Formation of a drug classification framework via data mining techniques

Virtual screening of candidate drug molecules via machine learning methods plays a key role in pharmaceutical industry to prevent adverse effects of the drugs. Computational classification methods can distinguish approved drugs from withdrawn ones. In this study, we focused on 3 various applications on drugs. We studied with different machine learning strategies to distinguish approved and withdrawn drugs. To begin with, 760 molecular descriptors such as ToxPrint Chemotype, global molecular, size and shape were calculated to study classification and feature selection problems for each drug molecule in this study. In first application, SVM and ensemble methods were applied on drug data sets to categorize more than 400 drugs belonging to nervous system and various disease groups as approved or withdrawn. Accuracy rates were found between 0.74 and 0.89 for data sets. Here, feature selection methods which were applied on drug data sets increased classification performance values. The number of total chemotypes, bond CN_amine_aliphatic_ generic, XlogP, aspheric: Cor3D:ori1ve Bonds descriptors were found as more significant descriptors to form model for nervous system drugs. Moreover, the fragmans located in minimum 60 % of nervous system withdrawn drugs were determined via application of gSpan algorithms on drug data sets. This is the first report that describes distinction of withdrawn and approved drugs pertaining to the spesific disease on the data sets. In the second part of study, 558 drugs with various disease groups were classified in 3 basic levels with hierarchical multi-label classification via Clus-HMC-Ens algorithms. While first level includes all drugs, second level consists of 3 groups of drugs. These are approved nerveous system drugs, approved drugs of various disease groups and withdrawn drugs. Last level has drugs of 5 different groups according to Anatomic Therapeutic Chemical classification of nerveous system drugs. In this application, some paremeters were selected for classification of drugs hierarchically. Selected paremeters such as FTest, w_0, k, classification treshold, m-estimate increased estimation performance of model. In last part of study, more than 1200 approved and withdrawn drugs were studied. Molecular identifiers that are effective in classification models have been identified by an effective feature selection strategy proposed in the thesis. ToxPrint chemotypes, effective descriptors, were used for determination of a number of rules in drug molecules. Available/unavailable chemotypes were analysed in approved/withdrawn drugs on drug data sets. While chemotypes such as bond:NN_hydrazine_alkyl_HH2 only presented in withdrawn drugs, ones such as bond:P=O_phosphorus_oxo,bond:PC_phosphorus_organo_generic,group:carbohydrate_aldohexose, group:carbohydrate_aldopentose, group:carbohydrate_hexopyranose _fructose, group:carbohydrate_hexopyranose_glucose etc. just examined in approved drugs. A model for classifier ensemble design was proposed on the unbalanced drug data set. Accuracy of 0.80 was obtained for the test set in order to classify the drugs as approved and withdrawn. Developed model in this study can be used as a simple filter in drug modelling to eleminate drug candidate molecules.