In this study, we developed a new pharmacophore-based interaction fingerprint (Pharm-IF) and examined its usefulness for in silico screening using machine learning techniques such as support vector machine (SVM) and random forest (RF) instead of similarity-based ranking. Using the docking results of PKA, SRC, cathepsin K, carbonic anhydrase II, and HIV-1 protease, the screening efficiencies of the Pharm-IF models were compared to GLIDE score and the residue-based IF (PLIF) models. The combination of SVM and Pharm-IF demonstrated a higher enrichment factor at 10% (5.7 on average) than those of GLIDE score (4.2) and PLIF (4.3). In terms of the size of the training sets, learning more than five crystal structures enabled the machine learning models to stably achieve better efficiencies than GLIDE score. We also employed the docking poses of known active compounds, in addition to the crystal structures, as positive samples of training sets. The enrichment factors of the RF models at 10% using the docking poses for SRC and cathepsin K showed significantly higher values (6.5 and 6.3) than those using only the crystal structures (3.9 and 3.2), respectively.