Prediction of protein-ligand binding affinity using neural networks

Abstract

A big problem in the life-sciences is the ability to calculate, in-silico, the binding affinity between a protein active site and a lead-ligand. This thesis introduces a new method to predict the binding affinity of a given drug ligand and active site, using backpropagation neural networks of 128 protein ligand complexes, with electrostatic, hydrogen bonding and molecular weight parameters. The parameters are given space and magnitude consideration, through the use of physico-chemical autocorrelation for the preparation of the input parameters. Self-Organizing Maps(SOM) are used as well to visualize the distribution of the input cases in similarity space. The results showed an improvement in accuracy over multiple regressive and the BLEEP method for calculation of binding affinity, using Root Mean Square, Relative Root Mean Square, Mean Absolute and Relative Mean Absolute Error calculations. The SOM additionally showed positive clustering of protein-ligand complexes, from similar families spread through the input space.