Classification and Similarity Analysis of Binding-Database: A Survey on Application of Multi-Class Classifiers for Deriving General Rules from Large Compound Databases

Abstract

Background: In this research, we extracted and modified features of active ligands related to specific biological targets with combination of data mining and classification methods to aid medicinal chemists in their drug discovery projects. Preparing an inactive ligand is the major problem for development of multi-class classifiers. Therefore, our models were developed based on only active ligands found in Binding-database (DB) without any needs for preparing inactive molecules. Methods: Our database consisted of 160372 ligands in 45 classes of common proteins and 1497 different features (topological, chemistry, physical, etc.) were calculated for each molecule. Then, the specific features of active ligands of any target were extracted based on combination of linear discriminate analysis and Apriori algorithm. Findings: Receiver operating characteristic (ROC) was a useful operator to analysis the accuracy and sensitivity of classification models and retrieving molecules from ZINC and Binding-DB databases. Area under curve (AUC) of this diagram was evaluated for analysis of each target in Zinc and Binding-DB and their results were 0.8341 ± 0.1495 and 0.8615 ± 0.1502, respectively. Conclusion: Specific features of active ligands could be found using the methodology described in this work and with these features, we can sort each database based on corresponding target. AUC shows that the present method is useful for virtual screening in big databases without survey on inactive ligands.

Keywords

• Virtual systems
• Multiple classification
• Data mining
• Database management systems
• Ligands