This work describes the computation and accurate reproduction of subtle shifts in reduction potentials for two mutants of the iron-sulfur protein Pyrococcus furiosus ferredoxin. The computational models involved only first-sphere ligands and differed with respect to one ligand, either acetate (aspartate), thiolate (cysteine), or methoxide (serine). Standard procedures using vacuum optimization gave qualitatively wrong results and errors up to 0.07 V. Using electrostatically screened geometries and large basis sets for expanding the wave functions gave quantitatively correct results, with errors of only 0.03 V. Correspondingly, only this approach predicted a change in the coordination mode of aspartate (i.e., a carboxylate shift) accompanying the reduction of the wild-type cluster, confirming results from synthetic models and explaining why electrostatic screening is necessary. Hence, the carboxylate shift appears to occur in the proteins from which data were collected. The results represent the most accurate predictions of shifts in reduction potentials for modified proteins, the success in part being due to the similar nature of the three amino acid ligands involved. The predicted carboxylate shift is expected to tune aspartate's degree of electron donation to the cluster's two oxidation states, thus making the reversible redox reaction feasible.