The Use of Self-organizing Neural Networks in Drug Design

Analusis, 1998

Self Organising Map (SOM), also known as Kohonen Neural Network, is tested as a non supervised procedure for comparing molecular databases. Each chemical compound being represented by a point in the hyperspace of the molecular descriptors, SOMs was used to reflect the multidimensional hyperspace onto a two dimensional (2D) map while preserving the order of distances between the points, but in a non linear way. The aim of this work was to apply SOM to the study of the overlapping of two databases in order to obtain information about the extent of their differences in regard to their molecular diversity. Firstly, the ability of SOM to discriminate between two virtual databases was investigated. The positions of these two virtual databases were made to vary from non-overlapping to overlapping ones. In any considered cases, all the individuals of these two databases are processed simultaneously to give one SOM. From this map it is possible to analyse and understand the structure of the original data. Secondly two chemical databases are compared. The first chemical database deals with the commercially available organophosphorous pesticides (OPC), the second one deals with more than two thousand OPC tested as potent pesticides. Given the biological data known for each compound, the second database was shown to bring an interesting supplement to the structural information nested in the first database taken as a reference. Furthermore, the results obtained indicate that SOM can be used for the search of new leads among available databases and the exploration of new structural domains for a given biological activity. Key words. Kohonen neural network − self organizing map − classification − chemical databases − virtual screening − pesticide − organophosphorous compounds.

Angewandte Chemie International Edition in English, 1993

The capabilities of the human brain have always fascinated scientists and led them to investigate its inner workings. Over the past 50 years a number of models have been developed which have attempted to replicate the brain's various functions. At the same time the development of computers was taking a totally different direction. As a result, today's computer architectures, operating systems, and programming have very little in common with information processing as performed by the brain. Currently we are experiencing a reevaluation of the brain's abilities, and models of information processing in the brain have been translated into algorithms and made widely available. The basic building-block of these brain models (neural networks) is an information processing unit that is a model of a neuron. An artificial neuron of this kind performs only rather simple mathematical operations; its effectiveness is derived solely from the way in which large numbers of neurons may be connected to form a network. Just as the various neural models replicate different abilities of the brain, they can be used to solve different types of problem: the classification of objects, the modeling of functional relationships, the storage and retrieval of information, and the representation of large amounts of data. This potential suggests many possibilities for the processing of chemical data, and already applications cover a wide area: spectroscopic analysis, prediction of reactions, chemical process control, and the analysis of electrostatic potentials. All these are just a small sample of the great many possibilities. AnKen,. C'hiw. Inr. Ed. EngI. 1993. 32, 503 -521 :(> VCH Veriugsgrce//.~th~Jf mhH, W-6940 Weinheim, 1993 0570-0#33/93/0404-05f)3 R 10.00+ .Xi(l Johann Gasteiger was born in Dachau in 1941. He studied chemistry at Munich and Zurich universities with a doctoral thesis (1971) on "Mechanistic Investigations into Reactions in the Cyclooctatetraene System" under R. Huisgen. In his post-doctorate studies (1971 -1972) he carried out ab initio calculations on carbanions under A . Streitwieser. Jr. at the University of California in Berkeley. In 1972 he transferred to the Technical University of Munich where he developed, along with a group led by I. Ugi, a protot.ype for a synthesis-planning program. He received his "habilitation" in chemistry in 1978 w>ith a study on models and algorithms for investigating chemical reactions and reactivity. Since 1989 he has been professor at the Technical University of Munich. His main area of research is the development of methods and computer programs ,for predicting reactions and planning syntheses, for evaluating and simulating mass spectra, andfor the three-dimensionalmodeling of molecules. From 1987-1991 he was the project manager for the Fachinformationszentrum Chemie in the implementation of a reaction database based on Chemlnform, and in 1991 was awarded the Gmelin-Beilstein memorial medal from the German Chemical Society for his achievements in the ,field of computer chemistry. Jure Zupan was born in Ljubljana, Slovenia in 1943. He studied physics at the University of Ljubljana, where he received his doctorate in 1972 under D. Hadzi with a thesis on the energy bands in boronitrides. Until 1973 he worked at the Josef Stefan Institute in the field of quantum chemistry and the magnetic properties ofceramic materials. Since 1974 he has led a group at the Institute,for Chemistry in Ljubljana. His areas of work include chemometrics, artificial intelligence, and the development qf expert systems and algorithmsfor chemical applications. In 1982 he was visiting professor at Arizona State UniversitylUSA, in 1988 at the Vrije Universiteit in Brussels, and 1990-1992 at the TU Miinchen in Garching. For his research work (approximately 150 original publications, 2 monographs, and 3 books) he received the highest Slovenian award for research in 1991. He obtained his habilitation in 1975 and has been Professor of Chemometry at the University of Ljubljana since 1988. 504

Russian Journal of Bioorganic Chemistry, 2001

A volume learning algorithm for artificial neural networks was developed to quantitatively describe three-dimensional structure-activity relationships using as an example N -benzylpiperidine derivatives. The new algorithm combines two types of neural networks, the Kohonen and the feed-forward artificial neural networks, which are used to analyze the input grid data generated by the comparative molecular field approach. Selection of the most informative parameters using the algorithm helped reveal the most important spatial properties of the molecules, which affect their biological activities. Cluster regions determined using the new algorithm adequately predicted the activity of molecules from a control data set.

Journal of Chemical Information and Modeling, 1997

A scheme of a neural device intended for searching direct correlations between structures and properties of organic compounds without preliminary computation of molecular descriptors (that are invariant with respect to renumbering atoms in a molecule) is suggested. The invariance of a property with respect to renumbering atoms in a molecule is ensured by the architecture of the neural device, which is constructed by analogy with biological vision systems. A model software of the neural device was tested on several examples. The descriptive and predictive performances of the device are shown to be comparable and even overcome the performances of using molecular descriptors, such as topological indexes and substructural descriptors, especially for analyzing heterogeneous data sets including inorganic compounds. The neural device can be advantageously used in the cases when more traditional approaches fail to work or "good" molecular descriptors have not been devised yet.

Current Computer Aided-Drug Design, 2005

We present self-organizing map or Kohonen network and counter propagation neural network as powerful tools in quantitative structure property/activity relationship modeling. Two areas of applications are discussed: estimation of toxic properties in environmental research and applications in drug research.

Journal of Chemometrics

Artificial neural networks (ANNs) are comparatively straightforward to understand and use in the analysis of scientific data. However, this relative transparency may encourage their use in an uncritical, and therefore possibly unproductive, fashion. The geometry of a network is among the most crucial factors in the successful deployment of network tools; in this review, we cover methods that can be used to determine optimum or near-optimum geometries. These methods of determining neural network architecture include the following: (i) trial and error, in which architectures chosen semirandomly are tested and modified by the user; (ii) empirical or statistical methods, in which an ANN's internal parameters are adjusted based on the model's performance; (iii) hybrid methods, such as fuzzy inference; (iv) constructive and/or pruning algorithms, that add and/or remove neurons or weights from an initial architecture, respectively, based on a predefined link between architecture and ANN performance;

Pharmaceutical Chemistry Journal, 2001

One of the basic prerequisites in drug design is the assumption that compounds possessing like structures exhibit similar types of biological activity. However, it is very difficult to strictly define what we mean by structural resemblance -and this is evidenced by the large number and diversity of methods and models used in establishing quantitative structure -activity relationships (QSAR). The approaches to solving QSAR can be classified depending on the methods used to reveal regularities and describe the structures [1]. The simplest models are formulated in terms of the particular structural fragments of molecules (descriptors) or certain physicochemical parameters of substituents (lipophilicity, charge, etc.) and topological indexes. In more involved 3D models, such as CoMFA or the Marshall method of active analogs [3], the main role belongs to characteristics reflecting the features of intermolecular interactions of the compounds studied .

Bioorganicheskaia khimiia

A volume learning algorithm for artificial neural networks was developed to quantitatively describe three-dimensional structure-activity relationships using as an example N -benzylpiperidine derivatives. The new algorithm combines two types of neural networks, the Kohonen and the feed-forward artificial neural networks, which are used to analyze the input grid data generated by the comparative molecular field approach. Selection of the most informative parameters using the algorithm helped reveal the most important spatial properties of the molecules, which affect their biological activities. Cluster regions determined using the new algorithm adequately predicted the activity of molecules from a control data set.

Russian Chemical Reviews, 2003

The published data devoted to the use of the neural The published data devoted to the use of the neural network approach in the simulation of structure ± property rela-network approach in the simulation of structure ± property relationships for organic compounds are reviewed. The basic princi-tionships for organic compounds are reviewed. The basic principles of the neural network simulation are discussed along with the ples of the neural network simulation are discussed along with the characteristic features of the neural network approach typical of characteristic features of the neural network approach typical of the representation and classification of structural chemical data. the representation and classification of structural chemical data. Brief information on neural network models of spectral character-Brief information on neural network models of spectral characteristics, reactivities, physicochemical properties and biological istics, reactivities, physicochemical properties and biological activities of organic compounds is presented. The bibliography activities of organic compounds is presented. The bibliography includes 159 references includes 159 references. .

2002

Abstract There are many examples where neural networks have been effectively used to predict protein secondary and tertiary structure from the primary sequence data. We describe the use of a Kohonen self-organizing map (SOM) to categorise proteins based on secondary structure, and attempt to relate this information to functional data