This rather lengthy summary describes a fascinating progress in the field of GRN modeling; it is a review of the article “Predictive computation of genomic logic processing functions in embryonic development”, which appeared few days ago in PNAS.
Gene regulatory networks (GRNs) are complex systems of interacting genes that deploy in both space and time and are the main regulatory logic behind the specification of cell fates during embryogenesis (Li et Davidson, 2009). Gene networks can be derived for any developmental process or structure and, in simple terms, consist of nodes (the genes themselves) connected by lines, which represent the causal relationships between the different genes. More or less, the GRNs unfold in time in a hierarchical mode where the products of earlier genes serve as regulatory factors (inputs), controlling the expression (output) of downstream targets. Provided that sufficiently detailed knowledge about the gene expression pattern and interactions for a given process is available, a computational model of the respective GRN can be constructed. The main objective of constructing a GRN model is to be able to causally explain each regulatory event in the system and, most importantly, to be able to predict the outcomes of experimental perturbations.
One of the the most complete, if not the best for any developmental process to date, is the GRN behind endomesoderm specification in the embryo of a sea urchin with the lovely name Strongylocentrotus purpuratus, famous for being the favorite model organism in Eric H. Davidson’s lab. Eric’s lab has been instrumental in establishing the endomesoderm network, which currenty contains some 50 genes, and their recent article describes the generation of a Boolean computational model, which beautifully and correctly reconstructs the spatiotemporal expression pattern of the interacting genes and allows for predictions to be made about the effects of changes in the system (Peter et al, 2012).
The computational model has been created by the incorporation of several sources of data: observations on the expression patterns of regulatory genes in space and time and results from trans and cis perturbation experiments. There are four spatial domains in the sea urchin embryo, based on differential gene expression, concerned with the endomesoderm specification -skeletogenic and non-skeletogenic mesoderm, and anterior and the future posterior endoderm. The data was used to construct an abstract representation of the interactions using Boolean logic. At the core of the model are the vector equations – one for every gene in the model. Each equation captures all inputs that sum at a respective node and give its output in Boolean terms, 1 – “on” or 0 – “off”. Understandably, the inputs are the outputs of other genes in the system and the relevant maternal inputs provide the initial triggers – the starting regulatory state.
A strong feature of the model is the incorporation of several other characteristics of the developing embryo, like the notion of intercellular signalling, embryonic geometry (which is the relative position of the four spatial domains in the embryo) and real-time kinetics of gene expression. Including such parametres in the model is important for the configuration of the spatial domains relative to each other changes with time and the kinetics of target gene transcription is a function of the temporal expression features of its regulators (like kinetics of RNA and protein synthesis and turnover). As for the latter, an earlier investigation by the same group has revealed a rather surprising phenomenon, that the step time (this is the “interval between activation of a given regulatory gene and the activation of an immediate downstream target…”) is rather uniform for most genes operating during early sea urchin development, and is approximately 3 hours. This characteristic was included in the model.
The usefulness of the computational model was first revealed by testing its ability to reproduce the regulatory state (the sum of all known regulatory genes expressed at a particular time interval) for each domain. Remarkably, there was a perfect match between the experimentally observed and computed patterns of expression of 33 genes, with only two exceptions – the bm1/2/4 and wnt8 genes. For these genes, the model resulted into a novel expression domain or prolonged expression, respectively, which is most likely due to the lack of knowledge about additional inputs operating in the embryo. Second, the incorporation of a uniform time step in the model successfully reproduced the observed dynamics of spatial gene expression, which is rather surprising, and further supports the idea that the regulatory genes of the S. purpuratus embryo operate with similar kinetics. Curiously, when the model was run with a different step time ( of 4 hours instead) the result was a catastrophic discrepancy in the expression kinetics of many genes.
The next task, and a most daunting one, was to challenge the model’s explanatory power by manually changing selected initial parameters (“removing” or “expressing ectopically” selected genes) and comparing the computational results with data derived from previous experimental perturbations of those same genes. For instance, in one of the settings the authors extinguished delta expression from the skeletogenic mesoderm, which in the model resulted in the loss of gatae and gcm expression in the non-skeletogenic mesoderm and the ectopic expression of V2 endoderm genes, like blimpb1, foxa and hox11/13b. So far so good – this was exactly in agreement with the experimental observations. However, the real predictive power of the model was revealed when the expression of additional genes appeared perturbed, genes that have not yet been tested experimentally! For example, the GRN model predicted that the absence of delta affects not only gatae and gcm, but also genex and j(suh). It would be interesting to experimentally investigate the expression profile of the latter two genes in embryos injected with delta morpholinos, to check whether their patterns conform to the model’s predictions. If they do, then “Hail The Model!”; if they do not, it would simply mean that there exist regulatory inputs beyond our current knowledge.
In a final and rather challenging test, the authors computationally mimicked an experiment from the classic days of transplantation studies by Hoerstadius, when donor skeletogenic micromeres were transplanted on the animal pole of a host embryo, which resulted in a complete second gut and associated mesodermal tissues. In modern terms, this implies that the donor micromeres were sufficient to induce the downstream GRN required for endomesoderm development. However, in their model the authors computationally endowed the animal pole cells (which normally do not form skeletogenic micromeres) with maternal factors sufficient to induce the micromere cell fate. Then, remarkably, the surrounding cells formed three successive spatial domains – ring 1, 2 and 3, which expressed the regulatory states of the mesoderm, veg1 and veg2 endoderm, respectively, precisely mimicking the experimental results.
In conclusion, this study demonstrates the great predictive importance of GRN modeling but also its usefulness for further in silico evaluation of current GRNs structure and discovery of knowledge gaps, or lacunae, as described by the authors. One type of lacunae appear as discrepancies between what is observed and what is computationally generated but, as this study reveals, their frequency is low. In fact, what seems as a rare “failure” of the model could potentially guide future investigations.
Peter IS, Faure E and Davidson EH, Predictive computation of genomic logic processing functions in embryonic development, PNAS Early Edition, Aug 27th 2012
Li E and Davidson EH, Building developmental gene regulatory networks, Birth Defects Research (Part C) 87:123–130, 2009