Evolving Turing Machines for Biosequence Recognition and Analysis

The evolutionary computation approach to motif discovery in biological sequences

“…Most of these have been concerned with multiple sequence alignments [1,12,13,17,18,25,27,32,36,37,42,44,[50][51][52][53][57][58][59]64,66,68,73], although a number have targeted other representations [20,28,29,33,34,39,40,56,60,63,65,71,72]. Most MSA approaches can be divided into two classes: those which directly evolve alignments e.g.…”

“…No direct comparison has yet been made against other promoter finding algorithms. A number of other researchers have also looked at how programmatic classifiers may be used for motif discovery [28,39,40,65]. Handley [28] has evolved GP expressions consisting of continuous numerical functions and left/right relative movement commands to recognise promoter regions in the E. coli genome.…”

“…Howard and Benson's GP-automata approach [33], for example, demonstrates a novel way of representing regulatory motifs and their interactions. However, more interesting are those approaches which break free of the regular expression mould, notably the use of Turing-complete classifiers by Koza et al [39] and Vallejo [65]. In this context, a Turing-complete classifier can be interpreted as a programmatic model of the biological interactions which lead to a particular motif or set of motifs; so, in essence, the GP approach is analogous to reverse-engineering the algorithm underlying the biological system.…”

Benchmarking a Recurrent Linear GP Model on Prediction and Control Problems

“…Same as the case of the Sun Spot benchmark, the typical approach for this problem is to build a predictor based on the sliding window [7]. Lorenz Chaotic time series is defined over three variables by the discrete differential system, [11] 0.082 0.086 0.35 TAR [9] 0.097 0.097 0.28 Recurrent NN [5] 0.1006 0.0972 0.4361 GP [10] 0.125 ± 0.006 0.182 ± 0.037 0.370 ± 0.06…”

“…Examples of this might involve encoding the temporal property of the problem using a sliding window (shift register) of some predefined depth and resolution. Such an approach has seen wide spread application to predictive problems [9], [10] and [11]. In the second case, a recurrent learning model is employed.…”