Modelos QSPR/QSAR/QSTR basados en sistemas neuronales cognitivos

Identifier:  TDX:287

Handle: https://hdl.handle.net/20.500.11797/TDX287

Authors:  Espinosa Porragas, Gabriela

Abstract:
One of the most attractive applications of computer-aided techniques in molecular modeling stands on the possibility of assessing certain molecular properties before the molecule is synthesized. The field of Quantitative Structure Activity/Property Relationships (QSAR/QSPR) has demonstrated that the biological activity and the physical properties of a set of compounds can be mathematically related to some 'simple' molecular structure parameters. <br/><br/>Artificial neural network (ANN) approaches provide an alternative to established predictive algorithms for analyzing massive chemical databases, potentially overcoming obstacles arising from variable selection, multicollinearity, specification of important parameters, and sensitivy to erroneous values. In most instances, ANN's have proven to be better than MLR, PCA or PLS because of their ability to handle non-linear associations. <br/><br/>In the last years there has been a growing interest in the application of neural networks to the development of QSAR/QSPR. The mayor advantage of ANN lies in the fact QSAR/QSPR can be developed without having to a priori specify an analytical form for the correlation model. The NN approach is especially suited for mapping complex non-linear relationships that exists between model output (physicochemical or biological properties) and input model (molecular descriptors). The NN approach could also be used to classify chemicals according to their chemical descriptors and used this information to select the most suitable indices capable of characterize the set of molecules. Existing neural networks based QSAR/QSPR for estimating properties of chemicals have relied primarily on backpropagation architecture. Backpropagation 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 fuzzy ARTMAP cognitive system, for unsupervised learning of categories, which represent structure and properties simultaneously. This class of neural networks uses a match-based learning, in that it actively searches for recognition categories or hypotheses whose prototype provides an acceptable match to input data. <br/><br/>The current study have been proposed a new QSAR/QSPR fuzzy ARTMAP neural network based models for predicting diverse physical properties such as phase transition temperatures (boiling and melting points) and critical properties (temperature and pressure) and the biological activities (toxicity indicators) of diverse set of compounds. In addition, traditional pre-screening methods to determine de minimum set of inputs parameters have been compared with novel methodology based in self organized maps algorithms. <br/><br/>The most suitable set of molecular descriptor was obtained by choosing a representative from each cluster, in particular the index that presented the highest correlation with the target variable, and additional indices afterwards in order of decreasing correlation. The selection process ended when a dissimilarity measure between the maps for the different sets of descriptors reached a minimum valued, indicating that the inclusion of more descriptors did not add supplementary information. The optimal subset of descriptors was finally used as input to a fuzzy ARTMAP architecture modified to effect predictive capabilities. <br/><br/>The proposed QSPR/QSAR model predicted physicochemical or biological activities significantly better than backpropagation neural networks or traditional approaches such as group contribution methods when they applied.