Cladistics. Phylogenetic analysis and alignment of behavioral sequences by direct optimization

Transcription

1 Cladistics Cladistics 22 (2006) /j x Phylogenetic analysis and alignment of behavioral sequences by direct optimization Tony Robillard 1,2 *, Fre déric Legendre 1, Laure Desutter-Grandcolas 1 and Philippe Grandcolas 1 1 Muse um national d Histoire naturelle, De partement Syste matique et Evolution, UMR 5202 CNRS, Case postale 50 (Entomologie), Paris cedex 05, France; 2 University of Missouri, Division of Biological Sciences, 215 Tucker Hall, Columbia, MO 65211, USA Accepted 28 June 2006 Abstract We propose a new approach to consider behavioral data in phylogenetic analyses. We show that behavior can be described as sequences of repeated units, and that these behavioral sequences can be analyzed under direct optimization in a way similar to molecular data. This approach provides repeatable hypotheses of homology for behavior when traditional criteria result in multiple alternatives or do not allow to propose any. We exemplify this approach by analyzing the calling songs of the North American Gryllus species under direct optimization. We first use two alternative coding schemes to describe the temporal patterns of the songs as sequences of repeated simple behaviors. We submit these behavioral data to phylogenetic analysis under direct optimization, first as separate analyses, and second in combination with molecular data and additional acoustic characters. The results show that the coding option that consists of discretizing the silent parts of the songs: (1) allows description of the songs in a more precise way; (2) discriminates further the songs between species; and (3) enhances the phylogenetic content of the behavioral sequences. Our study demonstrates that behavioral sequences can be transformed so that they can be used in genuine phylogenetic analysis, in isolation or combined with other data sets. We discuss how this approach may provide phylogenetic signal where none or little is usually available, and the applications to the study of the evolution of behavioral evolution. Ó The Willi Hennig Society There is a long and profitable tradition to consider that behavior is positively informative for the study of the evolution and the systematics of organisms (Lorenz, 1941; Roe and Simpson, 1958). In this traditional framework, careful comparative ethological studies were carried out to observe some stereotyped behaviors and to document behavioral homologies (e.g., Atz, 1970; Hodos, 1976; Lauder, 1986; Wenzel, 1992). In more recent times, this framework was sometimes considered with doubts in the context of phylogenetic analysis where supposedly adaptive and homoplastic traits are regarded with suspicion (de Queiroz, 1996; Marquès and Gnaspini, 2001). *Corresponding author: address: robillar@mnhn.fr Behavioral characters can be defined in the same way as any other characters, either morphological or molecular (Wenzel, 1992), following the classical rules of observation, definition and analysis (Cracraft, 1981; Greene, 1994; Noll, 2002; Desutter-Grandcolas et al., 2003; Desutter-Grandcolas and Robillard, 2003) and, empirically, most studies agree that the informativeness of behavioral characters is no less than for other characters (Brooks and McLennan, 1991, 2002; Coddington, 1986a,b; McLennan et al., 1988; de Queiroz and Wimberger, 1993; Proctor, 1996; Scharff and Coddington, 1997; Stuart and Currie, 2001; Stuart et al., 2002; McLennan and Greene, 2006; Robillard et al., 2006). As for all other characters, one of the most recurrent problems is the confusion between behavioral classes and behavioral characters (Wenzel, 1992; Proctor, 1996). Many studies consider behavior a posteriori to Ó The Willi Hennig Society 2006

2 T. Robillard et al. / Cladistics 22 (2006) a phylogenetic analysis and broad behavioral classes are optimized on the already reconstructed tree (e.g., Alexander, 1962; Huang et al., 2000). This method may be conceived as exploratory when behavior is studied not yet thoroughly enough but it can also be misleading when broad behavioral classes have no real characterization, thus leading to biased reconstructions (Desutter-Grandcolas and Robillard, 2003; Grandcolas and D Haese, 2004). Therefore, detailed behavioral hypotheses of homologies are needed prior to the reconstruction of the phylogenetic tree, which are submitted subsequently to the test of congruence during the tree reconstruction. In this respect, most of our tools deal with the longstanding criteria of homology mostly position, special quality and connection through intermediate forms from Remane (1952), as reviewed by Wenzel (1992). These criteria can be applied to behavior as long as it is properly described (Price and Lanyon, 2002; Stuart and Currie, 2002; Branham and Wenzel, 2003; Desutter-Grandcolas and Robillard, 2003), not by considering broad classes, but behaviors clearly observed and described as a succession of acts within sequences emitted in definite and particular contexts. If the behavior is standardized enough, the succession of acts in the sequence is stereotyped within a taxon and the criterion of position can be fully applied for comparison among different individuals and taxa (Baerends, 1958; Tinbergen, 1963; Wenzel, 1992). In this respect, some algorithms of sequence alignment have been applied tentatively for non-phylogenetic classification and comparisons of sequences (Abbott, 1995; Wilson et al., 1999; Hay et al., 2003; van der Aalst et al., 2003). However, as for characters in molecular DNA sequences (Wheeler, 1996), a problem very often arises where sequences of very different lengths evolved with many basic and repeated acts. In this case, as for DNA, making only one prior and unambiguous hypothesis of homology among these sequences is difficult or impossible. Many alternative alignments of the different sequences can be proposed, preventing any repeatable and reasonable phylogenetic analysis to be carried out with these data. Such ambiguously aligned sequences may be analyzed using direct optimization (Wheeler, 1996). This method performs a one-step minimization of total changes on a trial topology, given a matrix of costs for each kind of character states (insertion deletion and substitutions). Under direct optimization, primary homology statements are considered dynamically and do not imply fixed site-to-site correspondences, as in static alignments (Wheeler, 1996). Therefore, the obtained phylogenetic topology is the one that minimizes the numbers of insertion and deletion events as well as substitutions required by this topology. This method has been implemented first in the case of DNA sequences (e.g., Phillips et al., 2000) but can be applied to any other kind of characters. Few attempts have been made in this direction today, although Schulmeister and Wheeler (2004) have adapted it to analyze developmental studies. We claim here that direct optimization can be applied to phylogenetic analyses of behavioral sequences. In terms of homology for behavior, direct optimization can be used very profitably when multiple hypotheses of homology are competing; as for molecular data, it would provide the most parsimonious hypotheses of homology, dynamically obtained. In terms of phylogenetic analysis of behavior, direct optimization would first allow investigation of the phylogenetic signal within behavioral data and second combining them with other data sets. This procedure of dynamic homology and alignment could be applied as well to very similar sequences the length of which is not so different and therefore often a priori considered as unambiguously aligned (for example, coding DNA sequences). The result could be simply more obvious with very few indels and only substitutions. To show how direct optimization can be applied to stereotyped behavioral data, we will analyze the evolution of calling songs in a cricket clade (Insecta, Orthoptera, Grylloidea, Gryllus Linne ). Crickets sing by rubbing their raised forewings together. During one to-and-fro movement of the forewings, a burst of sound, called a syllable or pulse, is emitted during wing closure, i.e., when the teeth of the stridulatory file of the right forewing are hit by the plectrum of the left forewing; wing opening is silent although it represents an active behavioral act (Michelsen and Nocke, 1974; Bennet-Clark, 1989). Each syllable period is thus constituted of two elementary behaviors, the forewing closure corresponding to the syllable (noisy part of the syllable period), and the forewing opening corresponding to the silent part of the syllable period (Fig. 1). Syllables are arranged according to a definite temporal pattern, called a chirp (Ragge and Reynolds, 1998), and chirps are repeated at length during the emission of a song bout. Cricket calls thus show up to three main rhythmic levels (Fig. 1), which correspond successively to the syllable, the chirp and the chirp sequence within song bouts. The temporal patterns of both the syllable and chirp repetitions are most often species specific; both determine the diversity and efficiency of cricket calls, and both are actually assessed in intraspecific communication (Alexander, 1962; Nabatiyan et al. 2003; Otte, 1992; Pollack, 2001). Both the sound production mechanism and the stridulatory structures involved in sound production are identical throughout crickets (Bennet-Clark, 1989; Gerhardt and Huber, 2002; Desutter-Grandcolas,

3 604 T. Robillard et al. / Cladistics 22 (2006) song bout chirp interchirp silence intrachirp silence syllable = forewing closure silent part of the syllable period = forewing opening time axis syllable period khz khz khz khz A.domesticus G. assimilis G. bimaculatus G. campestris s s s s khz khz 20 khz 20 G. firmus G. fultoni G. integer G. lineaticeps khz s s s s khz khz 20 G. pennsylvanicus G. rubens G. texensis G. veletis 15 khz khz s s s s Fig. 1. Levels of temporal organization of a cricket song. The Sonagrams (down part) represent are second of the calling song of each species. 2003). This would not be true at the scale of Ensifera, as sound production in this clade involves different behaviors for sound production, the sound being produced during forewing opening, closure or both (e.g., Ewing, 1989), or using different emitting structures. In particular, the stridulatory file may be carried by different veins (Desutter-Grandcolas et al., 2003; Desutter-Grandcolas, 2003), or the resonant structures involve various parts of the forewings (Ewing, 1989; Desutter-Grandcolas, 2003). Syllables can then be compared among crickets on a homology basis involving both the emitting structure and elementary behavior, and not only considering end-products such as acoustic emissions (Stuart and Hunter, 1998; Stuart and Currie, 2001, 2002; Stuart et al., 2002; Robillard et al., 2006). However, comparing the temporal patterns of whole songs is much more problematical than just identifying a unit syllable, because it is not possible to identify one by one the syllables through the different syllable series, for lack of reference point, and because both chirps and chirp series are most often irregular in length (Desutter- Grandcolas and Robillard, 2003). A historical analysis of the complete temporal pattern of the cricket song thus cannot be performed through a classical cladistic analysis. In the present paper, we analyze the calling songs of Gryllus species through direct optimization. We selected species for which both molecular data (Huang et al., 2000) and calling song samples (Bonnet, 1995; Walker and Moore, 2000) were available. We first describe the temporal patterns of the songs as sequences of repeated simple behaviors. We then submit these data to phylogenetic analyses using direct optimization. We also define additional acoustic characters to further describe the acoustic properties of the songs, and perform combined phylogenetic analyses using direct optimization for molecular, acoustic and behavioral data. Our aim is not so much to resolve the phylogenetic relations of the Gryllus species, but to demonstrate that stereotyped

4 T. Robillard et al. / Cladistics 22 (2006) behaviors can be described and actually analyzed in a phylogenetic context using direct optimization. Materials and methods Studied taxa Among the 16 Gryllus species presently known in North America, 10 are considered here, together with two European Gryllus species and one outgroup, Acheta domesticus (Table 1). They have been chosen according to the availability of molecular data (mitochondrial 16S and cytb genes: Huang et al., 2000 and Table 1) and adequate song samples (Bonnet, 1995; Walker and Moore, 2000). One species, G. ovisopis Walker (1974) is mute. Acoustic samples Only one song sample per species has been considered for both acoustic and sequential behavioral data. The choice of the song bouts certainly influences the acoustic sequences optimized with POY, but in a way that is minor compared with the differences observed between the songs of different species (Fig. 3). When different song samples were available for one species (Walker and Moore, 2000), the selected recording matches the localities from where sequenced specimens were collected (Huang et al., 2000). Owing to the poor amount of acoustic data available per species, inter- and intra-individual variation of the songs was not taken into account here. This polymorphism should ideally be considered by taking several terminals documented for both acoustic and molecular data per species, to make sure that all variants are effectively terminals in the analysis. Characters of interest Behavioral sequences In a cricket song, syllables are clearly separated as sound units, and each syllable period can be referred to as one to-and-fro movement of the forewings. This constitutes a strong basis for homology statement, making the identification of each syllable by a definite character state in a song sequence quite straightforward. Apart from the silent part of the syllable period, which depends on the opening movement of the wings and is not independent from its noisy counterpart, the song includes other silences located between, and sometimes within, the chirp. These silences are much more problematical to take into account because no specific homology criterion can be directly applied to them. Yet, silences are critical for the definition of many acoustic features of the songs, such as chirp duration, duty cycle or repetition rate, these features being in turn important in terms of signal content, song recognition and evolution. There is not a single solution to this problem, and we consequently explored two coding options. First, we code each silence as one character, whatever its duration (option 1, hereafter referred to as DO1). This coding is quite clear-cut, but loses the information content of the upper levels of song temporal patterns. As an alternative (option 2, or DO2), we use the silent counterpart of the syllable period to characterize the silent parts of the songs: Silences are then divided into as many silent units as necessary to take their durations into account, after removal of the duration of a forewing opening corresponding to the first syllable of the next chirp. In both codings, we separate silences within and between chirps as different character states (Fig. 1). A song bout of 5 seconds has been considered for each taxon and analyzed using the oscillograph option Table 1 Cricket material included in the analysis, with GenBank references for molecular sequences, geographic origin of sequenced specimens, song source and origin of recorded specimens Cytb sequence 16S sequence Locality Song source (origin) Ingroup species assimilis AF AF Florida Walker and Moore (2000) (Florida) bimaculatus AF AF Cyprus Bonnet (1995) (France) campestris AF AF Germany Bonnet (1995) (France) firmus AF AF Florida Walker and Moore (2000) (Florida) integer AF AF California Walker and Moore (2000) (Nevada) texensis AF AF Texas Walker and Moore (2000) (Mississipi) fultoni AF AF Florida Walker and Moore (2000) (Virginia) lineaticeps AF AF California Walker and Moore (2000) (California) ovisopis AF AF Florida Mute species rubens AF AF Florida Walker and Moore (2000) (Florida) pennsylvanicus AF AF New York Walker and Moore (2000) (Tennessee) veletis AF AF Connecticut Walker and Moore (2000) (Michigan) Outgroup species Acheta domesticus AF AF Unknown Walker and Moore (2000) (Florida)

5 606 T. Robillard et al. / Cladistics 22 (2006) of the sound analysis software Avisoft-SASLab Pro, version 4.22 (Specht, 2003); this duration guarantees to keep a record of the information present in the higher levels of rhythmic patterns. Coding only one chirp of a song would have suppressed this crucial information. According to the character states currently used by the POY version (Wheeler et al., 2002) for the direct optimization procedure, the following coding characters were used: A, song syllable; G, interchirp silence; C, intrachirp silence; M, ambiguous character corresponding either to a syllable or an intrachirp silence (A or C); R, ambiguous character corresponding either to a syllable or an interchirp silence (A or G). Acoustic characters To further characterize the songs of the Gryllus species, their acoustic properties have been analyzed with Avisoft using four graphic representations of the sound: oscillograph, envelope amplitude, logarithmic power spectrum (Hamming) and sonagraph (Hann) according to Robillard (2004). Oscillographs (amplitude by time), envelopes (absolute value of amplitude by time) and sonagraphs (frequency by time with indication of power displayed) document the temporal structure and amplitude properties of the songs. Power spectra (acoustic power by frequency) and sonagraphs document the spectral properties of the song: we distinguished the fundamental frequency, or first harmonic (f1), i.e., the lowest frequency peak of the spectrum, and up to three successive integer multiples of the fundamental frequency, i.e., the second (f2), third (f3) and fourth (f4) harmonics when present (Fletcher, 1992). The dominant frequency of the song was thereafter identified as the most powerful component among f1, f2, f3 and f4. Acoustic characters have been chosen according to their biological significance and following the acoustic properties commonly considered in comparative studies of acoustic signals (Cocroft and Ryan, 1995; Price and Lanyon, 2002; Desutter-Grandcolas and Robillard, 2003; Robillard, 2004; Robillard et al., 2006). There inevitably exists a certain amount of dependency between some of these properties, as they all describe the same physical phenomenon, i.e., the sound. But dependency is a recurrent problem in phylogenetic analysis, which could more or less apply to any phylogenetic character (Grant and Kluge, 2003). Some dependency can also be suspected between the behavioral sequences and the acoustic characters, more specifically for those describing the temporal features of the songs (characters 1 6). However, this relationship does not bias the independence of the characters in the present analysis. These two character sets should in fact be seen as complementary rather than redundant, as the acoustic characters describe general properties of the song, whereas the behavioral sequences characterize the effective succession of syllables and silences within a song bout. As an example, the phylogenetic information carried by the chirp rate (acoustic character 4) is not explicitly coded in the sequence of behavioral characters, in particular because our coding options (DO1, DO2) do not take into account the differences in syllable durations between the species. In addition, the behavioral sequence is not described by the few temporal characteristics covered by the acoustic characters. Cladistic analysis Molecular sequences represent the almost complete cytochrome b gene (about 1036 bp) and a fragment of about 500 bp of the 16S gene (Table 1). In the acoustic analysis, continuous data have been separated into classes according to observed natural gaps within the data (Stevens, 1991), and minimizing the number of uninformative states. The number of states per character has been limited to 5, according to present-day POY requirements. The matrix for acoustic characters has been edited with WinClada version (Nixon, ), considering all characters equally weighted. Inapplicable and missing data were coded as?. Some authors have strongly advocated the use of non-additive characters because they argue that assigning costs to some transformations implies strong evolutionary assumptions (Hauser and Presch, 1991; Hauser, 1992; Scotland and Williams, 1993). In contrast, other authors (e.g., Lipscomb, 1992; Wilkinson, 1992; Slowinski, 1993; Wiens, 2001) stress out that the information on relative degrees of similarity should be used to decide relative costs between different states in the same manner as it is used to decide primary homology. Mathematically, and for strictly continuous characters (non-discretized states), taking the distance between the character states to weight the transformations is appealing. However, studies about genetics and development clearly support the idea that slight changes of key components of the genetic developmental process can involve drastic changes in phenotypic characters (e.g., Emlen and Nijhout, 2000; Emlen, 2001; Emlen and Allen, 2004; Emlen et al., 2005). Consequently, ignoring the genetic and physiological control of the calling songs in crickets, there is no clue to say that intermediate states of song parameters are obligatory steps to reach high or low values. Furthermore, as one of our goals is to study the evolution of the song parameters, it is preferable to keep the character states neutral and let the congruence decide how to order the states without a priori assumptions. The acoustic characters were thus coded as non-additive. Phylogenetic treatments were done under direct optimization (Wheeler, 1996), as implemented in the program POY version (Wheeler et al., 2002). Gap, transition and transversion, or their equivalent in behavioral sequence analyses, had an equal weight of 1.

6 T. Robillard et al. / Cladistics 22 (2006) Characters were polarized through outgroup comparisons (Nixon and Carpenter, 1993). A first analysis was performed using molecular data (both genes) and acoustic characters. Then, each coding option of behavioral sequences was submitted to a separate analysis, and their phylogenetic performance compared through the properties of the implied alignments (Wheeler, 2003). The most informative coding option in terms of behavioral characters was then combined with molecular and acoustic data in order to characterize the evolution of cricket songs. In total four analyses were performed: Analysis 1. Molecular and acoustic data. Analysis 2. Behavioral sequences without discretizing silences (DO1). Analysis 3. Behavioral sequences with discretized silences (DO2). Analysis 4. Combined analysis of molecular, acoustic and behavioral data sets (DO2). Congruence between data sets was examined by visual inspection of trees derived from separate subsets of data and by the incongruence length difference (ILD) test (Farris et al., 1994) implemented in Winclada NONA. Briefly, the ILD test provides a statistical evaluation of the incongruence of phylogenetic signal in different data sets by comparing the length of trees from user-defined subsets [ex: molecular and acoustic data (Analysis 1) versus behavioral sequences with discretized silences (Analysis 3)] with tree lengths calculated from randomized subsets of these data. If the data have strongly differing signals, then it is likely that trees from randomly partitioned data will be longer than the trees calculated from the original partitions (Larson, 1994). In any case, the result of the ILD test was not used to assess the combinability of the data sets (Barker and Lutzoni, 2002; Darlu and Lecointre, 2002). When necessary, character optimizations have been performed with Winclada, considering unambiguous transformations only. The outgroup was not considered for optimizations of acoustic characters because of the variability of cricket songs (Grandcolas et al., 2004). The ancestral song of Gryllus has been reconstructed in a two-step process, using the optimizations of the acoustic characters and diagnosing the ancestral behavioral sequence using POY (command diagnose). Both hypotheses were then checked for compatibility to one another and with the song of extant species. Procedures The fragment of cytb, which lacks gaps, and the acoustic characters were treated as prealigned. The 16S gene and the behavioral sequences were neither prealigned, nor split into several fragments. POY analyses were performed on a personal computer with processors cadensed at 2.80 GHz (IntelPentium 4). The following POY command line was used for each analysis: -poy Filename(s) -norandomized outgroup -fitchtrees -terminalsfile Filename -minterminals 0 -replicates 50 -noleading -molecularmatrix 111.txt -repintermediate -buildsper replicate 5 -buildspr -buildtbr -buildmaxtrees 1 -sprmaxtrees 1 -tbrmaxtrees 2 -numdriftchanges 30 -driftspr -numdriftspr 10 -drifttbr -numdrifttbr 10 -maxtrees 20 -treefuse -fuselimit 10 -fusemingroup 5 -check slop 20 -indices -diagnose -time -impliedalignment -phastwincladfile Wincladafile.ss > Outfile.out 2 > Logfile.log Consistency (CI, Kluge and Farris, 1969) and retention (RI, Farris, 1989) indexes have been calculated in POY. Results Acoustic characters and behavioral sequences The acoustic analysis of the songs allows comparison of their acoustic properties and description of them either by numerical values or by descriptive alternatives. Hypotheses of homology have been assessed on the basis of the criteria of relative position and special quality, as they are commonly applied to morphology. The 16 acoustic characters describe the temporal patterns (7), frequency spectra (4) and amplitude variations (5) of the songs (Tables 2 and 3). Character 5, chirp duration over chirp rate, partly depends on the characters chirp duration (character 2) and chirp rate (character 4); we tentatively used this ratio to take into account the phylogenetic information emerging from the interaction between these two variables. The optimizations of the acoustic characters confirm that the transformations of character 5 are not entirely explained by the transformations of characters 2 and 4 (Fig. 3). The only property of the songs that cannot be approached by the acoustic characters is the effective sequence of emission of syllables and chirps along the song bout, for lack of reference point between species. This behavioral sequence was analyzed by translating the song samples into successions of stereotyped behaviors to be analyzed under direct optimization with POY. The behavioral sequences obtained after each coding option of the 5-second song bouts are listed in the Appendix 1. Their unequal length is related to the intrinsic temporal properties of the song of each species (number of syllables per chirp, chirp repetition rate, intra and interchirp silences).

7 608 T. Robillard et al. / Cladistics 22 (2006) Table 2 List of acoustic characters describing the songs of 12 Gryllus species and one outgroup Temporal patterns 1. Number of syllables per chirp: 2 3 (0); 4 5 (1); 6 8 (2); (3); > 150 (4). 2. Chirp duration (minimal interval: 30 ms): < 50 ms (0); ms (1); ms (2); ms (3); > 1500 ms (4). 3. Duration of interchirp silence: < 200 ms (0); ms (1); ms (2); > 1000 ms (3). 4. Chirp rate (in chirp per minute): < 45 (0); (1); (2); (3); > 500 (4). 5. Chirp duration over chirp rate: < 0.1 (0); (1); (2); 1 3 (3). 6. Syllable duration within a chirp, not considering the first or first two syllables (minimal class interval: 3 ms): ms (0); ms (1); ms (2); ms (3); > 27 ms (4). 7. Syllable duration increasing through the chirp (not considering the first or first two syllables) (0); syllable duration stable (1). Frequency 8. Calling song fundamental frequency: < 4000 Hz (0); > 4000 Hz (1). 9. Harmonic peak f3 more powerful than f2 (0); less powerful than f2 (1). 10. Frequency clearly increasing on at least some of the syllables of the chirp, the last syllables having higher frequency values (0); frequency value stable along the chirp (1). 11. Frequency values decreasing at the end of each syllable (0); stable (1). Amplitude 12. Amplitude of the first or first two syllables of the chirp much lower than that of the other syllables (1); amplitude homogeneous along the whole song (0). 13. Profile of amplitude of the chirp: increasing (1); homogeneous (0); not considering the first or first two syllables. 14. Profile of amplitude of the syllables: smooth (0); irregular (1) (see characters 15 and 16). 15. Maximal amplitude of the syllables: at the beginning of the syllable (0); at the end of the syllable (0). 16. Syllable with regular and deep amplitude modulations (0); unmodulated (1). Table 3 Matrix of 16 acoustic characters for 12 cricket species and one outgroup assimilis bimaculatus campestris firmus fultoni integer lineaticeps ovisopis pennsylvanicus rubens texensis veletis A. domesticus As an example of the different coding options, we compare below the songs of G. assimilis and G. pennsylvanicus, which are both made of chirps comprising from four to seven syllables, regularly repeated at different rates: G. assimilis Do1 AAAAAAGAAAAAAAGAAAAAAAGAAAAAG Do2 AAAAAAGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGAAAAAAAGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGAAA AAAGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGAAAAAGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGGGGG G. pennsylvanicus Do1 AAAAAGAAAAAGAAAAAGAAAAAGAAAAGA AAAAG Do2 AAAAAGGGGGGGGGGGGGGGGGGGGGAAA AAGGGGGGGGGGGGGGGGGGGGGGGGGAA AAAGGGGGGGGGGGGGGGGGGGGGGGGGG

8 T. Robillard et al. / Cladistics 22 (2006) GGGGGGGGGGGAAAAAGGGGGGGGGGGGG GGGGGGGGGGGGAAAAGGGGGGGGGGGGG GGGGGGGGGGAAAAAGGGGGGGGGGGGGG GGGGGGGGGGG The 5-second sample of G. assimilis song is described by a sequence of 29 characters (DO1) or 406 characters (DO2). The chirps are described the same way in both sequences, i.e., as a short succession of syllables, but the silent parts of the song are described by one character in DO1 versus characters in DO2. In the same way, the song of G. pennsylvanicus is described by 35 (DO1) or 185 (DO2) characters, thus strongly resembling the song of G. assimilis with DO1, but showing clearly a different temporal structure when silences are discretized. The coding DO2 thus allows account to be taken not only the syllables and the syllable content of the chirps, but also of the chirp repetition rate, i.e., all the three temporal levels of cricket song structure. Phylogenetic analyses Analysis 1. Molecular and acoustic data The analysis is based on 1553 characters, 16 acoustic and 1537 molecular, among which 233 (15%) are parsimony informative. One tree results from this analysis (791 steps, CI 61, RI 50; Fig. 2). The monophyly of the ingroup is not tested by the analysis because we use only one outgroup species; the topology shows a basal trifurcation between A. domesticus (G. bimaculatus G. campestris) and all the other species, the European species being separated from the American ones. Within this clade, the Caribbean G. assimilis is the sister group of all the remaining species, with the following relationships [G. assimilis (((G. fultoni (G. integer G. veletis))(((g. lineaticeps (G. texensis G. rubens)) (G. firmus (G. ovisopis G. pennsylvanicus))))))]. This topology is similar to that obtained for molecular data by Huang et al. (2000) in their maximum parsimony analysis, except for the basal trifurcation; these authors obtained a sister group relationship between European and American Gryllus species. Unambiguous optimizations of acoustic characters show that these data bring little phylogenetic information. Very few character changes support nodes, as for example states 2(3) and 5(3) for [G. firmus (G. ovisopis G. pennsylvanicus)] or state 9(1) for (G. campestris G. bimaculatus). Most changes are either autapomorphic or homoplastic, with up to four transformations per character. Analysis 2. Behavioral sequences without discretizing silences (DO1) The analysis based on behavioral sequences alone results in one tree (139 steps, CI 94, RI 60; Fig. 2). Four species are found at the unresolved base of the tree, A. domesticus, G. veletis, G. bimaculatus and G. campestris. The following node shows G. rubens as the sister group of a trifurcation made of (G. firmus (G. integer G. fultoni)), G. texensis and (G. pennsylvanicus (G. assimilis G. lineaticeps)). The implied alignment shows little overlap between the songs of the different species, especially because of the long-chirped song of G. texensis, with its autapomorphic state representing intrachirp silences, that of G. integer and the trill of G. rubens (Appendix 1 and Table 4A). The song of G. lineaticeps is also poorly aligned, perhaps because the chirps of this species typically have more syllables. This involves numerous insertion events throughout all the sequences (Table 5) and many uninformative characters. Actually, among the 774 characters of the implied alignment, only 84 (10.9%) are defined for at least two species, and only 73 (9.4%) are parsimony informative. Sequence overlaps occur only at the level of shorter chirps (Table 4B): only seven sections made of four to 25 characters show a hypothesis of alignment for more than two taxa. Insertions are located both between and within the chirps. Substitutions also occur and involve either the lack or the gain of a syllable. The intrachirp silence within the song of G. texensis is always aligned with insertions, either within the songs of other species, or within long insertion sequences not aligned with other songs. Analysis 3. Behavioral sequences with discretized silences (DO2) Only one tree was obtained with behavioral sequences alone (982 steps, CI 81, RI 78; Fig. 2). A basal trifurcation separates A. domesticus, G. veletis and all the other species. Among these, G. fultoni separates first, followed by the clade (G. integer G. texensis). All the other species are arranged in a paraphyletic series as follows [G. assimilis (G. lineaticeps (G. pennsylvanicus (G. rubens (G. bimaculatus (G. campestris G. firmus)))))]. The implied alignment (937 characters) shows numerous insertions, located both between and within the chirps. According to the species, DO2 implies more or fewer insertions, but the percentage of insertions related to the number of characters is always smaller than in DO1 (Table 5). Large proportions of the sequences of G. rubens are still not aligned compared with the others (up to 149 sites in one sequence portion), but this is no longer the case for G. integer and G. texensis, whose songs are fully or mostly aligned on the others; G. lineaticeps shows small non-overlapping portions, for both silent bouts and chirps. The chirps of most species are overlapping and aligned via insertions and substitutions (Table 6): 475 of the 937 characters (50.7%) are defined for more than one species, and 422 (45%) are actually informative. Sequences are aligned over very long portions

9 610 T. Robillard et al. / Cladistics 22 (2006) A. domesticus A. domesticus G. bimaculatus G. veletis G. campestris G. assimilis G. fultoni G. integer G. bimaculatus G. campestris G. rubens G. firmus G. veletis G. integer A G. lineaticeps G. rubens G. texensis G. firmus B G. fultoni G. texensis G. pennsylvanicus G. ovisopis G. assimilis G. pennsylvanicus G. lineaticeps A. domesticus A. domesticus G. veletis G. assimilis G. fultoni G. integer G. integer G. texensis G. fultoni G. veletis G. assimilis G. rubens G. lineaticeps G. texensis G. pennsylvanicus G. lineaticeps G. rubens G. ovisopis C G. bimaculatus D G. pennsylvanicus G. campestris G. firmus G. bimaculatus G. firmus G. campestris Fig. 2. Phylogeny obtained for 12 Gryllus species and one outgroup under direct optimization of: (A) molecular data (cytb, 16S) and 16 acoustic characters (Analysis 1: one tree, 791 steps, CI 61, RI 50); (B) behavioral sequences coded with non-discretized silences (DO1) (Analysis 2: one tree, 139 steps, CI 94, RI 60); (C) behavioral sequences coded with discretized silences (DO2) (Analysis 3: one tree, 982 steps, CI 81, RI 78); and (D) molecular (cytb, 16S), acoustic characters and behavioral sequences coded with discretized silences (DO2) (Analysis 4: one tree, 1865 steps, CI 76, RI 67). comprising several tens of characters (Appendix 2). Many character changes are autapomorphic, but synapomorphic changes support most of the nodes of the tree (Table 6). Substitutions concern the acquisition of a syllable instead of a silence, or the reverse. The intrachirp silence within the song of G. texensis is aligned either with an insertion, or with an interchirp silence. According to the different properties of the implied alignments of both coding options, discretizing the silent parts of the songs could thus prove more informative for phylogenetic analyses of song temporal structures.

10 T. Robillard et al. / Cladistics 22 (2006) G. assimilis 5 khz 0.1 s G. integer 2 4 G. fultoni 13 G. veletis 3, 6 G. rubens 2, 3 4 1, 7 G. texensis G. lineaticeps G. ovisopis mute 11 G. pennsylvanicus 4 G. firmus 2, 5, G. bimaculatus G. campestris 4, 7 Fig. 3. Pattern of song evolution in reference to the topology obtained with all data sets, including the behavioral sequences coded with discretized silences (Analysis 4). Sonagraphs represent one second of the song of each species. Acoustic characters have been optimized on the tree using the unambig. command of Winclada (numbers above arrows refer to the characters described in Table 2). Symbols: dotted line, mute species (loss of acoustic communication in G. ovisopis); black arrows, transformations of temporal characters (characters 1 7), the direction indicating either an increase (arrow pointing up) or a decrease (arrow pointing down); white arrows, decrease in fundamental frequency (character 8); white ellipsoid, increase of power of harmonic f3 (character 9); black rectangle, decrease of frequency at the end of each syllable (character 11); black diamond, loss of the increasing profile of amplitude of the chirp (character 13). Analysis 4. Combined analysis of molecular data, acoustic characters and behavioral sequences with discretized silences (DO2) Combining behavioral, molecular and acoustic data yielded only one tree (1865 steps, CI 76, RI 67; Fig. 2). A basal trifurcation separates A. domesticus from [G. assimilis (G. integer G. fultoni)] and the rest of the species. This last clade is fully resolved and shows G. veletis as the sister group of [(G. texensis G. rubens)

11 612 T. Robillard et al. / Cladistics 22 (2006) Table 4 Examples of character states alignment resulting from direct optimization of the behavioral sequences with non-discretized silences (Analysis 2 with DO1). See the text for descriptions of character states; - represent indels A. domesticus assimilis bimaculatus campestris firmus fultoni integer A A G A A G lineaticeps pennsylvanicus rubens A A A A A texensis A C A A C veletis A. domesticus A A G A A G assimilis A A A A G A A A A bimaculatus G A A A A A A A G A A A A G A campestris G A A A A A A A G A A A A G A firmus A A A A G A A A A A G A A fultoni A A A G A A A G A A A G A A G A integer A A A G A A A G A A A G A A G A lineaticeps A A A A G A A A A A pennsylvanicus A A G A A A A A G A rubeens A A A A A A A A A A A A A A texensis A A A A A A A A A A A A A A veletis G A A A A A A A G A A A A G A Table 5 Properties of implied alignments resulting from analyses # 2 (DO1), # 3 (DO2) and # 4 (TOT, total analysis) DO1 DO2 TOT N char. N ins. % ins. N char. N ins. % ins. N char. N ins. % ins. A. domesticus assimilis bimaculatus campestris firmus fultoni integer lineaticeps pennsylvanicus rubens texensis veletis N char., number of behavioral character per taxon (A, C, G, R or M); N ins., number of insertion deletion per taxon; % ins., proportion of indels over the total sequence length (N char + N ins.) (G. lineaticeps (G. ovisopis (G. pennsylvanicus (G. firmus (G. campestris G. bimaculatus)))))]. The topology obtained with the combined analysis differs markedly from the tree resulting from Analysis 1, as some species that separate basally in one analysis are apical in the other, and vice versa. This is particularly true for the clade (G. campestris G. bimaculatus) and for G. assimilis, which from the most two basal positions are now either apical, or included in a separate clade, respectively. Some clades are found in both analyses, such as (G. campestris G. bimaculatus) and (G. texensis G. rubens), which correspond to the best supported nodes. Several taxa also remain close in both topologies, such as G. integer and G. fultoni on one hand, and G. lineaticeps, G. firmus, G. ovisopis and G. pennsylvanicus on the other. Optimizations of all the acoustic

12 T. Robillard et al. / Cladistics 22 (2006) Table 6 Examples of character states alignment resulting from direct optimization of the behavioral sequences with discretized silences (Analysis 3 with DO2). See the text for descriptions of character states; represent indels A. domesticus G G G G G G assimilis G G G G G G G G G G G G G bimaculatus campestris firmus fultoni G G G G G G G G G G G G A integer G G A A G G G G G A A A A A G G A lineaticeps G G G G G G G G G G G G G pennsylvanicus rubens texensis A A A A A A A A c A A A A A A A A veletis G G G G A A A. domesticus A A G G assimilis G G G G G G G G G G G G G G G bimaculatus G G G G A A A A A A A campestris G G G G A A A A A A A firmus G G G G A A A A A A A fultoni A A A G G G G G G G G G G G G integer G G G G A G G G G G G G G A A A G lineaticeps G G G G G G G G G G G G G G G pennsylvanicus G G G G G G G G G G G G G rubens A A G G G G r A R A A A A texensis A A C A A A A A A A A C A A A A A veletis A A G G A A characters were performed to visualize the unambiguous changes of the song parameters (Fig. 3). According to the implied alignment, this analysis is based on a total of 2596 characters, including 1537 molecular characters, 16 acoustic characters and 1043 characters in behavioral sequences. Here the behavioral data represent 40.2% of the whole character set. Among all the behavioral sites, 643 (24.8%) are parsimony informative, to be compared with the 15% of informative characters in the Analysis 1. Through the behavioral sequences, 465 characters (44.6%) have a definite state for at least two species, and so are potentially informative for song evolution. This is less than in the DO2 analysis (50.7%), but still much higher than in DO1 (10.9%). Among these 465 characters, 410 are parsimony informative, which represents 39.3% of behavioral characters, and 15.8% of the whole character set. Globally, the alignment resulting from this analysis is thus based on a higher proportion of informative characters compared with previous analyses. The implied alignment for behavioral data still shows some portions of non-overlapping sequences, but only for one chirp of G. texensis, whose song is characterized by long chirps with autapomorphic intrachirp silences, and for several interchirp silences in G. lineaticeps. These non-overlapping parts are, however, much smaller than in previous analyses, covering from 31 to 52 contiguous character states. Apart from these, all the chirps of all the species, with the exception of one chirp of G. campestris, are part of aligned portions. This may explain the increase in character number and insertions in behavioral sequences compared with DO2, as more insertions may be necessary to improve the alignment of the chirps of all studied taxa. The implied alignment shows that both substitutions and insertions occur between the songs of the species, including the transformation of a silence into a syllable, or the reverse (Table 7). The intrachirp silence within the song of G. texensis is always aligned with insertions, never with a silent interchirp character state. Table 7 shows the number of character transformation per type of process of song evolution with Slow optimization. We analyzed the whole set of temporal characters and made comparisons with the subsets of informative and non-informative characters. According to the results, insertions of silence ( > G) are predominant in non-informative characters (55.82%), whereas they are as common as substitutions of a silence by a syllable in informative characters. In noninformative characters, insertion is thus the process favored for alignment, which represents addition of elements not shared by taxa, i.e., autapomorphies, but mostly of silences ( > G). For informative

13 614 T. Robillard et al. / Cladistics 22 (2006) Table 7 Character transformations in the implied alignment of discretized behavioral sequences (DO2) in the total analysis. Only Slow optimization has been considered; see the text for descriptions of character states; represent indels. Percentages are related to the total number of transformations Transformations All characters from implied alignment of behavioral sequence (1043) Informative characters (410) Non-informative characters (633) > A 306 (21.87%) 137 (17.75%) 169 (26.95%) > C 66 (4.72%) 0 66 (10.53%) > G 534 (38.17%) 184 (23.83%) 350 (55.82%) G > A 216 (15.43%) 194 (25.13%) 22 (15.46%) A > G 57 (4.07%) 50 (6.48%) 7 (3.51%) A > 91 (6.50%) 91 (11.79%) 0 G > 121 (8.657%) 110 (14.25%) 11 (1.75%) G > C 8 (0.57%) 6 (0.78%) 2 (0.32%) Deletions (A,C,G > ) 212 (15.15%) 201 (26.04%) 11 (1.75) Insertions ( > A,C,G) 906 (64.76%) 321 (41.58%) 585 (93.33%) Substitutions (A,C,G > A,C,G) 281 (20.09%) 250 (32.38%) 31 (4.94%) characters, homologies made by direct optimization as part of the most parsimonious alignment show more or less a balance between deletions (A,C,G > ) and substitutions (A,C,G > A,C,G), insertions ( > A,C,G) remaining slightly more common. Congruence of phylogenetic signal was examined by comparing the implied alignments obtained from the analysis of the molecular and acoustic data and the analysis of the behavioral sequences with discretized silences (DO2). The ILD test shows that there is a significant incongruence between these data sets (P ¼ ), which confirms the observed incongruence between the topologies (see below). Ancestral calling song of Gryllus The optimizations of the 16 acoustic characters (Table 2) were used to reconstruct the ancestral song of the ingroup by considering all the possible most parsimonious scenarios and all the possible ancestral states for each character. According to this data set, the ancestral song of Gryllus consists of a chirp of 2 8 syllables (character 1: 0 1 2), for a duration of ms (character 2: 1); the duration of interchirp silences is ancestrally below 200 ms (character 3: 0), and the chirp rate is chirps per minute (character 4: 2). The ancestral syllable duration is low ( ms; character 6: 0 1) but slightly increases along the chirp (character 7: 0). The ancestral fundamental frequency is stable within the syllables (character 11: 1) and its value is over 4000 Hz (character 8: 1) and the second harmonic is more powerful than the third (character 9: 1). The amplitude profile of the chirp is ancestrally increasing (character 13: 1) and that of the syllable is smooth (character 14: 0). In parallel, the topology resulting from Analysis 4 was used to perform the diagnosis of the behavioral sequences at the basal node of the ingroup. This procedure permitted inference of the ancestral sequence of syllables and silences corresponding to the ancestral song of Gryllus reconstructed by direct optimization. This putative ancestral sequence is given in Fig. 4 and consists of characters (10 uncertain elements involving potential ancestral gaps). Only temporal????? A A R R G G G G G G G G G G G G G G G G A A A G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G A A A G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G A A A G G G G G G G G G G G G G G G G G G G G G A A G G G G G G G G G G G G G G G G G G G G G G G G G R R A A A G G G G G G G G G G R G G G G G G G G G G G G G G G G G G?? G G G G G G G G G G G G G G G G G G G G G G G G G R G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G G??? A A G G G G G G G G G G G G G G G G G G G G G G G A A A G G G G G G G G Fig. 4. Putative ancestral song of Gryllus reconstructed by direct optimization with discretized silences (DO2 coding, see text for more details). Each rectangle represents a time equal to the duration of one syllable period in the ancestral song (see text for details). Symbols: black rectangles (A), syllables; white rectangles (G), interchirp silences; dotted rectangles (?), ambiguous ancestral characters involving either a gap or another state; gray rectangles (R), ambiguous character involving either a syllable or a silence.