Dual Self-Adaptive Intelligent Optimization of Feature and Hyperparameter Determination in Constructing a DNN Based QSPR Property Prediction Model

The octanol-water partition coefficient (K-ow) is an extremely important and widely used parameter for the study of the distribution and balance of organic pollutants in environmental media. Therefore, the theoretical determination and prediction of such a property are vital in environmental chemistry. The past research on the use of models based on the quantitative structure- property relationship (QSPR) for the estimation of K-ow has significant problems such as data redundancy and computational complexity in the molecular description. In this work, the genetic algorithm is coupled with the random forest algorithm to select the most suitable molecular feature combination from a vast number of features. As a consequence, the number of descriptors for developing the model recommended in available models is significantly reduced. Moreover, a combination of the backpropagation neural network and Bayesian optimization allows the development of an intelligent procedure for tuning the relevant model parameters. On the basis of comparison of the obtained estimations to the results of the available QSPR models in the literature, the developed model in this work shows considerably higher accuracy and predictability.

Automated summary

This study combines genetic algorithm with random forest algorithm to select suitable molecular feature combination, significantly reducing the number of descriptors required for model development, and utilizes a combination of backpropagation neural network and Bayesian optimization for intelligent tuning of relevant model parameters.