Our analysis indicates that multicellularity is a phylogenetically conservative character (p-value < 0.01). If the terminal taxa of the Bayesian consensus tree are randomly re-shuffled, a count through 1,000 re-shuffled trees gives an average of 20 transition steps. However an average of only nine parsimonious transitions was observed in a count through 10,000 randomly sampled trees of our ancestral character state reconstruction.

Results of the character state reconstruction using the AsymmMK model with transition rates estimated by Mesquite 2.71 60 are displayed in Figure 5. Using maximum likelihood analysis, average frequencies of the characters were counted across 10,000 trees randomly sampled from the two Metropolis-coupled Markov Chain Monte Carlo (MC ³) searches of the Bayesian tree reconstruction.

Cyanobacteria share a unicellular ancestor, but multicellularity evolved early in the cyanobacterial lineage. We identified multicellular character states for three basic ancestors leading to clades E, AC and C in our tree. Together, these clades encompass the entirety of the morphological sections II, III, IV and V. Additionally character states were reconstructed using maximum likelihood analysis and fixed transition rates to analyze properties of the data set. Transition rates are presented in Table 2. Probabilities for character states at nodes 3, 4 and 5 were examined in detail (Table 3). A multicellular ancestry is very likely for these three nodes. For node 3 the relative probabilities of a multicellular ancestor range from 0.79 to 1.00, depending on the probability of the transition rates. For node 4 with varying transition rates, the relative probabilities of a multicellular ancestor range from 0.83 to 1.00. For node 5 the probabilities for multicellularity range from 0.90 to 1.00.

The maximum likelihood analysis is not contradicted by a Maximum Parsimony optimization (Table 3 and Additional File 5). Applying maximum parsimony as a reconstruction method, the uniquely best states were counted across 10,000 trees randomly sampled from the two (MC ³) runs of the Bayesian tree reconstruction. The relative probabilities for a multicellular ancestor at nodes 3, 4 and 5 are 0.68, 0.68 and 0.69, respectively. In contrast, the relative probabilities for a unicellular ancestor at nodes 3, 4 and 5 under parsimony reconstruction are 0.0013, 0.0014 and 0.0014, respectively.

Using Bayesian methods, a similar pattern is observed for these nodes. As an evolutionary model, BayesFactors revealed that a "hyperprior" approach with exponential prior distributions, whose means were sampled from a uniform distribution between 0 and 10 gave the best fit. Transition rates were estimated to be almost equal. Figure 6 displays the posterior probability distributions of character states at these three nodes as they were estimated over 10,000 randomly sampled trees. At nodes 3 and 5 posterior probabilities of a multicellular character state display values above 0.90 for most of the trees. At node 4 a multicellular state is more likely as well. Posterior probabilities at node 4 are above 0.75 for most of the trees.

At least five reversals to unicellularity occurred in the tree, three of them within clade AC. The first transition occurred on a branch which led to a group of thermophilic cyanobacteria: Acharyochloris sp., Synechococcus lividus C1 and Thermosynechococcus elongatus. Posterior probabilities (PP) and bootstrap values (BV) for this group are 0.99/73%, whereas the sister group within AC is supported by 0.96/66% (PP/BV). The second transition within clade AC led also to a thermophilic cyanobacterium Synechococcus C9. Sister relation of this species to a filamentous thermophilic cyanobacterium is supported by 1.0/99% (PP/BV). The last transition in clade AC occurred within the group including the marine pico-phytoplankton genera Synechococcus and Prochlorococcus. The filamentous Prochlorothrix hollandica is supposed to be the closest relative to the group that includes marine pico-phytoplankton, supported by 1.0/61% (PP/BV). Clade AC has a PP of 0.99, while its BV is below 50%. Although bootstrap support is below 70% for clade AC and some groups within it, posterior probabilities show a very high support (> 0.95). Simulation studies have shown that posterior probabilities approach the actual probability of a clade 61 62 63 . Bootstrapping tends to underestimate the actual probability of a true clade. Although, posterior probabilities tend to be erroneous if the model of evolution is underparameterized, overparameterization has only a minor effect on the posterior probabilities. Therefore, using a complex model of evolution, such as the "general time reversible with gamma distributed rate variation"(GTR+G), is recommended 62 63 . We used the GTR+G+I model for our analysis, and assume that nodes with a PP higher than 0.95 are reliable.

It is very likely that at least one additional reversal to unicellularity occurred in clade E1, but phylogenetic support is not high enough to locate the exact position of this transition. Similarly, support for the nodes where the other transition to multicellularity within clade E occurred is missing. The exact locations of reversals within clade E therefore are not certain and a scenario where multiple reversals occurred cannot be excluded. In clade E, there is also a reversal to multicellularity observed in Spirulina sp. PCC 6313. The location of this transition is supported by posterior probabilities of 0.99 at two ancestral nodes.

Stucken et al. 64 compared gene sets of multicellular cyanobacteria and found that at least 10 genes are essential for the formation of filaments. Besides genes previously thought to be correlated with heterocyst formation (hetR, patU3 and hetZ) they found seven genes coding for hypothetical proteins. The species they compare are all located within clade E in our tree, most of them being differentiated. Unfortunately no genomes from multicellular species in more basal clades are available at present. But genome projects of Phormidium sp. ISC 31 and Plectonema sp. ISC 33 are presently being conducted http://www.ncbi.nlm.nih.gov/genomes/lproks.cgi. If these species turn out to group with Phormidium mucicola IAM M-221 and Plectonema sp. F3 from the basal clades C and AC in our study, this could provide important information on the original metabolic pathways in ancient multicellular cyanobacteria and on possible advantages of multicellularity.

The majority of cyanobacteria living today are described as successful ecological generalists growing under diverse conditions 20 . Our analysis indicates that this diverse range of cyanobacterial morphotypes found in various habitats today —whether multicellular or unicellular— has evolved from multicellular ancestors.

In eukaryotes, simple multicellular forms build the foundation for the evolution of complex multicellular organisms. Although complex multicellularity exhibiting more than three cell types is presumably missing in prokaryotes, bacteria invented simple multicellular forms possibly more than 1.5 billion years earlier than eukaryotes 24 25 26 65 . Multicellularity has been described as one of several major transitions that occurred in the history of life. These transitions between different units of selection 66 resulted in changes in the organizational confines of the individual. Maynard Smith and Szathmary 67 (1995, p.6) summarize eight major transitions in the evolution of life after which, "entities that were capable of independent replication before the transition can replicate only as part of a larger whole after it". These transitions can create new units of selection at a higher level of complexity 68 . Origin of chromosomes, origin of the eukaryotic cell, origin of multicellular organisms and the origin of eusocial communities are some major transitions that redefine the degree of individuality 66 67 69 70 . Some transitions are thought to be unique, such as the evolution of meiosis or the evolution of the genetic code. Other major transitions occurred several times independently, such as the evolution of eusociality 71 72 and multicellularity 10 66 73 74 75 . There is a tendency to assume that these transitions occur in a progression that leads to an increase in complexity. However, it seems that in cyanobacteria this is not the case. Anatomical complexity has been lost during their evolution several times (Figure 5). In a similar fashion, a complex character such as eusociality has been lost several times in halictid bees 72 76 . Conversely the phylogeny indicates that multicellularity re-evolved in Spirulina. Regaining complex characters has been observed in other studies as well 77 78 79 . Nonetheless, some studies state that re-evolution of a complex character after a previous loss is not possible 80 81 . Such studies argue that according to 'Dollo's law', a loss of complexity is irreversible 82 , a statement that is not supported in the cyanobacterial case. Repeated transitions in either direction are possible.

Various claims for life during the early Archean Eon, more than 3.00 billion years ago exist. Most of them from two regions: the Berberton Greenstone Belt, South Africa (around 3.20-3.50 billion years old) and the Pilbara Craton, Western Australia (around 2.90-3.60 billion years old). For some of these "fossils" a biological origin is questioned 26 27 83 , but for others biogenicity is very likely 23 25 26 84 85 86 87 . These candidates for early life have clear age constraints and there is no non-biological explanation for these structures. The ages and possible metabolic features of seven fossils of proposed biological origin are plotted in Figure 7 (1-7) 23 25 26 84 85 86 87 . Some of these fossils are assumed to have been photosynthetic and mat builders, characteristics that can be identified in cyanobacteria as well. One of the oldest fossils recorded, 3.45 billion year old prokaryotic remains found in the Panorama Formation, East Pilbara Craton, Western Australia exhibit a filamentous morphotype and possibly carried out anoxygenic photosynthesis 25 26 .

Some late Archean fossils show an oscillatorian or chroococcacean morphotype (Figure 7: 8, 9). 2.52 and 2.56 billion year old oscillatorian-like fossils 24 88 89 could possibly represent close relatives of cyanobacterial ancestors. 2.72 billion year old filamentous bacteria 24 could potentially represent one of the first multicellular cyanobacteria detected. For single celled forms, 2.56 billion year old unicellular fossils 89 90 91 92 could likely represent chroococcacean fossils, relatives of ancestral Gloeobacter violaceus or Synechococcus sp. P1 (Figure 7).

The first conclusive cyanobacterial fossils from all five sections have been reported from around 2.15 billion year old rocks. In 1976, Hofmann described Microfossils from stromatolitic dolomite stones in the Kasegalik and McLeary Formations of the Belcher Supergroup in Hudson Bay, Northern Canada. Among these fossils are Halythrix which seems to belong to the order Oscillatoriales (section III), Eosynechococcus and Entophysalis both presumably order Chroococcales (section I) and Myxococcoides fossils (section II). In 1997 similar fossils were described by Amard and Bertrand-Safarti in paleoproterozoic cherty stromatolites from the "Formation C (FC)" of the Franceville Group in Gabon, dating back 2.00 billion years. They also characterized chroococcalean fossils, particularly Eosynechococcus and Tetraphycus, and filamentous bacteria (Gunflinta) which could likely resemble cyanobacteria and Myxococcoides fossils. Furthermore, large microfossils (so called Archaeoellipsoides elongatus), with akinetes similar to the ones from Anabaena-like species were found 4 19 . Akinetes are resting cells which are only present in differentiated cyanobacteria from sections IV and V. As it has been confirmed in several studies, sections IV and V share a most recent common ancestor 4 33 36 . Therefore these fossil akinetes document the existence of differentiated cyanobacteria 2.00 billion years ago. Given that differentiation in cyanobacteria is evolutionary stable only in a multicellular setting 93 , this again supports the notion that multicellular species belonging to the cyanobacteria must have existed earlier than 2.0 billion years ago.

Several studies have assessed prokaryotic history using phylogenetic dating methods 50 52 . In these studies the origin of cyanobacteria has been estimated around the time of the "Great Oxygenation Event" of 2.20-2.45 billion years ago 2 7 . Other studies have reported elevations of oxygen levels before the great rise of atmospheric oxygen 7 94 . Using small and large ribosomal subunit sequences, Blank and Sanchez-Baracaldo 39 estimated the origin of cyanobacteria between 2.7 and 3.1 billion years ago. They also try to address the evolution of cyanobacterial traits and assess that multicellular cyanobacteria did not originate before 2.29-2.49 billion years ago. In the study of Blank and Sanchez-Baracaldo 39 , a smaller set of cyanobacterial taxa was used, with some basal multicellular species that are present in clade C of our analysis missing. These taxa could have an essential effect on the timing of the first multicellular cyanobacteria. To resolve this issue further dating analyses would be needed. Clearly, as Blank and Sanchez-Baracaldo point out, for such analyses to ultimately resolve the cyanobacterial history, a larger number of cyanobacterial genome data would be needed to represent all the morphological and genetic diversity within this phylum.

A total of 2,065 16S rRNA gene sequences from the phylum cyanobacteria were downloaded from GenBank. Unidentified and uncultured species were excluded. With this large dataset phylogenetic reconstructions were conducted as described in the next section. Aside from cyanobacteria, the dataset included six chloroplast sequences and six eubacterial sequences: Beggiatoa sp., Thiobacillus prosperus, Agrobacterium tumefaciens, Chlorobium sp., Candidatus Chlorothrix halophila and Escherichia coli HS.

From this large tree a subset of 58 cyanobacterial sequences were selected for further analyses. Accession numbers are provided in Table 1. Species from all five sections described by Castenholz et al. 9 were included. Taxa were chosen to represent a 1:1 ratio of unicellular and multicellular species. The final data set contained 22 single-celled taxa from section I, 7 single-celled taxa from section II, 21 multicellular taxa from section III, 5 multicellular, differentiated taxa from section IV and 3 differentiated, branching taxa from section V as described by Castenholz et al. 9 .

An outgroup for further analyses was chosen from a set of eubacterial, non-cyanobacterial species whose 16S rRNA gene sequences were downloaded from GenBank (Table 4). Species were sampled to cover a wide range of different phyla. Aside a set of species from phyla represented in the "tree of life" 51 , species from additional phyla as described on NCBI http://www.ncbi.nlm.nih.gov/Taxonomy/ - Taxonomy Browser: Bacteria) were selected for analyses.

Character state reconstructions were performed using maximum parsimony (MP; Additional File 5) and maximum likelihood criteria as implemented in Mesquite 2.71 60 . 5,000 trees from each MC ³ run were randomly chosen from the post burn-in Bayesian sample and combined. Discrete characters were coded into multicellular or unicellular states. The results over 10,000 Bayesian trees were summarized and displayed on the consensus tree of the Bayesian analysis. For maximum likelihood estimates, both the "Markov k-state 1 parameter model" (MK1 model) and "Asymmetrical Markov k-state 2 parameter model" (AsymmMK model) were applied. Rate of change is the only parameter in the MK1 model. The AsymmMK model exhibits two parameters, describing the forward and backward transitions between states. Phylogenetic conservativeness of multicellularity was tested by comparing the observed distribution of parsimony steps across 10,000 randomly chosen trees from the Bayesian analysis against the distribution from 1,000 trees modified from the Bayesian consensus by randomly shuffling the terminal taxa, while keeping the relative proportion of states unaltered. The root was assumed to be at equilibrium. Transition rates for the MK1 and AsymmMK model were estimated by the program. Rates for the latter models presented in Table 2 were estimated for the consensus tree. To explore properties of the data set, character states were additionally reconstructed with manually fixed transition rates (F1-F6; Table 2). The state of the outgroup was excluded from the analyses to avoid biased inferences within the ingroup.

The character states of nodes 3, 4 and 5 of the Bayesian consensus tree were additionally estimated using a reversible jump MCMC search as implemented in BayesTraits 117 . MCMC was run for 30 million iterations, and a burnin set to 50,000. The analysis was run several times with parameters of the evolutionary model being chosen from different prior distributions. In order to determine which model offered the best fitting priors, models were tested using Bayes Factors. A hyperprior approach with mean-values of the exponential priors derived from a uniform distribution between 0 to 10 was determined to fit best the data. The results of the analysis were visualized in Tracer v1.5 115 .