Metal ions are indispensable cofactors for many proteins and play crucial roles in numerous biological processes, yet the computational design of metalloproteins and the optimization of metal coordination sites poses significant challenges due to the complex electronic structure of metals. Many standard protein design tools such as Rosetta do not take electronic effects for metal ions correctly into account (especially for transition metals) or do not consider metal ions at all. Widely used structure prediction methods such as AlphaFold2 consider metals only implicitly and often predict the amino acids in the metal binding conformation, however, without the metal ion being present. In this thesis, we developed and benchmarked several new methods that can be used to predict the stability of proteins in general and metalloproteins in particular, as well as for the localization and identification of metal ions in proteins. We found that existing thermostability datasets still present many biases and that combination of a physics-based potential with a self-supervised deep learning model can accurately identify stabilizing mutations in presence of metal ions. We also highlight that explicit treatment of metal ions is indispensable for correctly predicting the stability of a metalloprotein. We show that the methods for metal ion location and identity prediction we developed in this work are more powerful than general models for structure prediction such as AlphaFold3 and RoseTTAfold-AllAtom for the specific subtask of metal ion location prediction. The methods developed in this work are blind predictors, which do not require explicit specification of the stoichiometry and identity of metal ions bound to a given structure. AllMetal3D and Metal3D also are superior to other specialized metal ion location predictors or metal classification tools. AllMetal3D and Metal3D provide spatially precise locations often below 0.5 Å compared to the experimental location. The models are confident for physiological sites but can also recognize more transient metal binding sites. We also adapt the Metal3D framework for prediction of water binding sites. The resulting Water3D model is able to predict water binding sites close to metal ions, which is highly relevant for the design of functional metalloproteins. In the final chapter, we merge the efforts to design stable proteins with the simultaneous optimization of proteins for metal binding and provide proof-of-principle for fixed backbone design of novel metalloproteins with a defined coordination geometry. EVOLVE-metal employs genetic algorithm optimization allowing to optimize for stability, metal binding, metal geometry and identity to accommodate a metal binding site in a given scaffold.