Small metal ions play critical roles in numerous biological processes. Of particular interest is how metalloenzymes are allosterically regulated by the binding of specific ions. Understanding how ion binding affects these biological processes requires atomic models that accurately treat the microscopic interactions with the protein ligands. Theoretical approaches at different levels of sophistication can contribute to a deeper understanding of these systems, although computational models must strike a balance between accuracy and efficiency in order to enable long molecular dynamics simulations. In this study, we present a systematic effort to optimize the parameters of a polarizable force field based on classical Drude oscillators to accurately represent the interactions between ions (K+, Na+, Ca2+ and Cl-) and coordinating amino-acid residues for a set of thirty biologically important proteins. By combining ab initio calculations and experimental thermodynamic data, we derive a polarizable force field that is consistent with a wide range of properties, including the geometries and interaction energies of gas-phase ion/protein-like model compound clusters, and the experimental solvation free-energies of the cations in liquids. The resulting models display significant improvements relative to the fixed-atomic-charge additive CHARMM C36 force field, particularly in their ability to reproduce the many-body electrostatic non-additivity effects estimated from ab initio calculations. The analysis clarifies the fundamental limitations of the pairwise additivity assumption inherent in classical fixed-charge force fields, and shows its dramatic failures in the case of Ca2+ binding sites. These optimized polarizable models, amenable to computationally efficient large-scale MD simulations, set a firm foundation and offer a powerful avenue to study the roles of the ions in soluble and membrane transport proteins.