---

I call this experiment 'replaying life's tape.' You press the rewind button and, making sure you thoroughly erase everything that actually happened, go back to any time and place in the past... Then let the tape run again and see if the repetition looks at all like the original.

The bad news is that we cannot possibly perform this experiment.

S.J. Gould (1989), Wonderful Life: the Burgess Shale and the Nature of History.

How repeatable is evolution? There are at least seven ways to evolve an eye. By contrast, there appear to be fewer ways for humans to evolve lactose tolerance or for stickleback fish to adapt to evade predators in Canadian lakes. When birds make it to remote islands, they often lose the ability to fly—much to their regret if predators are later introduced.

Understanding the factors determining the dynamics and repeatability of evolutionary outcomes is more than an academic exercise. Evolution happens all around us, continually buffeting us with its effects. Microorganisms adapt to resist antibiotics that once kept them in check. Similar processes allow them to evade host immune responses and cause disease, and even shift between host species, evidenced by the jump of HIV from simians to humans. Predicting next year's prevailing influenza virus is a multi-million dollar business for vaccine producers. Evolution is also a key part of industrial processes, for example, allowing selection for increased production of valuable biomolecules and therapeutics. In all these cases, an improved ability to predict the path of evolutionary change—and with it the ability to help or hinder an evolving population—would have important implications.

There are two general explanations for the observation that evolutionary outcomes sometimes appear repeatable, and sometimes do not. The evolutionary process may be intrinsically 'noisy' and therefore unpredictable. This view recognizes that evolution results from selection of a series of randomly occurring mutations, the effects of which will often depend on those that have already occurred. This kind of dependence can act to amplify differences between populations as evolution progresses. Alternatively, it may be that at each new evolutionary step, a single mutation is heavily favored, either because it occurs at a high rate or because it provides an unusually large benefit and will therefore outcompete co-occurring mutations. These factors create a bias, such that, all else being equal, evolution will tend to arrive at similar outcomes in similar ways. In this view, different evolutionary outcomes result from the influence of changing environmental factors—migration between populations, environmental fluctuations, the effects of parasites—that conspire to ensure that all else is not equal.

One way to examine the factors determining the repeatability of evolution is, in the words of Stephen Jay Gould, to 'replay life's tape'. Gould imagined a world in which replicate evolution experiments are possible. Genetically identical populations would be evolved in identical environmental conditions, and their evolutionary responses compared. By removing 'chance' factors—different genetic starting points and fluctuating environmental conditions—this experiment would allow the effect of the remaining factors—principally, mutation and selection—on the outcome of evolution to be isolated. In principle, understanding these factors would allow us to 'tame' the unpredictability of evolution.

On the scale that he imagined, Gould's experiment is, of course, impossible. Nevertheless, he would have been pleased at the progress that experimental evolutionists have made. Indeed, several laboratory experiments, the longest going for -45,000 generations (the equivalent of nearly a million human years!), have been designed and carried out to examine the dynamics and repeatability of evolution. In these experiments, replicate populations are started from the same ancestor, propagated in identical environments, and analyzed in great detail to compare their evolutionary trajectories. The results of these analyses have been intriguing: replicate populations often follow similar, but not identical, evolutionary paths as they adapt to a new environments.

New technology allows this analysis to go a step further. The ability to sequence large regions of DNA has enabled researchers to determine the genetic causes of adaptive changes. Applying these tools to controlled evolution experiments reveals that, even when populations adapt via similar physical changes, the underlying genetic causes of these changes can be different.

Predicting when similar genetic changes underlie adaptation is important. The ability to intervene effectively to control bacteria resistant to a particular antibiotic will usually require knowing not just that it is resistant, but also the nature of the genetic changes that have conferred the resistance. We can only expect independently arising resistant bacteria to respond similarly to some alternative treatment if their resistance has the same underlying cause. Understanding the precise genetic cause of adaptations is also crucial for predicting future evolutionary opportunities. Recent work has demonstrated that the probable outcome of viral therapeutics can differ dramatically depending on the presence or absence of a small number of mutational changes. In short, predicting evolution requires understanding not only the physical outcomes of adaptation, but also how these outcomes occur.

We propose to develop a general theoretical framework to predict the evolution of populations. Progress on such a framework has been slow. Predicting evolution requires some way to model the complex network of interactions that link individuals differing by single mutational steps and to follow the movement of a population across this network as it is pushed and pulled by the fitness effect conferred by each mutation. Because individuals with high fitness genotypes will, on average, leave more descendants than those with low fitness, the population will tend to move toward fitter genotypes—the process Darwin called natural selection. Further, the huge number of potential genetic changes map to a much smaller number of fitness values—indeed, many genetic changes will have no measurable effect on fitness. Together these factors build up a picture of a landscape: genotypes of high fitness represent peaks, separated from one another by valleys caused by genotypes of low fitness. Developing a framework able to incorporate these features, and predict the evolution of populations across landscapes has proven to be a huge analytical and computational challenge.

Large deviation theory (LDT) is an approach that has been used in engineering and physics to model the transition of systems between stable states. The theory can identify regions of possible system states that are stable—and therefore attractive. Moreover, it can predict the path a system will follow as it moves from an arbitrary position to a stable state, or between these states following some perturbation of the system, such as the failure of some component. Our insight is that this framework can also take into account key aspects of the evolutionary landscape described above. An evolving population would take the place of an engineered machine, mutations the place of machine failures, and fitness the place of system stability. Other aspects are also analogous, movement of an entity across a landscape tending toward regions of high attraction, be that high stability or fitness. The first step of our proposal is to formalize this analogy and develop a LDT-based theoretical framework that can be applied to analyze and predict the evolution of populations across adaptive landscapes.

An important extension to any theoretical model is the use of an experimental system to test the predictions it makes. To test the predictive power of the LDT framework, we require a sufficiently tractable experimental system. Ideally this system would allow replicate populations to be evolved in constant environments. Resulting genetic and phenotypic changes would then be identified and, ideally, tracked during the period of evolution allowing theoretical and empirical estimates of evolutionary trajectories to be compared. Microbial systems meet these criteria. To complement the microbial evolution experiments we will also use computational evolving systems.

Combining theoretical and experimental approaches will open new types of evolutionary questions. Under what circumstances will evolution be repeatable? Even if the end point is repeatable, is the path taken by a population to get there the same? How sensitive will populations be to perturbation? Our goal is to develop and test a theory that allows these questions to be addressed.