Molecular Systematics of the African Electric Fishes (Mormyroidea: Teleostei) and A Model for the Evolution of Their Electric Organs

The superfamily Mormyroidea (families Mormyridae + Gymnarchidae) sensuNelson (1994) is a large group of freshwater fishes endemic to Africa, comprising more than 200 recognized species placed in 19 genera. They are distributed over most of the continent, with the exception of the Sahara, northernmost Mahgreb and southernmost Cape provinces (Lowe-McConnell, 1987; Roberts, 1975), and are most diverse in the river systems of Central and West Africa. Mormyroids are by far the most speciose extant lineage of the ancient teleost order Osteoglossiformes and represent a remarkable radiation within this otherwise relictual group. The monophyly of the mormyroids is supported by the derived presence of electric organs, matched electroreceptors and a greatly enlarged cerebellum in all taxa, amongst other characters (Taverne, 1972; Lauder and Liem, 1983).

While mormyroids were known to the ancient Egyptians, who accurately depicted them on the walls of their tombs (Brewer and Friedman, 1989), the discovery that these fishes produce and sense weak electric signals was made less than 50 years ago (Grundfest, 1957; Lissmann, 1951, 1958; Lissmann and Machin, 1958). Since then, mormyroids have become a model system for the study of electrogenesis, electroreception and electrocommunication (for recent reviews, see Bullock and

Heiligenberg, 1986; Moller, 1995; Turner et al., 1999). Although advancements in our understanding of mormyroid systematics have lagged behind progress in these other areas, a modern phylogenetic hypothesis for these fishes is a prerequisite for placing the growing body of knowledge about them into an evolutionary context. In this paper, we present the most complete phylogenetic study of mormyroids to date. We have used maximum parsimony methods to analyze 3270 characters from the mitochondrial and nuclear genomes of 41 mormyroid species belonging to 18 genera. We use our phylogeny for the mormyroids to investigate the evolution of the electric organ in these fishes.

Table 1 lists the specimens used in this study from 41 mormyroid species belonging to 18 genera. The only recognized genus absent in our dataset is the questionable monotypic Heteromormyrus Steindachner 1866 from Angola. Heteromormyrus pauciradiatus is known only from a single specimen, which is now lost (Daget et al., 1984), and the validity of this genus is uncertain (Taverne, 1972). We use the notopterids Chitala ornata and Xenomystus nigri and the osteoglossid Pantodon buchholzi as outgroups for this study. According to Li and Wilson (1996) and Lauder and Liem (1983), the Notopteridae are the sister group to the mormyroids, while Pantodon represents a more distant outgroup. Most of these specimens were collected on field trips to Gabon (16 taxa), the Central African Republic (12 taxa), Mali (eight taxa), Ivory Coast (one taxon) and Ghana (one taxon). J. Snoeks provided Hippopotamyrus discorhynchus from Lake Malawi, and we obtained Brienomyrus brachyistius, Campylomormyrus sp.1, Campylomormyrus tamandua (1), Gymnarchus niloticus, Chitala ornata, Xenomystus nigri and Pantodon buchholzi from the aquarium trade. We identified specimens to species whenever possible by consulting original species descriptions, and we designated undescribed species or species of uncertain identification with numbers.

We killed the fishes using MS222 in accordance with NIH guidelines as supervised by the Cornell University Animal Care and Use Committee. Fresh tissue was preserved in the field in 90 % ethanol or a dimethylsulfoxide/EDTA saturated salt solution (Seutin et al., 1991). In most cases, DNA was extracted from the tissue samples using the QIAamp tissue kit (Quiagen Inc., Valencia, CA, USA).

This study directly builds upon that of Alves-Gomes and Hopkins (1997), who used partial 12S and 16S mitochondrial sequences to examine relationships among 12 mormyroid species, and Lavoué et al. (1999), who used partial cytochrome b sequences in a phylogenetic analysis of 27 mormyroid taxa. We have again used the 12S, 16S and cytochrome b genes, but expanded taxonomic sampling and used longer sequences in the case of cytochrome b.

Primer sequences for the mitochondrial genes were taken from Palumbi (1996). These sequences are as follows: 12S forward, 5′-AAA CTG GGA TTA GAT ACC CCA CTA T-3′ (L1067); 12S reverse, 5′-GAG GGT GAC GGG CGG GCG GTG TGT-3′ (H1478); 16S forward, 5′-CGC CTG TTT ATC AAA AAC AT-3′ (L2510); 16S reverse, 5′-CCG GTC TGA ACT CAG ATC ACG T-3′ (H3080); cytochrome b forward, 5′-TGA TAT GAA AAA CCA TCG TTG-3′ (L14724); and cytochrome b reverse, 5′-CTC CAG TCT TCG rCTT TAC AAG-3′ (H15930). Certain taxa failed to amplify with primer L14724. For these, we used an alternative forward primer: 5′-TAC CTA TAC AAA GAA ACm TGA AA-3′ (L15047). Although some of the taxa sampled have been sequenced before for these same regions (Alves-Gomes and Hopkins, 1997; Lavoué et al., 1999), we obtained new sequences for all individuals included in this study with the exception of Pollimyrus isidori, for which we use the sequence of Lavoué et al. (1999).

To obtain an estimate of phylogeny from a locus independent of the mitochondrial genome, we also sequenced a portion of the nuclear recombination activating gene 2 (RAG2), which has recently been used in a study of beloniform fishes (Lovejoy, 1999). This gene and the closely linked RAG1 encode components of an enzyme involved in the recombination of immunoglobin and T-cell receptor genes. These genes appear as single copies in all vertebrates examined to date (Hansen and Kaattari, 1996; Willett et al., 1997).

Two forward and two reverse primers taken from Lovejoy (1999) were used in polymerase chain reaction (PCRs), certain combinations working best with different templates. The forward primers are 5′-TTT GGr CAr AAG GGC TGG CC-3′ (F1) and 5′-ArA CGC TCm TGT CCm ACT GG-3′ (F2). The reverse primers are 5′-GTr GAr TAG TAG GGC TCC CA-3′ (R4) and 5′-TGr TCC ArG CAG AAG TAC TTG-3′ (R6). These primers correspond to base pairs 90–110, 112–132, 1287–1307 and 1425–1446, respectively, in the zebrafish (Danio rerio) RAG2 gene (Willett et al., 1997).

Approximately 100–500 ng of total genomic DNA was used as template for 50 μl PCR reactions containing 1.25 units of Perkin Elmer AmpliTaq Gold and the following ingredients at the indicated final concentrations: Perkin Elmer GeneAmp PCR buffer II at 1×, each amplification primer at 0.2 μmol l⁻¹, each dNTP at 200 μmol l⁻¹ and MgCl₂ at between 1.5 and 3 mmol l⁻¹.

Amplification conditions consisted of an initial 95 °C denaturation step for 10 min, 35 cycles of 94 °C for 1 min, annealing for 1 min (at 55–60 °C for the 12S and 16S fragments, at 42 °C for cytochrome b and at 53 °C for RAG2), extension at 72 °C for 1.5 min, followed by a final extension at 72 °C for 7 min. PCR products were purified using the Promega Wizard PCR Preps DNA purification kit (Promega, Madison, WI, USA).

We sequenced the double-stranded PCR products directly in both directions with the primers used for amplification on an Applied Biosystems 377 automated sequencer. We edited the sequences using the Sequencher software package (Gene Codes Corp., Ann Arbor, MI, USA). All sequences are available in GenBank (accession numbers AF201483– AF201660).

We aligned our 12S and 16S sequences using CLUSTAL W (Thompson et al., 1994) with three different parameter settings, which adjust the gap opening and gap extension costs relative to the base change cost. These were gap opening/gap extension: 10/5 (default), 7/5 and 10/10. The ‘transitions weighted’ option was used on all three alignments (identical bases receive a score of 3, transitional mismatches a score of 1, other mismatches a score of 0). Bases whose positions differed in the three alignments were excluded from the phylogenetic analyses to avoid errors in positional homology (Gatesy et al., 1993). We used the edited 12S and 16S datasets to create a matrix of gap characters using PAUPGap 1.0 (Cox, 1997). We preferred this method to coding gap positions as ‘fifth bases’ in PAUP so that shared gaps of any size receive the same weight in the analysis (i.e. a shared gap of five bases is treated as one character, not five). The matrix of two-state gap characters was appended to the NEXUS file consisting of the aligned sequences. These gap characters were given the same base weight as the nucleotide characters in those phylogenetic analyses in which they were included.

The NEXUS file containing the aligned sequences used for phylogenetic analysis is available at http://www.nbb.cornell.edu/neurobio/hopkins/hopkins.html or upon request from the authors.

We estimated the phylogenetic signal in the aligned sequence data using maximum parsimony (MP) analyses in PAUP^* version 4.0b2 (Swofford, 1999). In all cases, the most parsimonious trees were obtained using 100 heuristic searches with random addition of the taxa to produce the starting tree. The default settings were used for all other parameters. We conducted separate MP analyses for the combined 12S and 16S data, the cytochrome b data and the RAG2 data. Relative support for internal nodes of these trees was estimated using bootstrap analysis (Felsenstein, 1985) consisting of 1000 pseudoreplicates in PAUP^*. Bootstrap values are interpreted here as an estimate of relative support for nodes on a tree produced from a particular dataset, but are not equivalent to statistical probability values. We looked for evidence of mutational saturation for each class of nucleotide substitution and within each codon position of the cytochrome b and RAG2 genes by plotting a metric of the observed number of nucleotide differences in pairwise comparisons of taxa (the ‘adjusted character distance’ of PAUP^*) against a metric of the corresponding number of substitutions inferred to have taken place on the branches joining each pair of species on a most-parsimonious tree (the patristic distance of PAUP^*). This tree was produced from a preliminary unweighted MP analysis. Saturation is estimated by the extent to which the slope of a linear regression departs from a value of 1 (Hassanin et al., 1998).

In cases of saturation of transitions, we examined the effect of down-weighting them relative to other character state changes through the use of step matrices and compensatory static character weights in PAUP^*. For example, to down-weight certain transitions by half relative to all other character state changes, we used a step matrix that gave transversions a weight of 2 and transitions a weight of 1 at those sites. In addition, we placed a compensatory static weight of 2 on all characters not subject to the step matrix. In this way, transitions were down-weighted with respect to all other character state changes without otherwise altering weight relationships in the dataset.

We used the partition homogeneity or ‘combinability’ test of Farris et al. (1995), as implemented in PAUP^*, to test for conflict of signal among the informative characters of the three datasets before pooling them into a combined dataset.

On the combined data we performed an unweighted MP analysis and, in addition, performed other MP analyses in which classes of mutation for which we found evidence of saturation were differentially down-weighted relative to other character state changes. There is no a priori objective method for selecting a single best weighting scheme for a parsimony analysis or of choosing a single best tree among those produced in the differently weighted analyses. For this reason, we adopted an approach similar to the successive approximations or ‘a posteriori reweighting’ method of Farris (1969) in which a character’s weight in a weighted parsimony search is set equal to the maximum rescaled consistency index (RCI) of that character on the set of trees obtained from an initial unweighted search. If necessary, this procedure is repeated until a single most parsimonious tree or set of trees is found. In our procedure, we expanded the set of candidate trees used to calculate each character’s maximum RCI by including all shortest trees found in six separate MP analyses (five in which different a priori weighting schemes were applied and one unweighted). Our intention was to prevent a limited set of tree topologies from biasing the reweighted analysis.

To estimate support for nodes in this combined analysis, bootstrap values and decay indices were calculated with PAUP^* and the software AutoDecay (Eriksson, 1997) respectively. In addition, we compared the nodes recovered in the a posteriori reweighted analysis with those present in the shortest trees produced in the original unweighted and five a priori weighted MP analyses.

We examined electrocyte morphology using tissue fixed in 4 % buffered formaldehyde embedded in JB-4 plastic (Polysciences), cut in sagittal plane at 6–7 μm and stained with Toluidine Blue, as described by Hopkins (1999a,b) and other sources (Bass, 1986b,c; Bass and Hopkins, 1983, 1985). Electrocytes were classified as to the presence or absence of penetrating stalks and the side of innervation using light microscopy. We explored the evolution of the mormyrid electrocyte on our preferred phylogenetic tree by searching for the most parsimonious reconstruction(s) of electrocyte character states using the software application MacClade 3.07 (Maddison and Maddison, 1997) combined with considerations of electrocyte ontogeny.

For the analysis of the 12S and 16S sequences, we used Chitala ornata and Xenomystus nigri as the sole outgroups since alignment was appreciably improved when the more distant Pantodon buchholzi was excluded from the dataset. After the removal of Pantodon buchholzi, the alignment of all 395 bases of the 12S gene fragment was identical under the three different parameter settings in CLUSTAL W. Of these sites, 100 (25 %) are variable, with 75 (19 %) being parsimony-informative. In the 16S fragment, 45 characters changed position among the three CLUSTAL W alignments and were excluded from later analyses, leaving 571 aligned bases. Of these sites, 151 (26.5 %) are variable in the dataset, with 113 (20 %) being parsimony-informative. From the aligned and edited 12S and 16S datasets, 29 gap characters were identified and coded using PAUPGap 1.0. Of these, 20 are parsimony-informative.

Nucleotide composition (Table 2) was very similar between the 12S and 16S fragments. Considering only informative sites, Gymnarchus niloticus and the notopterid outgroup taxa are relatively lower in G and higher in A content than the others. This pattern is the same as that reported by Alves-Gomes and Hopkins (1997). Nevertheless, a χ²-test to detect non-homogeneity of base frequencies across taxa did not approach significance on the informative-characters-only dataset.

In Fig. 1, the 12S, 16S and cytochrome b uncorrected genetic distances (p-distances) calculated in all possible pairwise comparisons of the taxa used in this study are plotted against the corresponding values for the RAG2 gene. Similar p-distances in corresponding pairwise comparisons for the 12S and 16S fragments indicate similar evolutionary characteristics. For this reason and because of similar base composition and a non-significant P-value of a partition homogeneity test in PAUP, the 12S and 16S datasets were pooled for subsequent analysis. Parsimony analysis of this pooled dataset with and without the gap characters revealed that their inclusion allows the resolution of several additional nodes. When the 29 gap characters were included and given the same weight as the nucleotide data, a heuristic search yielded 10 most parsimonious trees of 834 steps with a Consistency Index (CI) of 0.38 and a Retention Index (RI) of 0.62. Fig. 2 shows a consensus of the results for the combined 12S, 16S and gap data when three different weighting schemes are employed. All nodes recovered above the 50 % bootstrap level in the three analyses were compatible.

The results of the 12S and 16S parsimony analysis support the monophyly of the ingroup (mormyroid) taxa, with Gymnarchus niloticus (Gymnarchidae) forming the sister group to the all the mormyrid taxa. Within the mormyrids, all four included Petrocephalus species form a clade that is the sister group to the remaining taxa. Within this latter group, Myomyrus macrops forms the sister group to the other mormyrids. Most other higher-level relationships remain unresolved. However, the genera Stomatorhinus, Pollimyrus, Mormyrops and Mormyrus, which are represented by more than one taxon, all appear monophyletic with strong support. A clade of four species of Campylomormyrus plus Gnathonemus petersii is strongly supported. This clade is nested within a larger clade including Genyomyrus donnyi and species of Marcusenius and Hippopotamyrus. Paramormyrops gabonensis is nested within a clade of Brienomyrus species from Gabon. Included in this well-supported clade is Marcusenius conicephalus, which failed to group with other Marcusenius species. Brienomyrus brachyistius, the type species of the genus, does not form a clade with the Brienomyrus species from Gabon, but instead is the sister taxon to Isichthys henryi. Brienomyrus niger clusters neither with B. brachyistius nor with the Gabon Brienomyrus species. Ivindomyrus opdenboschi and Boulengeromyrus knoepffleri, two monotypic species both endemic to the Ivindo/Ntem system of Gabon and Cameroon, form a strongly supported clade.

With the exception of Hippopotamyrus pictus, for which cytochrome b primers failed to amplify a fragment, cytochrome b sequences were obtained for all taxa used in the 12S and 16S dataset. We obtained 1141 nucleotides of this gene for 37 taxa and 800 nucleotides for seven taxa (indicated in Table 1) which failed to amplify with the L14724 cytochrome b primer. For these seven taxa, we used the L15047 forward primer. Finally, for Pollimyrus isidori, which failed to amplify with any of our primers, we used a 500 base sequence from Lavoué et al. (1999). Alignment of the cytochrome b sequences was straightforward: no gaps/insertions were observed. Codon position was easily determined, and no stop codons were observed within any of the sequences.

In our plot of pairwise-adjusted character distances versus corresponding pairwise patristic distances inferred from a preliminary MP tree, we observed significant evidence of saturation in third-position transitions at greater genetic distances. Saturation of third-position transversions was only evident in pairwise comparisons between Pantodon buchholzi (the most distant outgroup taxon sequenced) and other taxa. For this reason, we opted to exclude Pantodon buchholzi from the cytochrome b dataset in further analyses. We rejected the alternative of retaining Pantodon buchholzi but excluding third-position sites, since these account for approximately 75 % of the informative characters in cytochrome b. Saturation of third positions was less severe for the remaining outgroups Xenomystus nigri and Chitala ornata.

Over the 1141 base fragment, 529 (46 %) of nucleotide positions are variable and, of these, 451 (39.5 %) are parsimony-informative (363 variable/336 informative in the third position, 126 variable/90 informative in the first position and 40 variable/25 informative in the second position). These ratios are similar to those reported by Lavoué et al. (1999).

We found nucleotide composition in the third-position sites (Table 2) to be strongly biased towards C and A while poor in T and G content. In addition, base composition varied greatly across taxa. A χ²-test of base frequencies across taxa was highly significant (P<0.0001), indicating a lack of homogeneity. An excess of T and a lack of C compared with most other taxa is apparent in Gymnarchus niloticus (21.5 % T, 32 % C) and in the notopterid outgroup taxa (15–17.5 % T, 38–41 % C). A base composition that differs significantly across taxa, as we found here, can impact on maximum parsimony and other phylogenetic reconstruction methods which may incorrectly group taxa with convergently or primitively similar base compositional biases (Swofford, 1999). Strong bias in nucleotide composition is also evident in the informative sites of the first two positions, although χ²-tests for heterogeneity across taxa did not approach significance (Table 2).

Considered as a unit, cytochrome b appears to evolve at approximately three times the rate found in 12S and 16S at lower genetic distances and closer to twice the rate of these genes at higher genetic distances in the taxa we studied (Fig. 1). The apparent relative slowing of cytochrome b compared with 12S and 16S at increasing genetic distance could be due to mutational saturation of third-position transitions (noted above) obscuring true genetic distance. Third-position saturation is undoubtedly in part the result of the overabundance of A and T (together comprising 86 % of all bases), which greatly increases the likelihood of homoplasious change.

A heuristic search on the complete unweighted cytochrome b dataset produced two most parsimonious trees (2453 steps, CI excluding uninformative characters 0.32, RI 0.47). Because evidence of saturation in the cytochrome b fragment was present only for the third position sites (see above), only third-position transitions were down-weighted relative to other changes in the weighted analyses. Single most parsimonious trees were obtained when third-position transitions were given weights of 0.5× and 0.25× relative to other characters, although several nodes differed in these two trees. Notably, in none of these trees does Gymnarchus niloticus form a clade with the other ingroup taxa, but instead appears as the sister taxon to Xenomystus nigri. Only when third-position transitions are down-weighted to one-tenth (or less) of the weight of all other character state changes does Gymnarchus niloticus cluster with the ingroup taxa. However, this extreme weighting also affects relationships in more terminal portions of the tree. We interpret the failure of the ingroup taxa to appear as a clade in analyses in which third positions are not heavily down-weighted as the effects of mutational saturation and the similar base compositional biases in the third positions of the Gymnarchus and notopterid cytochrome b sequences, noted above.

Fig. 3 shows the results of a combinable component consensus of three MP bootstrap trees obtained from the cytochrome b data under three different weighting regimes. Excepting the failure of Gymnarchus niloticus to group with the mormyrid taxa, every supra-generic clade recovered above the 50 % bootstrap level in the combined 12S and 16S bootstrap analysis is also recovered in the cytochrome b bootstrap analysis (allowing for the absence of Hippopotamyrus pictus from the cytochrome b dataset). Several additional clades are recovered as well. The first of these is a clade of the four Campylomormyrus species. These species plus Gnathonemus petersii form a clade with a Marcusenius moorii plus Marcusenius sp. 1 sister pair and Genyomyrus donnyi. In addition, Marcusenius conicephalus forms the sister taxon to the Brienomyrus/Paramormyrops clade. Some intrageneric relationships in the genera Petrocephalus, Mormyrops, Pollimyrus and Stomatorhinus and within the Gabon Brienomyrus/Paramormyrops clade differ between the MP bootstrap analyses of this dataset and the 12S/16S dataset. Despite these differences, a partition homogeneity test between the two datasets did not approach significance.

We obtained 1143 bases of the RAG2 gene for all 45 taxa in Table 1. Insertions/deletions (indels) were observed only between the Pantodon buchholzi sequence and the others. These are four indels varying from one to three codons in length which were easily aligned by eye by maintaining codon structure and identifying flanking codons for amino acid residues identical to those in the other sequences. Of the 1143 bases, 578 (50.6 %) are variable and 356 (31.1 %) of these are parsimony-informative (335 variable/237 informative in the third position, 148 variable/76 informative, in the first position and 95 variable/43 informative in the second position).

Base composition for each position of the RAG2 fragment is shown in Table 2. χ²-tests for non-homogeneity of bases across taxa did not approach significance for any codon position in RAG2.

We observed no evidence of saturation in third-position transitions when we plotted PAUP’s pairwise-adjusted character distances against the patristic distances calculated from a tree produced by a preliminary maximum parsimony analysis. RAG2 uncorrected p-distance in pairwise comparisons of closely related taxa is approximately half that of the 12S and 16S p-distance and less than 20 % of that of cytochrome b p-distance in the same comparisons (Fig. 1). The apparent evolutionary rate of RAG2 increases relative to that of the mitochondrial genes in pairwise comparisons of more distantly related taxa such that, at intermediate overall genetic distance, RAG2 p-distances are roughly equal to those of the 12S and 16S fragments. In pairwise comparisons between outgroup and ingroup taxa involving greater genetic distance, RAG2 p-distances surpass those of the 12S and 16S fragments and are roughly equal to those of the cytochrome b gene. We interpret this pattern to be the result of increasing mutational saturation of the mitochondrial genes at greater genetic distances relative to RAG2, which shows no evidence of saturation. Also notable is the large genetic distance in the RAG2 data separating Myomyrus macrops from other species within Taverne’s subfamily Mormyrinae (Fig. 1).

A heuristic MP search on the entire RAG2 dataset produced 144 equally parsimonious trees, each 1226 steps long (CI excluding uninformative characters 0.56, RI 0.72). A strict consensus of these trees is shown in Fig. 4. Bootstrap support values are indicated to the left of each node. This tree is better resolved than either of the mitochondrial bootstrap consensus trees and is consistent with the suprageneric relationships recovered in them.

In the RAG2 tree, Brienomyrus niger and Hippopotamyrus pictus are resolved as sister taxa that, together with Hyperopisus bebe, are placed in an unresolved polytomy at the base of the Campylomormyrus/Gnathonemus/Genyomyrus/ Marcusenius/Hippopotamyrus clade. This large clade is sister to another that is resolved into a polytomy of three subclades. Two of these consist of the species of Pollimyrus and Stomatorhinus, respectively. The third contains the clade consisting of several Brienomyrus species from Gabon plus Paramormyrops gabonensis. Again, as in the mitochondrial trees, Marcusenius conicephalus forms the sister taxon to this clade. The sister group to this larger clade is Ivindomyrus opdenboschi plus Boulengeromyrus knoepffleri. Several additional clades not obtained in the mitochondrial analyses are resolved in the RAG2 tree as successive outgroups to the taxa discussed above. These are (in descending order down the tree) the two species of Mormyrus, the sister pair of Brienomyrus brachyistius and Isichthys henryi, and the clade comprising the three Mormyrops species. As in the mitochondrial trees, Myomyrus macrops appears as the sister taxon to all other Mormyrinae, the species of Petrocephalus are the sister taxon to all the Mormyrinae, and Gymnarchus niloticus is the sister taxon to all Mormyridae. With Pantodon buchholzi included as the designated outgroup, the two notopterid taxa appear as sister taxa to each other and as the sister group to the mormyroids.

We pooled the three datasets after partition homogeneity tests among them failed to detect incompatibility. The notopterids Xenomystus nigri and Chitala ornata were included as the only outgroups. We retained Hippopotamyrus pictus in this analysis despite the absence of a cytochrome b sequence for this taxon (all cytochrome b characters were coded as ‘missing’). This dataset contains 3270 characters, 969 of which are parsimony-informative. A heuristic search on the unweighted dataset produced 19 equally parsimonious trees (4308 steps, CI excluding uninformative characters 0.38, RI 0.57). Suprageneric relationships in the strict consensus of these 19 trees are consistent with the well-supported relationships in the analyses of the individual datasets. However, several nodes inside the sister group to the Mormyrus clade, which had been resolved in the MP analysis of the RAG2 data alone, are lost in this combined analysis.

To establish character weights for an a posteriori reweighted MP analysis of the combined data, we calculated the rescaled consistency index (RCI) of each of the 969 informative characters on the above 19 trees plus 13 other shortest trees produced from five weighted parsimony searches. Each of these employed different a priori down-weighting of transitions in the 12S/16S datasets and in the third-position sites of the cytochrome b dataset (see legend to Fig. 5). Each character was then assigned a weight equal to the highest RCI found for that character on one or more of these 32 trees: 184 of these characters received a weight of 1, 212 characters received a weight of 0 and 785 characters received a fractional weight between 0 and 1. The mean weight of all informative characters was 0.321, and the median weight was 0.20. A parsimony analysis on this a posteriori reweighted dataset yielded a single most parsimonious tree of 1179.61 steps (Fig. 5). We assigned nodes recovered in the a posteriori reweighted tree to three subjective confidence classes: high, moderate and low (indicated by red, green and blue, respectively, in Fig. 5) on the basis of their performance in bootstrap and decay analyses and their presence or absence in the five a priori weighted MP analyses and the equally weighted MP analysis. Specifically, a node was given a rating of high confidence if it had a bootstrap value on the a posteriori reweighted tree of 90 % or higher, a decay index of 2 or above, and if it was present in all most parsimonious trees recovered in the one unweighted and five a priori weighted parsimony analyses. Nodes with lesser performance were rated moderate or low confidence. We rank only six of the 40 nodes on this tree as low-confidence nodes. Only one of these weak nodes, node f, affects a hypothesis of a suprageneric relationship. We regard the relationships between Genyomyrus donnyi, Marcusenius moori and Marcusenius sp.1 to be uncertain with respect to the well-supported Campylomormyrus + Gnathonemus clade.

An alternative topology to one clade we classify as ‘moderate confidence’ based on a high bootstrap value and decay index appears in three of five a priori weighted MP analyses (analyses 3, 4 and 5 in Fig. 5) and deserves mention. In these analyses, node dd is absent, and the two species of Mormyrus form a clade with the Brienomyrus brachyistius and Isichthys henryi sister pair.

All suprageneric relationships depicted in the a posteriori reweighted tree of the combined data are compatible with those recovered above the 50 % bootstrap level in the MP analyses of the individual datasets.

Because of the remarkable diversity of electric organ discharge waveforms among mormyroids (Hopkins, 1980, 1981, 1986, 1999b; Hopkins and Bass, 1981) and the corresponding diversity of electric organs, we examined the evolution of the electric organ in mormyroids using a simplified version of our tree (see Fig. 7). We incorporated into this tree all high or moderate confidence suprageneric nodes, collapsing the low-confidence nodes e and f in Fig. 5. In addition, for the sake of simplicity, we collapsed clades of single genera that were found to be monomorphic with respect to electrocyte anatomy into single terminals (e.g. the clades of Petrocephalus, Mormyrus, Stomatorhinus and Pollimyrus species). Furthermore, we removed Mormyrops masuianus from this tree since its position within the Mormyrops clade was not strongly resolved and its absence would have no effect on the reconstruction.

We observed six types of electrocyte among the species included in this study. These data are summarized in Table 1. Type S electrocytes found in Gymnarchus niloticus are stalkless. There are two major categories of stalked electrocytes. In the first, stalks arising from the posterior face of the electrocyte are non-penetrating and receive innervation on the posterior side (non-penetrating stalk with posterior innervation, type NPp, shown in Fig. 6E for Petrocephalus bovei). All fish with electrocytes of this type produce simple biphasic electric organ discharge (EOD) waveforms. In the second category, the stalk penetrates through the electrocyte. Several different forms within this category are observed. In the simplest of these, the stalk penetrates through the electrocyte to the opposite side, where it receives innervation. Electrocytes with stalks arising from the posterior face that penetrate through to receive innervation are called type Pa (penetrating stalk with anterior innervation, shown in Fig. 6I 

for Myomyrus macrops). Those with stalks arising from the anterior face that penetrate through to receive innervation from the posterior face are called type Pp (penetrating stalks with posterior innervation). In Brienomyrus brachyistius and Brienomyrus niger, the stalk penetrates through the electrocyte a second time to receive innervation on the face from which it originated (doubly penetrating stalk with posterior innervation, type DPp). Species of only one genus, Pollimyrus, have both doubly penetrating and non-penetrating stalks (type DPNP). In contrast to Bass (1986a,c) and Alves-Gomes and Hopkins (1997), we find all species of Stomatorhinus have type Pa electrocytes. The EOD waveform produced by each electric organ morphotype is a function of the direction of current flow through the stalk and the relative order in which the anterior and posterior electrocyte faces depolarize (see Bass, 1986a,c). At the base of the mormyroid tree, we use the parsimony criterion in conjunction with available developmental data to choose a most likely scenario for electric organ evolution. The sister group to the mormyrids, Gymnarchus niloticus, has stalkless (S-type) electroctyes and an electric organ fundamentally different in structure and organization from the mormyrid adult electric organ (Dahlgren, 1914; Denizot et al., 1978, 1982; Kirschbaum, 1987). While the Gymnarchus niloticus organ and the mormyrid adult organ are clearly homologous at some level, we believe the most immediate homologue to the electric organ in Gymnarchus niloticus is the larval electric organ in mormyrids, which develops soon after hatching and degenerates as the adult organ develops (Denizot et al., 1978, 1982; Kirschbaum, 1987, 1995). Both are present in the medial part of the deep lateral muscle rostral to the caudal peduncle. In both, the electrocytes are arranged myotomically, with myofibrils present, and in both only the caudal electrocyte face is electrically active, so the EOD waveforms are monophasic. This is in contrast to the adult electric organ of mormyrids which is restricted to the caudal peduncle, in which the myotomic arrangement has been lost, myofibrils are largely absent, both electrocyte faces are electrically excitable and EOD waveforms have both positive and negative phases.

Given this, there exist two possible reconstructions for the electric organ at the base of the mormyroid tree (node 1, Fig. 7). First, the common ancestor of all mormyroids may have had an electric organ and electrocytes much like that of Gymnarchus niloticus. In this case, the developmentally separate adult electric organ present in living mormyrids would have evolved in the immediate common ancestor of mormyrids (node 2, Fig. 7). Alternatively, the common ancestor of all mormyroids could have possessed separate larval and adult electric organs, as do extant mormyrid taxa, and a subsequent paedomorphic loss of the adult organ took place in the lineage leading to Gymnarchus. We favor the first scenario for the evolution of electric organs. If the origins of the larval and adult electric organs are considered as two separate evolutionary steps, the first hypothesis represents the most parsimonious reconstruction.

In the subfamily Mormyrinae, one form of non-penetrating stalked electrocyte (NPp) and four forms of penetrating stalk electrocyte (Pa, Pp, DPp and DPNP) occur in extant taxa. All electrocytes examined from Petrocephalus species in the sister subfamily Petrocephalinae are of type NPp. In addition, ontogenetic study of the electrocytes in Brienomyrus brachyistius, which have penetrating stalk electrocytes, has demonstrated that the developing, but functional, electrocyte first passes through a stage identical to the NPp condition before penetrations develop (C. D. Hopkins, unpublished observations). Similar observations in Hyperopisus bebe and Mormyrops deliciosus were made by Szabo (1960). For this reason and because penetrating stalk electrocytes are absent in the basal-most lineage of mormyrids, we hypothesize that NPp is the primitive condition for the mormyrid electrocyte (node 2, Fig. 7). Above this node, MacClade generates two equally parsimonious reconstructions when the electrocyte is coded as a binary character with states corresponding to non-penetrating and penetrating electrocyte stalks (all varieties of penetrating stalk are treated as a single character state). Both reconstructions require eight steps. In the first of these, penetrating stalks originate four times from ancestors with non-penetrating stalks, and there are four reversals to the non-penetrating condition. In the second reconstruction (shown in Fig. 7 and equivalent to a reconstruction using the ACCTRAN algorithm), the penetrating stalks originate once in the common ancestor of the Mormyrinae, with seven reversals to non-penetrating stalks at the nodes numbered 5, 6, 8, 10, 11, 13 and 14. We favor this latter reconstruction since we regard multiple independent origins of a particular modification to an ancestral and ontogenetically antecedent morphology as less likely than a single origin followed by multiple paedomorphic reversals.

Superimposing the four forms of penetrating stalk electrocyte (Pa, Pp, DPp and DPNP) onto this reconstruction, additional patterns emerge. Electrocyte type DPNP is shown to have originated once in the genus Pollimyrus, while type DPp has evolved twice, in the distantly related Pa-type ancestors of Brienomyrus niger and in Brienomyrus brachyistius.

The remaining form of penetrating stalk electrocyte, type Pp, is known only from several species of Mormyrops (e.g. M. zanclirostris and M. masuianus in this study). This electrocyte is simply reversed in anterior/posterior polarity from type Pa. Within the genus Mormyrops, other species possess Pa-type electrocytes, including Mormyrops deliciosus (Gosse and Szabo, 1960) and Mormyrops curviceps (Moller and Brown, 1990), while still others have type NPp (e.g. M. nigricans in this study).

The only available higher-level hypothesis of mormyroid interrelationships has been that of Taverne, who organized the modern taxonomy of the group in the course of his osteological studies (Taverne, 1967, 1968, 1969, 1971b, 1972; Taverne and Géry, 1968). Taverne’s classification divided the group into two families, the Gymnarchidae, consisting of the monotypic genus Gymnarchus, and the Mormyridae, comprising the remaining genera. Within the Mormyridae, he recognized two subfamilies: the Petrocephalinae (genus Petrocephalus) and the Mormyrinae (the remaining genera). Taverne diagnosed the Mormyrinae by the loss of the basisphenoid bone. The results of our phylogenetic analysis of molecular data support these taxonomic divisions. Within the Mormyrinae, Taverne proposed a large subdivision diagnosed by the loss of the lateral ethmoid bone, although he did not assign this group any formal taxonomic category. These taxa are not recovered as a monophyletic group in our study, and we infer that the lateral ethmoid has in fact been lost several times within the Mormyrinae. The phylogenetic tree of Taverne (1972) depicts additional relationships within the Mormyrinae that are not explicitly supported by character data and many of which differ substantially from relationships recovered in our data analysis.

Only recently has mormyroid systematics begun to benefit from cladistic studies. Agnèse and Bigorne (1992) studied enzyme variability at 16 presumptive loci among 22 populations of 11 species belonging to five mormyrid genera. In one analysis, the presence and absence of alleles were coded as characters in a Wagner parsimony analysis. Phylogenetic conclusions from the unrooted network produced from these data were limited. However, despite the authors’ claims that their analysis ‘genetically characterizes’ the three species of the genus Marcusenius used, their network instead stipulates their paraphyly, regardless of the position of the root. Different taxonomic sampling in that study and the present study makes comparisons difficult, although our data also show the genus Marcusenius to be an artificial group.

Van der Bank and Kramer (1996) examined the relationships between eight mormyrid species using cladistic methods (among others) to analyze a dataset consisting of allozyme, morphological, behavioral and ecological characters and reported a sister-group relationship between genera Petrocephalus and Pollimyrus. Some of the character analysis used to create their dataset has since been criticized (Alves-Gomes, 1999; Alves-Gomes and Hopkins, 1997; Lavoué et al., 1999). In our analysis, species of Petrocephalus form the sister group to all other mormyrid species, and species of Pollimyrus are nested far within this latter group.

Alves-Gomes and Hopkins (1997) produced a phylogeny of seven mormyroid taxa plus two outgroups from maximum parsimony and maximum likelihood analysis of 932 bases from the mitochondrial 12S and 16S genes and presented a model for the evolution of the electric organ in these taxa. This study found Gymnarchus niloticus to be the sister taxon to the other included mormyroids, and one species of Petrocephalus to be the sister group to the other included mormyrids. Furthermore, the genus Brienomyrus was found to be non-monophyletic. These results are consistent with those from our data analysis. In the most inclusive study to date, Lavoué et al. (1999) sequenced a 588-base-pair portion of the cytochrome b gene to infer relationships among 27 mormyroid taxa, including representatives of all nominal genera with the exception of Isichthys, Stomatorhinus and Heteromyrus. Their results supported the monophyly of the Mormyridae (exclusive of the Gymnarchidae) and a sister-group relationship between species of Petrocephalus and all remaining mormyrids, consistent with Taverne’s classification and the findings of Alves-Gomes and Hopkins (1997). In addition, three clades within the Mormyridae were supported: (1) a clade including species of the genera Gnathonemus, Marcusenius (minus Marcusenius conicephalus), Hippopotamyrus, Genyomyrus and Campylomormyrus; (2) a sister-group relationship between the two monotypic genera Ivindomyrus and Boulengeromyrus; and (3) a clade consisting of two species of Brienomyrus from Gabon and Paramormyrops gabonensis to which Marcusenius conicephalus formed the sister group. In addition, these authors tentatively placed the genus Myomyrus as the sister group to the remaining Mormyrinae.

These results are consistent with those we report here.

Our analysis of the data presented here provides particularly strong resolution of basal nodes in the mormyroid tree: the position of Gymnarchus niloticus as the sister taxon to all other mormyroids, the position of Petrocephalus species as the sister group to all other mormyrids and the position of Myomyrus as the sister taxon to the remaining Mormyrinae. This position for Myomyrus, weakly supported by data presented by Lavoué et al. (1999), is supported with high confidence in our analysis, although it has never been suggested from morphological evidence. Above Myomyrus, there is very strong support for a sister-group relationship between the species of Mormyrops and the remaining Mormyrinae. Genera represented by more than a single species and that appear as monophyletic groups in our analysis are Petrocephalus, Mormyrops, Mormyrus, Pollimyrus, Stomatorhinus and Campylomormyrus. Nominal genera that do not appear as monophyletic groups in our analysis are Marcusenius, Hippopotamyrus and Brienomyrus. Lavoué et al. (1999) found the genus Marcusenius to be polyphyletic. In our larger study, both the genera Marcusenius and Hippopotamyrus are rendered polyphyletic. We think it is likely that these two genera have been defined with characters plesiomorphic for the larger clade, which includes Genyomyrus, Gnathonemus and Campylomormyrus. Marcusenius conicephalus, however, is found not belong to this clade, but instead is strongly supported as the sister taxon to a clade consisting of Brienomyrus species from Gabon and Paramormyrops gabonensis, also from Gabon. This result was also reported in Lavoué et al. (1999). To explain it, Lavoué et al. (1999) suggested that a disjunction between the mitochondrial phylogeny and the true species phylogeny could have resulted from hybridization and introgression of a foreign mitochondrial genome into an ancestor of M. conicephalus. However, additional support for this relationship from the independent nuclear RAG2 locus allows us to reject this hypothesis. Closer scrutiny of morphological characters in Marcusenius conicephalus is needed.

Another strong, but non-intuitive, result of this study, not reported elsewhere, is the sister-group relationship between Brienomyrus brachyistius and Isichthys henryi. While dissimilar in appearance, these taxa share nearly identical distributions throughout the coastal drainages of the Upper and Lower Guinea ichthyofaunal provinces of West and Central Africa (type 7 distribution; see Lévêque, 1997). A third species, unavailable for this study, but morphologically similar to Brienomyrus brachyistius and sharing this distribution, is Brienomyrus longianalis. This species probably represents a third member of this clade.

Lavoué et al. (1999) reported a sister-group relationship between Ivindomyrus opdenboschi and Pollimyrus marchei, both species from the Ivindo River of Gabon. We recovered identical sequences for all gene fragments from our specimen of Ivindomyrus opdenboschi and specimens tentatively identified as Pollimyrus marchei. Because of uncertainty in specimen identification, we decided to exclude all specimens of Pollimyrus marchei from this study.

Taverne (1971b) erected the genus Brienomyrus and designated Marcusenius brachyistius Gill 1863 the type species of the genus. Nominal species of this genus do not form a clade on our tree. Brienomyrus niger, which Taverne (1971a) placed into the separate subgenus Brevimyrus, appears at the base of the Marcusenius/Hippopotamyrus/Gnathonemus/ Campylomormyrus clade along with Hippopotamyrus pictus and Hyperopisus bebe. A third clade of nominal Brienomyrus species appears as the sister group to Marcusenius conicephalus in our study, as noted above. Paramormyrops gabonensis is nested within these species of Brienomyrus. Relationships among these and additional taxa belonging to this clade are the subject of additional study by the authors. All available data point to the Lower Guinea ichthyofaunal province and, in particular, the Ogooué River Basin as the center of diversity for this group. (Although not included here, other putative members of this clade occur in the Congo River Basin.) Two successive outgroups to this clade, Marcusenius conicephalus and the Boulengeromyrus knoepffleri/ Ivindomyrus opdenboschi sister pair, are endemic to the Ivindo River, an Ogooué tributary (and to the Ntem River, a separate coastal drainage whose headwaters co-mingle with those of the Ivindo). We take this combination of phylogenetic and distributional data to be evidence that this region is, in addition, the center of origin for the Brienomyrus/Paramormyrops clade. The results of our phylogenetic analysis of molecular data demonstrate a need for taxonomic revisionary work and renewed study of morphological characters within mormyroids with the aim of establishing genera that reflect natural groups. In addition, if the novel relationships within these fishes suggested here and by Lavoué et al. (1999) are accurate, these should be supported by morphological, as well as molecular, synapomorphies.

Our tree-based hypothesis (Fig. 7) suggests that penetrating stalked electrocytes evolved once early in the history of modern mormyrids and that the occurrence of electric organs with non-penetrating electrocytes within the Mormyrinae is the result of multiple independent reversals to the ancestral condition. To confirm that penetrating stalks have only evolved once in these fishes would require a more detailed analysis of phylogenetic relationships within genera that contain both species with penetrating and species with non-penetrating stalked electrocytes, such as Brienomyrus, Marcusenius, Mormyrops and Campylomormyrus. To date, our analysis has sampled too few species from each of these large clades to do so.

Our phylogenetic tree does allow us to formulate a likely scenario for the process of electric organ evolution in mormyrids, although we may only speculate on the adaptive significance of penetrating stalks or the reasons for the multiple independent reversals to non-penetrating stalked electrocytes. If we assume that the first mormyrids used their electric organ discharges both for communication and for electrolocation, we can imagine three important advantages for penetrating stalks over non-penetrating stalks. First, fish with penetrating stalks might be very effective in reducing the direct current (DC) components of their electric organ discharges. Bennett (1971) and Bennett and Grundfest (1961) and others have noted that fish with penetrating stalks all produce an initial head-negative phase in their discharges. This pre-pulse might facilitate balancing the positive and negative phases of the rest of the electric discharge, thereby eliminating undesirable DC components to the EOD waveform. By so doing, a fish would be able to generate an electrolocation signal that would not jam its own DC-sensitive ampullary receptors, which are used for passive sensing of weak low-frequency signals from external sources such as prey or predators. Reducing the DC component of the EOD would also reduce its electrical conspicuousness to catfish predators, which also carry DC-sensitive ampullary electroreceptors and might use this component of the discharge as a homing beacon for prey.

Second, fish with penetrating stalks might be able to produce EOD waveforms with higher spectral frequencies than fish without them. The penetrating stalk should extend the bandwidth of the spectrum of the EOD waveform to higher frequencies by adding an extra modulation into the waveform without significantly increasing the duration of the overall pulse. When combined with a tuned electroreceptor, such as the Knollenorgan, acting as a matched filter (see Bass and Hopkins, 1984; Hopkins, 1983), these fish might make use of the more private unused high-frequency electrical bandwidth for intraspecific communication, especially when several species of mormyrids are already in sympatry. This would be an advantage for extending the distance of electrical communication in the presence of jamming EODs from heterospecifics, because it would reduce the noise in the signal channel.

Third, fish with penetrating stalks might be able to produce complex EOD waveforms that could serve in species recognition and reproductive isolation. Complex EOD waveforms might be especially useful when several similarly discharging species occur in sympatry.

There must also have been reasons why selection favored reversal to non-penetrating stalked electrocytes in certain clades. We have noticed that a majority of the fish with EODs of very long duration (over 2 ms in duration) tend to have non-penetrating stalked electrocytes. Aside from the ancestral condition seen in Petrocephalus, in which all the known EODs are short in duration and rather simple in waveform, the longest-duration EODs reported for the more derived mormyroids are from Paramormyrops gabonensis (>10 ms), males of Isichthys henryi (>4 ms), males of Mormyrus rume (>20 ms) and Campylomormyrus numenius (>10 ms), all species with non-penetrating electrocytes. In clades of mormyrids in which penetrating stalked electrocytes are ubiquitous, a reversal to non-penetrating stalks may have allowed the evolution of long-duration EODs that might have been impossible with penetrating stalks.

Indeed, long-duration EODs may arise by the forces of sexual selection under the influence of female choice, as has been argued by Hopkins et al. (1990) for South American pulse gymnotiform fishes (Hopkins, 1999b). Since long-duration EODs may be more costly to produce compared with shorter EODs of the same amplitude, females may select mates that are capable of generating these costly discharges because they are reliable and ‘honest’ indicators of the overall health and quality of a potential mate. Thus, while penetrating stalks might have pushed the signal bandwidth to higher frequencies, non-penetrating stalks may allow the signal bandwidth to relax to lower frequencies, especially if the fish had already evolved the molecular machinery necessary both to slow the duration of the action potentials and to modulate the duration of the EOD seasonally under the influence of steroid hormones, as have many of the genera of mormyrids apart from Petrocephalus (for a review, see Bass, 1986a). The forces of sexual selection, operating on overall pulse duration, and the requirements for unique EOD waveforms for species recognition may have been factors in the reversal of the electric organ from penetrating stalked to non-penetrating stalked electrocytes in several of the more speciose clades of mormyroids.

This research was supported in part by grants to C.D.H. from the National Geographic Society (5801-96), the NSF (INT-9605176) and the NIMH (Grant MH37972). We thank Garry Harned for help with histology, Melanie Stiassny for helping fund J. Sullivan’s collecting trip to the Central African Republic. Paul Posso, IRET, Jean-Daniel Mbega and Jean Hervé, Mve IRAF, provided logistic help in Gabon. Emanuel Vreven, African Museum, Belgium, helped with field work.

S.L thanks the Museum National d’Histoire Naturelle for support for field work and a travel grant to Ithaca, New York. J. Snoeks provided mormyrid specimens from Lake Malawi. We gratefully acknowledge the Cornell University Evolutionary Genetics Core Facility for use of facilities and we thank K. Zamudio and G. Harned for comments on our manuscript.