We present a statistical mechanical approach for quantifying thermodynamic properties of proteins in mixed solvents. This approach, based on molecular dynamics simulations which incorporate all atom models and the theory of preferential binding, allows us to compute transfer free energies with experimental accuracy and does not incorporate any adjustable parameters. Specifically, we applied our approach to the model proteins RNase A and T1 and the solvent components water, glycerol, and urea. We found that the observed differences in the binding of glycerol and urea to RNase T1 and A are predominantly a consequence of density differences in the first coordination shell of the protein with the cosolvents, but the second solvation shell also contributes to the overall binding coefficients. Our approach allows us to determine the contributions of individual sites on a protein to preferential binding. One conclusion from determining these contributions is that hydrophobic amino acids in RNase T1 tend to bind less water and more cosolvent molecules than hydrophilic amino acids. The success of this approach in modeling preferential binding indicates that it incorporates the important underlying physics of proteins in mixed solvent systems and that the difficulty in quantitative prediction to date can be surmounted by explicitly incorporating the complex protein-solvent and solvent-solvent interactions.