2-D Molecular Descriptor Screening and Quantitative Structure-Activity Relationship Modeling for Estrogen Receptors Alpha Bioactivity

With the continuous development of data mining in the field of computational biology and bioinformatics in recent years, it is possible to develop better quantitative prediction models based on these methods. Quantitative structure-activity relationship (QSAR) of compounds can be modeled using molecular descriptors and bioactivity data of compounds, so as to predict new compound molecules with better bioac-tivity, or guide the structural optimization of existing active compounds. In the molecular descriptors of compounds, 2-D descriptors can be easily obtained through COTS toolbox, so as to avoid some complex descriptor calculation methods. However, 2-D descriptors also have the problems of over fitting and poor model interpretation. In this paper, the recursive feature elimination based on random forest (RFE-RF) is used to sort 2-D molecular descriptors. In each round of recursion, RF is used to eliminate the features, and the molecular descriptors that have the greatest impact on bioactivity are selected round by round. RFE-RF significantly reduces the types of descriptors, not only improves the calculation speed of the prediction model, but also effectively avoids the common over fitting problem in computational biology. The selected molecular descriptors are used to model, train and verify RF model. The regression model can predict the bioactivity of compounds against Estrogen receptors alpha, and Estrogen receptors alpha is an important target for breast cancer treatment. In this paper, the predicted data and experimental data are compared and analyzed, and the correctness and applicability of the modeling are evaluated from the aspects of variance, standard deviation, mean absolute error and root mean square error.