|
Internet Electronic Journal of Molecular Design - IEJMD, ISSN 1538-6414, CODEN IEJMAT
| ABSTRACT - Internet Electron. J. Mol. Des. October 2006, Volume 5, Number 10, 515-529 |
Artificial Immune System Prediction of the Human Intestinal Absorption of
Drugs with AIRS (Artificial Immune Recognition System)
Ovidiu Ivanciuc
Internet Electron. J. Mol. Des. 2006, 5, 515-529
|
Abstract:
Artificial immune systems (AIS) are machine learning procedures
inspired by the structure and function of the biological immune
system. In this paper we present the first application of the artificial
immune recognition system (AIRS) to the modeling of the human
intestinal absorption (HIA) of drugs. The learning task was to classify
a dataset of 196 drugs into a subset of 131 that penetrate the human
intestine and a subset of 65 drugs that do not penetrate the intestine.
The chemical structure was encoded with 159 structural descriptors
from five classes, namely constitutional, topological indices,
electrotopological state indices, quantum descriptors, and geometrical
indices. Eight user defined parameters influences the classification
performance of the AIRS algorithm, namely affinity threshold scalar,
clonal rate, hypermutation rate, number of nearest neighbors, initial
memory cell pool size, number of instances to compute the affinity
threshold, stimulation threshold, and total resources. In order to
explore the AIRS sensitivity to these parameters, leave-20%-out (five-fold)
cross-validation predictions were performed over a wide range of
values for the AIRS parameters. The affinity threshold scalar has the
highest influence on the prediction quality, whereas the remaining
parameters have only a marginal effect. The best predictions of the
AIRS algorithm (selectivity 0.794, specificity 0.615, accuracy 0.735,
and Matthews correlation coefficient 0.406) surpass those obtained
with several well-established machine learning methods. In a
comparison with 13 machine learning algorithms, AIRS predictions
were better in seven cases (Bayesian network, naïve Bayes classifier,
updateable naïve Bayes classifier, logistic regression, Gaussian radial
basis function network, decision tree with naïve Bayes classifiers at
the leaves, and random tree), and worse in six cases (K* instance-based
classifier, alternating decision tree, C4.5 decision tree, logistic
model trees, random forest, and fast decision tree learner). The results
obtained suggest that classifiers based on artificial immune systems
may be successful in structure-activity relationships (SAR),
quantitative structure-activity relationships (QSAR), drug design, and
virtual screening of chemical libraries.
|