This work evaluates usage feature selection methods to reduce the number of features required to predict docking results between Indonesian medicinal plant compounds and HIV protease. Two feature selection methods, Recursive Feature Elimination (RFE) and Wrapper Method (WM), are trained with a dataset of 7,330 samples and 667 features from PubChem Bioassay and DUD-E decoys. To evaluate the selected features, a dataset of 368 Indonesian herbal chemical compounds labeled by manually docking to PDB HIV-1 protease is used to benchmark the performance of linear SVM classifier using different sets of features. Our experiments show that a set of 471 features selected by RFE and 249 by WM achieve a reduction of classification time by 4.0 and 8.2 seconds respectively. Although the accuracy and sensitivity are also increased by 8% and 16%, no meaningful improvement observed for precision and specificity.