THE ORIGIN OF SULAWESI TAXA
The divergence time estimates for Sulawesi taxa can be used to test vicariance hypotheses by contrasting the TMRCA with the temporal constraints on vicariance events derived from geology (Fig. 1). The TMRCA for a Sulawesi taxon must predate (but compare discussion below) or coincide with a vicariant event such as the opening of the Makassar Strait or the separation of the relevant areas of the Sula Spur from New Guinea (compare Fig. 1, red arrows labeled 1 and 2), otherwise a vicariance hypothesis is falsified (but see below for caveats). For the purpose of this discussion, the oldest time estimate has always been used, even though there is some evidence that the alternative rate may be the better choice, see discussion on the reliability on the molecular clock analyses below. Similarly, in case of multiple colonizations, the older TMRCA was used (see Fig 5D). Confidence intervals of divergence time estimates have been considered in all cases here, that is, the interpretation is not only based on the mean TMRCA (compare Table 1 and Figs. 2–4). By applying these constraints, a different pattern emerges for taxa of Asian and Australian origin, respectively.
All post-Eocene TMRCA estimates of Sulawesi taxa postdate the opening of the Makassar Strait about 45 Mya (see Introduction) and are thus not compatible with the assumption of a vicariance hypothesis for an origin of Sulawesi taxa from the West (Fig. 5A and Table 1). Three of the four studies on Sulawesi taxa of Asian origin that have considered a vicariance hypothesis involving the opening of the Makassar Strait have also found a post-Eocene TMRCA for the respective taxon (grasshoppers, Walton et al. 1997; ricefish, Takehana et al. 2005; cockroaches, Maekawa et al. 2001), and support for a vicariance scenario was solely based on a somewhat relaxed interpretation of the time frame for the separation of Borneo and Sulawesi. Trans-Makassar Strait vicariance is only supported for one of the four taxa for whom a respective hypothesis has been suggested, namely mite harvestmen (Clouse and Giribet 2010), as the TMRCA estimates of about 51 and 61 Mya found in our analyses for the two lineages on Sulawesi fit the geological time frame (Fig. 2 and Table 1). However, poor node support and low ESS values were found in both the strict and relaxed clock analyses, which might indicate general issues with the dataset as analyzed.
Vicariant “tectonic dispersal” hypotheses to account for the origin of taxa from the East on the Sula Spur have been proposed or at least discussed for three of the four Sulawesi taxa with Australian affinities considered here (pachychilid snails, Glaubrecht and Rintelen 2003; von Rintelen and Glaubrecht 2005; Köhler and Glaubrecht 2010; sailfin silversides, Sparks and Smith 2004; phalangerids, Ruedas and Morales 2005). The TMRCA for all three taxa either clearly predates (pachychilid snails) or rather closely matches (sailfin silversides, phalangerids) the beginning separation of the Sula Spur parts that collided and remained with Sulawesi from New Guinea-Australia about 15 Mya (compare Fig. 1, red arrow no. 2; Figs. 2–4 and Table 1). For the latter two taxa, this only holds true for one of the two substitution rates used for estimating divergence times, though, as the second rate yields a younger TMRCA in both cases (if confidence intervals are considered, the divergence time of sailfin silversides matches the timing of Sula-Australia split also assuming the faster rate). The TMRCA of the fourth Sulawesi taxon with closest relatives in Australia, the megapodes, clearly postdates the relevant vicariant events. However, the Sulawesi megapodes are probably the best dispersers among these taxa, because they are capable of flying at least short distances, depending on the taxon (Dekker 2007).
In summary, the divergence time estimates derived by application of a standardized molecular clock approach are not compatible with a vicariance hypothesis in 80% of the taxa analyzed and suggest dispersal as the predominant process in the origin of Sulawesi taxa. This pattern becomes even more pronounced if the faster of two external rates used for some datasets should turn out to be more appropriate for the respective taxon (Fig. 5B, here called rate 2; compare also discussion below). However, our data are prima facie also compatible with vicariance hypotheses in 20% of the analyzed taxa. This proportion is 75% for taxa of Australian origin/affinities, but this can hardly be regarded as representative given the low number of respective datasets included.
Vicariance hypotheses involving Sulawesi taxa are at odds with the current paleogeographic reconstructions of Wallacea, that is, the past distribution of land and sea during the Cenozoic. Recent updates of the respective models suggest that both in West Sulawesi and on the Sula Spur larger areas may have been subaerial since the Eocene than previously assumed (compare Hall 2009b; Lohman et al. 2011). There is as yet no evidence for the existence of a subaerial link between the Sula Spur and Australia-New Guinea during the critical time period from the Oligocene to the early Miocene, though. Such a link is an essential assumption for matching a vicariance hypothesis with the respective dated phylogenies, for example, for the pachychilid snails. However, as emphasized above, reconstructions of the past distribution of land based on geological data tend to be biased towards detecting larger areas of land (Hall 2009b), and it should also be noted that the intervals of 5 or 10 My for which reconstructions are available for the last 40 My are rather coarse. We thus suggest not to reject a vicariance hypothesis for explaining the origin of Sulawesi taxa merely because it is not consistent with paleogeographic reconstructions, which are subject to at least as much uncertainty as the biological data. It seems fair to assume “bottlenecks” in the area of available land both in Southwest Sulawesi and on the Sula Spur, though, under a vicariance scenario. Interestingly, the two taxa where there seems to be least temporal uncertainty about the compatibility with a vicariance scenario, the mite harvestmen and pachychilid snails, show an extreme time lag between their respective TMRCA and the first speciation event on Sulawesi (Fig. 5A). This pattern matches the assumption of the respective lineage persisting on very little land during the Miocene both in Southwest Sulawesi at around 20 Mya and on the Sula Spur until about 20 Mya, which would presumably constrain the potential for speciation in both taxa by limiting population size, habitat diversity, and the possibility for allopatry. However, a long time lag between the TMRCA and speciation on Sulawesi might also be, for example, caused by insufficient taxonomic coverage or extinction. A role of the latter is possibly supported by the lower TMRCA found for the older Sulawesi lineage of mite harvestmen assuming a BD process, which also considers extinction in node age estimation. Nevertheless, a sampling bias in terms of taxa and area coverage is quite likely as well for a number of the invertebrates considered here, given the low number of sample sites on Sulawesi, but also other areas in some cases, for example, the mite harvestmen (compare Clouse and Giribet 2010; Fig. 1).
The earliest speciation event on Sulawesi in any taxon included in our study dates from the early Miocene about 20 Mya (Fig. 5A). This might be linked to two geological/climatological events: (1) the availability of more land on Sulawesi from about that time as a consequence of the collision of the Sula Spur with the North and West arms of the island (Hall 2009b; Spakman and Hall 2010; compare Fig. 1) and (2) the climate-induced start of periodic sea-level lowstands in the Mid-Miocene with inverse peaks in the late Miocene and Pliocene (see Haq et al. 1987; De Graciansky et al. 1998; Fig. 5C: pink arrow). Both factors would not only increase the potential of speciation on Sulawesi by providing, for example, opportunities for creating new ecological niches or allopatric differentiation, they would also increase the likelihood for chance dispersal, for example, across Makassar Strait being successful by providing a larger “target” at a lower distance from the source area. A link between Miocene colonization of Sulawesi and sea-level changes was already proposed for squirrels by Mercer and Roth (2003). The effect of sea-level lowstand(s) on decreasing the maximum open water distances between Sulawesi and the potential source areas would probably have been greatest from the Pliocene onwards (Hall 2009b). The splits of Sulawesi taxa from their sister groups and the earliest speciation events on Sulawesi are particularly densely clustered in the Pliocene/Pleistocene. If this is not merely a sampling artifact, it would support the assumption of an important role for sea-level lowstands coupled with the increasing extent of land area on Sulawesi for dispersal onto the island.
Several taxa have colonized Sulawesi more than once (see Table 1), and the time between independent colonization events ranges from almost instantly in geological terms (≤ 0.5 My) to about 10 My (Fig. 5E). The timing of the two colonization events is not correlated to the two geological/climatological events discussed above, that is, both colonization events in each taxon occurred either in the Miocene or Pliocene. The water beetles (black bars in Fig. 5E) are an exception here, but this might be an effect of substitution rate issues (see below) rather than indicative of very differently timed colonizations of Sulawesi. There is also no clear pattern of exclusive distribution of the two Sulawesi lineages in the respective taxa, which might be taken as indicative of the colonization of different potentially separated paleoislands of Sulawesi. In some taxa, different colonizations resulted in a largely or completely overlapping distribution of the Sulawesi clades, for example, in fanged frogs (Evans et al. 2003a) or shrews (Ruedi et al. 1998), whereas the two lineages of mite harvestmen, for example, are found in Southwest and North Sulawesi, respectively (Clouse and Giribet 2010). However, here as well as in some other cases (Ruedi et al. 1998), the geographic coverage of sampling on Sulawesi is poor. In taxa such as the macaques or the wild boar, the different Sulawesi lineages do not occur sympatrically, but it is doubtful whether monophyly of the Sulawesi taxa can be ruled out, either because there is insufficient support of the respective nodes (wild boars, Larson et al. 2005) or because nuclear data are in conflict with the mtDNA gene trees on this point (macaques, Tosi et al. 2003; Evans et al. 2010). Not only is there no obvious respective pattern in the distribution of lineages that have colonized Sulawesi independently, there is in our opinion also little hard evidence for any Sulawesi clade being confined to part of the island only, which might suggest a link of its origin on Sulawesi and tectonic processes. For most invertebrates, the poor sampling prevents any respective statement, and for vertebrates we are not aware of any recent example (and the fossil record is rather limited, see van den Bergh et al. 2001). Michaux (2010) has proposed that Sulawesi should not be employed as an area of endemism in biographic analyses in its entirety, but that rather parts of different geological or tectonic origin such as West Sulawesi or the Banggai-Sula block should be considered separately. This is not supported by our data.
However, subsequent to colonization of the island, geology, or climate-related changes in paleogeography seem to have played an important role in driving diversification processes on Sulawesi. A pronounced and at least partly congruent geographic structure of Sulawesi populations or species has been found in all taxa that have been sufficiently extensively sampled on the island (see, e.g., Bridle et al. 2001; Evans et al. 2003c, 2008; Larson et al. 2005; Brown et al. 2010; T. von Rintelen et al. unpubl. data). Although not the focus of this study, it would be highly interesting to test for temporal congruence in intra-Sulawesi diversification as well.
The datasets analyzed in this study comprise (mostly nonflying) terrestrial and limnic taxa. The distribution of TMRCA estimates of limnic taxa does not deviate from the general pattern (Fig. 5D), suggesting that freshwater organisms are not necessarily suffering “unique biogeographic constraints” (Unmack 2001). It rather seems that the capability for dispersal varies as much as in terrestrial taxa, and the limnic candidate taxa for a vicariant origin on Sulawesi, for example, comprise taxa absolutely confined to freshwater such as viviparous snails (the pachychilid Tylomelania, von Rintelen and Glaubrecht 2005) and a group of secondary freshwater fish with brackish water species (telmatherinids, Sparks and Smith 2004).
RELIABILITY OF DIVERGENCE TIME ESTIMATES
The validity of our results and their interpretation in terms of biogeographic hypotheses is strongly dependent on the quality of the divergence time estimates. This is not an exclusive issue here, as the usage of substitution rates or fossil calibrations for divergence time estimations is crucial in all molecular clock analyses (e.g., Graur and Martin 2004; Pulquério and Nichols 2007; Ho et al. 2008; Wilke et al. 2009; Wertheim and Sanderson 2010). Fossil record data might be imprecise and only give minimum estimates for the age of a group (e.g., Benton and Donoghue 2007; Donoghue and Benton 2007); the application of substitution rates derived from nonclosely related organisms and/or for different genes is fraught with difficulties (Thomas et al. 2006; Wilke et al. 2009). The latter point is of particular relevance here, because only a subset (n= 10) of the datasets in this study could be analyzed using a taxon-specific substitution rate or calibration point, respectively (see Table 1 for details and references). The majority of analyses were conducted using a rather unspecific substitution rate (compare Table 1), for example, for insects (general mtDNA rate), mammals (general mtDNA rate), fish (ND2 rate for cyprinodont fish), and frogs/toads (ND2 rate for leptodactylid frogs).
Precise calibration points and specific substitution rates are strong priors for the estimation of divergence times and if a large error in either calibration method is unrecognized this makes the resulting divergence time estimates unreliable to an unknown degree (Pulquério and Nichols 2007). Our study provides striking examples of the issues involved. For the ricefish, for example, Takehana et al. 2005 suggested a TMRCA of 29–32 Mya for the Sulawesi lineage compared to the 3–7 Mya estimated here. This discrepancy is largely due to their use of a substitution rate for Antarctic notothenioid fish (Bargelloni et al. 2000), which is an order of magnitude slower than the rates usually reported for vertebrates (Near 2004) and does seem an unusual choice for small tropical freshwater fish. Divergence time estimates also almost always differed considerably if different genes were analyzed for the same taxon (Fig. 5F and Table 1). The water beetle dataset is rather illustrative here, comprising data from three different gene fragments, which were both independently and jointly (see below) analyzed here (datasets 2, 9, 22) using the same rate (“general” insect mtDNA rate of 2.3%/My; Brower 1994). This approach resulted in extremely varying divergence times for the same Sulawesi taxon (the older of the two lineages found) from its sister group ranging from the Oligocene to the Pleistocene for the individual genes (datasets 2, 9, 22, summarized in Fig. 5F). In a recent study by Pons et al. (2010), beetle-specific rates for protein-coding mitochondrial genes have been estimated using a calibration derived from fossils. The rates for some individual genes differ considerably from the 2.3% average of the general insect mtDNA rate. For cytochrome c oxidase I (COI, dataset 9, Fig. 2), the specific rate is more than eightfold higher than the general insect rate and for cytochrome b the rate is more than 30% faster than the general rate (Cyt b, dataset 2, Fig. 2). More specific rates thus do not necessarily result in better congruence among the divergence times estimated from different genes (see Table 1). Using the new rates, the age of the studied water beetle lineages does decrease considerably for COI, shifting the TMRCA from the Mid-Miocene to the Pleistocene (Figs. 2–4 and Table 1). Pons et al. (2010) also show that the average rates between beetle taxa vary considerably, indicating that a beetle rate may be little better than a general insect rate for estimating divergence times in some lower rank taxa of beetles, which is perhaps not totally surprising in this old and megadiverse “superradiation” (Hunt et al. 2007). However, other recent studies have demonstrated a good fit between divergence time estimates derived from multiple calibration points and a general rate (Jønsson et al. 2011). In our study, the results from some vertebrate datasets suggest that the use of unspecific rates is not necessarily misleading. For the squirrels (datasets 12, 13), for example, the calibration point used by Mercer and Roth (2003) was employed and the resulting estimated clock rates (Table 2) approximately fit the faster of two published rates for mammals. These were used for all other mammal datasets in this study, such as the tarsiers (dataset 3). Here, the slower mammalian mtDNA rate 1 seems more appropriate, with the mean estimated TMRCA of 8.3 My (see Table 2) roughly (considering confidence intervals) fitting the split from T. bancanus about 11 Mya derived from fossil data (Merker et al. 2009), whereas the faster rate (rate 2 in our study) yields a much younger TMRCA. The substitution rates in primates have long been known to be lower than those of other mammals, particularly rodents (Li et al. 1987) and among primates the lemurs, which are comparatively closely related to tarsiers, have the lowest rates (Hasegawa et al. 1990), so these differences in rate applicability between mammal datasets should perhaps be expected. The application of external, that is, non-taxon specific rates in the absence of fossil calibrations is not necessarily misleading, but as it is hard to assess the reliability without external reference, the results must be interpreted with extreme caution.
With the exception of the mite harvestmen, datasets comprising multiple gene fragments were not only analyzed separately for each gene but also as a concatenated alignment using one rate (fanged frogs, macaques, ricefish, water beetles). As expected, the resulting TMRCA was intermediate to the estimates gained from the separate analyses (see Table 1). For the macaques and, to a lesser degree, also the water beetles, the TMRCAs from the separate analyses differ strongly and the TMRCA estimated from the concatenated dataset is very far from the mean of the two single gene TMRCAs, even though the same external rate was used in all analyses (Table 1). We suspect that this might be an effect of length differences between the individual gene fragments. For the macaques, for example, the TMRCA of the combined analysis is 18.6 Mya, which is quite close to the 19.5 Mya estimated for the NADH dataset alone, which has a length of 1325 bp and very different from the 3.7 Mya for the 12S dataset with only 413 bp length.
Although it has been suggested that a strict clock must be employed in molecular dating analyses using external rates (e.g., Wilke et al. 2009), the latter can also be employed under relaxed clock models as implemented in BEAST (Drummond and Rambaut 2007). The effect of the choice of molecular clock model (strict vs. relaxed clock) on the estimated TMRCAs seems secondary, though, if compared to the differences stemming from applying different substitution rates to the same gene or the same rate for different genes (see results; compare Tables 1 and S2). Interestingly, the differences between node age estimates are not linked to the rejection of a strict clock by the Bayes factor analyses in both approaches (compare Tables 2 and S2); for the tarsiers, for example, a strict clock was not rejected by the Bayes factor values of the DHM analysis or was the Bayes factor (BEAST) remarkably high (Table 2). The choice of clock model clearly matters and a strict clock model should not be used when rejected by Bayes factors, but we suggest that the choice of substitution rate and calibration is generally more important with respect to the robustness of divergence time estimates. Remarkably, the biogeographic implications of our analyses remain valid under each model or rate used.
BIOGEOGRAPHIC INTERPRETATION OF MOLECULAR DIVERGENCE TIME ESTIMATES
Our study essentially relies on comparing divergence times of Sulawesi taxa with vicariant events derived from geology (Fig. 1). This approach has lead to a “resurrection of oceanic dispersal” (de Queiroz 2005; Cowie and Holland 2006; Heaney 2007) in historical biogeography by using divergence time estimates to falsify vicariance hypotheses by demonstrating that the phylogenetic splits in question postdate a vicariance event (e.g., Knapp et al. 2005; Heinicke et al. 2007; Michalak et al. 2010; Schweizer et al. 2010; Bartish et al. 2011). Crisp et al. (2011) have recently argued to use calibrated trees in biogeography in a more explicitly hypothesis-driven way instead of an inductive “pattern first, process later” procedure of fitting some geological data to the tree. Overall, our study provides ample evidence in justification of their argument. Crisp et al. (2011: Box 1) have also suggested that a vicariance scenario is only supported by the data if the timing of respective phylogenetic splits and their confidence is overlapping with the time frame of the vicariant events. Although we acknowledge that this approach is rigorous and objective, we would suggest that a TMRCA predating a vicariant event, as in the case of the mite harvestmen (Fig. 2, dataset 1) in this study, does not automatically falsify an involvement of vicariance in causing the phylogenetic split. The mite harvestmen are probably representative for many groups of small tropical invertebrates in having a high likelihood of being both taxonomically and geographically undersampled (compare Clouse and Giribet 2010; Fig. 1). Extinction could also cause a time lag between the split recovered in a phylogeny of extant taxa and the vicariant event. It might be argued that the risk of extinction is particularly high in Wallacea and adjacent regions where ongoing tectonic processes entail frequent local submergence and uplift events.
Although stem age estimates are widely used to test vicariance hypotheses in biogeography (see Crisp et al. 2011 and above), this approach has recently been heavily criticized (Heads, 2005, 2009, 2011; Nelson and Ladiges 2009), particularly the use of divergence time estimates as maximum ages for phylogenetic splits. If a molecular clock has been calibrated using fossils or external rates ultimately derived from fossil calibrations, any resulting age estimate will be a minimum age (e.g., Donoghue and Benton 2007). Using this age estimate as a de facto maximum age that is then employed in testing congruence between the age of a vicariant event and a specific phylogenetic split can indeed create problems if the TMRCA of a lineage is much older than the fossil used to define its age or if the difference between the age geological vicariant event and the phylogenetic split is small (though confidence intervals in a properly conducted molecular clock analysis would then probably overlap with the time range assigned to the vicariant event). Heads (2009, 2011) suggested to rely on tectonic calibrations and use well-established vicariance events instead of fossils for fixing node age. We consider this proposal even more problematic than the issue it tries to overcome, mainly for two reasons: (1) age estimates for vicariant events can be imprecise as well, as also acknowledged by Heads (2011), for example, for the Panama Isthmus, and, more importantly, (2) using them would either lead to circular reasoning if the aim of a study is to test for a role of vicariant events in the evolution of a taxon (see also Kodandaramaiah 2011) or imply the a priori acceptance of a vicariance model, which would make biogeography less of a science and more of a religion. Although naive dispersal concepts have been rightly criticized (e.g., Croizat et al. 1974), which paved the way for vicariance biogeography (e.g., Wiley 1988), it seems futile to deny a role of either dispersal (long-distance dispersal) and vicariance in biogeography (for details see Thornton 1983).
Sulawesi has a fauna of mixed geographic and temporal origin. Given the patterns discussed here, it is little surprising that Wallace had difficulties in understanding the biogeography of this “anomalous island” (Wallace 1876). Dispersal seems the primary mechanism of bringing taxa to the island and a standardized molecular clock approach has let to the falsification of vicariance hypotheses for some Sulawesi taxa of Asian origin. The divergence time estimates presented here are compatible with a vicariance hypothesis in 10–20% of the studied taxa, though, depending on the choice of substitution rate in two cases. A vicariance scenario is dependent on accurate tectonic and paleogeographical data, but evidence for vicariance can at the same time inform geology and, for example, increase the probability for assuming land in certain parts of Wallacea or suggest where geological evidence for this might be sought. Sulawesi has a rich fauna (and flora), and we have used but a few taxa for which the respective data were available for this review of the biogeography of the island. Despite the methodological problems inherent to a molecular clock approach and the by necessity poor taxon coverage, we believe our results are generally robust. However, more data are needed to gain a truly comprehensive understanding of Sulawesi's biogeography, including intra-island diversification. Biogeographers in the pre-molecular and precladistic age compiled many fascinating distribution patterns involving Sulawesi, and the simplified “origin from Asia” or “origin from Australia” dichotomy this study is largely confined to will surely become more complex. Recently, for example, the dating of phylogenetic splits in two genera of nymphalid butterflies has revealed sister groups of the Sulawesi lineages from Timor and the Moluccas, respectively (Müller and Beheregaray 2010; Müller et al. 2010). The time frame of Sulawesi colonization for the butterflies (Miocene) and speciation on the island (Miocene?/Pliocene) is fully compatible with the data presented here, though, and suggests that Sulawesi's fauna was decisively shaped by events in the Miocene and Pliocene. The need for new data also applies to geology. Biogeographers must acknowledge the limitations of tectonic or paleogeographic reconstructions (Hall 2009a), but they should also use them explicitly to derive temporal constraints for biogeographic hypotheses of vicariance, as for example, exemplarily done for the Madagascan fauna by Warren et al. (2010). This will only work in conjunction with molecular divergence time estimates. Again, these are not the Holy Grail but biological data with a confidence margin and should be treated accordingly, but ignoring them with a vague reference to paleoendemics (Michaux 2010) is not an option in our opinion.