|
Internet Electronic Journal of Molecular Design - IEJMD, ISSN 1538-6414, CODEN IEJMAT
| ABSTRACT - Internet Electron. J. Mol. Des. September 2006, Volume 5, Number 9, 488-502 |
Artificial Immune System Classification of Drug-induced Torsade de
Pointes with AIRS (Artificial Immune Recognition System)
Ovidiu Ivanciuc
Internet Electron. J. Mol. Des. 2006, 5, 488-502
|
Abstract:
Artificial immune systems (AIS) represent a family of machine learning
algorithms that use immune system components and mechanisms as
templates in modeling information processes, such as pattern recognition
and classification. This paper demonstrates the first application of the
artificial immune recognition system (AIRS) algorithm in modeling
structure-activity relationships (SAR). A dataset of 349 drugs was used in
the evaluation of the AIRS algorithm. The learning task was to classify
these chemicals into a subset of 106 drugs that induce torsade de pointes
(TdP) and a subset of 243 drugs that do not induce TdP. The chemical
structure was described with five linear solvation energy relationships
descriptors, namely the excess molar refraction, the combined
dipolarity/polarizability, the overall solute hydrogen bond acidity, the
overall solute hydrogen bond basicity, and the McGowan's characteristic
volume. The classification performance of the AIRS algorithm depends on
a large number of parameters: affinity threshold scalar, clonal rate,
hypermutation rate, number of k-nearest neighbors, initial memory cell
pool size, number of instances to compute the affinity threshold,
stimulation threshold, and total resources. The cross-validation predictions
were investigated over of a wide range of values for these eight AIRS
parameters. The best leave-10%-out cross-validation predictions of the
AIRS algorithm (selectivity 0.783, specificity 0.893, accuracy 0.860, and
Matthews correlation coefficient 0.671) surpass those obtained with 11
other machine learning algorithms, namely logistic regression, Bayesian
network, naïve Bayesian classifier, alternating decision tree, C4.5 decision
tree, logistic model trees, decision tree with naïve Bayesian classifiers at
the leaves, fast decision tree learner, random trees, random forests, and K*
instance-based classifier. The results obtained suggest that classifiers
based on artificial immune systems may be successful in structure-activity
relationships, drug design, and virtual screening of chemical libraries.
|