Mining potential proteins as drug targets using ensemble models

Mining drug targets from database is a problem of challenge but significance, which has received much attention from academia and industry. As a main source of drug targets, proteins have abundant sequence properties analyzed by the modern measuring techniques and those properties provide a new perspective for detecting the potential proteins as drug targets. In our research, two typical ensemble models, Bagging and AdaBoost of decision trees, were applied for predicting potential DTPs. To tackle the uncertainty of non-drug targets proteins, random sampling was used to construct the negative dataset. The experiments showed both of two ensemble models were insensitive to the feature selection and ratio of training dataset. The average of recall ratios for proteins as drug targets maintained a high level, over 90% in two models. In one turn of parallel experiments, Bagging and AdaBoost predicted 257 potential drug target proteins in common. All of these have confirmed the availability of sequence information and applicability of data mining techniques. The process of retrieving drug targets would be accelerated with the booming development of new data mining technics.