Abstract:
Drug-resistant human immunodeficiency virus type 1 (HIV-1) strains have
developed under the selective pressure of antiretroviral treatment. Genotypic resistance testing
is a powerful tool in identifying such resistant viruses in individual patients during or prior to
therapy and it has been used to improve drug efficiency and to design subsequent drug
therapies. Although genotypic testing is easier to perform, previous clinical trials have shown
that patients receiving genotypic testing do not respond to treatment due to multiple drugs
resistant patterns that are increasingly complex and often difficult for clinicians to interpret. In
an attempt to overcome the drawbacks of genotypic resistance testing, this study presents a
neural network system that can predict the HIV-1 phenotypic fold change value from
genotypic results.
The neural network is an artificial intelligence tool for implicitly identifying
any complex nonlinear relationships from experimental data and it has the ability to detect all
possible interactions between all input variables. In this study, 598 HIV-1 protease sequences
and their corresponding phenotypic fold change values (in 50% inhibitory concentration) for
six drugs were retrieved from the Stanford HIV RT and Protease Database. The genotypephenotype
data were divided into a training and test set. Neural network models were
developed from training set data and the performance of the models was determined from the
test data set. The prediction results were expressed as a logarithm of fold change that could be
defined interms of susceptibility classes using cutoff values. The results from the neural
network prediction system were compared with those from the rule-based method provided
with the Stanford HIV RT and Protease Database and the support vector machine method of
the Geno2Pheno interpretation system.
The neural network system predicted values showing a high correlation
coefficient of 0.96 and high accuracy of 95%, both of which were higher than the other two
systems, when compared with experimental phenotypic testing values. Regarding consensus
based prediction, neural network system predicted values also showed better results (97%)
than the other two systems. The correlation of neural network predicted values with clinical
outcome was 78%. So the neural network system proved to be acceptable and more efficient
for designing drug therapies.