3D Structural Information: From Property Prediction to Substructure Recognition with Neural Networks

Abstract

Two applications of neural networks in molecular recognition, incorporating at different levels the structural information, are presented. Interatomic distances are the basis of the search for a given 3D substructure, with a Hopfield network using either a Boltzmann machine or a « mean field annealing » algorithm (according to a model we previously proposed by analogy with the « travelling salesman problem »). Besides atom spatial locations, the model can incorporate characteristic points featuring selected electronic or steric features, and add supplementary constraints on the nature of these points or some property value on them. For model compounds, this approach retrieves the correct (flipped) orientations in binding the adenosine A1 receptor. In QSAR field, we use a three layers feed forward neural network to predict the activity of polychlorinated dibenzofurans toward the AcH receptor. Due to the high homogeneity of the studied population input data only consist here of a topological descriptor, refined by a cross validation process. Results compete favorably with the previous approaches, with no need for complex field calculations.

Key Words:

• Hopfield network
• molecular recognition
• 3D substructure
• activity prediction
• chlorinated dibenzofurans