Polymer and Separations Research Laboratory

(PolySep)

Last update:

08/30/2006

Combinatorial Cognitive Neural Networks for Drug Discovery and Chemical Property Estimations

The development of new chemicals and synthetic drugs require understanding of the role of various chemical functional groups on the activity of various drugs and the impact of such functional groups on the physicochemical properties of such chemicals and drugs (e.g., aqueous solubility and lipid phase partitioning). The ability to effectively utilize and develop libraries of chemicals for drug discovery relies to a large extent on the ability to estimate the likely physicochemical and bioactivity of potential drugs. Also, synthetic chemicals that are being manufactured is growing exponentially and it is infeasible to experimentally determine all of the required physicochemical properties of these chemicals for process design applications and environmental assessment of the likely impact of such chemicals. Therefore, over the last two decades, various methods have been developed to estimate physicochemical properties and bioactivity of synthetic chemicals and therapeutic drugs. In particular, in recent years there has been a growing interest in the application of neural networks to the development of quantitative structure property relations (QSPRs) for the correlation and estimation of physical properties of organic compounds (1, 2). The premise of QSPRs is that physicochemical properties can be quantitatively related to  geometric and electronic molecular descriptors. The major advantage of neural networks lies in the fact that QSPRs can be developed without having to a priori specify an analytical form for the correlation model. The neural network approach is especially suited for mapping complex non-linear relationships that exist between model output (i.e., physicochemical properties) and model input (i.e., molecular descriptors). The neural network (NN) approach could also be used to classify chemicals according to their propensity for persistence in the environment.

Existing neural network based QSPRs for estimating physicochemical properties have relied primarily on back-propagation architecture. Back-propagation neural networks are an error-based learning system in which adaptive weights are dynamically revised so as to minimize estimation errors of target values. However, since chemical compounds can be classified into various structural categories, it is also feasible to use cognitive classifiers, such as the Fuzzy ARTMAP network, for rapid unsupervised learning of categories which represent structure and physicochemical properties. This class of neural networks uses a match-based learning, in that it actively searches for recognition categories or hypotheses whose prototype (expectations) provides an acceptable match to input data.

In this present project, a fuzzy ARTMAP and back-propagation neural network will be used to expand previous efforts in our laboratory to develop functional QSPRs to estimate a variety of physicochemical properties of organic compounds including, viscosity, surface tension, vapor pressure, octanol/water partition coefficients  The development of these QSPRs will rely on the generation of topological and quantum chemical descriptors from graph theory and molecular modeling based on PM3/MO-theory calculations. Comparisons of the neural network-based QSPRs with other published estimation methods will be carried out to demonstrate the power of the approach. The outcome of this research will be of significance to the field of combinatorial chemistry with applications to data mining, pattern recognition and chemometric analysis in drug discovery.  

Fundamental areas: neural networks, cognitive systems, molecular modeling, combinatorial chemistry, bioinformatics.

References:
1. Espinosa, G., D. Yaffe, Y. Cohen, A. Arenas and F. Giralt, "Neural Network Based Quantitative Structural Property Relations (QSPRs) for Predicting Boiling Points of Aliphatic Hydrocarbons," Journal of Chemical Information and Computer Sciences, V40(N3):859-879 (2000).

2. Espinosa, G., D. Yaffe, A. Arenas, Y. Cohen and F.  Giralt, "A Fuzzy ARTMAP based Quantitative Structure-Property Relationships (QSPRs) for predicting Physical Properties of Organic Compounds", I&EC Research, 40, 2757-2766 (2001).

3. Giralt, F., A. Arenas, G. Espinosa, Z. Girones, E. Besalu, R. Caro-Dorca and Y. Cohen, "QSAR Based on Cognitive Neural Systems for Anti-HIV Activity Derivatives", AIChE Meeting, Indianapolis, November 3-8. 2002). 

Vapor Pressures a Heterogeneous Set of Organics: Experimental and Neural Network Predictions:

________________________________________________________________

________________________________________________________________________________________

Solubility Predictions Using Different Structure-Based Approaches:

___________________________________________________________________

_____________________________________________________________