Biology Reference
In-Depth Information
computer programs that best perform a user-defined task.
248
Through a
Darwinian natural selection process, GP is capable of evolving computer
programs that solve, or approximate the solution, of a variety of prob-
lems. GP focuses on exploiting the regularities, symmetries, homo-
geneities, similarities, and patterns of problem environments by
means of automatically defined functions (ADFs). An ADF is a function
(i.e. subroutine, procedure, or module) that is dynamically evolved dur-
ing a run of GP. ADFs were first conceived and developed by James P.
Rice and John R. Koza of the Knowledge Systems Laboratory at
Stanford University.
249
Klien and colleagues (2000)
247
applied GP meth-
ods to the same data set used by Brusic (1998(b), 121-130). For MHC
binding predictions by GP, they used nine amino acid descriptors based
on physicochemical properties (hydrophobic, positive, negative, polar,
charged, small, tiny, aromatic, and aliphatic). GP instructions set in the
virtual machine used a straightforward collection of Boolean functions,
which allowed expression of complex associations of amino acid residues,
locations, and properties. They implemented a sliding window of length
nine to view the peptides and moved the window one residue at a time
to look for any possible sequence of nine residues within each peptide.
Peptide windows that did not include a known anchor residue in posi-
tion 1 were not tested. Instead of using mathematical instructions, the
researchers used only logical instructions and those that queried the type
of amino acid in a specific position or the properties of an amino acid in
a specified position. Their system implemented genetic operators for
mutation, reproduction, and crossover. However, they reported a lower
predictive performance (75% accuracy) compared to the
80% achieved
by Brusic (1998(b), 121-130) using a hybrid method. Details of their
GP implementation and source code are available at http://www.dd.
chalmers.se/~f 97johan/Projects/EvComp/index.html.
>
Probabilistic Graphical Models
Probabilistic graphical techniques result from the merging of proba-
bility and graphical theories. A wide variety of probabilistic graphical
methods have been developed, including the hidden Markov models
(HMM) and Bayesian networks (BN). In fact, systems that use the
