Prediction of Binding Affinities for Protein-Ligand Complexes with Neural Network Models

Abstract

Molecular mechanisms of drug action are often based on an interaction of them with target macromolecules, such as proteins and nucleonic acids. The formation of ligandtarget complexes is typical for biologically active compounds, including activators and inhibitors of various enzymes. Prediction of the dissociation constant (Kd) of protein-ligand complex is often used as a scoring function for the modeled complexes and there are many approaches in the field of prediction of such constant [1]. In the present work various parameters of protein-ligand complexes were used to predict Kd. These parameters can be quickly calculated immediately during docking procedure, which we usually used for complexes modeling. The artificial feedforward neural networks (AFNNs) were used as a mathematical approach to prediction of protein-ligand complexes Kd. In practice, neural networks are especially useful for classification and function approximation problems, which have a lot of training data. Neural networks are often used in situations where you do not have enough prior knowledge to set the activation function, as in case of the prediction of the proteinligand complexes dissociation constant. The Kd values for 83 various complexes of biological molecule [2] were used in the present work. All of these complexes have crystallographic coordinates of 3D structures. This set of complexes was divided randomly into two subsets. The training set includes 68 points and the test one includes 15 points. Hereinafter Kd will appear as the predicted values. The crystallographic data for all complexes passed preliminary handling according the uniform scheme by using program suite Sybyl [3]:• rebuild hydrogen atoms in molecules;• remove crystallographic water molecules (except the case of HIV protease, where one molecule of water was accepted as an element of a ligand); • check and correct types of atom and bond;• solvate the complexes;• optimize the structure of complexes in the water environment. Estimation of following parameters were done for all of 83 complexes: 1. The number of atoms in target and ligand part of complex. 2. The value of energy due to electrostatic interactions [3]. 3. The attitude of the closed surface in a complex to the full surface which is accessible for water (sphere radius 1.4 Å) in unbound molecule. These parameters were estimated both for ligand and target parts of complex [4]. 4. Analogous parameters estimated by using sphere with 0.5 Å radius [4]. 5. The changes of integral parameters of hydrophily and lipophilicity and the changes S.