Inhibitor prediction by machine learning approaches

YAO LIXIA

Publication

Inhibitor prediction by machine learning approaches

YAO LIXIA

Abstract

In this thesis, I explored three widely used algorithms from machine learning community to facilitate inhibitor prediction for three pharmacologically important proteins. The aim is to evaluate the feasibility of introducing these machine learning approaches to lead identification and its ADME/toxicity properties analysis.Specifically, I have worked on the inhibitor/antagonist prediction for therapeutic targets--5-HT2 and cholinesterase, and an ADME associated protein--CYP3A4. The machine learning approaches I use include decision tree, k-nearest neighbor and support vector machine, and preprocessing techniques such as normalization and principal component analysis.The entire flowchart includes five steps. First, the examples of the inhibitors/antagonists of the three protein targets are collected from the available references. Then the 3D structures of the compounds are transformed using QSAR molecular descriptors to numerical vectors, which are recognizable to the machine learning procedures. After some data preprocessing techniques, different machine learning models are derived from the data sets. In the end, the prediction capacity of these models is evaluated.Experimental results on all the three data sets demonstrate support vector machine beats decision tree and k-nearest neighbor and gives good results. Therefore, support vector machine may be a promising approach to analyze pharmaceutical properties of chemical compounds, and may lead to a new practical tool for drug design in near future. In addition, principal component analysis also shows its usefulness in dimensionality reduction, when used with support vector machine.