1st International and 10th National Iranian Conference on Bioinformatics

Drug-Target Interaction Prediction with Deep Learning and Recommender Systems

Paper ID : 1193-ICB10

Authors:

Seyed Amirhossein Nosrati *¹, Seyed AmirAli Ghafourian Ghahramani², Kaveh Kavousi³

¹MSc Student of Computer Engineering at Sharif University of Technology, International Campus (SUTIC)

²Assistant Professor of Computer Science, Department of Computer Engineering, Sharif University of Technology, International Campus (SUTIC)

³Assistant Professor of Systems Biology and Bioinformatics, Institute of Biochemistry and Biophysics (IBB), University of Tehran

Abstract:

Drug designing is a complex, costly, and time-consuming process with high failure chance [1]. Therefore, drug repurposing is gaining importance. To this aim, one method is to identify new interactions between chemical compounds and protein targets called predicting Drug-Target Interactions (DTIs) [2]. DTIs are usually represented by bi-partite networks where predicting drug-target interaction can be formulated as the link prediction. Recently, Graph Neural Networks (GNNs) have achieved tremendous success on machine learning tasks defined over graph-structured data such as node classification and link prediction [3]. They can extract informative features of the input graph and improve the performance of down-stream tasks.

We propose a GNN-based framework to predict drug-target interactions. To this aim, we represent the drug-target interactions as a bi-partite network and we construct a protein-similarity network between proteins based on their structural similarities. The proposed framework learns informative representations for both drugs and target proteins in an end-to-end fashion. The learned representations are used to predict interactions. We prepared two data sets. The first one extracted from DrugBank [4] containing relevant information about drugs, target proteins, and their interactions. As the second data set, we used the Yamanishi benchmark dataset [5] containing interactions between drugs and different groups of enzymes, ion channels, GPCRs, and nuclear receptors.

The results indicate that the proposed framework exhibits acceptable performance and can get better results compared to some proposed methods in the literature. On the first data set, our method achieved 92% and 85% of accuracy over training and test sets. For the second data set, accuracies are 89%, 86%, 82%, and 80%, respectively, on four classes of targets.

The proposed GNN-based framework starts with random representations for drug and proteins and learns highly informative embeddings to predict the possible interactions. Results, indicate the high abilities of the method in predicting DTIs.

Keywords:

Graph Neural Network; Drug-Target Interaction; Drug Repurposing; DTI Classification;

Status : Paper Accepted (Poster Presentation)