Diabetes-compound Relationship Identification based on Complex-valued Flexible Neural Tree and Negative Sample Selection Algorithm

Abstract

Background: Virtual screening (VS) could select possible effective candidates from a large number of organic compounds, which plays an important role in network pharmacology. Virtual screening is a very important step in network pharmacology. Objective: The accuracy of screening compounds directly determines the subsequent network construction, target determination and pathway analysis. In order to improve the accuracy of screening the important compounds in herbs for treating diabetes, a novel methodology based on complex-valued flexible neural tree (CVFNT) model and negative sample selection algorithm is presented. Methods: In our method, diabetes-related targets were obtained by literature search. According to diabetes-related targets, active compounds were searched from the public database. The negative sample selection algorithm based on Tanimoto index was proposed to establish inactive compound set. The CVFNT model optimized was utilized to screen effective candidate compounds. Result: Our proposed method performs better than eight classical classifiers in terms of TPR, FPR, Precision, Specificity, F1, AUC and ROC curve. Our method could also predict 18 compounds from Liangxue Sanyu Decoction, which are involved in the treatment of diabetes.