BOND: Benchmarking based on heterogeneous biOmedical Network and Deep learning novel drug-target associations

Project Summary/Abstract The applicant?s goals are to develop the necessary skills to become an independent translational biomedical informatics researcher in the area of computational drug repurposing. Exploring novel drug-target interactions (DTI) plays a crucial role in drug development. In order to lower the overall costs and uncover more potential screening targets, computational (in silico) methods have become popular and are commonly applied to poly-pharmacology and drug repurposing. Although machine learning-based strategies have been studied for years, there is no standardized benchmark that provides large-scale training datasets as well as diverse evaluation tasks to test different methods. Furthermore, the existing methods suffer from remarkable limitations, where 1) results are often biased due to a lack of negative samples, 2) novel drug-target associations with new (or isolated) drugs/targets cannot be explored, and 3) the comprehensive topological structure cannot be captured by feature learning methods . Therefore, in the era of big data, the applicant proposes a study to tackle the challenges by achieving two aims. ? Aim 1 (K99 Phase): Develop a large scale benchmark for evaluating drug-target prediction based on the generation of a multipartite network from heterogeneous biomedical datasets. ? Aim 2 (R00 Phase): Adapt a deep learning model to build an accurate predictive model based on a novel feature learning algorithm that mines the multi-dimensional biomedical network (multipartite network). In the mentored phase, the applicant will integrate heterogeneous biomedical datasets and build a benchmark for evaluation of the drug-target prediction based on well-designed strategies. The applicant will receive training in standardization tools for data integration, tools, and skills for data management, evaluation methods for drug-target predictions, and state-of-the-art machine learning/deep learning methods in computer-aided pharmacology. Complementary didactic, intellectual, and professional training will help prepare the applicant for the R00 phase where he will develop a deep learning-based predictive model and multi-dimensional graph embedding methods for feature learning. Together, these novel studies will advance the current computational drug repurposing by providing 1) comprehensive benchmarking for testing and evaluation, and 2) a scalable and accurate predictive model based on a biomedical multi-partite network. The applicant will be mentored by senior, established investigators with substantial expertise in Semantic Web, computational biology, cancer genomics, drug development, and machine learning/deep learning. Importantly, this project will provide a foundation for the applicant to establish independent research programs in 1) computational drug repurposing in real cases, 2) investigation of the diverse hidden associations in system biology (e.g., associations between drugs, genetics, and diseases), and 3) precision medicine aimed applications leveraging biomedical knowledgebases and electronic health records.