Prediction Of ADME/T Properties With Biological Relevance Spectrum

Pharmacokinetic properties and possible toxicity are crucial for investigated new drug to get into market successfully. Traditionally, ADME/T did not draw enough attention, resulting in high failure rate of drug candidates in clinical trials. Theoretically predicting the ADME/T properties of drug candidates can improve the success rate of drug development and reduce spending and time wasted. Currently, based on computer modeling, several methods have been employed in ADME/T properties prediction, e.g. Quantitative Structure-Activity Relationships (QSAR) and machine learning.In this thesis, a multiple dimensional molecular descriptor, named biological relevance spectrum (BRS), was proposed to encode the molecular structures. BRS gives the distribution (coordinates) of the objective compound in biologically relevant chemical space, which is related to the biological activity, toxicity and other properties. Supporting Vector Machine (SVM) method was employed to establish the prediction models for human intestinal absorption (HIA), blood-brain barrier (BBB) penetration, sub-cellular distribution, acute toxicity and druggability of small compound.Firstly, prediction of pharmacokinetic properties. In this part, HIA and BBB classification models were built with SVM method, the overall prediction accuracies of models for testing set were97.96%and97.40%, and the overall prediction accuracy for the external sets were93.69%and68.29%. Then, drugs located in the nucleus and periplasm were used to develop a model for sub-cellular distribution, the overall prediction accuracies of this model for testing set was81.82%.Secondly, acute toxicity prediction models. With acute toxicity data of mouse by the intramuscular and rabbit by intravenous system in ToxicFinder datebase, acute toxicity prediction models were created. The coefficient of determination (R2) of the models for testing set were0.7110and0.7701.Thirdly, druggablity prediction. Two druggability prediction models were constructed with the approved drugs from Drugbank as positive samples and ACD compounds or drug candidates that failed to survive the drug development pipeline as negative samples. The overall prediction accuracies of models for testing set were81.42%and75.57%. The druggability of compounds in different drug discovery phases were predicted with the models. The results indicated that compounds at later development phases possess higher druggability, this is well dovetailed with the practical situation.In summary, BRS perform well in ADME/T properties prediction. We believe that BRS can be used for various other purposes, such as bioactivity estimation and other ADME properties.