The scoring functions for protein-ligand interactions plays central roles in computational drug design, virtual screening of chemical libraries for new lead identification, and prediction of possible binding targets of small chemical molecules. An ideal scoring function for protein-ligand interactions is expected to be able to recognize the native binding pose of a ligand on the protein surface among decoy poses, and to accurately predict the binding affinity (or binding free energy) so that the active molecules can be discriminated from the non-active ones. Due to the empirical nature of most, if not all, scoring functions for protein-ligand interactions, the general applicability of empirical scoring functions, especially to domains far outside training sets, is a major concern. In this review article, we will explore the foundations of different classes of scoring functions, their possible limitations, and their suitable application domains. We also provide assessments of several scoring functions on weakly-interacting protein-ligand complexes, which will be useful information in computational fragment-based drug design or virtual screening.