Biology Reference
In-Depth Information
used to search for predictive alignments as a pre-processing step
before ANN analysis. The population in this case consisted of a set of
alignment matrices that were transformed into a new population by
means of a GA and a selection mechanism for fitness improvement.
The constraint applied to direct the evolutionary search, derived from
knowledge of primary anchor positions reported in binding motifs,
consisted of fixing position 1, corresponding to the primary anchor in
each matrix, while the rest of the matrix was subjected to the GA. The
size of the population of matrices was arbitrary selected and the selec-
tion technique used was elitist in that each parent matrix produced
two children, an identical copy of itself and a mutant copy, and passed
the offspring with the higher fitness value to the next generation. All
matrices of the final generation were used to score peptide alignments
by assigning a score to each putative binding core within each bind-
ing peptide. In each simulation, the alignment scored as highest by
the majority of the final generation matrices was selected and passed
to the final stage, i.e. ANN training. For chromosome encoding the
researchers used real numbers instead of binary representations. They
included mutation and reproduction evolutionary operators, but
opted to exclude the crossover operator because it resulted in a pop-
ulation of matrices with high similarities that promoted a linear model
of peptide alignment. The measure of the predictive power of an
alignment matrix was used to define its fitness. Such a fitness function
favored matrices that correctly classified non-binders and resulted in a
population of matrices in which individual matrices captured disjoint
regions in the solution space. The termination criteria (selection of
parent for mutation and reproduction operations) was determined in
a separate experiment that measured sensitivity, specificity, and aver-
age fitness function and in which each population was evolved up to
10
6
generations. The final generation of matrices was used to score
potential alignments, from which the highest scoring alignment of
binding peptides was selected. As previously described, this align-
ment, along with putative non-binders, was used for ANN training.
Other types of EAs, such as genetic programming (GP), have also
been used to address the problem described by Brusic (1998(b),
121-130). GP is a subclass of GA that is an automated procedure to find
