Evolutionary theory explains the origin of complex organismal features through a combination of reusing and extending information from less-complex traits, and by needing to exploit only one of many unlikely pathways to a viable solution. While the appearance of a new trait may seem sudden, we show that the underlying information associated with each trait evolves gradually. We study this process using digital organisms, self-replicating computer programs that mutate and evolve novel traits, including complex logic operations. When a new complex trait first appears, its proper function immediately requires the coordinated operation of many genomic positions. As the information associated with a trait increases, the probability of its simultaneous introduction drops exponentially, so it is nearly impossible for a significantly complex trait to appear without reusing existing information. We show that the total information stored in the genome increases only marginally when a trait first appears. Furthermore, most of the information associated with a new trait is either correlated with existing traits or co-opted from traits that were lost in conjunction with the appearance of the new trait. Thus, while total genomic information increases incrementally, traits that require much more information can still arise during the evolutionary process.

Phylogenetic trees group organisms by their ancestral relationships. There are a number of distinct algorithms used to reconstruct these trees from molecular sequence data, but different methods sometimes give conflicting results. Since there are few precisely known phylogenies, simulations are typically used to test the quality of reconstruction algorithms. These simulations randomly evolve strings of symbols to produce a tree, and then the algorithms are run with the tree leaves as inputs. Here we use Avida to test two widely used reconstruction methods, which gives us the chance to observe the effect of natural selection on tree reconstruction. We find that if the organisms undergo natural selection between branch points, the methods will be successful even on very large time scales. However, these algorithms often falter when selection is absent.