CNN in drug design — Recent developments

We describe a method for construction of specific types of Neural Networks composed of structures directly linked to the structure of the molecule under consideration. Each molecule can be represented by a unique neural connectivity problem (graph) which can be programmed onto a Cellular Neural Network. The idea was to translate chemical structures like small organic molecules or peptides into a self learning environment which is CNN based. In the case of small molecules, each cell of the CNN stands for one atom of the molecule under consideration. But in contrast to the standard CNN architecture where each cell is connected to the neighboring cells, only those cells of the feature net are connected for which there also exists a chemical bond in the molecule under consideration. This implies that the feature net topology varies from molecule to molecule. In the case of peptides, the amino acids that form the building blocks of the peptide are reflected by the CNN cells wherein the amino acid sequence defines the network topology. Unlike the standard CNN used for image processing, there are no input values like the input image that are fed into the feature net. Instead, all information about the input molecule is supplied to the feature net by means of the topology. The output of several feature nets is fed into a supervisor neural network which computes the final output value. The combination of several feature nets and a supervisor networks constitutes the Molecular Graph Network (MGN). The designed networks are used for selection of molecules representing wanted properties such as activity against specific diseases, interactions with other compounds, toxicity etc. and possibly being candidates to be tested further as new drugs.