A major limitation of this study, with respect to phylogenetic implications, is the sparse taxonomic sampling. As in most studies, there is a trade-off in terms of taxa and genes or proteins. In this case, we have emphasized a large number of proteins to increase the statistical resolution of relationships and time estimates at the expense of taxonomic representation. Nonetheless, our results agree with most previous molecular phylogenetic analyses in supporting a close relationship between insects and crustaceans (Pancrustacea) and between myriapods and chelicerates. Of the two groups, Pancrustacea has received the strongest support in the past 2021222324. Nonetheless, a myriapod-chelicerate grouping has been found previously with mitochondrial genes 232425, nuclear ribosomal genes 21, and nuclear protein-coding genes 26. Additional evidence has come from hemocyanine structure 27.

We propose the name Myriochelata (in allusion to the joining of Myriapoda and Chelicerata) for the group containing myriapods and chelicerates, which otherwise is unnamed 28. Although we are unable to identify any morphological trait diagnostic of this clade, some trends are evident that might reflect the morphological or ecological nature of the ancestral myriochelatan. For example, many species of extant myriochelatans (e.g., spiders, scorpions, centipedes) inject and immobilize prey with a poison, albeit with structures that are not homologous. Envenomation of prey is also found among pancrustaceans, but it is less broadly distributed in that group. Certainly, envenomation has arisen multiple times in arthropods, associated mostly (but not exclusively) with terrestrial predation. The significance of this trait in arthropod evolution must await sequence evidence from a greater diversity of taxa (e.g., pycnogonids, remipedes) than is currently available, and a careful examination of the early fossil record of animals (especially from the Cambrian). In general, the difficulty in finding morphological characters diagnostic of these major clades of arthropods is probably the result of deep branching of the lineages and an early fossil record that shows great morphological diversification (the Cambrian Explosion) and some important gaps 262930.

Among animals, arthropods have been considered to be the earliest colonizers of land based on fossil evidence 567. However, it is possible that other animal phyla colonized land even earlier. Among them, nematodes, tardigrades, and annelids are likely candidates given their current exploitation of terrestrial environments, yet these groups have relatively poor fossil records.

Determining the number of such colonization events requires a consideration of phylogeny, the fossil record, and morphological traits associated with terrestriality. For arthropods, at least four major colonization events are inferred, leading to the chilopods + diplopods, insects, arachnids, and isopod crustaceans. Additional events may have occurred in smaller lineages 242531. Moreover, the recent discovery of a basal marine hexapod fossil from the Devonian 32 suggests that some hexapod traits previously believed to have evolved as adaptations to land may have first appeared in a marine setting. Along the same lines, it is also possible that millipedes and centipedes colonized land independently. However, because the earliest fossils of those groups are presumably terrestrial 6 and our molecular time estimate is only 5% earlier than the age of those fossils, our assumption of a terrestrial common ancestor of millipedes and centipedes has little affect on the time of colonization.

Our relatively young time estimates for the millipede-centipede (442 ± 50 Ma) and xiphosuran-arachnid (475 ± 53 Ma) divergences contrast with the much older time estimate for the deuterostome-arthropod divergence (~1100-900 Ma) (Tables 1 and 2). The first two are close to the corresponding fossil record estimates whereas the third one is ~400 million years earlier than the earliest fossil evidence for animals 12. Most molecular clock analyses in the last three decades, including those using many genes, have resulted in similarly deep time estimates for the arthropod-deuterostome divergence 33. Currently a debate exists as to whether divergences among animals are best represented by molecular clock dates or the fossil record 3334353637. The molecular clock dates suggest that early animals fossilized poorly, possibly because they were small and soft-bodied. Alternatively, others have argued that molecular time estimates are older than the true times because of statistical biases, rate changes, and calibration biases 3638, although replies to those criticisms have been made 37.

The results of this study address several of these criticisms of molecular clocks in the following ways. (1) The time estimates reported here are not uniformly discordant with the fossil record suggesting that if statistical biases are present, they are not causing a directional and proportional bias in all time estimates. (2) Our time estimate for the deuterostome-arthropod divergence using the arthropod fossil record (exclusively) was similar to time estimates obtained in previous studies by using the vertebrate fossil record, suggesting that the vertebrate calibration is not obviously biasing time estimates. (3) Our use of tunicates instead of vertebrates for representing deuterostomes resulted in similar time estimates, further indicating that vertebrates (per se) are not causing a systematic bias in the time analysis. (4) Our use of a diversity of clock methodology, including Bayesian inference and likelihood-smoothing methods, did not alter the conclusions, indicating that the global clock methods used in previous studies were not responsible for biased time estimates. (5) The use of concatenated alignments (supergenes) yielded similar results to non-concatenated (multigene) analyses, indicating that the multigene analyses of previous studies were not responsible for the discordance between molecular clocks and the fossil record.

In summary, our molecular clock analysis resulted in some time estimates (e.g. millipede-centipede) in agreement with the fossil record and others (e.g., insect-crustacean) much earlier than fossil evidence (assuming that the crustaceans are, in fact, monophyletic). However, all studies, including this one, are limited by the small size of the available sequence data (5 – 10 proteins across major groups of arthropods). In the future, it should be possible to use several hundred proteins across a diversity of arthropod taxa, and such an analysis should greatly increase the precision of phylogenetic and molecular clock analyses.

We sequenced the protein-coding region of two nuclear glycolytic enzymes, glyceraldehyde 3-phosphate dehydrogenase (G3PDH) and enolase, in an annelid (Nereis macrydi), a mollusk (Marisa sp.; not sequenced for enolase), and eight species of arthropods: a centipede (Lithobius sp.), millipede (Diplopoda sp.), blue crab (Callinectes sapidus), brine shrimp (Artemia sp.), water flea (Daphnia magna; not sequenced in G3PDH), muscle shrimp (Ostracoda sp.; possibly Cypridopsis sp.), Atlantic horseshoe crab (Limulus polyphemus), tarantula (Phormictopus sp.), and a scorpion (Centruroides sp.). Reverse-transcriptase PCR was used to amplify and sequence (both complementary strands) a total of approximately 1 100 base pairs (~363 amino acids) of the coding region of enolase and 890 base pairs (~294 amino acids) of the coding region of G3PDH. Additional sequence data were obtained from the public databases (GenBank). Genes were selected if there were nucleotide sequences available from an annelid (outgroup) and at least one representative of the following arthropod taxa: Branchiopoda (e.g., brine shrimp), Malacostraca (e.g., blue crabs), Diplopoda (e.g., millipedes), Chilopoda (e.g., centipedes), Insecta (e.g., insects), Xiphosura (e.g., horseshoe crabs), and Arachnida (e.g., spiders and scorpions). Sequences were aligned with Clustal X 39. Additional methods and sequence accession numbers are presented elsewhere (Additional file 1).

The alignments of all nine nuclear and 15 mitochondrial genes were concatenated (Additional file 1) because individual gene data sets were, for the most part, insufficient for statistical resolution of arthropod phylogeny (Additional file 2). This required the use of exemplars (representatives) from each of the major groups. Preliminary analyses using 16 – 54 taxa (Additional file 1) were used as a guide to choosing exemplars so that fast- or slow-evolving species were not selected.

Analyses of the concatenation of the nine nuclear genes, as well as a full data set of 24 genes, were carried out using minimum evolution, maximum parsimony, maximum likelihood, and Bayesian inference. Minimum evolution analyses were performed using Kimura two-parameter distances with transversions-only, gamma corrected Kimura two-parameters distances with transversions-only, and paralinear distances. For the paralinear distance analyses, the proportion of invariable sites in the considered alignment was estimated using likelihood and assuming a HKY85 + gamma + proportion of invariable sites model of DNA evolution. Identical sites were then removed proportionally to the base frequencies estimated from all sites (PAUP default settings). Maximum parsimony analyses were carried out with an equal weighting and a 2:1 weighting scheme favoring transversions over transitions. The model used for the maximum likelihood analyses (GTR + gamma+ proportion of invariable sites) was selected using Modeltest 19, whereas Bayesian inference was carried out under mixed models, and full parameter estimation was performed during tree search for each gene.

All minimum evolution analyses were performed using MEGA 2.1 40 and PAUP 41, while likelihood and parsimony analyses, as well as the estimation of the specific gamma parameters for the minimum evolution analyses were carried out using PAUP. Bayesian analyses were performed with MrBayes 3.0 42. The branch and bound algorithm was used in the parsimony analyses, while in the likelihood analyses 100 heuristic searches were performed. In the latter, starting trees were obtained using random sequence addition and swapped using the tree bisection reconnection algorithm. In the Bayesian analysis, 300 000 generations were run and trees were used only after convergence was reached.

Robustness for the nodes in the minimum evolution, maximum parsimony, and maximum likelihood trees was evaluated using the bootstrap. PAUP settings for the bootstrap analyses were as follows: 2 000 replicates with full heuristic search (one random addition sequence) and the multiple tree option turned off. Trees were swapped using the TBR algorithm. Support for the groups recovered in the Bayesian analyses were expressed as their posterior probabilities. Gapped sites were removed prior to all analyses.

Molecular clock analyses were performed using a diversity of methods 37: multigene global least squares (MGG_LS) 4344, multigene local least squares (MGL_LS) 45, supergene global least squares (SGG_LS) 46, supergene local least squares (SGL_LS), supergene local divtime (SGL_DT; DivTime5b) 47, supergene local multidivtime (SGL_MDT; Multidivtime) 48, and supergene local penalized likelihood (SGL_PL; r8s version 1.5) 4950. Least squares methods are distance based, SGL_PL is a semi-parametric likelihood-smoothing method, and SGL_DT and SGL_MDT are Bayesian methods. Multigene methods use the mean or mode 374451 of time estimates from individual proteins whereas supergene methods derive a single time estimate from the simultaneous analysis of all available proteins. SGL_DT and SGL_MDT are two different implementations of the Bayesian method of Thorne et al. 47. The difference between the two being that with SGL_DT protein sequences are concatenated and considered as a single entity, whereas with SGL_MDT each considered protein maintains its individuality during the analyses.

We used the divergence of centipedes (Chilopoda) and millipedes (Diplopoda) as the minimal (most recent) time of arthropod terrestrialization and the divergence of myriapods with their closest relatives (in this case, chelicerates) as the maximal (earliest) time for myriapod terrestrialization. The maximal time for arthropod terrestrialization was considered as the divergence between insects (Insecta) and crustaceans (Crustacea). This does not preclude the possibility that land was colonized even earlier by extinct species or groups of arthropods (or other animals).

Divergence times were estimated using two calibration points from the arthropod fossil record (Additional file 1): Xiphosura-Arachnida (~480 Ma) 52 and Chilopoda-Diplopoda (423 Ma; see Additional File 1 for explanation of date) 5354. A third calibration point, the divergence of deuterostomes (vertebrates) and arthropods based on a molecular clock study using 50 proteins and calibrated with the vertebrate fossil record (993 Ma) 33, was used in some analyses and tested (with the arthropod fossil record and by substituting vertebrates with a tunicate, Ciona intestinalis) in others (see below). Analyses were performed using calibration points separately and simultaneously, for comparison. The millipede-centipede fossil divergence was not used to estimate the molecular clock time of that divergence.

GenBank was screened for nuclear encoded proteins (>100 amino acids) in which divergence times could be estimated between two or more arthropod groups, resulting in 84 proteins. For each, reciprocal BLAST analyses were used to identify and assemble sets of available sequences, and orthology was investigated through phylogenetic analyses. In those analyses, 61 proteins (27 291 amino acid positions) were selected for further analysis after determining that orthologous relationships (divergences corresponding to speciation events) could be distinguished from paralogous relationships (divergences corresponding to gene duplication events).

The least squares-based global and local clock methods (MGG_LS, SGG_LS, MGL_LS, SGL_LS) were performed on 49 proteins (Additional file 1) that passed the relative rate test 55 as implemented in PHYLTEST 56. Gamma-corrected Poisson distances were used for these analyses and specific gamma parameters were estimated with PAML 57. The Bayesian and the likelihood-smoothing local clock methods (SGL_DT, SGL_MDT, and SGL_PL) were used with all 61 proteins that passed or did not pass the rate tests. Branch lengths for the penalized likelihood analyses were estimated with PAML (assuming a Gamma corrected JTT model of amino acid substitution). For the Bayesian analyses, branch length estimation was performed using the software AUTOestbranches 47 in the case of SGL_DT, and Estbranches 48 in the case of SGL_MDT. Standard errors for SGG_LS and SGL_LS were calculated from among-gene comparisons (i.e., MGG_LS and MGL_LS) and were not available for SGL_PL.