Evolution of networks for body plan patterning; interplay of modularity, robustness and evolvability
Figure 8 displays another representation of the evolutionary process. Here we depicted those agents along the ancestral lineage in which a major evolutionary innovation arose, i.e. an increase in segment or domain numbers. Note that we only show a subset of selected innovations. As in Figure 7A, we see that segment and domain numbers increase more or less simultaneously and that the number of positive feedback loops present in the core network increases. Furthermore, the number of regulatory inputs to the segmentation gene (gene 5) also increases during evolution. Finally, we see that over evolutionary time there is little conservation of the structure of the core network.
Figure 7B displays the SF networks evolutionary dynamics. We see the fast initial increase of segment numbers and a subsequent more gradual increase of domain numbers during evolution typical for this type of evolutionary trajectory. As before, the increase in domain numbers is correlated with an increase in attractors. However, we also observe that increases in attractor numbers not always lead to increases in domain numbers (around time 2500). This can be understood from the fact that attractors should be reachable through the developmental process in order to increase domain numbers.
Similar to before, we observe no clear correlation between genome size and increases in segment and domain numbers and instead see intermittent periods of genome expansion and contraction. However, here there is a strong correlation between evolutionary increases in segment and domain numbers and increases in size of the core genome (especially clear in the inset). Similarly, the genome size of the minimum segmentation, respectively minimum domain network are correlated with segment, respectively domain numbers. So, in contrast to what we saw before, here the size of the minimum genome needed to encode the necessary information does increase with segment and domain numbers.
In Figure 9 we again show those agents along the ancestral lineage in which an innovation arose. We see that first bistability, than oscillations and subsequently faster oscillations are invented (see minimum segment network) generating first 2, then 8/9 and finally 12 segments (see developmental space time-plots). Only later on in evolution the number of domains increases. We can see that part of this increase occurs without the number of positive feedback loops increasing and hence presumably results from network rewiring increasing the independence of already present positive loops (see minimum domain network).
If we compare the minimum segment and domain networks present in the different phases of evolution, we see that previously invented parts are often maintained while new parts are being added. Thus, not only is the final evolved network functionally modular, but these modules are also constructed during evolution in an incremental fashion. This contrasts with the changing nature of the core network we observed for the SS strategy.
In this paper we investigated the in-silico evolution of complex body plans that are both segmented and show anterior-posterior differentiation. An implicit assumption of our study thus is that extensive body plan differentation and segmentation tend to evolutionary co-occur. We base this on the fact that most unsegmented, relatively simple animals such as cniderians possess only a small number of different Hox genes and body domains. In contrast, more complex animals with a larger set of Hox genes and more extensive anterior posterior patterning are either segmented, or show signs of past segmentation [55]–[58]. Note that we made no further assumptions on the order in which segmentation and differentiation evolved, or on whether they evolved once or multiple times [58]–[64].
However, the main aim of the current study was not to settle any of the above issues, but rather to use this setup to study whether or not modular developmental networks evolved. We furthermore investigated how evolution of developmental network modularity depends on indirect selection for robustness. In addition, we studied whether evolved modularity and robustness influence future evolvability. Indeed, we could have used a much more general fitness criterion for body plan patterning, for example maximizing the number of celltypes [65]–[67] or the amount of positional information [68], to study these issues. Instead, we decided to use a more specific fitness criterion that ‘invites’ modularity to evolve, by independently selecting for two functions, segmentation and differentiation. Furthermore, we wished to study segmentation and differentiation as these are considered two major innovations in bilaterian body plan patterning and thus have been extensively studied both experimentally and theoretically.
Evolution was successful in generating body plans that were both significantly segmented and differentiated in 60% of our simulations. This demonstrates two things. First, complex body plan evolution is possible but not trivial. Second, this evolution can be achieved without any coding sequence evolution, by allowing evolution to rewire the regulatory interactions between a simple set of developmental toolkit genes and to duplicate and reuse these genes. Our results thus agree with the argued importance of regulatory evolution [1], [2], [4]–[8] and duplication and divergent usage of existing gene categories [10]–[12] in body plan evolution.
Interestingly, we found that our successful simulations could be divided into only 2 distinct evolutionary scenarios. In 66% of successful simulations segment and domain numbers increased more or less simultaneously during evolution. The evolved developmental networks produced a complex gene expression transient that upon passage of the wavefront was translated into a stable, spatially differentiated expression pattern producing both segments and domains. In the other 33% of successful simulations, first the number of segments increased substantially before the number of domains increased. The evolved SF networks generate gene expression dynamics consisting of a combination of regular oscillations and a complex time transient. The oscillatory dynamics are responsible for producing segments, whereas the complex transient generates domains. Under default parameter settings the segments simultaneous evolutionary strategy is dominant. However, we find that adding noise, thus producing indirect selection for robustness, causes the segments first evolutionary strategy to become the dominant strategy. We furthermore demonstrate that the SF networks also have a higher evolutionary potential for evolving new segments and domains.
Based on the observed differences in expression dynamics, robustness and evolvability we hypothesized that SF networks may also be more modular than SS networks. However, when applying commonly used, purely architectural modularity algorithms similar modularity scores were found for SS and SF networks. Furthermore, these scores were below those of neutrally evolved networks and very close to those of random networks, indicating that no selection for the type of modularity measured by these algorithms occurred.
Only by using our functional knowledge of the networks (they should generate both segments and domains), and taking both functional (different network parts should independently generate either segments or domains) and architectural (these network parts should be largely non-overlapping) aspects of modularity into account could we establish differences in modularity between SS and SF networks. We found that SS networks generated segments and domains in a rather integrated manner, while SF networks operate in a more modular fashion. However, the found modularity was not 100%. Indeed, the SF subnetworks needed to generate either segments or domains share a small subset of their genes and regulatory interactions. Furthermore, a subset of the domains can only be generated in a segment dependent manner. Still, SF networks are considerably more modular than SS networks.
Our results agree with the often heard suggestion that selection for robustness favors modular GRNs and that these modular GRNs tend to be more evolvable [2], [13]–[16]. Furthermore, our findings demonstrate the importance of considering functional aspects of biologically relevant network modularity [36]–[39].
We observed two additional interesting differences between the SS and SF evolutionary strategies. First, while genome size is uncorrelated with body plan complexity for the SS networks, for SF networks not total but core genome size is correlated with organismal complexity. Second, we observed that the complexity and functionality of SF networks changed during evolution in a much more incremental fashion than did the SS networks. Both these differences are likely to contribute to the larger robustness and evolvability of SF networks.
We never observed a domains first segments later evolutionary strategy. In hindsight this is easy to understand. Segments can be generated through two alternative mechanisms. The first, applied in SF networks, uses a segmentation gene oscillator to produce regular segments independent of any domains. The second, used in the SS networks, creates segments by linking segmentation gene expression to the expression of domain forming genes. In this latter case, once a differentiation gene has a spatially varied expression pattern, evolution of a single regulatory link to the segmentation gene suffices to produce segments. Because of this easiness of using domains to make segments, we never observe early evolution of domains with a later evolution of segments.
Previous simulation studies on the evolution of body plan patterning have modeled the evolution of either segmentation [41], [42], [69] or differentiation [43], [65]–[67], [70] alone. The major aim of these studies was to gain an understanding of how natural developmental mechanisms might have evolved. As a consequence these studies focused on the resemblance between in-silico evolved network architectures and those found in nature [41]–[43]. Below we compare our results both to the findings of these earlier studies and to developmental networks found in nature. It should however be kept in mind that in our study this resemblance was neither an explicit aim nor part of our model design.
As discussed above, SS networks generate a single complex gene expression transient that produces both segments and domains. In contrast, SF networks generate both oscillatory dynamics and a complex time transient, the first responsible for producing segments and the second responsible for generating domains. The translation of oscillatory dynamics by a wavefront into a regular segmentation pattern is called the clock-and-wavefront mechanism for segmentation. It was first suggested by Cooke and Zeeman [71] and has been extensively modeled [72]–[76]. This mechanism is responsible for vertebrate somitogenesis [77]–[81], arthropod short germband segmentation and annelid segmentation [64], [82]–[84]. It is suggested to be the ancestral mode of segment formation [60], [62], [85].
Recently, Francois and co-workers [41] found that selection for body plan segmentation in the presence of a propagating morphogen wavefront always leads to the evolution of a clock-and-wavefront type mechanism. In contrast, we find that under selection for both segmentation and differentiation either a clock-and-wavefront type segmentation mechanism or a mechanism in which segmentation depends on the expression of domain forming genes may evolve. In the latter case, segments arise downstream of the differentiation process, with different segments arising from different combinations of domain forming genes. This mechanism very crudely resembles the long germband, Drosophila type of segmentation [86]–[88]. However, in our model segments are formed sequentially rather than simultaneously. The fact that we do not observe a hierarchy of mutual repressors as has been observed in simulations of long germband type patterning [42], [43] is most likely due to this sequential rather than simultaneous patterning. Our results suggest that key to understanding Drosophila segmentation is not just considering that the process occurs simultaneously rather than sequentially, but to also take into account that the segmentation and differentiation processes are tightly integrated.
We found that both SS and SF networks use a complex gene expression transient to produce different domains, and in case of the SS network also different segments. In addition, we found for the SF network that the domains produced were of a continuous staggered nature, somewhat similar to the Hox gene anterior posterior expression domains. In a previous study, Francois and Siggia [43] explicitly selected for such a Hox like differentiation pattern. They found that in case of a propagating morphogen wavefront, a special timer gene was needed to control the order and location in which genes were switched on. The expression level of this timer gene slowly accumulated in the time preceding the passage of the wavefront, thus allowing a translation of wavefront passage time into timer gene expression level and finally expression of a different set of downstream genes. In contrast, in our study we obtained anterior-posterior differentiation without the need for such a timer gene, by combining the presence of alternative attractors with a long and complex time transient. Together this ensures convergence to different attractors at different times of wavefront passage, thus also producing sequential spatial differentation.
Experimental data suggest that the initial Hox gene activation occurring during the primitive streak phase is temporally colinear and may involve timing mechanisms such as chromosomal looping, ordered opening of chromatin domains and cluster level activator and repressor regions [89]–[94]. In contrast, the Hox gene activity in the presomitic mesoderm and during somite formation appears to be under more individual gene level regulatory control [93], [95], [96] and coordinated with the somitogenesis clock and morphogen wavefront [94], [96]–[103]. Indeed, in our segment first simulations we find that the segmentation and patterning processes both depend on the morphogen wavefront (Figure 6B, middle and bottom row), and that they require some coordination (see Figure S10 in Text S1). This resemblance to vertebrate axial patterning evolved for free, as it was neither part of our fitness criterion nor of the model design and is a side effect of considering the combined evolution of segmentation and differentiation. Furthermore, it demonstrates that the evolution of natural developmental mechanisms such as vertebrate axial patterning is neither a very unlikely event nor a completely random outcome, but a type of solution that can be expected.