Deep learning for prediction of drug-target interaction space and protein functions

Rifaioğlu, Ahmet Süreyya
With the advancement of sequencing and high-throughput screening technologies, large amount of sequence and compound data have been accumulated in biological and chemical databases. However, only small number of proteins and compounds have been annotated by wet-lab experiments due to the huge compound and chemical space. Therefore, computational methods have been developed to annotate protein and compound space. In this thesis, we describe the design and implementation of several methods for accurate drug-target interaction prediction and functional annotations of proteins within the framework of Comprehensive Resource of Biomedical Relations with Deep Learning and Network Representations (CROssBAR) project whose aim is to integrate biological and chemical data scattered in different sources and to create prediction methods for drug discovery based on deep learning. The first method, DEEPred is a sequence based automated protein function prediction method that employs a stacked multi-task deep neural networks based on Gene Ontology (GO) directed acyclic graph hierarchy. The performance of DEEPred was compared with state-of-the-art methods and its source code is available at DEEPScreen is the second method and it is a drug-target interaction (binary) prediction method. In DEEPScreen, the idea is to learn compound features automatically using compound images via convolutional neural networks. DEEPScreen was trained for 704 target proteins and the input compounds predicted as active or inactive against trained targets. The performance of DEEPScreen was compared with the state-of-the art methods using different benchmarking datasets. The source code is available at The third method is called MDeePred which is a binding affinity prediction method. MDeePred is a chemogenomic method where both protein and compounds features were fed to a hybrid pairwise deep neural network structure. The main difference between MDeePred and DEEPScreen in terms of features is that MDeePred employs compound-target feature pairs whereas in DEEPScreen only compound features were used. The main novelty of MDeePred is the proposed multi-channel featurization approach for protein sequences where each channel represents a different property of input protein sequences. The performance of MDeePred was calculated on multiple benchmarking datasets and compared its performance with the state-of-the-art methods. The source code for MDeePred is available at The fourth method is called iBioProVis which is an online interactive visualization tool for chemical space. The main purpose of iBioProVis is to embed and visualize compound features on 2-D space. It relies on the assumption that topologically and chemically similar compounds have similar bioactivity profiles. The inputs for iBioProVis are target protein identifiers and optionally, SMILES strings of user-input compounds. The tool then generates circular fingerprints for active compounds of targets and user-input compounds and then, t-Stochastic Neighbor Embedding (t-SNE) method is used to embed compounds on 2-D space. The tool also provides cross-references for well-known databases for input targets and compounds. iBioProVis is available at
Citation Formats
A. S. Rifaioğlu, “Deep learning for prediction of drug-target interaction space and protein functions,” Thesis (Ph.D.) -- Graduate School of Natural and Applied Sciences. Computer Engineering., Middle East Technical University, 2020.