RANDOM WALK APPLIED TO HETEROGENOUS DRUG-TARGET NETWORKS FOR PREDICTING BIOLOGICAL OUTCOMES

Abstract

Prediction of unknown drug target interactions from bioassay data is critical not only for the understanding of various interactions but also crucial for the development of new drugs and repurposing of old ones. Conventional methods for prediction of such interactions can be divided into 2D based and 3D based methods. 3D methods are more CPU expensive and require more manual interpretation whereas 2D methods are actually fast methods like machine learning and similarity search which use chemical fingerprints. One of the problems of using traditional machine learning based method to predict drug-target pairs is that it requires a labeled information of true and false interactions. One of the major problems of supervised learning methods is selection on negative samples. Unknown drug target interactions are regarded as false interactions, which may influence the predictive accuracy of the model. To overcome this problem network based methods has become an effective tool in predicting the drug target interactions overcoming the negative sampling problem. In this dissertation study, I will describe traditional machine learning methods and 3D methods of pharmacophore modeling for drug target prediction and will show how these methods work in a drug discovery scenario. I will then introduce a new framework for drug target prediction based on bipartite networks of drug target relations known as Random Walk with Restart (RWR). RWR integrates various networks including drug– drug similarity networks, protein-protein similarity networks and drug- target interaction networks into a heterogeneous network that is capable of predicting novel drug-target relations. I will describe how chemical features for measuring drug-drug similarity do not affect performance in predicting interactions and further show the performance of RWR using an external dataset from ChEMBL database. I will describe about further implementations of RWR approach into multilayered networks consisting of biological data like diseases, tissue based gene expression data, protein- complexes and metabolic pathways to predict associations between human diseases and metabolic pathways which are very crucial in drug discovery. I have further developed a software tool package netpredictor in R (standalone and the web) for unipartite and bipartite networks and implemented network-based predictive algorithms and network properties for drug-target prediction. This package will be described.