Designing QSAR Models for Promising TLR4 Agonists Isolated from Euodia Asteridula by Artificial Neural Networks Enhanced by Optimal Brain Surgeon
Abstract
Undoubtedly, computer aided drug design (CADD) has gained important position in medicinal chemistry thanks to balancing random approaches to discovery of new drugs by prioritizing rational insight into the development process. From many CADD methods, quantitative structure activity relationships (QSAR), which are able to exploit chemical and biological information hidden in chemical structures through utilization of numerous machine learning and artificial intelligence methods, are expected to provide the necessary assistance in mechanistic interpretation and prediction of biological activities.
In the present work, 56 derivatives of a natural adjuvant euodenine A, which occurs in Euodia asteridula, were selected for a QSAR study with the use of artificial neural networks (ANN). Since building of robust QSAR models is still a challenging research area, several methods had to be utilized to achieve a robust solution.
Among various backpropagation based algorithms, much effort has been devoted to research of an optimal brain surgeon (OBS) method, which attempts to prune unimportant ANN elements according to the second derivation of the output signal error with respect to the weights. Herein, the performance of OBS in QSAR analyses is discussed and compared with other ANN learning methods.