This post is written by MSU faculty Mark Reimers and Arend Hintze
Let us marvel about the complexity of life for a moment. We have DNA transcribed into mRNA, just to get that translated into proteins, which metabolize, catabolize, or process many other molecules and are responsible for the form and function of each cell. But it doesn’t end there, cells form aggregates and make tissues, which make organs, which form organisms, which are controlled by a complex neural network, just so that we can have social interactions and form communities. And again it doesn’t end there. Each of these steps, in themselves dependent on all other layers of interactions. You see an attractive mate, your brain processes this information, it secretes hormones, which trigger signal transduction cascades, which express genes, regulate other functions, and as usual, the story told much later was: “Well, one thing led to each other, and Mommy and Daddy fell in love, and then you were born!” – evolution in action!
But it also makes us wonder if our computational models keep up with this nested complexity. In pretty much every system we use to study evolution, we have some form of genome that gets translated into something who’s performance we evaluate, and maybe we have agents interacting, but these at most three steps of modeling is far from the apparent complexity of nature – but does it matter?
We think it is more than fair to say that we learned very much about evolutionary processes and evolution in general from using computational models, and quite frankly, that was possible because they were simple. Every time you add another complication to your model, you potentially open a can of worms, and you need to control for yet another factor. Simplicity is the key to successful research.
Most simulation studies at BEACON assume a one-to-one correspondence between ‘genes’ and traits. This strategy makes sense for simulating evolution of bacteria, whose business is biochemistry, and where many phenotypes depend on one gene (or one operon); the E coli studies and single objective evolutionary algorithms were the two key inspirations for BEACON. However we argue that this approach is insufficient for studying metazoan evolution, because animals are constructed through interaction of many components specified by genes. Each trait or body feature is then affected by many genes; and most genes affect several distinct traits, although some of these may be revealed only under a life stress not typically encountered or hard to test in the lab.
At this point we want to broaden our understanding, and use computational models to understand more complex biological processes, in particular those where the system itself changes over time, typically: neuronal or developmental plasticity. In both cases, the genome isn’t translated into the final structure, but the genome encodes a process that controls an ever changing system. In terms of developing a computational model, it becomes less about specifying a form, but about specifying the rules that control a dynamic system. But even that would not really capture natures complexity, the challenge becomes to specify rules that specify rules, which control rules, which are all codependent and interacting. Or maybe this isn’t necessary, and simple models are already sufficient and there is no additional effect on evolutionary dynamics or adaptation. Quite frankly, we don’t believe that. What controls rate of adaptation, and at what fixed point one arrives depends strictly on how the fitness landscape is explored, and it is mutations that facilitate that. If mutations all have a direct effect on how they move the organism around in the FLS then you will have a local exploration. If the effects are random, the organisms would jump around randomly in the FLS. Consequently, a complex nested system that has neural or developmental plasticity will not only move around the fitness landscape in strange ways, it will also start at one point in the landscape, and due to it’s lifetime adaptability move a different point over it’s lifetime, and the genes control how this lifetime movement happens.
If we want to study how animals and their traits evolve, we need to consider how genes affect developmental processes, and model such processes in our simulations. Development proceeds by signals between cells (or other components); the strength of these signals is specified by genes, but the consequences for traits depend on the interaction of any modified signal with all the others active in the same place at the same time. Of course we need to abstract from the complexity of nature.
Other considerations suggest that we should explore this. Evolution typically selects via very many criteria simultaneously. Although research over the past few decades has shown how evolutionary algorithms work for a single criterion and to some extent for two, we have little idea about how to select for many criteria simultaneously. However this is exactly what happens in animal evolution. As Gerhart and Kirschner argue in The Plausibility of Life, summarizing the work of many evo-devo labs, the flexibility of a metazoan to adapt simultaneously to many different criteria and changing selective forces depends on indirect and emergent mechanisms generated by exploratory processes, and ‘weak linkages’, both specified by genes. If we want our simulations to be relevant to the relation between molecules and animal forms or behavior, we should simulate such mechanisms.
We think this approach will likely also shed light on one of the major issues in human molecular genetics. Despite the promise of the Human Genome Project to identify the genetic variants that contribute to complex human diseases and thus to clarify the molecular processes that drive such diseases, little actionable knowledge has been accumulated nearly 20 years on. What evidence we have about haplotype blocks and purifying selection suggests that many disease-related variants have actually been selected for, rather than against. Such observations cannot be explained by the kind of ‘one gene/one trait’ models often studied, but they are entirely consistent with the many and varied selective pressures that are brought to bear on a complex long-lived animal in changing circumstances.
Similar considerations hold for evolving behavioral traits. It is easier in simulations to specify discrete behaviors through distinct genes rather than to simulate the processes that produce behavior, and so such simulations are a natural first step. But as we want our simulations to be more relevant to animal behavior, then we need to consider the development of the nervous system, and the history of learning, both whose processes, but not outcomes, are specified by genes.
In the last few years we at BEACON have made considerable progress in understanding how complexity can emerge through the evolution of simple traits; we have expanded our repertoire of computational tools; and we learned to work closer and better across disciplines. Now we contemplate the prospect of researching the layers of complexity possible through evolving emergent systems; it is mind boggling, as we open a door to glimpse at what nature holds in store for us. We look forward to being able to open this door even further with the support of BEACON and critically discussing insights with our community!