Digital experiments of how the process of evolution works—also known as in silico (experimental) evolution or simply evolutionary simulations—have been explored and studied since the early days of computing. Nowadays, while using the computer to simulate evolution is a well-accepted approach, not everyone is clear about what it teaches us. As I use such simulations in my studies, I figured it would not hurt to spend a few sentences to argue the relevance, benefits, and drawbacks.
What do we learn?
- Digital evolution deepens our understanding of how biological systems become organised due to evolution,
- It discloses how the (evolved) structure of a biological system biases future evolution,
- It highlights how biological systems maintain their function and how they may fail to do so (over evolutionary time-scales),
- And, finally, it helps to understand the dynamics of evolution, not just the end-points.
I started my PhD thesis with the statement that we often do not know the consequences of evolutionary processes, and in my opinion that is still very much the case. An intricate configuration may seem to have “evolved for a specific function”, but that structure may very well be a by-product or side-effect of the evolutionary process.
The above arguments also hold to a large degree for experimental evolution in the wet-lab, e.g. Lenski’s long term evolution experiment. Below I offer four benefits of the in silico variety that those experiments cannot offer. Note: I do not remember from whom I copied these arguments, I just know I do not want to claim them as the product of my own thinking.
- Digital evolution allows us to do long-term experiments relatively quickly, 100 000 generations is not an issue on today’s computers. This is not easy or even possible to do in any other way.
- It can keep track of a perfect fossil record by storing parent–child relationships. This enables us to detect when fitness changes, which mutation caused it, and how this mutation changed the working of the individual.
- Almost every parameter can be checked for its influence on the evolutionary process.
- Simulations are repeatable.
The last two arguments hold, of course, for many numerical simulation techniques.
To keep in mind
Off the top of my head, I can think of three points of critique regarding the use of in silico evolution. Whether they are relevant or not strongly depends on the research question that one is asking.
First of all, digital evolution can only exploit what is available in the model. Let me explain what I mean with an example. Around the turn of the century, researchers at the University of Sussex evolved electronic circuits to emit an oscillating electronic signal. They used programmable hardware for this, in contrast to the default approach of using a simulation environment in the computer. Surprisingly, they discovered that some successful circuits exploited the electromagnetic waves emitted by a nearby computer. They had by accident evolved a radio. In a simulation environment, this would never have happened.
The only way to mitigate this issue to a certain extent, is by doing “constructive” evolution where you provide the process with many building blocks and their interactions, but you do not fix these structures. Instead you let evolution shape it. Currently, to the best of my knowledge no-one has managed to define truly open-ended evolutionary simulations.
Second, evolution biases which parts of genotype space are visited. It is not an unbiased method like Monte-Carlo sampling or a full enumeration of all possibilities (the atlas or catalogue approach). I would argue that it is kind-of the point of doing evolutionary simulations to see which parts of genotype space are favoured by evolution.
And, third, results of evolutionary simulations depend on how the genotype is encoded and which mutations operate on this encoding. It is often helpful to try different encodings and mutational operators, and see what effect these have on the results.