NOVEL ALGORITHMS AND TOOLS FOR LIGAND-BASED DRUG DESIGN

摘要

Computer-aided drug design (CADD) has become an indispensible component in modern drug discovery projects. The prediction of physicochemical properties and pharmacological properties of candidate compounds effectively increases the probability for drug candidates to pass latter phases of clinic trials. Ligand-based virtual screening exhibits advantages over structure-based drug design, in terms of its wide applicability and high computational efficiency. The established chemical repositories and reported bioassays form a gigantic knowledgebase to derive quantitative structure-activity relationship (QSAR) and structure-property relationship (QSPR). In addition, the rapid advance of machine learning techniques suggests new solutions for data-mining huge compound databases. In this thesis, a novel ligand classification algorithm, Ligand Classifier of Adaptively Boosting Ensemble Decision Stumps (LiCABEDS), was reported for the prediction of diverse categorical pharmacological properties. LiCABEDS was successfully applied to model 5-HT1A ligand functionality, ligand selectivity of cannabinoid receptor subtypes, and blood-brain-barrier (BBB) passage. LiCABEDS was implemented and integrated with graphical user interface, data import/export, automated model training/ prediction, and project management. Besides, a non-linear ligand classifier was proposed, using a novel Topomer kernel function in support vector machine. With the emphasis on green high-performance computing, graphics processing units are alternative platforms for computationally expensive tasks. A novel GPU algorithm was designed and implemented in order to accelerate the calculation of chemical similarities with dense-format molecular fingerprints. Finally, a compound acquisition algorithm was reported to construct structurally diverse screening library in order to enhance hit rates in high-throughput screening.