Topology Preserving Neural Networks for Peptide Design in Drug Discovery

Abstract

We describe a construction method and a training procedure for a topology preserving neural network (TPNN) in order to model the sequence activity relation of peptides. The building blocks of a TPNN are single cells (neurons) which correspond one-to-one to the amino acids of the peptide. The cells have adaptive internal weights and the local interactions between cells govern the dynamics of the system. The TPNN can be trained by gradient descent techniques, which rely on the efficient calculation of the gradient by back-propagation.1 Introduction Building artificial neural networks with problem specific topology has a long tradition. In particular for pattern recognition the topology of the retina served as a template for the famous Perceptron (Rosenblatt [7][8][9]. In a MGN a compound can be described as graph where each atom is a node and each chemical bond is an edge. In this way a chemical structure is translated into cellular network topology: Each atom becomes a cell, each bond a local interaction between the cells and the weights depend on element and bond types. The MGN works in this case as a discrete time cellular neural network [10]. In the current work we propose a simpler strategy that is dealing with the particular structure of peptides. Instead of building an MGN based on the atoms of the peptide, we use the amino acids as building blocks. This has two reasons: First of all, it reduces the complexity of the problem. MGNs work fine for small molecules but in the case of peptides there are on average 16 Atoms per amino acid which leads to large MGNs even for small peptides. As a consequence learning rates are quite poor and the adaption of weights is slow and time consuming. The second reason is the intrinsic structure of the problem. Our method should optimize the properties of peptides with respect to several pharmacological aspects. We suggest several new peptides to be synthesized and tested during the modelling process described in this paper. In this terms it is straightforward to build models from amino acids rather than from atoms of molecules. The TPNN should work as a tool for peptide design in drug discovery. Therefor it is necessary, that it catches some fundamental items of the underlining structure activity relation. It is very difficult to run any statistical method that could learn a correlation from a few datapoints in such a high dimensional space (there are for example 10 20 different possible decapeptides from the 20 natural amino acids). Many attempts have been made to utilize Artificial Neural Networks in order to generate predictive models of quantitative sequence activity relationships between a set of molecular descriptors and activity. An overview could be found in Weekes and Fogel [11] or in Terfloth and Gasteiger [12] Earlier works from Schneider and Wrede [13] proposed the use of neural networks and computerbased evolutionary search for peptide design. Their work was criticized by Darius and Rojas [14]

De Novo Generation of Hit-like Molecules from Gene Expression Signatures Using Artificial Intelligence

“…Techniques such as virtual screening aim to identify hits from virtual libraries containing large number of molecules usually by similaritybased searches or by molecular docking 4,6,7 . Another technique is automated molecular generation or automated de novo design where new molecules with specific properties are automatically generated by methods such as inverse QSAR or genetic algorithm [8][9][10][11] . Recently, artificial intelligence, in particular generative models, has been extensively used for molecular de novo design, compound optimization and hit identification 12,13 .…”