Abstract: A model to identify P-glycoprotein(P-gp) substrate was constructed with a random forest method based on open source software CDK(Chemistry Development Kit) descriptors and a training data set which contained 170 compounds(96 P-gp substrates). The study on the relationship between CDK descriptors and P-gp substrates indicates that sum of the atomic polarizabilities and charged partial surface area play important roles in identifying P-gp substrates. An external test data set containing 42 compounds(24 P-gp substrates) was employed. The correct classification rate on the training set is 99.42% and the correct classification rates for P-gp substrates, non-substrates and the total compounds on the test set are 87.50%, 83.33% and 85.71%, respectively. Leave-One-Out cross-validation correct classification rate(212 compounds) was 77.4%.

Key words: P-glycoprotein(P-gp), Random forest, Pattern recognition