Final Report Summary - SELECTIONINFORMATION (Limits to selection in biology and in evolutionary computation)
Natural selection is the central concept in biology, and selection is widely used to solve difficult computational problems. This project aimed to deepen our understanding of selection, in the context of both evolutionary biology and evolutionary computation, and to help bring these fields together. On the one hand, population genetics can show how to optimise genetic algorithms, and can inspire new algorithms. On the other, the central problem in evolutionary computation is to optimise the "evolvability" of the algorithms - an issue that has only recently become prominent in biology. The over-arching aim is to understand how selection can so effectively gather information from the environment, so as to construct extraordinarily complex organisms.
The project focussed on the factors that limit natural selection: lack of recombination, interaction between genes, and spatial subdivision. Progress was made by combining several techniques: multilocus algebra, branching processes, an analogy with statistical mechanics, and a new model for population structure. This analysis was applied to biological and computational problems in parallel, focusing on how recombination aids selection; how interactions between gens (epistasis) may evolve to facilitate adaptation; and how selection acts in structured populations subject to frequent extinction and recolonisation. Methods were developed for analyzing DNA sequence data, and applied to diverse datasets.
The work covered a wide range, but key results include showing that:
• The rate of accumulation of favourable mutations is ultimately limited by recombination.
• Whether a species can adapt to a steep environmental gradient depends on two parameters – the steepness of the gradient relative to the strength of selection, and the strength of selection on each gene, relative to random fluctuations.
• The stochastic dynamics of high-dimensional systems can be accurately approximated by just a few macroscopic variables, using an analogy with thermodynamics.
• The size of a gene regulatory network is limited by crosstalk between its components, and the specificity of the recognition sequences.
• Complex models of population structure can be efficiently inferred from whole-genome data.
• The “infinitesimal model” represents complex quantitative traits, even in the presence of gene interactions.
• Evolutionary algorithms from computer science and models of natural populations can be represented in the same mathematical framework.
The project focussed on the factors that limit natural selection: lack of recombination, interaction between genes, and spatial subdivision. Progress was made by combining several techniques: multilocus algebra, branching processes, an analogy with statistical mechanics, and a new model for population structure. This analysis was applied to biological and computational problems in parallel, focusing on how recombination aids selection; how interactions between gens (epistasis) may evolve to facilitate adaptation; and how selection acts in structured populations subject to frequent extinction and recolonisation. Methods were developed for analyzing DNA sequence data, and applied to diverse datasets.
The work covered a wide range, but key results include showing that:
• The rate of accumulation of favourable mutations is ultimately limited by recombination.
• Whether a species can adapt to a steep environmental gradient depends on two parameters – the steepness of the gradient relative to the strength of selection, and the strength of selection on each gene, relative to random fluctuations.
• The stochastic dynamics of high-dimensional systems can be accurately approximated by just a few macroscopic variables, using an analogy with thermodynamics.
• The size of a gene regulatory network is limited by crosstalk between its components, and the specificity of the recognition sequences.
• Complex models of population structure can be efficiently inferred from whole-genome data.
• The “infinitesimal model” represents complex quantitative traits, even in the presence of gene interactions.
• Evolutionary algorithms from computer science and models of natural populations can be represented in the same mathematical framework.