We present self-consistent reaction field (SCRF) calculations, utilizing correlated ab initio quantum mechanics, of aqueous solvation free energies for a large data base of molecular solutes. We identify a subset of chemical functional groups for which there are systematic deviations in the comparison of theory and experiment; furthermore, for one case which has been extensively investigated, methylated amines, similar deviations appear in explicit solvent free energy perturbation calculations employing several commonly used molecular mechanics potential functions. By carrying out high-level ab initio quantum chemical calculations of hydrogen-bonding energies of the solutes to a water molecule, we arrive at a coherent explanation of the disagreements between theory and experiment, namely, that hydrogen-bonding energies are in some cases poorly correlated with classical electrostatic interaction energies. We show that the deviation in hydrogen-bonding energies of a solute from a reference molecule (for which there is good agreement between the SCRF calculations and experiment) is an excellent predictor of the errors made for that solute in the SCRF calculations. A new SCRF model is developed in which short-range empirical corrections, based upon solvent accessibility, are made for these chemical functional groups; this reduces the mean error of the calculated solvation free energies for the entire data base by a factor of ∼2, to 0.37 kcal/mol. These results have significant implications for the accuracy of explicit solvent potential functions as well as dielectric continuum models. Finally, we also identify cases where the observed discrepancies in solvation free energies cannot be explained by pair hydrogen-bonding results and suggest problems here that may be specific to dielectric continuum theory.

We explore and discuss several important issues concerning the derivation of many-body force fields from ab initio quantum chemical data. In particular, we seek a general methodology for constructing ab initio force fields that are ''chemically accurate'' and are computationally efficient for large-scale molecular dynamics simulations. We investigate two approaches for modeling many-body interactions in extended molecular systems. The interactions are adjusted to reproduce the many-body energy in small molecular clusters. Subsequently, the potential parameters affecting only pair interactions are then varied to reproduce the ab initio binding energy of dimers. This simple procedure is demonstrated in the design of a new polarizable force field of water. In particular, this new model incorporates the usual many-body interactions due to electrostatic polarization and a type of nonelectrostatic many-body interactions exhibited in bifurcated hydrogen-bonded systems. The static and dynamical properties predicted by the new ab initio water potential are in good agreement with the successful empirical fluctuating-charge potential of Rick et al. and with experiment. The aforementioned ''cluster'' approach is compared with an alternative method, which regards many-body interactions as manifestations of the electrostatic polarization properties of individual molecules. The effort required to build ab initio databases for force field parametrization is substantially reduced in this alternative method since only the monomer properties are of interest. We found intriguing differences between these two approaches. Finally our results point to the importance of discriminating ab initio data for force field parametrization. This is essentially a consequence of the simple functional forms employed to model molecular interactions, and is inevitable for large-scale molecular dynamics simulations.