The Evolution of Complex Features

When Darwin first proposed his theory of evolution by natural selection, he realized that it had a problem explaining the origins of the vertebrate eye [Darwin1859]. Darwin noted that `In considering transitions of organs, it is so important to bear in mind the probability of conversion from one function to another.'' That is, populations do not evolve complex new features de novo, but instead modify existing, less complex features for use as building blocks of the new feature. Darwin further hypothesized that `Different kinds of modification would [...] serve for the same general purpose'', noting that just because any one particular complex solution may be unlikely, there may be many other possible solutions, and we only witness the single one lying on the path evolution took. As long as the aggregate probability of all solutions is high enough, the individual probabilities of the possible solutions are almost irrelevant.

Substantial evidence now exists that supports Darwin's general model for the evolution of complexity (e.g., [Dawkins86,Jacob77,Newcomb97,Nilsson94,Wilkins02]), but it is still difficult to provide a complete account of the origin of any complex feature due to the extinction of the intermediate forms, imperfection of the fossil record, and incomplete knowledge of the genetic and developmental mechanisms that produce such features. Digital evolution allowed us to surmount these difficulties and track all genotypic and phenotypic changes during the evolution of a complex trait, with enough replication to obtain statistically powerful results [Lenski03]. We isolated the computation EQU (logical equals) as a complex trait, and showed that at least 19 coordinated instructions are needed to perform this task. We then performed an experiment that consisted of 100 independent populations of digital organisms being evolved for approximately 17,000 generations. We evolved 50 of these populations in a control environment where EQU was the only task rewarded; we evolved the other 50 in a more complex environment where an assortment of 8 simpler tasks were rewarded as well, to test the importance of intermediates in the evolution of a complex feature.

Results: In 23 of the 50 experiments in the complex environment, the EQU task was evolved, whereas none of the 50 control populations evolved EQU, illustrating the critical importance of features of intermediate complexity (P = 4.3 × 10^-9, Fisher's exact test). Furthermore, all 23 implementations of the complex trait were unique, with many quite distinct from each other in their approach, indicating that, indeed, this trait had numerous solutions. This is not surprising since even the shortest of the implementations found were extraordinarily unlikely (approximately 1 in 10²⁷). We further analyzed these results by tracing back the line of decent for each population to find the critical mutation that first produced the complex trait. In each case, these random mutations transformed a genotype unable to perform EQU into one that could, and even though these mutations typically affected only 1-2 positions in the genome, a median of 28 instructions were required to perform this complex task---a change in any of these instruction would cause the task to be lost, thus it was complex from the moment of its creation. Also of note is the fact that in 20 of the 23 cases the critical mutations would have been detrimental if EQU were not rewarded, and in three cases the prior mutation was actively detrimental (causing the replication rate for the organisms to drop by as much as half), yet turned out to be critical for the evolution of EQU; when we reverted these seemingly detrimental mutations, EQU was lost.