Accurate prediction of drug-protein interactions is crucial for drug discovery. Due to the bottleneck of traditional scoring functions, many machine learning scoring functions (MLSFs) have been proposed for structure-based drug screening. However, existing MLSFs face two challenges: small data limitations and poor interpretability. To address these challenges, we have proposed a physics-based small data machine learning framework for interpretable and generalizable prediction of drug-protein interactions on the target with scarce positive data through a strategy of three training phases with three (score, weight, and ranking) loss functions, named DrugBaiter. DrugBaiter has been extensively evaluated on the 102 targets of DUD-E and 81 targets of DEKOIS 2.0 for drug screening, and compared with 14 other MLSFs. It is shown that our DrugBaiter model can significantly improve the drug screening performance even if few actives are known for a target. In addition, DrugBaiter is interpretable in describing the interactions at the atomic level. The power of DrugBaiter is also confirmed by a drug screening application on the SARS-Cov-2 main protease target. It is anticipated that DrugBaiter will serve as a general machine learning scoring model for screening novel drugs on new targets with scarce known actives. DrugBaiter is freely available at http://huanglab.phys.hust.edu.cn/DrugBaiter.