In a recent article, Behrens and Vingron (JCB 17, 12, 2010) compute waiting times for k-mers to appear during DNA evolution under the assumption that the considered k-mers do not occur in the initial DNA sequence, an issue arising when studying the evolution of regulatory DNA sequences with regard to transcription factor (TF) binding site emergence. The mathematical analysis underlying their computation assumes that occurrences of words under interest do not overlap. We relax here this assumption by use of an automata approach. In an alphabet of size 4 like the DNA alphabet, most words have no or a low autocorrelation; therefore, globally, our results confirm those of Behrens and Vingron. The outcome is quite different when considering highly autocorrelated k-mers; in this case, the autocorrelation pushes down the probability of occurrence of these k-mers at generation 1 and, consequently, increases the waiting time for apparition of these k-mers up to 40%. An analysis of existing TF binding sites unveils a significant proportion of k-mers exhibiting autocorrelation. Thus, our computations based on automata greatly improve the accuracy of predicting waiting times for the emergence of TF binding sites to appear during DNA evolution. We do the computation in the Bernoulli or M0 model; computations in the M1 model, a Markov model of order 1, are more costly in terms of time and memory but should produce similar results. While Behrens and Vingron considered specifically promoters of length 1000, we extend the results to promoters of any size; we exhibit the property that the probability that a k-mer occurs at generation time 1 while being absent at time 0 behaves linearly with respect to the length of the promoter, which induces a hyperbolic behaviour of the waiting time of any k-mer with respect to the length of the promoter.

8 figures