Drug repositioning or
repurposing refers to identifying new indications for existing drugs and
clinical candidates. Predicting drug-target interactions (DTIs) is of great
challenge in drug repositioning. This challenge can be divided into two
sub-problems. Firstly, the exponential growth in the volume of data available
on drugs and proteins. Secondly, the number of known interacting
drug-target pairs is much smaller than that of non-interacting drug-target
pairs or unlabeled samples which they have not been
experimentally verified to be true non-interactions. Many computational methods have been proposed to address this
problem. However, they suffer from the high rate of false positive predictions
leading to biologically interpretable errors. To cope with these limitations, we propose in this paper a
new computational method based on semi-supervised deep learning (SS-DeepL-DTIs)
using the
pre-trained unsupervised stacked autoencoder model on unlabeled data to initialize
the supervised deep neural networks model. It aims to predict potential drug targets
and new drug indications for drug reprofiling using a large scale chemogenomics
data while improving the performance of DTIs prediction. The proposed approach
has been compared to five state-of-the-art machine learning algorithms applied
all on the same reference datasets of DrugBank. Experimental results have
shown that our approach outperforms other techniques. Its overall accuracy
performance is more than 98%.