Automatic feature subset selection for decision tree-based ensemble methods in the prediction of bioactivity

a r t i c l e i n f o In the structure–activity relationship (SAR) study, a learning algorithm is usually faced with the problem of selecting a compact subset of descriptors related to the property of interest, while ignoring the rest. This paper presents a new method of molecular descriptor selection utilizing three commonly used decision tree (DT)-based ensemble methods coupled with a backward elimination strategy (BES). Our proposed method eliminates descriptor redundancy automatically and searches for more compact descriptor subset tailored to DT-based ensemble methods. Six real SAR datasets related to different categorical bioactivities of compounds are used to evaluate the proposed method. The results obtained in this study indicate that DT-based ensemble methods coupled with BES, especially boosting tree model, yield better classification performance for compounds related to ADMET. In modern pharmaceutical industry, structure–activity relationship (SAR), an important area of chemometrics, is urgently needed for predicting ADMET (absorption, distribution, metabolism, excre-tion and toxicity) properties to select lead compounds for optimization at the early stage of drug discovery and to screen drug candidates for clinical trials [1–8]. The aim of SAR is to search information relating the molecular structure to biological activity. In this process, molecular structures are usually represented by a variety of molecular descriptors which are easily calculated by some software packages. In most cases, the number of compounds with the biological activity values available is usually small compared with the number of descriptors. This may lead to several difficult problems for constructing a SAR model undoubtedly. Among these problems we could list: (1) In most cases only a small number of descriptors have substantial influence on the property of interest. A lot of descriptors usually include ones unrelated to the property of interest. The use of such descriptors likely generates noise in an established model, which may affect the prediction accuracy of that model. (2) It is really difficult to determine which descriptors or combinations are responsible for the property of interest. The identification of important descriptors is of fundamental and practical interest. Research in biology and medicine may benefit from the examination of the top ranking descriptors to confirm recent discoveries in new drug research or suggest new avenues to be explored. (3) Much time may be needed to conduct a learning algorithm with a very large number of descriptors. (4) Most of modeling methods may not work well or even may be …