Ethics Statement

DNA and tissue samples were obtained from pre-existing collections as donations from The Australian Centre for Ancient DNA (University of Adelaide), The Research School of Biology (Australian National University), The Department of Environment and Conservation, Western Australia and as loans from The Australian National Wildlife Collection, Canberra – in each case with permission from the relevant authorities within these institutions. One additional frozen tissue sample was purchased from a local butcher (EcoMeats) in Canberra. No live animals were sampled and none of the DNA/tissue collection or handling procedures required either approval or a permit from a review board or ethics committee. Sequences published previously by other groups were obtained from GenBank.

Taxon sampling and DNA sequencing

In order to reconstruct a mitochondrial tree for kangaroos and wallabies we targeted three protein-coding genes, NADH1, NADH2 and Cytb along with the 12S and 16S ribosomal RNA genes. Taxon sampling focused on ten Macropus species and the monotypic Wallabia bicolor. The affinities of the three extant Macropus species not included here are uncontroversial [6], [8], [48] and add little to the sampled diversity. M. parma and M. parryi are nested within M. (Notamacropus) and M. bernardus groups with the other wallaroos within M. (Osphranter). Outgroup sampling includes the macropodids, Lagorchestes (Lagor. hirsutus and Lagor. conspicillatus), Dendrolagini (Petrogale xanthopus and Dendrolagus dorianus) and Lagostrophus fasciatus, in addition to the potoroids, Aepyprymnus rufescens and Potorous tridactylus.

Mitochondrial DNA was sequenced from DNA previously extracted at University of Adelaide (A. rufescens, D. dorianus, P. xanthopus, Lagor. conspicillatus, W. bicolor and M. fuliginosus) and Australian National University (M. eugenii and M. rufogriseus). For the remaining new sequences, DNA was extracted from tissue samples. These were provided by collections at Murdoch University (M. rufus), The Department of Environment and Conservation, WA (M. irma) and The Australian National Wildlife Collection (M. dorsalis) or purchased from EcoMeats in Canberra (M. giganteus). In addition, we sequenced two nuclear genes (IRBP and ApoB) from M. irma and W. bicolor to validate the provenance of our samples, given that their mtDNA placements differed from Meredith et al. [7]. Using our IRBP and ApoB sequences in place of Meredith et al. 's varied the maximum likelihood bootstrap support on the nuclear data for the placements of M. irma and W. bicolor by <1.5%. As a default however, we preferentially use the previously available W. bicolor sequences, which derive from the same individual as each of the other nuclear loci. For M. irma we use our IRBP and ApoB sequences, which cover a 28 bp sequencing gap and resolve for six ambiguity codes in the previously available sequences.

DNA extraction for M. rufus and M. irma was carried out at Murdoch University using a Qiagen DNeasy kit (Qiagen Sciences, MD, USA) and at Australian National University for all other taxa, using the salting out method (following [49]). DNA was amplified using standard PCR protocols on a Corbet Research iPAQ thermocycler (NSW, Australia). Primers and amplification conditions are provided in Table S1. All amplicons were sequenced by Macrogen (Seoul, South Korea). Nuclear sequences have been submitted to GenBank for M. irma (IRBP; JN967008, ApoB; JN967009) and for W. bicolor (IRBP; KC429577, ApoB; KC429578). GenBank accession details for the mt sequences are provided in Table S2.

Data matrices

The primary mitochondrial dataset (Mt₁₆) combines the NADH1, NADH2 and Cytb protein-coding genes with the 12S and 16S rRNA genes. The sequences were initially aligned in ClustalW2 [50] with penalties of 5 for gap opening and 0.2 for gap extension. Manual adjustments were then made in Se-Al 2.0a [51], where sites with ambiguous homology were excluded, leaving a 5,593 bp mtDNA alignment. An expanded matrix (Mt₁₇) includes M. dorsalis; although poor tissue preservation resulted in low DNA yields and the sequence is 72% complete, missing 147 bp of NADH2, all of Cytb and 269 bp of the rRNA genes. The only mt sequence available on GenBank for M. antilopinus was included in a 1,146 bp Cytb₁₈ matrix (with the taxa from Mt₁₆ and an additional W. bicolor sequence) and was sufficient to confidently place M. antilopinus as sister to M. robustus, in agreement with nuclear genes (Figure 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Phylogenetic analysis of kangaroos and wallabies.

Maximum likelihood phylogenies inferred from the (A) mitochondrial (Mt₁₆) and (B) nuclear (Nuc₁₇) concatenated datasets, with RAxML bootstrap values (BP_ML) above branches and MrBayes Bayesian posterior probabilities (BPP) below branches. The mt placement of M. dorsalis is derived from the reduced-length Mt₁₇ and the mt placements of M. antilopinus and W. bicolor (NSW, New South Wales) are derived from the Cytb₁₈ alignment. Support for grouping M. eugenii and M. agilis increases (BP_ML = 88; BPP = 0.98) for Mt₁₆, which excludes M. dorsalis, but increases sequence length. Asterisks indicate full support. Clades including members of Macropus are shaded.

https://doi.org/10.1371/journal.pone.0057745.g002

Previous research has demonstrated a requirement to ameliorate nucleotide compositional biases among marsupial mt genomes for phylogenetic inference of ordinal level relationships [23], [24]. By contrast, the present focus on closely related genera is relatively shallow. Our composition homogeneity χ² testing on Mt₁₆ in PAUP* 4.0b10 [52] with uninformative and gapped sites excluded offers little evidence for base compositional non-stationarity among the Macropus and Wallabia ingroup (protein 1^st codons: P = 0.4222, 2^nd codons: P = 0.9708, 3^rd codons: P = 0.5810, RNA stems: P = 0.9941, RNA loops: P = 0.3400).

We analyse the mt sequences alongside the nuclear dataset of Meredith et al. [7], which includes protein-coding segments from BRCA1 (exon 11, breast and ovarian cancer susceptibility gene), ApoB (exon 26, Apolipoprotein B), IRBP (exon 1, interphotoreceptor retinoid binding protein gene), RAG1 (intronless recombination activating gene-1) and vWF (exon 28; vonWillebrand factor gene). Aligning the nuclear sequences followed the procedure described above for the mtDNA. Two 5,988bp matrices were constructed, Nuc₁₆, with taxon sampling matching Mt₁₆ and also Nuc₁₇, which further includes M. antilopinus. Combined analyses (MtNuc₁₆) concatenated the Mt₁₆ and Nuc₁₆ matrices.

Phylogenetic inference of mitochondrial and nuclear gene trees

Kangaroo phylogeny was inferred under maximum likelihood (ML) and Bayesian inference from the mitochondrial and nuclear sequences separately and concatenated, as well as for the individual nuclear genes. Substitution model categories for each data partition employed the more general of the jModelTest 0.1.1 [53] hLRT or AIC recommendations (Table S3) or the next most general available for each phylogenetic inference program. Substitution was modelled separately among the mt protein-coding codon positions and RNA stem and loop sites.

Our initial efforts to reconstruct kangaroo phylogeny employed Bayesian inference in MrBayes 3.1.2 [54] and ML in RAxML vGUI093 [55]. MrBayes analyses ran two independent sets of two MCMC chains for 6,000,000 (Nuc₁₇, MtNuc₁₆) or 4,000,000 (Mt₁₆, Mt₁₇, Cytb₁₈ and individual nuclear genes) generations, with trees sampled every 2,500 generations. Burn-ins varied from 500,000 to 1,200,000 generations, and were chosen to ensure that –lnL had plateaued, clade frequencies had converged between runs and estimated sample sizes for substitution parameters were >200 (using Tracer v1.5 [56]). ML analyses in RAxML carried out 500 full bootstrap replicates. Branch-length multipliers and substitution models were partitioned among protein codons and RNA stems and loops for each of the ML and Bayesian analyses, with MtNuc₁₆ further partitioned between mt and nuclear sites.

Support among alternative topologies was further examined with the approximately unbiased (AU) test [57], using the RELL method (100,000 replications) within CONSEL [58]. Site likelihoods employed in CONSEL were inferred in PAUP*, with all substitution parameters and branch-lengths ML optimized separately for each of the protein codon and RNA structural partitions, for each tree hypothesis. Maximum likelihood trees conforming to the alternative Mt₁₆ and Nuc₁₇ placements of W. bicolor and M. irma (Figure 2) were identified for each gene in PAUP* with 20 random addition heuristic searches. Support among these individual genes for the alternative placements was compared with SH tests [59], which as pairwise comparisons reduce to equivalency with AU and KH [60] tests. ML bootstrapping (500 replicates) for each gene was also performed in PAUP* with the substitution model parameters estimated in the earlier heuristic searches.

We estimated a mitochondrial timescale for kangaroo evolution using BEAST v.1.6.1 [61] with Mt₁₆ partitioned as per the phylogenetic analyses. An uncorrelated relaxed clock model was used with rates among branches distributed according to a lognormal distribution. Note that likelihood ratio tests in PAUP* rejected strict clocks for both the mt and nuclear sequences (P<0.01). Four independent runs totalling 40,000,000 MCMC generations ensured estimated sample size values >100 (as estimated in Tracer v1.5) for all node height, prior, posterior, −lnL, tree, and substitution parameters. Chains were sampled every 5,000 th generation after burn-ins of 1,000,000 generations.

Four fossil-based priors were used to calibrate the BEAST analysis. (i) Potoroidae/Macropodidae (15.97–28.4 Ma), with the minimum based on the Early Miocene macropodid, Ganguroo [2] and the maximum covering putative Late Oligocene macropodids [62] and potoroids [63]. (ii) Macropodidae (11.6–23 Ma), with the Middle Miocene macropodid, Wanburoo [2] providing the minimum bound and the maximum acknowledging that earliest Miocene macropods fall outside of the macropodid crown. Ganguroo is also a candidate for calibrating Macropodidae, however, its placement within this crown clade is not well resolved [2]. (iii) Dendrolagini (4.46–16.0 Ma), with the minimum based on the Hamilton fauna Dendrolagus [64] and the maximum bound recognising that all middle Miocene macropods fall outside the Dendrolagini crown. (v) Macropus/Lagorchestes (4.46–16.0 Ma), with the minimum based on Hamilton fauna Macropus [2] and the maximum bound recognising that all middle Miocene macropods fall outside of this crown clade.

Palaeontological data do not clearly favour any particular timing within the given bounds for calibrations (i), (iii) and (iv) and hence, flat priors were employed. A normal prior was employed for calibration (ii), in line with the recommendation of Ho and Phillips [65] for when the balance of evidence [2], [7], [62] suggests crown divergences fall well within the bounds. The normal prior was applied conservatively (90% of prior probability between the bounds).

Partition homogeneity testing

To identify incongruence between partitions we performed the incongruence length difference test [66] in PAUP*. This test however, can be biased in cases where parsimony is statistically inconsistent [67]. To overcome this concern we also perform likelihood-based parametric bootstrap tests. For these we infer one ML score (ML_F) with branch lengths and the models (as shown in Table S3) partitioned across genes, but assuming a single topology (T*) and another ML score (ML_V) for which the topology is allowed to vary across genes. The difference between ML_F and ML_V provides a critical value for testing the null hypothesis that all genes evolved on the same phylogeny. Next we used Seq-Gen 1.3.2 [68] to simulate 200 datasets partitioned into the original mtDNA and five nuclear gene sequence lengths and evolved on topology T* with the original branch lengths and model parameters for each gene. The distribution of ML_F - ML_V differences from the 200 simulated datasets was then compared with the critical value from the original dataset.

Species tree reconstruction

*BEAST analysis [27] within BEAST v1.6.1 employed the multi-species coalescent to infer the species tree underlying the mt and five nuclear gene trees coestimated from MtNuc₁₆. The mtDNA was further partitioned into protein codon positions and RNA stems and loops for substitution modelling. Separate mt and nuclear uncorrelated lognormal relaxed clock models were used, with a Yule species tree prior and differential ploidy (autosomal nuclear and haploid mitochondrial). Eight independent runs totalling 80,000,000 MCMC generations ensured estimated sample size values >100 (estimated in Tracer v1.5) for all node height, prior, posterior, −lnL, tree, and substitution parameters. Chains were sampled every 5,000 th generation after burn-ins of between 1,000,000 and 4,000,000 generations. Given our focus on phylogeny rather than dating, we did not calibrate the *BEAST analysis, and so avoid potentially misleading influences of calibration priors on clade posterior probabilities [65]. Instead we provided a nominal mean substitution rate of 1.00 with unspecified time units for the nuclear data and allowed the mt rate to vary relative to this.

Minimizing deep coalescences, MDC [28] trees were inferred under the dynamic programming mode in PhyloNet 2.4 [69] from the mtDNA and five nuclear gene trees (Figure 2A, Figure S1), which were each estimated under ML bootstrapping in PAUP*. We collapsed branches that received <50% bootstrap support in these source trees to reduce the influence of stochastic artefacts among the individual genes on MDC tree building.

Bayesian concordance analysis within BUCKy [29] was run on 500 bootstrap replicate trees inferred in RAxML, for each of the six loci. Bootstrap distributions typically reflect stochastic variation in the gene tree estimates more closely than do Bayesian posterior distributions [24], [70]. Otherwise, BUCKy analyses employed default parameters, except where stated.

Coalescent simulation

In order to better understand whether ILS could plausibly account for incongruence among gene trees we simulated the evolution of the five nuclear genes and the mtDNA under a coalescent process in MCcoal [71] within BPP 2.1 [72]. Two alternative guide trees were used for MCcoal, the combined data and nuclear-only *BEAST species trees. In the former case M. irma was excluded from the *BEAST analysis because of the concern that its mtDNA derives from introgression (see Discussion), which violates the assumptions of *BEAST. Instead, M. irma was grafted onto the tree – its temporal placement along the stem lineage from the other Notamacropus members was scaled in proportion to the nuclear-only *BEAST tree. For comparability the combined data and nuclear-only guide trees were both scaled to a root height of 20 Ma, closely matching both the mt BEAST estimate (21.3 Ma) and Meredith et al. 's [7] estimates from the concatenated nuclear genes (ave. BEAST estimate, 20.0 Ma).

The model that MCcoal simulates under (JC+Γ) is less complex than the models selected in jModelTest for each locus. Therefore we used a two-step simulation process (illustrated in Figure 3A). First, MCcoal was run on the species guide tree to provide simulated coalescent trees. Gene sequences were then simulated on these coalescent trees under their respective substitution models (Table S3) in Seq-Gen. All model parameters were estimated from the original data and the simulations maintained the aligned sequence length for the mtDNA and each nuclear gene.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Macropodid clade support from datasets simulated under coalescence.

(A) Simulation workflow. (B) Mean number of the five nuclear genes supporting each clade in maximum likelihood analyses of 200 simulations of the combined data *BEAST species tree for N_e values of 1,000 (triangle), 10,000 (open circle), 100,000 (square) and 1,000,000 (filled circle). For comparison, the grey bars show the number of genes supporting each clade on the observed data. (C) Percentage of ML analyses supporting each clade among 200 mtDNA simulations on the nuclear-only *BEAST species tree for N_e values set to mitochondrial equivalency for the same populations (one quarter of the corresponding nuclear values). Abbreviations: Lagor.; Lagorchestes, Wall.; Wallabia, M. (Notamac.); M. (Notamacropus), M. (Osphran.); M. (Osphranter).

https://doi.org/10.1371/journal.pone.0057745.g003

MCcoal requires a population dynamics parameter θ = 4N_eμ (2N_eμ for mtDNA), where N_e is the effective population size and μ is the mutation rate per site per generation. Mutation rates per site per year were obtained by scaling PAUP* ML treelengths for each locus to the *BEAST timetree length. Generation time across macropods is not well studied, but we used an average of 7 years in consideration of life history data from most macropodid species [73]. The influence of effective population size was evaluated with N_e varied from 1,000 to 1,000,000.

Phylogenetic inference from separate mitochondrial and nuclear sequences

Our analyses of the mt and nuclear sequences agree on grouping Macropus and Wallabia to the exclusion of the consecutive outgroups, Lagorchestes, Dendrolagini, Lagostrophus and Potoroidae (Figure 2). The inclusion of the mtDNA greatly enhanced phylogenetic resolution. Whereas four clades received 36–70% ML bootstrap support on the nuclear data alone, on the combined data all but one clade received ≥90% ML bootstrap support (Figure 4A). Only the relative affinities of the three Macropus sub-genera (Macropus, Notamacropus and Osphranter) remained poorly resolved. However, the combined result hides conflict between the mt (Figure 2A) and nuclear (Figure 2B) trees for the placements of M. irma and W. bicolor.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Macropodid species tree estimates from the combined mitochondrial and nuclear sequences (MtNuc₁₆).

(A) concatenated sequences, showing BPP/BP_ML (@ = 0.89/59), (B) *BEAST partitioned between the mtDNA and five nuclear genes, showing BPP values. (C) both MDC and BUCKy, which recovered the same tree from the mt and five nuclear gene trees. (D) Meredith et al. (2008) with BP_ML values included for comparison. Several supraspecific clades that were identical across all reconstructions were collapsed for visualization convenience. Relationships within each of the collapsed clades were as inferred in Figure 2. Asterisks indicate full BPP or BP_ML support.

https://doi.org/10.1371/journal.pone.0057745.g004

The nuclear data favours M. irma and W. bicolor as consecutive sister groups to the wallabies we refer to as core-Notamacropus, which here includes M. rufogriseus, M. eugenii, M. agilis and M. dorsalis. Placing M. irma with core-Notamacropus receives high ML bootstrap support (BP_ML, 93%) and Bayesian posterior probability (BPP, 1.00). Further expanding this clade to include Wallabia is only supported modestly (BP_ML 58%, BPP 0.89). These results closely mirror Meredith et al. [7]. Our mitochondrial trees strongly conflict with these placements, instead favouring M. irma as sister to the wallaroos (M. robustus, M. antilopinus) (BP_ML 88%, BPP 1.00) and placing W. bicolor outside a monophyletic Macropus (BP_ML 100%, BPP 1.00).

Maximum likelihood AU testing reveals strong incongruence between the nuclear and mt data for the placements of both W. bicolor and M. irma. Table 1A shows that for Mt₁₆ the favoured nuclear placement for W. bicolor as sister to the subgenus M. (Notamacropus) is rejected at P = 0.008 and reciprocally, AU testing on Nuc₁₇ rejects the favoured mt placement for W. bicolor as sister to all Macropus at P = 0.011. Similarly for M. irma (Table 1B), the favoured mt placement as sister to M. robustus is rejected with the nuclear data (P<0.001) and reciprocally, the favoured nuclear placement as sister to core-Notamacropus is rejected with the mtDNA (P = 0.031).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Approximately unbiased (AU) test results.

https://doi.org/10.1371/journal.pone.0057745.t001

Turning to the individual nuclear genes, support for the favoured mt versus nuclear placements reveals distinctly different patterns for M. irma and W. bicolor (Table 2). All of the nuclear genes favour an M. irma relationship with M. (Notamacropus) over the mt relationship with M. (Osphranter) – except BRCA1, for which both of these relationships were equally likely. In contrast, the overall nuclear placement of W. bicolor as sister to M. (Notamacropus) is only favoured over the mt placement by BRCA1 and vWF. Another gene (IRBP) instead favours the mt placement of W. bicolor outside Macropus, while the ML analyses for ApoB and RAG1 find the nuclear and mt hypotheses for W. bicolor affinities to be equally likely.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Individual nuclear gene –lnL differences and SH test results.

https://doi.org/10.1371/journal.pone.0057745.t002

Our analyses of Mt₁₇ show that M. dorsalis groups with M. eugenii and M. agilis (BP_ML = 81%, BPP = 1.00; Figure 2A), with the latter two wallabies favoured as sister taxa (BP_ML = 63%, BPP = 0.88). AU testing (Table 1E) echoes these results, favouring M. dorsalis as sister to M. eugenii and M. agilis, although with other placements of M. dorsalis within core-Notamacropus rejected only at modest significance levels (P values from 0.079–0.334).

Kangaroo species tree inference

Partition homogeneity testing performed in PAUP* identified significant incongruence between the mt and nuclear datasets (P = 0.027) and between the five nuclear genes (P = 0.003). Nuclear gene trees are shown in Figure S1. These parsimony-based results are in agreement with the likelihood-based parametric bootstrap test. For the latter, the improvement in likelihood of partitioning over concatenation for the mt and nuclear sequences (MtNuc₁₆) and among the five nuclear genes (Nuc₁₆) was 30.44 and 82.23 –lnL units respectively. In both cases these critical values fall higher than the distribution of likelihood improvements from partitioning for each of the 200 simulated datasets, therefore rejecting homogeneity at P<0.005.

We employed four approaches to inferring the kangaroo species tree from the mtDNA and five nuclear genes. First the data were concatenated, with substitution models and relative rates partitioned between mt and nuclear sequences and within these, between the protein-coding codon positions and RNA stems and loops. The concatenated MtNuc₁₆ ML and Bayesian analyses provide a well resolved tree (Figure 4A) that combines the mitochondrial placement of Wallabia as sister to Macropus, with relationships among the Macropus species following the nuclear tree.

The second approach using MtNuc₁₆ applied the multi-species coalescent within *BEAST to allow for ILS among the mtDNA and the five nuclear genes. As shown in Figure 4B *BEAST reconstructed the same topology as the concatenated analysis, except with M. irma as sister to M. (Osphranter) rather than core-Notamacropus. We also ran the *BEAST analysis without the putative introgressive hybrid, M. irma. Macropus monophyly was retained (BPP = 0.99), although among the subgenera, M. (Osphanter) grouped (at BPP = 0.86) with M. (Macropus), instead of with M. (Notamacropus).

Among the two other species tree approaches, BUCKy carries out Bayesian concordance analysis, which requires a prior level of discordance (α) to be assigned. We ran separate analyses with α at 0.5, 1, 5 and 10, which provide for a range of prior expectations for the 6 loci representing one or two distinct trees up to representing five or six distinct trees. The same concordance tree was recovered under each of these levels and shares the same topology with both the concatenated analysis (Figure 4A) and the MDC tree (Figure 4C). On the nuclear data alone, each of the species tree methods followed the concatenated nuclear tree (Figure 2B), in placing W. bicolor with M. (Notamacropus). However, all methods used to combine the mt and nuclear sequences support Macropus monophyly.

Coalescent simulations

There is agreement between the mtDNA and the majority of nuclear genes on deep and shallow clades with stem lineages ≥1.5 Ma on the combined species tree. Almost all of the discordance arises among a tight cluster of six consecutive divergences covering 4.2–4.9 Ma (see Table 3B,C for divergences) from the Macropodini/Dendrolagini divergence up to the M. (Notamacropus) crown divergence. In fact most gene discordance within this cluster (Figure S1) and the lowest bootstrap support (<60%, Figure 2B) involves the sequential divergence of Lagorchestes, Wallabia and each of the three Macropus subgenera covering just 2.0–2.4 Ma (Table 3; Figure 1B, grey strip). Partition homogeneity testing and AU testing (Tables 1,2) suggest that the extent of the incongruence cannot be explained by stochastic error alone. However, another explanation, incomplete lineage sorting is also consistent with the association between short stem length and incongruence.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Macropodoid divergence time estimates in millions of years before present.

https://doi.org/10.1371/journal.pone.0057745.t003

Simulating the coalescent process over the combined data species tree allows inference of whether ILS provides a plausible explanation for the strong incongruence between the nuclear genes. In Figure 3B the grey bars show the number of nuclear genes supporting each of five short-stem-lineage clades within the diversification cluster. Phylogenetic analyses of the simulated datasets provide estimates for the probability of each gene supporting a given clade for alternative values of N_e. The sum of these values over the five genes is the mean expectation for the number of genes supporting each clade. Coalescent time is very short with N_e = 1,000 and as a result the coalescent simulations overestimate the number of genes supporting each of the five clades. Under this scenario ILS appears to be a poor explanation for the lack of concordance among the nuclear genes. Increasing N_e to 10,000 makes little difference.

The greater potential for ILS with N_e = 100,000 provides for a remarkably close match to the observed gene support among the clades (Figure 3B). The sum of squares difference between expected and observed support for the five clades decreases from 13.25 and 12.20 for N_e = 1,000 and 10,000 respectively, to 0.76 for N_e = 100,000. This improvement does not continue with N_e being further increased to 1,000,000 (sum of squares = 10.16). Coalescent simulations with such high N_e overestimate the extent of incongruence, with no genes expected to support any of the clades in 80% of the simulated datasets.

MCcoal simulations were also run with the nuclear-only species tree providing the guide phylogeny. Here the aim was to determine whether ILS can potentially explain the mtDNA supporting Macropus monophyly or placing M. irma distantly from other members of M. (Notamacropus). Effective population size was set to mitochondrial equivalency for the same populations (one quarter of the corresponding nuclear values). All simulated mtDNA sequences favoured the Wallabia/M. (Notamacropus) and M. (Notamacropus) groupings for each of the three lower N_e values (250, 2,500, 25,000). At N_e = 250,000 these fell to 46% and 66% respectively, although support among the simulated datasets for Macropus monophyly remained at 0% and only increased to 1.5% for M. irma falling within any grouping outside core-Notomacropus that is at least as shallow as M. (Osphranter) – where M. irma was placed on the observed mtDNA. Hence, the mitochondrial placements of Wallabia and M. irma are difficult to reconcile with ILS.

Mitochondrial sequences provide confirmation and incongruence

Molecular studies have consistently shown that Lagorchestes, Wallabia and the Macropus subgenera M. (Macropus), M. (Osphranter) and M. (Notamacropus) diverged from each other in rapid succession. We estimate that together, their consecutive divergences cover a temporal window of little more than 2 million years (Figure 1B, Table 3), in agreement with Meredith et al. [7]. The short internal branches may provide low phylogenetic resolution due to stochasticity associated with few substitutions along branches and conflicting signals attributable to incomplete lineage sorting (ILS) among genes. This expectation was borne out for the nuclear dataset, with all relationships among the five groups poorly resolved (Figure 2B, BP_ML from 36–59%).

Adding the mitochondrial (mt) sequences to the nuclear dataset substantially enhances resolution (Figure 4A). All groupings on the tree receive ≥90% BP_ML and are consistent with the supertree (Figure 1A) modified from Cardillo et al. [6], with the exception of a near-trichotomy among the Macropus subgenera. As strong as these results are, there is significant incongruence between the five individual nuclear genes (Figure S1). Moreover, mtDNA discordance with the combined nuclear sequences (see Figure 2) necessitates caution, especially for inferrig the affinities of the swamp wallaby (W. bicolor) and the black-gloved wallaby (M. irma).

The relationship of Wallabia to Macropus

All concatenated and species tree analyses of the combined mtDNA and nuclear genes recover W. bicolor as sister to Macropus (Figure 4), consistent with recent morphological analyses [2], [12], [13]. If we accept this relationship, it implies that the nuclear (Nuc₁₇) placement of W. bicolor within Macropus is an artefact, potentially of ILS among the individual genes. This interpretation is consistent with extreme incongruence among the nuclear genes on both MP and ML-based partition homogeneity tests (P≤0.005). Moreover, support for the overall nuclear placement with M. (Notamacropus) derives from only BRCA1 and vWF, while IRBP favours the mtDNA placement as sister to Macropus (Table 2A).

We examined the incongruence further with coalescent simulations and show that with small populations the models overestimate support among the nuclear genes for clades within the diversification cluster. However, simulating deeper ILS consistent with N_e = 100,000 on the species tree closely matches observed support for these clades (Figure 3B). Neaves et al. [74] recently estimated N_e of similar magnitude for M. fuliginosus. If indeed W. bicolor does fall outside Macropus, then it might appear anomalous that only IRBP among the five nuclear genes favours Macropus monophyly. This result however, matches the coalescent simulations on the species tree under the best fitting N_e; for 74% of the simulated datasets Macropus monophyly was recovered for only one (or none) of the nuclear genes. In contrast, none of the corresponding mtDNA coalescent simulations (Figure 3C) on the nuclear-only species tree favour Macropus monophyly. Hence, the observed mtDNA support for Macropus monophyly is unlikely to be an artefact of incomplete mitochondrial lineage sorting.

Strong mtDNA support for Macropus monophyly and apparently extensive ILS among nuclear loci within the diversification cluster around the base of Macropus caution against Meredith et al. 's [7] recommendation to subsume Wallabia within Macropus. Nevertheless, statistical support among our species tree analyses for excluding W. bicolor from Macropus is not conclusive. Indeed, coalescent simulations (Figure S2) on the species tree suggest that 30 or more nuclear loci may be required to confidently resolve relationships among the diversification cluster. However, considering our present phylogenetic results alongside the distinct browsing ecology and associated morphology/behaviour of W. bicolor [2], [4], [14], [75] and its unique 2n = 10(♀)/11(♂) karyotype, we believe that Wallabia currently warrants separate generic status.

Deep mitochondrial introgression in Macropus irma

Incongruence between the mt and nuclear placements for the black-gloved wallaby (M. irma) differs in several respects from that concerning W. bicolor. It is the nuclear placement of M. irma with M. (Notamacropus) that concurs with morphology and none of the nuclear genes prefer the mt placement with M. (Osphranter) (Table 2). A further point of difference is that the concatenated (MtNuc₁₆) analysis and two species tree reconstructions (MDC and BUCKy) support the nuclear placement for M. irma.

The mt placement of M. irma with wallaroos is difficult to reconcile with incomplete mitochondrial lineage sorting, being nested within M. (Osphranter) and because the path-length to its species tree position at the base of M. (Notamacropus) is 3–8 million years. Mitochondrial introgression may provide a more plausible explanation. Methods are being developed to directly test for introgressive hybridization, although these are not feasible without many loci or strong priors on the probability of hybridization [33], [34]. Nevertheless, our coalescent simulations (Figure 3C) further suggest that the aberrant mt placement of M. irma is not an artefact of incomplete lineage sorting. Even with the most extreme deep coalescence (ILS) scenario fewer than 2% of the simulated mtDNA datasets favoured M. irma falling outside its species tree grouping and into clades at least as shallow as M. (Osphranter). Hence, on the available evidence the more likely explanation is that M. irma obtained its mt genome from introgressive hybridization with an ancestor of the wallaroos, the deepest such event yet hypothesised among marsupials.

Previous examples of hybridization or introgression among wild macropodids (from Petrogale and Macropus) involve closely related parent species [32], [76]. The absence of evidence for introgression among more distantly related macropodids may reflect sparse sampling. Certainly, captive-bred hybrids include more distantly related Macropus crosses as well as Macropus×Wallabia and reports of Macropus×Thylogale [77], [78]. The results of Neaves et al. [33] may also be relevant here. Despite finding evidence for introgression in 17 of 223 grey kangaroos in the M. giganteus/M. fuliginosus sympatry zone, no F1 individuals were identified. The authors interpreted this result as suggesting a role for selection in accelerating introgression of some loci into the gene pool at well beyond the actual rate of hybridization.

Adaptive introgression of nuclear loci, as proposed for the grey kangaroos has also been suggested for mitochondria in wild goats [79] and Hares [80]. It is also possible that introgression of M. (Osphranter) genes into M. irma has adaptive significance and may not be limited to mtDNA. Most notably, Milne and O'Higgins [81] found that skull shape principle components that were correlated with vegetation cropping and mastication grouped M. irma within M. (Osphranter), thus matching the mtDNA and further suggesting adaptive convergence. This is consistent with Christensen [82] regarding M. irma as somewhat transitional in diet and locomotion between M. (Notamacropus) wallabies and the larger wallaroos and kangaroos of M. (Osphranter) and M. (Macropus). In particular, M. irma favours more open habitats [82] and may rely on grazing and poorer quality plant material more than most of its wallaby relatives [83].

If W. bicolor is sister to Macropus, then sharing deep gene coalescences with the more ecologically similar M. (Notamacropus) wallabies rather than with the larger, grazing M. (Macropus) and M. (Osphranter) might also point towards adaptive significance. The influence of selection on patterns of ILS and introgression is poorly understood at present and clarification of our intimations concerning M. irma and W. bicolor requires more thorough genomic sampling and analysis of functional correlations.

Variation among species tree inferences

Each of the MtNuc₁₆ concatenated and species tree analyses recovered the same kangaroo phylogeny (Figure 4A–C), except for *BEAST placing M. irma as sister to M. (Osphranter), close to the mtDNA relationship. One shortfall in the present usage of *BEAST is that species were represented by only one individual and therefore N_e could only be estimated for internal branches. It is not clear however that this should have any specific impact on the placement of M. irma. Indeed, running the analysis without the sequence data indicates that the placement of this taxon was not attributable to any aspect of the tree prior. Furthermore, using single individuals did not promote any other topological differences from the concatenated tree. Instead, the *BEAST result is consistent with mitochondrial introgression being the source of the incongruence concerning M. irma. This violates the assumption of *BEAST that ILS is the only source of incongruence. In contrast, another species tree method, BUCKy does not assume any particular source of incongruence and recovered the expected placement of M. irma as sister to core-Notamacropus.

It is interesting that MDC recovered the expected placement for M. irma, despite also assuming that incongruence derives solely from ILS. The explanation may lie in MDC being a consensus method, such that no matter how strong the signal from the mtDNA, it will be overwhelmed by consistent signal among multiple independent nuclear loci. *BEAST is not a consensus method. Instead it models the multi-species coalescent to co-infer gene trees embedded in a species tree, which as Heled and Drummond [27] explain, effectively provides a “reverse auction”, where the lowest bidder can set the limit. Consequently, *BEAST is a powerful tool for identifying true species relationships and divergence times when incongruence derives from ILS. However, the reverse auction might often leave *BEAST less robust than consensus methods to introgression or paralogy. Nevertheless, we believe that multi-species coalescent methods such as *BEAST are an important advance for phylogenetics. Allowing for limited post-speciation gene flow [84] will improve their reliability and will provide a valuable test for distinguishing incongruence from ILS.

Our partition homogeneity testing and analysis of coalescent simulations on the species tree (Figure 3B) are consistent with widespread ILS across the cluster of six rapidly diverged lineages, Dendrolagini, Lagorchestes, Wallabia and the three Macropus subgenera. The apparent introgression of M. (Osphranter) mtDNA into M. irma is the sole instance of significant mitochondrial discord with the species tree. These patterns of incongruence, although too few to draw strong conclusions on, nevertheless fit the expectations set out earlier. Specifically, that ILS will be more common among nuclear loci, consistent with longer coalescence times than for mtDNA, which in turn will be more susceptible to introgressive selective sweeps associated with fitness differences across populations, promoted by Muller's ratchet.

Overall our results suggest that sampling multiple mt genes is well suited to providing a first estimate for species-level phylogenies among marsupials and in combination with the five nuclear genes, substantially enhances phylogenetic resolution. Moreover, concerns that mt signal from three-fold as many parsimony-informative characters would swamp nuclear signal were unfounded. Combined analyses generally favoured the nuclear placements for M. irma and among the Macropus subgenera over their mt placements. The combined data only favoured the mt placement of Wallabia, for which the nuclear loci themselves were incongruent and contradicted morphology. However, larger scale nuclear genomic sampling will ultimately provide a more comprehensive understanding of evolutionary history, including for whether selective advantages contribute to patterns of ILS and introgression in W. bicolor and M. irma.