Computational techniques based on continuum electrostatics treatments have been successful in predicting and interpreting the pK a values of ionizable amino acids in folded proteins. Despite this progress, efforts to reproduce the pH-dependence of protein stability have met with only limited success: agreement with experimental results has been only qualitative. It has been argued previously that the most likely reason for discrepancies is the presence of residual electrostatic interactions in the unfolded state, which cause pK a values to be shifted from their model compound values. Here we show that by constructing atomistic models of the unfolded state with a simple molecular mechanics protocol that uses the native state as a starting point, much improved reproduction of pH effects on protein stability can be obtained. In contrast, when a fully extended model of the unfolded state is used, no such improvement is obtained, a result that suggests that local interactions with residues nearby in the sequence are not sufficient to properly account for the pK a shifts in the unfolded state. In comparison to model compound values, the pK a values of acidic residues in “native-like” unfolded states are typically found to be shifted downwards by ∼0.3 pH unit, in good agreement with the average downward shift deduced from experimental measurements. Given its success in the present situation, the protocol employed here for developing simple models of the unfolded state may prove useful in other computer simulation applications.