GESSE: Predicting Drug Side Effects from Drug–Target Relationships

The in silico prediction of unwanted
side effects (SEs) caused
by the promiscuous behavior of drugs and their targets is highly relevant
to the pharmaceutical industry. Considerable effort is now being put
into computational and experimental screening of several suspected
off-target proteins in the hope that SEs might be identified early,
before the cost associated with developing a drug candidate rises
steeply. Following this need, we present a new method called GESSE
to predict potential SEs of drugs from their physicochemical properties
(three-dimensional shape plus chemistry) and to target protein data
extracted from predicted drug–target relationships. The GESSE
approach uses a canonical correlation analysis of the full drug–target
and drug–SE matrices, and it then calculates a probability
that each drug in the resulting drug–target matrix will have
a given SE using a Bayesian discriminant analysis (DA) technique.
The performance of GESSE is quantified using retrospective (external
database) analysis and literature examples by means of area under
the ROC curve analysis, “top hit rates”, misclassification
rates, and a χ2 independence test. Overall, the robust
and very promising retrospective statistics obtained and the many
SE predictions that have experimental corroboration demonstrate that
GESSE can successfully predict potential drug–SE profiles of
candidate drug compounds from their predicted drug–target relationships.