‘it seems to fit the data well’. Since any number of ‘well-fitting’ classifications can be constructed based on different criteria, preference for one pattern over all others needs explicit justification. The hierarchical system is preferred as the ‘natural’ system because it reflects real historical relationships produced by evolution, whereas the alphabetical system is not natural because it reflects the arbitrary decision of a small group of Westernized humans that scientific names must begin with one of only 26 symbols.
2 ‘The pattern can be estimated by sampling characters and finding replicated, internested sets of synapomorphies’. The view that the natural hierarchy can be discerned via parsimony analysis, rather than for example phenetic clustering or maximum likelihood, almost certainly makes further assumptions about the rate and pattern of evolutionary change, although the extent and nature of these assumptions remains debated (e.g. Felsenstein, 1978; Farris, 1983; Friday, 1987; Sober, 1988, 1989; Goldman, 1990; Penny et al., 1994 ; Yang, 1996). If these assumptions are violated, then other methods might perform just as well, or even better, at retrieving the hierarchy. For instance Patterson (1988), in a paper written when molecular clocks had not yet been widely questioned, argued that phenetic clustering of molecular data, just like cladistic analysis of morphological data, would reveal the true hierarchy of life. The justification for the validity of the phenetic analysis of molecular data was explicitly evolutionary: molecular changes were assumed to be random and neutral, resulting in constant rates of divergence in all lineages. Paradoxically, however, it was simultaneously asserted that no evolutionary justification was necessary for cladistic analysis of morphological data. Rather, it was argued that cladistic analysis was logically prior to any evolutionary models.
The first two of Platnick (1979) therefore only appear to be justifiable through appeals to evolutionary theory. However, while it is now generally acknowledged that cladistic analysis cannot be totally divorced from evolutionary theory, there is a widespread view that such analyses will be more accurate and objective if evolutionary assumptions are at least minimized. Thus, the remaining axiom of the pattern cladist approach might be more reasonable – it at least has widespread current acceptance. This axiom is:
3 ‘Knowledge of evolutionary history … is derived from the hierarchic pattern’ retrieved by cladistic analysis and thus should not contribute to analysis of that pattern ( Platnick, 1979, p. 538).
Accordingly, most recent cladistic analyses of morphological data, and a large (but decreasing) proportion of molecular studies, employ equal weighting of all characters with no constraints on the possible directions of character state change (e.g. see Philippe et al., 1996 ). Apart from the minimal assumptions that anagenesis and cladogenesis have occurred to generate a hierarchical branching pattern, and the (debated) assumptions justifying parsimony analysis, no other evolutionary assumptions are explicitly incorporated. More specific evolutionary mechanisms, such as directional selection, adaptation, or speciation modes, are not considered in character analysis or cladogram construction.
The rationale for this approach is that observation of patterns such as character distribution is logically prior to inference about processes such as character evolution ( Platnick, 1979; Patterson, 1980, 1988; Nelson & Platnick, 1981; Rosen, 1984; Brady, 1985, 1994; Schoch, 1986; Rieppel, 1988). Character distribution can be observed directly, and used to reconstruct phylogenies. Conversely, long-term evolutionary phenomena (e.g. the reversibility of characters) cannot be observed directly; rather, their nature can only be inferred based on character distributions and phylogenetic patterns. Accordingly, cladistic analyses should minimize such assumptions about the evolutionary process as far as possible, although they cannot be eliminated altogether. To incorporate additional evolutionary assumptions into cladistic analysis is unnecessary. It is also dangerously circular since these assumptions have no independent justification but are themselves derived largely from phylogenetic investigations. This view is here termed ‘strict cladistics’, since it does not admit extraneous assumptions that might cause one to modify a straight parsimony analysis. Based on the number of published studies which employ simple unweighted parsimony analysis with no other constraints, this approach appears to be currently dominant, at least among morphological systematists.
The opposite view asserts that inferences of evolutionary process should be an integral part of phylogenetic reconstruction (e.g. Hecht & Edwards, 1976; Bock, 1977, 1981; Arnold, 1981; Fisher, 1981; Kemp, 1985, 1988; Szalay & Bock, 1991; Penny et al., 1994 ; Naylor & Brown, 1998). Knowledge of how characters evolve and interact will improve the accuracy of phylogenetic reconstruction. For instance, highly labile and/or correlated characters can be downweighted, and characters more likely to change in one direction rather than the reverse (e.g. because of constraints or directional selection) can be treated accordingly. If these initial assumptions are reasonably accurate – and this is a big ‘if’– then an analysis which includes such refinements will yield more accurate trees than an analysis which naively assumes all characters are independent, fully reversible and equally informative. However, most modern cladistic analyses have been reluctant to incorporate such complex evolutionary principles (e.g. see Lee & Doughty, 1997), and this viewpoint appears to be held by relatively few working systematists. The approach which advocates incorporating evolutionary principles directly into phylogenetic reconstruction is here termed ‘evolutionary cladistics’
Most systematists accept that cladograms should be interpreted as theories of evolutionary relationships, and as such are summaries of patterns of common ancestry among taxa ( Hennig, 1966; Bock, 1977, 1981; Gaffney, 1979; Wiley, 1981; Hill & Crane, 1982; Ball, 1983; Kemp, 1985, 1988; Hill & Camus, 1986; Ridley, 1986; de Queiroz & Donoghue, 1988; de Queiroz, 1988; Hull, 1989; Arnold, 1990; Scott-Ram, 1990; Panchen, 1992; Fisher, 1995; Kluge, 1997). If this view is accepted, cladistic analysis should strive to generate cladograms which match as closely as possible the actual historical patterns of evolutionary descent.
If there was some way of ascertaining, without any prior knowledge of phylogeny, exactly how certain characters evolved (e.g. the relative transition probabilities between states; the correlation between different characters), there can be little argument that incorporation of such information into cladogram construction would certainly improve their accuracy (i.e. how closely they reflect actual evolutionary relationships). However, the character transformations of cladistic analyses usually occur over time scales too great to be observed by humans, and there is debate over whether such knowledge is obtainable. The strict cladist view asserts that such inferences about long-term evolutionary processes (‘evolutionary principles’) can only be derived after phylogenetic patterns are established. Their inclusion in the initial phylogenetic analysis would be viciously circular, and cladistic analysis should avoid these unjustifiable assumptions as much as possible. However, the evolutionary cladist view argues that many long-term evolutionary principles can indeed be derived without any recourse to a pre-existing phylogenetic framework. If so, their inclusion in the initial phylogenetic analysis would be justifiable. Crudely stated, if one could ‘know’ beforehand how certain characters evolve, refusing to incorporate this into a phylogenetic analysis amounts to discarding important information.
Here, I argue that both viewpoints are partially correct. There are some evolutionary principles which are derived almost entirely from the topology of phylogenetic trees and distribution of characters. Such principles have no independent standing and should not be incorporated into cladistic analysis. However, other principles can be derived without any knowledge of phylogeny or character distribution. Such principles clearly have independent standing, and there is no methodological problem with incorporating them into cladistic analyses. There are, however, operational problems which might prevent one from integrating some of the principles into phylogenetic analysis. As before ( Lee & Doughty, 1997), the focus is on morphological studies since that is my area of greatest familiarity. However, the arguments developed below apply equally to other phylogenetically informative systems (molecules, behaviour, ecology) as well.
Phylogenetic assumptions and evolutionary principles
- Top of page
- Phylogenetic assumptions and evolutionary principles
- Problems with incorporating phylogeny-neutral principles into systematics
A ‘phylogeny-neutral’ evolutionary principle is one derived independently of any particular hypothesis of phylogeny. For instance, several characters might be correlated because they are under common genetic control (pleiotropy), or because of developmental constraints, such as a suite of linked changes associated with paedomorphosis (e.g. Gould, 1977; MacNamara, 1997). In taxa known to exhibit these genetic and/or developmental systems, such characters would be expected to change simultaneously, and thus to not provide independent evidence of relationships. A strong functional argument might also be used to infer such character correlation. Limb reduction results in compensatory body elongation in lizards ( Lande, 1978; Greer et al., 1998 ): if the animal is to remain mobile, the decreasing efficiency of limb-driven locomotion must be compensated via an increasing reliance on axial undulation. Similarly, it might be argued that once a complex character is lost, the elaborate genetic and developmental machinery to make it soon decays, and the complex chain of mutation and selection required to rebuild these systems is unlikely to reoccur ( Dollo, 1893; Muller, 1939). Thus, for complex characters such as the pentadacyl limb of tetrapods ( Lande, 1978), wings of insects ( Goloboff, 1997) or colour dichromatism of certain birds ( Omland, 1997), convergent loss could be argued to be more likely than convergent gain (see de Pinna, 1991).
Selectionist arguments have been proposed to explain why some evolutionary changes, such as the evolution of either all-female parthenogenesis, polyploidy or haplo-diploidy from typical diploid sexual organisms, should be irreversible (e.g. Bull & Charnov, 1985). Functional (‘adaptive’) analysis has also been used to infer polarities between character states ( Bock, 1981; Fisher, 1981; Gallis, 1996), e.g. a transition from a well-developed state to an apparently functionless vestigial state is more likely than the reverse ( Ridley, 1986; Omland, 1997). Experimental (manipulative) studies of phenotypic plasticity might reveal that expression of alternative states of a character is mostly environmentally rather than genetically determined; this would be a strong reason for downweighting such characters ( Schoch, 1986). In such cases, the entire reaction norm might be a better character. In molecular studies, there are functional arguments (involving stabilizing selection) why transitions are more likely to occur than transversions, and why third codon sites should change more rapidly ( Brown et al., 1982 ; Kimura, 1983), and empirical evidence for such patterns largely independent of any particular phylogenetic hypotheses (e.g. Yang et al., 1995 ).
Patterson (1994) validly points out that inferring greater frequencies of transitions based on the observed differences between sequences assumes that evolution has occurred and organisms are somehow related. The same might be said of other principles discussed above, such as the greater probability of convergent loss rather than gain of complex characters. However, the crucial point is that derivation of these principles does not invoke a particular phylogeny, although it makes the assumption that some phylogeny exists. Phylogenetic analyses similarly do not attempt to ascertain whether or not a phylogeny exists; rather, they are concerned with determining which of the possible phylogenies is likely to be correct. Thus, the above principles can be considered to be independent of phylogenetic (cladistic) analysis in the sense that their derivation is not predicated upon any particular result of such analyses.
Some of these principles are undoubtedly more robust than others: the argument for treating characters as correlated because of pleiotropy appears to be strong, but the assumption that complex characters are more likely to be convergently lost than gained appears rather weak without any explicit genetic or developmental evidence. More complex phylogeny-neutral evolutionary principles are also relevant to the problem of constructing, or at least evaluating, cladograms ( Panchen & Smithson, 1987): for instance, certain traits cannot co-occur in any implied ancestor because of functional incompatibilities leading to maladaptation (e.g. small body size, nakedness and endothermy in a terrestrial vertebrate: Kemp, 1988; Lee & Doughty, 1997).
Incorporation of phylogeny-neutral principles into cladistic analysis is not problematic. As discussed above, derivation of these principles does not invoke any particular pattern of hierarchical relationship, i.e. they do not rely for support from any particular outcome of a cladistic analysis. Thus, one is not incorporating the results of a cladistic analysis back into phylogenetic reconstruction as an initial (and thus circular) assumption. For instance, if one knows that two characters are developmentally correlated in all the organisms of interest, there can be little objection to incorporating this information in a cladistic analysis by either deleting one of the redundant characters or downweighting both characters. There is no circularity involved and the flow of information is clearly unidirectional. Developmental genetics yield insight about correlation of characters, this information is incorporated into a phylogenetic analysis via character weighting, and influences the topology of the resultant tree:
Thus, in such instances, a priori weighting of characters can be justified on an evolutionary basis and supported by (often strong and rigorous) external evidence. Recent workers have either rejected this approach (e.g. Smith, 1994; de Salle & Brower, 1997), or have tended to advocate it for different reasons. In particular, hypotheses of homology for characters with low ‘information content’ (e.g. simple characters, losses) are more likely to be wrong, and such characters should be downweighted ( Neff, 1986; Bryant, 1989). Both criteria might be valid: characters might be less informative for intrinsic evolutionary reasons (e.g. character correlation), and also for extrinsic, observational reasons (mistaken homologies). Both factors are valid reasons to downweight such characters. This realization blunts arguments that a priori weighting is either circular because it requires some preconceived notion of phylogeny (e.g. Patterson, 1980; Schoch, 1986; Smith, 1994) or arbitrary in that the criteria employed are ad hoc assumptions with little empirical support (e.g. Farris, 1983; de Salle & Brower, 1997).
Indeed, some of these arguments against a priori weighting are internally inconsistent. de Salle & Brower (1997) suggest that extrinsic evolutionary criteria (e.g. inferred rates of change or homoplasy, or variability within terminal taxa) can be used for ‘a priori selection of characters’, where ‘characters we suspect of not providing enough phylogenetic information to be worth the effort … are avoided at the outset’ (p. 758). This is obviously a very extreme form of a priori weighting where certain characters are weighted zero. However, they simultaneously suggest that differential a priori weighting is undesirable, as it ‘requires ad hoc assumptions about the nature of the empirical evidence’ (p. 751), and that all characters admitted into the analysis should be weighted equally. This approach seems paradoxical: a priori weighting based on external evidence is accepted as valid in its most extreme form (i.e. rejection of certain characters), but not in a milder form (i.e. differential weighting of remaining characters). Either a priori weighting is invalid (in which case all characters must be used) or it is valid (in which case it must apply to all characters).
A ‘phylogeny-dependent’ evolutionary principle is one derived largely or entirely on the basis of knowledge of tree topology and character distribution. One might observe that two characters invariably change together on a phylogeny and that this correlation is statistically significant. One might conclude that these characters are somehow linked, despite little knowledge of the underlying causal factors for this correlation (genetic, developmental or functional). Many aspects of the comparative method (e.g. Harvey & Pagel, 1991) are designed to find such correlations. Similarly, one might observe that a particular character only evolves once on a cladogram but is repeatedly lost. This would imply that such a character is more likely to be convergently lost than gained, and that the former transformations are less phylogenetically informative than the latter ( Goloboff, 1997; Lee & Shine, 1998).
Incorporation of phylogeny-dependent principles into cladistic analysis is potentially problematic, since such principles (by definition) are themselves products (or artefacts) of such analyses. For instance, some have recommended placing heavy emphasis on key ‘homologous’ characters, assumed a priori to be unique and unreversed, in constructing phylogenies (e.g. Hecht & Edwards, 1976). Less dramatically, Mayr et al. (1953 ) stated that ‘reliable’ characters should be used to construct phylogeny, with recognition of such traits stemming partly from the ‘art’ (p. 107) or ‘intuition’ (p. 106) of the taxonomist (they did, however, also recommend that reliability be tested more rigorously by statistical methods). These are examples of the circularity which results when phylogeny-dependent principles are employed as a priori assumptions ( Hull, 1967; Szalay, 1981; Neff, 1986; Schoch, 1986; Scott-Ram, 1990). Reliable traits are defined in terms of a preconceived phylogeny in which they are highly congruent; weighting these traits heavily in any subsequent phylogenetic analysis will tend to result in the ‘discovery’ of very similar phylogenies.
A less obvious example is as follows. Based on an initial, tentative phylogeny of reptiles we might conclude that viviparity is unlikely to reverse since it originates approximately 50 times and is only lost five times, and this bias in direction of change is statistically significant (e.g. see Lee & Shine, 1998). Obviously, if we now code this character such that origins are more likely (i.e. less costly in terms of tree length) than losses, and re-run the analysis, circularity is inevitable. Unlike the previous example, the evolutionary principles are not derived from an external source: there are no explicit functional or physiological models justifying this assumption ( Lee & Shine, 1998). Rather, a particular phylogeny is being used to justify particular evolutionary principles, and these principles are then being fed back into phylogenetic analysis.
This, of course, is a classic case of a circular argument (see Hull, 1967; Sober, 1989; de Queiroz, 1996), where certain assumptions are used to form a conclusion, and the conclusion in turn is used to justify those same assumptions. Here, an assumed phylogeny is being used to make certain conclusions about evolutionary processes, and these principles are then used to construct the same (or very similar) phylogeny.
This danger of circularity (or at least unwarranted extrapolation) persists even if principles derived from a phylogenetic analysis of one group are employed in an analysis of a completely different group. As before, based on an initial, tentative phylogeny of reptiles we might conclude that viviparity is unlikely to reverse since it originates approximately 50 times and is only lost five times. In an analysis of elasmobrachs (sharks and rays), we might then code this character such that origins are five times more likely (i.e. less costly in terms of tree length) than losses. However, even if there are strong (but as yet unknown) reasons why viviparity is unlikely to reverse in reptiles, there is no justification for assuming that these constraints exist in elasmobranchs. Whatever factor discourages loss of viviparity in reptiles might not be present in elasmobranchs. Panchen (1992) has noted that most biological ‘laws’ are merely ‘taxonomic statements’ that only apply to particular portions of phylogeny. An irreversible or reliable character in one group might be fully reversible or unreliable in another ( Kluge, 1997). Incorporation of an evolutionary principle characterizing reptiles into a phylogenetic analysis of elasmobranchs is therefore an unwarranted extrapolation. It would also be partly circular. Again, an assumed phylogeny would be used to infer evolutionary principles, and these are then used to construct phylogenies of other groups. These other phylogenies will of course tend to be consistent with both the original phylogeny and its evolutionary implications and thus offer (circular) support for them. Coding viviparity as unlikely to reverse in all these analyses would only be justified if there was some phylogeny-neutral evidence that it was irreversible, e.g. a physiological constraint present in all relevant taxa. However, although viviparity is often assumed to be irreversible, compelling functional arguments have yet to be articulated ( Lee & Shine, 1998).
It is here proposed therefore that certain evolutionary principles can validly be incorporated into phylogenetic analysis without circularity, but that other principles should be purged.
Many previous arguments have failed to make the distinction between phylogeny-neutral and phylogeny-dependent principles and have thus erroneously assumed they are all one type, i.e. either phylogeny-neutral or phylogeny-dependent. The strict cladist view accepts the pattern cladist axiom that documentation of systematic pattern is logically prior to inferences about evolutionary process, e.g. ‘pattern analysis is … primary and independent of theories of process, and is a necessary prerequisite to any analysis of process’ ( Nelson & Platnick, 1981 p.35), ‘any hypothesis of transformation remains empirically empty if not constrained by a rigorous phylogenetic framework’ ( Rieppel & Grande, 1994, p. 245), and ‘sound [evolutionary] models demand either some knowledge of phylogeny, or assumptions about it’ ( Patterson, 1994, p.173). This view incorrectly assumes that all principles are phylogeny-dependent. The suggestion that evolutionary processes acting over long time scales cannot be observed directly, but must be inferred from observed patterns, might be correct. However, phylogenetic patterns are not the only empirically observable patterns that can be used to infer such processes. Functional patterns, or patterns of genetic and epigenetic interactions within individual organisms, can also be used to infer the probable evolution of characters over large time scales. Of course, such inferences about character evolution might prove incorrect or at least dubious – a point discussed below – but so might inferences drawn from phylogenetic patterns if the phylogenies are incorrect. To assert that only the latter method can yield valid insights into the nature of long-term evolution appears to be an overly narrow epistemological view.
The strict cladist view thus fails to acknowledge that certain inferences about long-term character evolution can indeed be made without any recourse to phylogeny – such as the correlated evolution of two characters that have been observed to be controlled by the same gene in all taxa under investigation. Scott-Ram (1990 p.177), in his extensive treatment of the various manifestations of cladistics, made the same observation:
‘The transformed [pattern] cladists are wrong if they assert that this [cladistic analysis] is the only way of constructing processes, because results from other domains of inquiry can yield theories of process’.
However, he did not proceed to discuss the implications of this far-reaching observation for the integration of evolutionary theory and systematics. Here, it is argued that incorporation of such phylogeny-neutral evolutionary principles into systematics is not circular. It is surprising that pattern cladists, who have repeatedly argued for the importance of ontogeny for understanding the polarity of characters (e.g. Nelson, 1978; Patterson, 1996), have also often denied that such information can be used in other aspects of character analysis, such as determination of character correlation, multistate transformation series, or weighting. If ontogeny can yield insights into the relationships between character states between ingroup and outgroup taxa, surely it can be used to understand the relationships between character states restricted to ingroup taxa only. In contrast, Rieppel & Grande (1994, p. 247) state that developmental constraints (e.g. correlated evolution of a suite of paedomorphic characters) can only be inferred on the basis of observing character distributions on a phylogeny. Presumably, they mean that the repeated co-occurrence of the same traits in different taxa provides evidence of developmental constraints. However, this statement is dubious for two reasons. Firstly, the repeated co-occurrence of these traits on a phylogeny might reflect functional correlation, or pleiotropy, rather than developmental integration. Secondly, developmental constraints can be inferred without any recourse to phylogeny, by observing and experimentally manipulating the ontogeny of individual organisms and noting the patterns of character correlation within their ontogenies. If these results are conclusive, a priori downweighting of such correlated (and thus nonindependent) characters is justified.
Pattern cladists have themselves acknowledged that evolutionary theory might be incorporated into phylogenetic reconstruction: Nelson & Platnick (1981, p.164) state that ‘the processes and mechanisms of genetics, heredity, and evolutionary change …. might alter, or otherwise impact upon, the empirical basis of synapomorphy resolution’. However, they immediately dismiss this: ‘yet this seems to have happened only to a limited extent, if at all. It is not yet generally possible, for example, to resolve synapomorphies by study of the genetic processes that produce them’. If the reasoning in this paper is correct, this dismissal was premature. Study of such genetic processes can reveal (for instance) patterns of character correlation, and this, if incorporated into a cladistic analysis via character weighting, can indeed contribute substantially to ‘the empirical basis of synapomorphy resolution’ ( Shaffer, 1986).
Of course, many principles fall in the grey area, and are inferred partly from phylogeny-dependent, and partly from phylogeny-neutral sources. For instance, wings might be expected to be more easily lost than regained because (1) they are complex integrated structures and (2) various phylogenies (e.g. of insects and birds) imply a single origin and numerous reversals ( Goloboff, 1997). Uncritical incorporation of these ‘semineutral’ principles into a phylogenetic analysis is problematic, since the practice partly adds independent empirical information, and is also partly circular. Rather, the decision to incorporate, or not incorporate, should rest on the strength of only the valid (phylogeny-neutral) evidence. The invalid (phylogeny-dependent) evidence should be ignored during this part of the analysis. Valid evidence is neither strengthened nor weakened by additional invalid evidence.
Problems with incorporating phylogeny-neutral principles into systematics
- Top of page
- Phylogenetic assumptions and evolutionary principles
- Problems with incorporating phylogeny-neutral principles into systematics
There is thus no inherent circularity or any other methodological problem with incorporating phylogeny-neutral principles into systematics (e.g. Hecht & Edwards, 1976; Bock, 1977, 1981; Arnold, 1981; Fisher, 1981; Kemp, 1985, 1988; Szalay & Bock, 1991; Penny et al., 1994 ; Naylor & Brown, 1998). In many cases, there is also no operational reason: these principles can be readily modelled using standard options in modern parsimony programs such as PAUP* ( Swofford, 1999). For instance, genetic, developmental or functional correlation of traits can be incorporated in cladistic analysis by downweighting all affected characters, or deletion of certain characters. If some transitions between character states are more likely than other transformations, this can be modelled using the step matrices function. The amount of weighting can be approximately determined by the extent of genetic correlation or the difference in transition probabilities (e.g. Shaffer, 1986). Such procedures are already being employed in parsimony analyses of molecular data which classify substitutions into various categories (transitions or tranversions; A–T, A–C, A–G, T–C, T–G or C–G changes; first, second or third codon positions) and weight each category according to their presumed frequencies (e.g. Williams & Fitch, 1990; Hillis et al., 1996 ). However, it will usually be much more difficult to estimate transition probabilities for morphological characters. Such characters cannot be readily classified into neat categories, their underlying genetic bases are usually poorly known, and the number and boundaries of character states for a single morphological character are difficult to define (e.g. Stevens, 1991).
Such principles can be incorporated into a cladistic analysis as a priori assumptions. Obviously, as in any scientific endeavour, if these initial, lower-level assumptions are wrong, the results are also likely to be wrong. Thus, if dubious phylogeny-neutral principles are incorporated into a cladistic analysis, and these principles are false, then the resultant cladogram is likely to be wrong. Note that this error is still not circularity, since the (erroneous) principles were not derived from any cladogram in the first place. Rather, the error is accepting a false premise and building an argument upon it. Thus a dilemma arises – if an evolutionary principle incorporated into a cladistic analysis is correct, it should improve our phylogenetic inferences, but, if incorrect, it might lead to a very wrong tree.
It might superficially appear that the most ‘objective’ or ‘conservative’ option would be expel all such principles from phylogenetic analysis as unwarranted assumptions. However, this in itself is an assumption: the principles are assumed to be incorrect. If there is strong evidence that some characters are less informative (i.e. more likely to change) than others, choosing to ignore this information is hardly objective or conservative ( Swofford & Maddison, 1992; Goloboff, 1997; Omland, 1997). For example, if genetic, developmental and functional studies all suggest certain characters are linked, downweighting these characters invokes the a priori assumption that these characters are indeed correlated, while equal weighting invokes the a priori assumption that these characters are independent. Clearly, the latter option cannot be argued to represent the more conservative option, or to minimize assumptions. In particular, failure to incorporate this character correlation does not produce a more conservative phylogeny (e.g. one where certain clades – which only appear when correlation is considered – are left unresolved). Rather, it will probably give us an equally resolved phylogeny constructed on the assumption that all characters are totally independent and equally informative. There is no simple answer to this dilemma, although a possible approach is as follows.
There should be no objection to incorporating strongly supported principles into an analysis. For example, Omland (1997) listed independent arguments based on several different lines of evidence as to why dichromatism in birds is more likely to be lost than gained: the developmental genetics of sex-linked traits, female preferences, sporadic presence of apparently functionless ‘vestigial’ dichromatism in otherwise monochromatic taxa, and the improbability of losing and re-evolving complex characters. As such principles have strong independent support, they are likely to be correct and their inclusion should improve the chances of inferring the correct tree. This then leaves a subjective and difficult decision regarding many other principles: how strongly supported does a phylogeny-neutral principle have to be before we can include it as a base level (a priori) assumption. There are undoubtedly many principles which might or might not be sufficiently strong to employ as a base-level assumption, depending on the opinion of individual workers.
For this grey area, the most conservative approach might be to perform separate analyses, excluding and incorporating the contentious principle, to determine if it affects the topology of the resultant tree. As an example, if one has only equivocal evidence that several characters are functionally correlated, it might be prudent to run the analysis with all characters equally weighted, and then again with the putatively correlated characters downweighted. If the phylogenies produced by the two studies are topologically identical, this would increase confidence in the results.
However, if incorporation of the evolutionary principles changes the tree, then one has to make the subjective decision whether the principle is strong enough that their effects can be accepted. Clearly, incorporation of phylogeny-neutral principles has changed tree topology. Whether the principles are robust enough to cause one to accept these new results over an unweighted parsimony analysis can only be decided on a case-by-case basis and necessarily involves a fair bit of subjectivity. For instance, an unweighted parsimony analysis could yield only marginal support for clade AB, with 10 characters supporting the clade and eight supporting a conflicting clade AC. Ontogenetic studies might show that characters 1–5 supporting AB are highly correlated with each other, and that characters 6–10 are also highly correlated with each other. An analysis that incorporates this information would treat these 10 characters as two independent characters, and instead yield a well-supported clade AC. If the evidence for character correlation is strongly supported from several independent areas such as genetics, ontogeny and functional morphology, it might be advisable to accept the results of the latter analysis.
Another analysis might yield strong support for clade XY in an unweighted parsimony analysis. However, a functional argument might suggest some of the supporting characters are correlated, and another analysis with differential character weighting might yield weak support for clade XZ. We then realize that the functional arguments for character correlation (and thus downweighting) are rather weak: there is no evidence for genetic or developmental correlation, and these characters are observed to evolve independently in other clades. In this situation, it might be prudent to accept – tentatively – the results of the unweighted parsimony analysis.
This approach is admittedly subjective and pluralist. However, it is potentially more enlightening than simply performing an unweighted parsimony analysis only and accepting its results uncritically. It allows exploration of the topological effects of various evolutionary principles ( Wheeler, 1994): whether these assumptions are strong enough to warrant acceptance of any new topologies can then be decided. Again, it should be stressed that unweighted parsimony analysis also makes many assumptions: it assumes other principles exist: that all characters are independent, and that forward and reverse changes are not merely possible but equally likely.
While all phylogeny-neutral principles can theoretically be included in systematic studies, in practice some are difficult to integrate into current phylogenetic analyses. An example, more fully discussed in Lee & Doughty (1997), is as follows. Kemp (1988), in his critique of Gardiner’s (1982, 1993) ‘Haematothermia’ hypothesis – where mammals and birds are sister groups – argued that the cladogram implied a functionally incongruous ancestor and was thus dubious. Optimization of characters on the cladogram implies that the common ancestor of mammals and birds must have been truly endothermic, and also small, uninsulated (naked) and with a low respiratory rate – a functionally incongruous combination of characters. According to Kemp, any cladogram that implies such animals is unacceptable. In theory, it should be possible to incorporate this evolutionary principle into a phylogenetic analysis, so that the algorithm used to generate phylogenies has a qualifier that ensures such incongruous character combinations do not occur in any part of the tree. For example, a parsimony analysis might be refined so that, if the ‘prohibited’ character combination appears on any node, the tree is rejected and successively less parsimonious trees (or optimizations) are investigated until the ‘forbidden’ combination disappears. In practice, however, it will be very difficult to incorporate such principles into phylogenetic analyses, especially if they are quite numerous.
Thus, the objection to incorporating such principles into phylogenetic reconstruction is not methodological, but operational. In theory they can be integrated into phylogenetic analyses, but in practice this will often make the analyses too complicated.
For such principles, the ‘consilience’ approach advocated by Lee & Doughty (1997; see also Hill & Camus, 1986; Kluge, 1997; Wilson, 1998) might be the most sensible. In such cases only, these spheres of investigation might have to be kept separate. Rather than incorporate these evolutionary principles directly into the initial phylogenetic analysis – which is too difficult – it might be best to temporarily omit them from the initial analysis, but to subsequently check that the resulting phylogeny is consistent with these principles. Thus, in the haematotherm example above, it might prove too difficult to incorporate into a parsimony analysis the principle that small, uninsulated endotherms with low respiratory rates are biologically implausible and must not be implied at any nodes. The best compromise might be to perform a standard parsimony analysis and then check whether the resultant cladogram implies any such organisms. If such organisms are not implied at any of the reconstructed nodes, the cladogram can be provisionally accepted. However, if such functionally incongruous organisms are implied, one should carefully evaluate the relative support for the cladogram, and the principle. If the cladogram topology is extremely well supported and the principle weak, the cladogram can be provisionally accepted and the principle re-evaluated. This would be equivalent to excluding a poorly supported principle from the set of a priori assumptions fed into a phylogenetic analysis. If, in contrast, the cladogram topology is poorly supported but the principle very strong, one should examine successively less parsimonious optimizations or topologies and accept the shortest tree which does not violate the principle. This would be equivalent to including a well-supported principle in the set of a priori assumptions.
While this paper has been mainly concerned with morphological parsimony analyses, these arguments also have implications for maximum-likelihood analysis of molecular sequences. Such analyses incorporate explicit evolutionary models, and this has been argued to be both a strength and a weakness. Again, it is the independence of these evolutionary models which is critical. Workers who argue that incorporation of these assumptions is desirable often stress that they are ‘robust’, i.e. they can be independently justified and are largely phylogeny-neutral (e.g. Bishop & Friday, 1985; Hillis et al., 1994 ). If so, their incorporation into phylogenetic analysis is valid and perhaps desirable. However, those opposed to this methodology argue that these assumptions either have poor empirical support or are phylogeny-dependent (e.g. Farris, 1983; Siddall & Kluge, 1997). If so, their incorporation into phylogenetic analysis is problematic. Therefore, a key question concerning the validity of likelihood methods is: ‘are the models used in constructing and evaluating tree topologies phylogeny-neutral, or dependent?’. The answer appears to lie between the two extremes. These models involve some parameters which can be estimated without precise assumptions regarding the correct phylogeny: for instance, in many real data sets all possible phylogenetic arrangements imply a transition bias, as do multiple pairwise comparisons which do not assume any particular phylogeny ( Yang et al., 1995 ). Other required information, such as branch lengths, however, is difficult to determine without some knowledge of phylogeny ( Yang et al., 1995 ). Thus, some assumptions appear to be relatively phylogeny-neutral (e.g. transition–transversion ratios), while others are phylogeny-dependent (e.g. branch lengths). Major methodological advances in likelihood models will be possible if ways can be found to reduce the importance of the latter parameters in influencing tree topology, or to estimate them without using phylogenetic information.