The prediction of the binding affinity of protein-ligand complexes is important for the molecular docking and rational drug discovery. In this study, we have analyzed the descriptors, which affect the binding affinities of protein-ligand complexes, from five dimensions, including protein-ligand interactions, protein properties, structural and physicochemical descriptors of ligands, metal-ligand bonding, and water effects. Based on these dimensions, we generated 87 descriptors and used stepwise regression to select seven of these descriptors to develop a new scoring function from 891 protein-ligand complexes. The seven selected descriptors include van der Waals contact, metal-ligand bonding, water effects, deformation penalties upon the binding process, and the number of highly conserved residues with hydrogen bonds. This new scoring function is evaluated on an independent set with 98 protein-ligand complexes and the correlation between predicted binding affinities and experimental values is 0.601. These results show that our new scoring function for the prediction of binding affinity is useful for molecular recognition and virtual screening for drug design.