Using Artificial Neural Networks to Predict Biological Activity from Simple Molecular Structural Considerations
Frank R. Burden · Quantitative Structure-Activity Relationships · 1996
Abstract Some simple molecular structural considerations relating to atom type were used as the independent variable inputs to an artificial neural network with the dependent variables consisting of the physicochemical parameters molecular refractivity and hydrophobicity. The low root mean squared error in each case was sufficiently low for the further mapping of biological activity to be attempted. A set of 236 dihydrofolate reductase inhibitors, for which the biological activity was known, was fitted in a similar manner and again producing a low root mean squared error. It is concluded that neural networks can be used to predict biological activity, within a series of closely related molecules, from molecular structural considerations alone so saving much effort in synthesis and in vivo testing with new candidate molecules.