Abstract

Quantitative structure-activity relationships (QSAR) are statistical models that may predict various physicochemical and biological properties of chemical compounds. Typical QSAR models use as input molecular descriptors computed from the chemical structure of the compounds in the dataset. QSAR models are based on algorithms or mathematical functions that correlate these molecular descriptors with the experimental property that is modeled. A different approach is explored in molecular graph machines (MGM) which represent a class of QSAR models that actively consider the molecular topology in the process of generating a structure-property model. After a review of the major MGM models, we present a detailed overview of the artificial neural network MolNet, which is a multilayer perceptron that encodes the molecular topology of each chemical presented to the network during learning or prediction. Each nonhydrogen atom in a molecule has a corresponding neuron in the input and hidden layers, whereas the output layer has only one neuron which provides the computed molecular property. The connections between the input and hidden layers encode the topological distance matrix of a molecule, whereas the connections between the hidden and output layers are classified according to atom types. Connection weights corresponding to the same topological distance or to the same atom type have a constant value for all chemicals in the training set. A MolNet application is presented for the glycogen synthase kinase-3β (GSK – 3β) inhibition by aloisines.

Keywords: Quantitative structure-activity relationships, QSAR, molecular graph machines, MGM, glycogen synthase kinase 3β, GSK, –, 3β, graph mining, topological indices, artificial neural network, Artificial neural networks, MolNet Input Indices, HOMO and LUMO energies