Deep neural network in QSAR studies using deep belief network

Abstract

There are two major challenges in the current high throughput screening drug design: the large number of descriptors which may also have autocorrelations and, proper parameter initialization in model prediction to avoid over-fitting problem. Deep architecture structures have been recommended to predict the compounds biological activity. Performance of deep neural network is not always acceptable in QSAR studies. This study tries to find a solution to this problem focusing on primary parameter computation. Deep belief network has been getting popular as a deep neural network model generation method in other fields such as image processing. In the current study, deep belief network is exploited to initialize deep neural networks. All fifteen targets of Kaggle data sets containing more than 70k molecules have been utilized to investigate the model performance. The results revealed that an optimization in parameter initialization will improve the ability of deep neural networks to provide high quality model predictions. The mean and variance of squared correlation for the proposed model and deep neural network are 0.618±0.407e−4 and 0.485±4.82e−4, respectively. The outputs of this model seem to outperform those of the models obtained from deep neural network.