Influence of cut-off truncation and artificial periodicity of electrostatic interactions in molecular simulations of solvated ions: A continuum electrostatics study Michael Bergdorf, Christine Peter, and Philippe H. Hu¨nenbergera) Laboratorium fu¨r Physikalische Chemie, ETH Zu¨rich, CH-8093 Zu¨rich, Switzerland ͑Received 13 May 2003; accepted 6 August 2003͒ A new algorithm relying on finite integration is presented that solves the equations of continuum electrostatics for truncated ͑and possibly reaction-field corrected͒ solute–solvent and solvent– solvent interactions under either nonperiodic or periodic boundary conditions. After testing and validation by comparison with existing methods, the algorithm is applied to investigate the effect of cut-off truncation and artificial periodicity in explicit-solvent simulations of ionic solvation and ion–ion interactions. Both cut-off truncation and artificial periodicity significantly alter the polarization around a spherical ion and thus, its solvation free energy. The nature and magnitude of the two perturbations are analyzed in details, and correction terms are proposed for both effects. Cut-off truncation is also shown to induce strong alterations in the potential of mean force for ion–ion interaction. These observations help to rationalize artifacts previously observed in explicit– solvent simulations, namely spurious features in the radial distribution functions close to the cut-off distance and alterations in the relative stabilities of contact, solvent-separated and free ion pairs. © 2003 American Institute of Physics. ͓DOI: 10.1063/1.1614202͔ I. INTRODUCTION Computer simulation with an explicit representation of the solvent molecules has become a standard tool for investigating the structure, dynamics, and function of ͑bio-͒molecules in solution.1–6 However, due to important computational costs, the system sizes that are accessible to such simulations remain truly microscopic (typically Ͻ1000 nm3 ). As a direct consequence, the longest-range (Ͼ5 nm) component of intermolecular interactions, which is generally dominated by electrostatics, cannot be computed in an exact manner. Unfortunately, because electrostatic interactions are of large magnitude, many simulated observables turn out to be highly sensitive to the treatment of these interactions and, due to their long range, to the boundary conditions used in the simulation ͑system size and shape, finite versus periodic system͒. For this reason, the approximate representation of long-range electrostatic interactions in explicit-solvent ͑bio-͒molecular simulations is probably nowadays one of the principal bottlenecks in the accuracy of these methods. Uncontrolled approximations can give rise to important artifacts ͑so-called finite-size effects͒, which may strongly impair the reliability of many current simulations. There is thus considerable effort in the scientific community towards the goal of improving the representation of electrostatics in computer simulations. The vast majority of explicit-solvent ͑bio-͒molecular simulations are carried out under periodic boundary conditions6,7 ͑PBC͒. In this case, the solute ͑bio-͒molecule is placed into a computational box ͑space-filling shape, e.g., rectangular͒, and the empty volume is filled by solvent molecules. The system considered in the simulation consists of the central box surrounded by an infinite array of periodic copies of itself, which has the advantage of removing any surface distortion associated with a solvent-vacuum boundary. There are essentially three methods to handle electrostatic interactions in simulations under PBC: ͑i͒ Straight truncation of the Coulomb interactions at a convenient cutoff distance;6,7 ͑ii͒ smooth truncation of the Coulomb interactions, e.g., by means of a switching or shifting function8–13 or by including a reaction-field correction;14–19 ͑iii͒ use of lattice-sum methods ͑Ewald,20 P3 M,21 or PME22,23 methods͒. Cut-off truncation reduces the computational costs and the effect of artificial periodicity in simulations. However, straight truncation ͑ST͒ represents a very severe approximation, leading to heating as well as important artifacts in simulated properties of liquids,16,24,25 solvated ions,26–35 ion pairs,9,36–44 and biomolecules.45–48 Smooth-truncation methods may be applied to reduce the heating caused by the application of a cut-off, but nevertheless retain ͑and sometimes amplify͒ a number of the undesirable effects of abrupt truncation.9,10,27,28,41,48–51 Furthermore, these methods are generally ad hoc and lack any physical basis. An exception is the inclusion of a Barker–Watts reaction-field correction14–16 ͑RF͒ to the cut-off truncation. This correction scheme approximately accounts for the mean effect of the medium beyond the cut-off sphere of each particle by assuming that this medium behaves as a homogeneous dielectric continuum of permittivity equal to that of the solvent. Due to the form of the added reaction-field term, the correction effectively acts as a ͑physically based͒ switching function in the limit of a͒ Author to whom correspondence should be addressed. Telephone: ϩ41 1 632 5503; Fax: ϩ41 1 632 1039. Electronic mail: phil@igc.phys.chem.ethz.ch JOURNAL OF CHEMICAL PHYSICS VOLUME 119, NUMBER 17 1 NOVEMBER 2003 91290021-9606/2003/119(17)/9129/16/$20.00 © 2003 American Institute of Physics Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp high solvent permittivities. Finally, lattice-sum ͑LS͒ methods20–23,52–54 rely on Fourier series to describe the longrange component of electrostatic interactions, i.e., they assume that these interactions are exactly periodic within the infinite system. Although LS and RF methods rely on more or less reasonable approximations for dealing with the longrange component of electrostatic interactions, an approximate treatment is certainly preferable to the plain omission of this long-range component, as done in the ST scheme. Nevertheless, some dependence of simulated observables on the cut-off distance or system size has also been evidenced for the LS,19,55–59 and RF27,28,60 methods. It is therefore of importance to carefully investigate and compare the properties of the three most common electrostatic schemes ͑ST, RF, and LS͒. A general strategy to analyze finite-size effects and improve electrostatic schemes for explicit-solvent simulations relies on the use of continuum electrostatics.26–28,51,52,56–64 In the continuum-electrostatics approach, the solute is treated as a low-dielectric cavity encompassing the solute atomic point charges, and embedded in a dielectric continuum of permittivity equal to that of the solvent. In the classical implementation of the method,65–68 the electrostatic potential in the system is computed by numerically solving the Poisson ͑or Poisson–Boltzmann, in the presence of implicit counterions͒ equation, giving access to the electrostatic solvation free energy of the solute. Although there is no choice of boundary conditions that adequately mimics an infinitely dilute solution in explicit-solvent simulations, this is not the case in continuum-electrostatics calculations. There, the boundary conditions to solve the Poisson equation are specified in the form of the potential at the surface of the computational box. For a reasonably large solute–wall distance, this potential is well approximated by the solvent-screened Coulomb potential of the solute charges.57 In this way, continuum electrostatics can be used to estimate, for a given solute configuration, the electrostatic solvation free energy corresponding to exact ͑nontruncated͒ Coulomb interactions ͑CB͒ under nonperiodic boundary conditions ͑NPBC͒, a good model for the ideal situation of a solute at infinite dilution. This suggests that artifacts linked with the use of approximate electrostatic interactions and periodic boundary conditions in explicit-solvent simulations could be investigated using continuum electrostatics, provided that the method is generalized to these modified interactions and boundary conditions. Such generalizations have recently been developed26–28,51,52,57–64 for nearly all types of relevant electrostatic interaction schemes ͑CB/LS, ST, or RF͒ and boundary conditions ͑NPBC or PBC͒, as summarized in Table I. By comparing, for a given solute configuration, the outcome of a continuum-electrostatics calculation based on modified interactions and boundary conditions with that of another calculation based on CB interactions under NPBC, it is possible to estimate the perturbation ⌬⌬Gsolv of the solvation free energy. The corresponding perturbation ⌬Edirect in the direct electrostatic interaction energy between solute atomic charges is straightforward to calculate. The sum ⌬⌬Gel of the two contributions represents the perturbation of the electrostatic free energy of the system due to the use of approximate electrostatics and boundary conditions in the simulation. This procedure is illustrated schematically in Fig. 1 for the specific case of an explicit-solvent simulation employing ST or RF electrostatics under PBC. The quantity ⌬⌬Gel ͑possibly evaluated for multiple solute configurations͒ gives the required information to investigate the nature TABLE I. Generalizations of the continuum-electrostatics approach to modified electrostatic interactions and boundary conditions. The methods have been classified using the codes 3D ͑problem solved in three dimensions͒, 1D ͑problem reduced to a one-dimensional equation by symmetry͒, analytical ͑analytical solution available͒, Poisson ͑based on solving the Poisson equation͒, FFT ͑based on the use of fast Fourier transforms͒, or direct ͑based on solving field equations analogous to Eqs. ͑1͒ and ͑6͒ in real- space͒. Electrostatics NPBC PBC CB/LSa 3D-Poisson ͑Ref. 95͒b 3D-Poisson ͑Ref. 57͒ 3D-FFT ͑Refs. 51, 63, and 64͒ ͑spherical ionc ͒ 1D-Analytical ͑Ref. 86͒ 1D-Analytical ͑Ref. 57͒ SC,RFd 3D-Directe 3D-Directe 3D-FFT ͑Refs. 51 and 60͒ ͑spherical ionc ͒ 1D-Direct ͑Refs. 26–28͒ - a Interactions follow from the Coulomb ͑CB͒ potential under NPBC or the Ewald lattice-sum ͑LS͒ potential under PBC. b First calculation on a biomolecule using a finite-difference algorithm ͑many alternatives have been proposed, including, e.g., finite-element, boundaryelement and inducible-multipole algorithms͒. c Solutions developed for the special case of a single spherical ion. d Interactions follow from the Coulomb potential truncated at a given cut-off distance, without ͑ST͒ or with ͑RF͒ the inclusion of a reaction-field correc- tion. e Developed in the present article. FIG. 1. Schematic illustration of the procedure used to assess, based on continuum electrostatics, artifacts linked with approximate electrostatics and boundary conditions in explicit-solvent simulations. Ideally, an explicitsolvent simulation aiming to describe a solute molecule ͑symbolized by a black sphere͒ at infinite dilution should be based on a quasi-macroscopic system under NPBC together with exact CB interactions ͑top left drawing͒. Due to computational limitations this is not feasible in practice, and one may simulate instead a system under PBC with ST ͑or RF͒ electrostatic interactions ͑top right drawing͒. The corresponding perturbation can be evaluated by considering the implicit-solvent analogs of the two cases. Using continuum electrostatics ͑for a given solute configuration͒, the solvation free energies and the direct interactions between solute charges can be computed both under CB/NPBC ͑bottom left drawing; based on a good approximation for the electrostatic potential at the surface of the computational volume͒ and under ST/PBC or RF/PBC ͑bottom right drawing͒. The free-energy difference ⌬⌬Gel represents the perturbation of the electrostatic free energy induced by the use approximate electrostatics and boundary conditions, and is a key quantity for the analysis of finite-size effects in the explicit-solvent simulation. 9130 J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Bergdorf, Peter, and Hu¨nenberger Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp and magnitude of the corresponding artifacts in explicitsolvent simulations. Since they ignore the discrete nature of the solvent, continuum-electrostatics models have some limitations, including an important sensitivity to empirical model parameters ͑atomic charges, atomic radii, solute permittivity, exact definition of the solute–solvent boundary͒, the neglect of nonlinear effects ͑electrostriction, dielectric saturation͒, and the neglect of the detailed solvent structure around the solute ͑structure of the first solvation shells, specific hydrogen bonds͒. Furthermore, the electrostatic contribution to the solvation free energy should be complemented by a nonpolar contribution, typically assumed to be proportional to the solvent-exposed surface area. However, since the present method relies on the comparison of two closely related continuum-electrostatics calculations involving the same parameters and solute configuration ͑Fig. 1, bottom drawings͒, it is likely that errors in the short-range description of solvation cancel out to a large extent. The difference will depend almost exclusively on long-range effects, for which continuum electrostatics can be expected to be accurate. Thus, electrostatic free energy differences from continuum electrostatics should be almost quantitatively transposable to interpret finite-size artifacts in explicit-solvent simulations. Inspection of Table I reveals one missing entry. There is currently no general continuum-electrostatics method to deal with truncated electrostatic interactions ͑ST or RF͒ under NPBC, although a method exists in the special case of a single spherical ion26–28 ͓one-dimensional ͑1D͒-Direct method͔. The goal of the present article is to describe and apply a general method based on field equations and a finiteintegration algorithm ͓three-dimensional ͑3D͒-Direct method͔. In addition to dealing with the NPBC case, this new method is also applicable to systems under PBC. However, it scales rather unfavorably with the system size ͑as Ng 6 , where Ng is the number of grid points along each Cartesian direction͒, and can only be used for small systems. Therefore, its application is restricted here to the investigation of the consequences of cut-off truncation and artificial periodicity of electrostatic interactions in molecular simulations of ionic solvation and ion–ion interaction. These systems are very important benchmark systems for evaluating the accuracy of electrostatic interactions in molecular simulations because ͑i͒ they offer the simplest context to investigate electrostatic finite-size effects, and ͑ii͒ despite the apparent simplicity of the problem, the accurate determination of ionic solvation free energies9,19,26–35,55,57,69–79 and ion–ion potentials of mean force9,36–44,57,80–84 has turned out to be a surprisingly difficult problem. In the present article, the algorithm is described in details and the influence of various parameters controlling its behavior is investigated. The accuracy of the algorithm is further tested by comparing solvation free energies computed for a single spherical ion to values estimated through the 1D-Direct26–28 ͑NPBC͒ or 3D-FFT51,60 ͑PBC͒ methods. Finally, the present 3D-Direct method is applied to investigate the effect of cut-off truncation and artificial periodicity in computer simulations of ionic solvation and ion–ion interac- tions. II. THEORY A. Continuum electrostatics In the continuum-electrostatics approach, a solute– solvent system is modeled ͑for a given solute configuration͒ as a set of solute atomic partial charges embedded in a polarizable medium of heterogeneous dielectric permittivity. Application of the laws of electrostatics within such a system leads to the following expression24,25,51,64,85 for the electric field E͑r͒ E͑r͒ϭV͑r͒ϩ ͵͵͵R3 d3 rЈTᠪ ͑rϪrЈ͒P͑rЈ͒, ͑1͒ where V͑r͒ is the vacuum field ͑electric field generated by the solute atomic partial charges in the absence of polarizable medium͒, P͑r͒ the polarization ͑dipole moment density͒, and Tᠪ (r) the dipole–dipole interaction tensor characterizing the solvent–solvent interactions within the system. In the application of Eq. ͑1͒, it will be assumed that both solute–solute and solute–solvent electrostatic interactions are described by truncated Coulomb interactions with a Barker–Watts reaction-field correction.14–16 In the Barker– Watts scheme, the potential generated at r by a unit charge at the origin is given by ␺BW͑r͒ϭ 1 4␲⑀o H͑RϪr͒ͩrϪ1 ϩ ␣r2 2R3 Ϫ ␣ϩ2 2R ͪ, ͑2͒ where H(r) is the Heaviside function ͓H(r)ϭ1 if rϾ0, H(r)ϭ0 otherwise͔, ⑀o the permittivity of vacuum, and R is the cut-off distance. The parameter ␣ ͑with 0р␣р1) determined by the relative dielectric permittivity ⑀Ј of the medium surrounding the cut-off sphere of each particle through ␣ϭ 2͑⑀ЈϪ1͒ 2⑀Јϩ1 . ͑3͒ The function ␺BW in Eq. ͑2͒ accounts both for the direct Coulombic potential generated by the charge (rϪ1 term͒ and for the polarization by the neighboring charges of the medium outside its cut-off sphere (r2 term͒. In the present work, the discussion of the general form of the Barker–Watts interaction function ͑BW͒ will essentially focus on the cases ␣ϭ0 (⑀Јϭ1), corresponding to straight truncation of the Coulomb interactions without reaction-field correction ͑ST͒, and ␣ϭ1 (⑀Ј→ϱ), corresponding to truncated Coulomb interactions with a reaction-field correction corresponding to a conducting medium ͑RF͒. When Eq. ͑2͒ is applied to the solvent–solvent interactions under nonperiodic boundary conditions ͑NPBC͒, the dipole–dipole interaction tensor reads Tᠪ NPBC͑r͒ϭ 1 4␲⑀o H͑RSSϪr͒ ϫٌ ٌͩrϪ1 ϩ ␣r2 2RSS 3 Ϫ ␣ϩ2 2RSS ͪ ϭ 1 4␲⑀o H͑RSSϪr͒ͩ3r rϪr2 1ᠪ r5 ϩ ␣1ᠪ RSS 3 ͪ, ͑4͒ 9131J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Cut-off and periodicity effects in simulations of ions Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp where RSS is the solvent–solvent cut-off distance and the notation a b has been introduced for the tensor with elements ␮␯ given by a␮b␯ . Under periodic boundary conditions ͑PBC͒, and provided that RSS is smaller than half the smallest dimension of the computational box ͑which will be assumed from here on͒, the dipole–dipole interaction tensor reads Tᠪ PBC͑r͒ϭ 1 4␲⑀o H͑RSSϪr¯͒ͩ3r¯ r¯Ϫr¯2 1ᠪ r¯5 ϩ ␣1ᠪ RSS 3 ͪ, ͑5͒ where r¯ is the minimum-image vector corresponding to r. Using the approximation of linear response, the reaction of the polarizable medium is linear in the local electric field,24,25,51,64,85 i.e., P͑r͒ϭ⑀o͓⑀͑r͒Ϫ1͔E͑r͒, ͑6͒ where ⑀(r) the relative dielectric permittivity of the medium, which may be heterogeneous in space. Typically, one distinguishes between solute and solvent regions, characterized by distinct homogeneous permittivity values. The dipole–dipole interaction tensor Tᠪ (r) defined by Eqs. ͑4͒ or ͑5͒ is singular at the origin. However, the singularity is integrable when applying Eq. ͑1͒. More precisely, defining ⍀(r;␳) as the sphere of radius ␳ and surface ⌺(r;␳) centered at r, one may write ͵͵͵R3 d3 rЈTᠪ ͑rϪrЈ͒P͑rЈ͒ ϭI͑r͒ϩ lim ␳→0 ͵͵͵R3‫گ‬⍀(r;␳) d3 rЈTᠪ ͑rϪrЈ͒P͑rЈ͒. ͑7͒ The first term can be evaluated as I͑r͒ϭ lim ␳→0 ͵͵͵⍀(r;␳) d3 rЈTᠪ ͑rϪrЈ͒P͑rЈ͒ ϭ 1 4␲⑀o ͫlim ␳→0 ͵͵͵⍀(0;␳) d3 sٌ ٌsϪ1 ͬP͑r͒ ϭϪ 1 4␲⑀o ͫlim ␳→0 ␳Ϫ2 ͵͵⌺(0;␳) d2 ␴sϪ2 s sͬP͑r͒ ϭϪ 1 2⑀o ͫ͵0 ␲ d␪ sin ␪ cos2 ␪ͬP͑r͒ϭϪ 1 3⑀o P͑r͒. ͑8͒ The second equality follows from inserting Eqs. ͑4͒ or ͑5͒, defining sϭrЈϪr, and noting that as ␳ tends towards zero: ͑i͒ The Heaviside function evaluates to one for any finite RSS ; ͑ii͒ the contribution proportional to the unit tensor vanishes; ͑iii͒ P(rЈ) may be approximated by P͑r͒ and factorized from the integral. The third equality follows from applying the gradient theorem and inserting ٌsϪ1 ϭϪsϪ3 s. The fourth equality follows from observing that, due to symmetry, the off-diagonal elements of the tensor vanish upon integration, and that the diagonal elements are all equal. The fifth equality follows from evaluating one of these diagonal elements ͑integrand sϪ2 sz 2 ϭcos2 ␪) in spherical coordinates. Combining Eqs. ͑1͒, ͑6͒, ͑7͒, and ͑8͒, the equation to be solved for the electrostatic field E͑r͒ reads E͑r͒ϭV͑r͒Ϫ 1 3 ͓⑀͑r͒Ϫ1͔E͑r͒ ϩ⑀o lim ␳→0 ͵͵͵R3‫گ‬⍀(r;␳) d3 rЈTᠪ ͑rϪrЈ͒ ϫ͓⑀͑rЈ͒Ϫ1͔E͑rЈ͒, ͑9͒ with Tᠪ (r) given by Eqs. ͑4͒ or ͑5͒. If this equation can be solved for E͑r͒, the free energy of interaction between the solute atomic point charges and the polarizable medium is given by ⌬GϭϪ 1 2 ͵͵͵R3 d3 rV͑r͒•P͑r͒. ͑10͒ In the special case of a nonpolarizable solute, this quantity represents the solvation free energy of the solute. B. Discretization To transform the solving of Eq. ͑9͒ into a computationally tractable problem, three approximations are made. First, the infinite integration domain is reduced to a finite region of space. More precisely, two types of computational domains are considered: ͑i͒ A spherical volume of radius S surrounded by vacuum under NPBC, or ͑ii͒ a cubic unit cell of edge L under PBC. In both cases, the restriction to a finite computational domain is expected to have limited consequences when SӷRSS under NPBC or Lӷ2RSS under PBC ͑because the truncation of solvent–solvent interactions largely reduces dipole–dipole correlations at large distances͒, provided that the vacuum field is only active over a small region within this domain ͑which will be the case due to truncation of the solute–solvent interactions͒. Second, the solute is assumed to be nonpolarizable and the solvent to be represented by a medium of homogeneous permittivity. Thus, the computational domain comprises two subdomains characterized by different homogeneous dielectric permittivity values: A ͑possibly discontinous͒ solute subdomain of relative permittivity one, and a solvent subdomain of relative permittivity ⑀s . Third, the problem is discretized by paving the computational domain using N grid cells, leading to piecewiseconstant representations Vi and Ei ͑with iϭ1,...,N) of the vacuum and electric fields. Within these approximations, Eq. ͑9͒ becomes Ei␮ϭVi␮Ϫ ⑀sϪ1 3 ␴iEi␮ ϩ⑀o͑⑀sϪ1͚͒jϭ1 N vj͑1Ϫ␦ij͒␴j ͚␯ϭ1 3 T␮␯͑riϪrj͒Ej␯ , ͑11͒ where the ␮ and ␯ indexes enumerate Cartesian components, ri and vi are the center coordinate and volume of grid cell i, the exterior function ␴i evaluates to one if ri is within the solvent subdomain and zero otherwise, and the matrix elements of the dipole–dipole interaction tensor ͓Eqs. ͑4͒ or ͑5͔͒ take the form 9132 J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Bergdorf, Peter, and Hu¨nenberger Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp TNPBC,␮␯͑r͒ ϭ 1 4␲⑀o H͑RSSϪr͒ͩ3r␮r␯Ϫ␦␮␯r2 r5 ϩ ␣␦␮␯ RSS 3 ͪ ͑12͒ or TPBC,␮␯͑r͒ ϭ 1 4␲⑀o H͑RSSϪr¯͒ͩ3r¯␮r¯␯Ϫ␦␮␯r¯2 r¯5 ϩ ␣␦␮␯ RSS 3 ͪ. ͑13͒ Defining the 3Nϫ3N-dimensional matrix Aᠪ as Ai␮j␯ϭͩ1ϩ ⑀sϪ1 3 ␴iͪ␦ij␦␮␯ Ϫ⑀o͑⑀sϪ1͒vj͑1Ϫ␦ij͒␴jT␮␯͑riϪrj͒, ͑14͒ Eq. ͑11͒ can be rewritten in matrix notation Aᠪ EϭV. ͑15͒ Because the size of the matrix Aᠪ is generally very large, it cannot be stored in memory and direct methods cannot be used to solve Eq. ͑15͒. Therefore, an under-relaxed Jacobi method is applied here to obtain successive approximate solutions for the discretized electric field. Given the approximate solution E(k) at iteration k, the grid E(kϩ1) at the next iteration is computed as E(kϩ1) ϭE(k) Ϫ␭Dᠪ Ϫ1 ͑Aᠪ E(k) ϪV͒, ͑16͒ where Dᠪ is the diagonal matrix defined by the diagonal elements of Aᠪ , and ␭ ͑with 0Ͻ␭р1) is a relaxation parameter. A reasonable initial guess for E(0) is provided by the vacuum field scaled by the solvent permittivity, i.e., E(0) ϭ⑀s Ϫ1 V. ͑17͒ It is easily seen that a self-consistent solution of Eq. ͑16͒ must satisfy Eq. ͑15͒. In order to assess the convergence of the numerical solution upon iterating, the residual ͑with units of an electric field͒ ␶(k) ϭ͚ͩiϭ1 N vi␴iʈ͑Aᠪ E(k) ϪV͒iʈ2 ͚iϭ1 N vi␴i ͪ1/2 , ͑18͒ is introduced as a measure of accuracy. After solving Eq. ͑15͒ for the discretized electric field, the solvation free energy can be evaluated as ͓Eq. ͑10͒; nonpolarizable solute͔ ⌬GsolvϭϪ 1 2 ⑀o͑⑀sϪ1͚͒iϭ1 N vi␴iVi•Ei , ͑19͒ where Eq. ͑6͒ was used. Note that the solvation free energy solely depends on the electric field within the solvent subdomain. Furthermore, due to the form of Eq. ͑14͒, the field Ei corresponding to a point i in the solvent subdomain does not depend on the field Ej at any point j within the solute subdomain. For this reason, increased computational efficiency can be achieved by omitting all grid points of the solute subdomain from the definition of the matrix Aᠪ and the determination of the solution of Eq. ͑15͒. C. Application to solvated spherical ions For a system consisting of a single spherical ion of radius RI and charge qI centered in the computational domain, the vacuum field corresponding to ion–solvent interactions described by the Barker–Watts scheme ͓Eq. ͑2͔͒ is given by V͑r͒ϭ qI 4␲⑀o H͑RISϪr͒ͩr r3 Ϫ ␣r RIS 3 ͪ, ͑20͒ where RIS is the ion–solvent cut-off distance. This equation is valid under both NPBC and PBC, provided that RIS is smaller than half the smallest dimension of the computational box ͑which will be assumed from here on͒. Because the ion is generally small compared to the size of the computational domain, while the variations of the electric field within the solvent are typically largest close to its surface, the accuracy of the results will depend crucially on the detailed representation of the ionic surface. For this reason, two levels of grid resolution are used. First, a coarse grid of spacing ⌬ is generated, that covers the entire computational domain ͑leading to grid-cell volumes viϭ⌬3 ). Second, all cells of the coarse grid with their center closer than ()/2)⌬ from the surface of the ion are further discretized, i.e., they are replaced by a set of finer grid cells of edge ␦ ϭnϪ1 ⌬ where n is a positive integer ͑leading to grid-cell volumes viϭ␦3 ). Any grid-cell center of the finer grid that is closer than RI from the ion center is discarded from the calculation ͑solute point͒. In this representation, the vacuum potential at any grid point i is evaluated in practice as Viϭ qI 4␲⑀o I͑RISϪri͒ͩri ri 3 Ϫ ␣ri RIS 3 ͪ, ͑21͒ where I(RISϪri) represents the fraction of the grid cell located within the ion–solvent cut-off sphere.64 After solving Eq. ͑15͒ for the discretized electric field, the radial polarization p(r) around the ion ͓Eq. ͑6͔͒ can be computed in the form of a histogram. To avoid artifacts linked with the use of two different grid spacings, this calculation is based on a uniform grid of NЈ points obtained by partitioning all cells of the coarse grid into finer grid cells of edge ␦ sharing a common value of the electic field. Under NPBC, the radial polarization is then computed as pNPBC͑rn͒ϭ⑀o͑⑀sϪ1͒H͑rnϪRI͒ ϫ ͚iϭ1 NЈ ␴iw͑ri ;rn ,⌬r͒ri Ϫ1 ri•Ei ͚iϭ1 NЈ ␴iw͑ri ;rn ,⌬r͒ , ͑22͒ where rnϭ͑nϩ 1 2͒⌬r, nϭ0,1,...,nmax , ͑23͒ ⌬r being the histogram width, and w͑r;rn ,⌬r͒ϭͭ1 if rnϪ 1 2 ⌬rрrϽrnϩ 1 2 ⌬r 0 otherwise. ͑24͒ Under PBC, the periodic copies of the central box must be taken into account and Eq. ͑22͒ is modified to 9133J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Cut-off and periodicity effects in simulations of ions Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp pPBC͑rn͒ϭ⑀o͑⑀sϪ1͒H͑rnϪRI͒ ϫ ͚l෈Z3͚iϭ1 NЈ ␴iw͑ri,l ;rn ,⌬r͒ri,l Ϫ1 ri,l•Ei ͚l෈Z3͚iϭ1 NЈ ␴iw͑ri,l ;rn ,⌬r͒ , ͑25͒ with ri,lϭriϩLl. In practice, the sum over l is restricted to vectors with integer components in the range ͓Ϫlmax ;lmax͔, ensuring a correct description of the polarization up to r ϭ(lmaxϩ1/2)L. These polarization histograms can be compared to the ideal ͑CB/NPBC͒ Born polarization86 pBorn͑r͒ϭ qI 4␲ ⑀sϪ1 ⑀s rϪ2 . ͑26͒ Combining Eqs. ͑19͒ and ͑21͒, the ionic solvation free energy is evaluated ͑based on the refined grid of NЈ points͒ as ⌬GsolvϭϪ qI 8␲ ͑⑀sϪ1͚͒iϭ1 NЈ vi␴iI͑RISϪri͒ ϫͩri ri 3 Ϫ ␣ri RIS 3 ͪ•Ei . ͑27͒ This value can be compared to the ideal ͑CB/NPBC͒ Born electrostatic solvation free energy86 ⌬Gsolv Born ϭϪ qI 2 8␲⑀o ⑀sϪ1 ⑀s RI Ϫ1 . ͑28͒ The application to two ͑or more͒ spherical ions is straightforward and only requires the following minor adaptations: ͑i͒ The quantity ␴i is zero ͑point discarded from the calculation͒ for all grid cells with centers located inside any ion, and one otherwise; ͑ii͒ the vacuum field V ͓Eq. ͑21͔͒ is expressed as a sum of contributions arising from each ion; ͑iii͒ the solvation free energy ⌬Gsolv ͓Eq. ͑27͔͒ is expressed as a sum of contributions arising from each ion; ͑iv͒ fine grids ͑spacing ␦Ͻ⌬) are used to describe the close neighborhood of all ions. III. COMPUTATIONAL DETAILS The solution of Eq. ͑15͒, restricted to the case of one or two spherical ions, was implemented in a C program. The single ion or the two ions are placed on the z axis of the coordinate system ͑single ion at zϭ0; two ions at zϭ Ϯd/2, where d is the interionic distance͒. Taking advantage of the symmetry of the problem, the storage of the discretized vacuum and electric fields is only required for one quadrant (x,yу0) of the computational domain. After an evaluation of the convergence properties of the algorithm, a set of computational parameters was selected and adopted for all subsequent calculations. The corresponding values are as follows ͑unless otherwise specified͒. The spacings corresponding to the coarse and fine grids were set to ⌬ϭ0.1 nm and ␦ϭ0.025 nm, respectively. The relaxation parameter ␭ was set to 0.4. The algorithm was terminated when the residual ␶(k) was either below 10Ϫ3 kJ molϪ1 nmϪ1 eϪ1 or reached a minimum value. All calculations under NPBC used a sphere of radius Sϭ4.0 nm as computational domain. To ensure that this domain is large enough to be representative of an infinite nonperiodic system, a number of single-ion and two-ion calculations were repeated with S ϭ5.0 nm. The observed differences in solvation free energy were in all cases below 0.1 kJ molϪ1 in magnitude. The method was first applied to solvated spherical ions. Ionic solvation free energies ͓Eq. ͑27͔͒ and radial polarization histograms ͓Eqs. ͑22͒ or ͑25͒ with ⌬rϭ0.025 nm and lmaxϭ2] were computed for all combinations of the following parameters: ionic charge qIϭ1 e, ionic radii RI ϭ0.2 nm ͑about the size of Naϩ ) or 0.4 nm ͑about the size of ClϪ ), cut-off radii RISϭRSSϭRCϭ0.8 or 1.2 nm, solvent permittivities ⑀sϭ2 ͑alkanelike solvent͒ or 78 ͑water͒, ␣ ϭ0 ͑ST scheme͒ or 1 ͑RF scheme͒, and NPBC or PBC ͑with Lϭ2.6 nm). To validate the method, the results are compared with those of calculations employing other methods ͑Table I͒, namely: ͑i͒ 1D-Direct27,28 ͑NPBC; bin size 0.005 nm, range 4.0 nm͒ or ͑ii͒ 3D-FFT60 ͑PBC; 180 grid points along each Cartesian axis͒. The method was then used to investigate the effect of periodicity on ionic solvation free energies in systems with truncated electrostatic interactions ͑ST or RF͒. To this end, ionic solvation free energies were computed under PBC using the above combination of parameters, for cubic box edges L ranging from 1.6 nm (RCϭ0.8 nm) or 2.4 nm (RC ϭ1.2 nm) to 8.0 nm. The effect on periodicity can be quantified by the relative periodicity-induced perturbation of the ionic solvation free energy ␥(L), defined as ␥͑L͒ϭ ⌬⌬Gsolv͑L͒ ⌬Gsolv NPBC with ⌬⌬Gsolv͑L͒ϭ⌬Gsolv PBC ͑L͒Ϫ⌬Gsolv NPBC . ͑29͒ Finally, the effect of cut-off truncation ͑ST or RF͒ and periodicity on the electrostatic solvation contribution ⌬Gsolv(d) to the potential of mean force for the interaction between two ions at distance d ͑under PBC, ions aligned along an axis of the cubic unit cell͒ was evaluated for the special case of charges qIϭϮqJϭ1 e, radii RIϭRJ ϭ0.4 nm, cut-off radii RISϭRSSϭRCϭ1.2 nm and for a cubic unit-cell of edge Lϭ6 nm ͑PBC͒. For validation, the results under PBC were compared with those of calculations employing the 3D-FFT60 method ͑180 grid points along each Cartesian axis͒. The corresponding overall electrostatic contribution ⌬Gel(d) to the potential of mean force was also evaluated as ⌬Gel͑d͒ϭ⌬Gsolv͑d͒ϩqIqJ␺BW͑d˜ ͒, ͑30͒ where d˜ϭd ͑NPBC͒ or d˜ϭmin͕d;LϪd͖ ͑PBC͒, and ␺BW is given by Eq. ͑2͒ with ␣ϭ0 ͑ST͒ or ␣ϭ1 ͑RF͒. These profiles can be compared with the expected long-range behavior of the electrostatic potential of mean force for a Coulombic interaction between the ions, namely 9134 J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Bergdorf, Peter, and Hu¨nenberger Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp ⌬Gel lr ͑d͒ϭ⌬Gsolv ion ͑qI ,RI͒ϩ⌬Gsolv ion ͑qJ ,RJ͒ ϩ qIqJ 4␲⑀o⑀s ,dϪ1 ͑31͒ where ⌬Gsolv ion (q,R) is the solvation free energy of an isolated ion of radius R and charge q under NPBC when applying the specific electrostatic scheme. IV. RESULTS A. Convergence properties The convergence properties of the under-relaxed Jacobi algorithm ͓Eq. ͑16͔͒ used to solve Eq. ͑15͒ are illustrated in Fig. 2 for a spherical ion of charge qIϭ1 e and radius RI ϭ0.4 nm in a solvent of permittivity ⑀sϭ78, based on the RF scheme with a single cut-off radius RCϭ1.2 nm and using three choices of the relaxation parameter ␭. Results for the ST scheme are qualitatively very similar ͑data not shown͒. Within few iterations, the residual ␶(k) decreases from about 20 to values below 3 kJ molϪ1 nmϪ1 eϪ1 ͓Fig. 2͑a͔͒. Convergence to zero residual only occurs when the solvent permittivity ⑀s is smaller than about 10 ͑data not shown͒. This limited convergence is probably related to the presence of a strong discontinuity in the system permittivity at the ion–solvent boundary in the case of high solvent permittivities. For ⑀s values larger than about 10, the residual reaches a minimum after a certain number of iterations ͑typically about 15–20 for ␭ϭ0.4) and slowly rises again afterwards. When this situation occurred, the algorithm was terminated at the minimum value ␶ of the residual. However, convergence of ␶(k) towards ␶ is associated with the simultaneous convergence of ⌬Gsolv (k) to a well-defined value ⌬Gsolv ͓Fig. 2͑b͒ and Table II͔. Because the values of both the minimum residual and the associated converged solvation free energy are essentially independent of the convergence parameter ␭, it appears that the method is nevertheless able to produce accurate results for high ⑀s values. This is also supported by the observation that for ⑀sϭ2, values of ⌬Gsolv (k) when ␶(k) ϭ3 kJ molϪ1 nmϪ1 eϪ1 typically differ from the corresponding converged values (␶(k) Ͻ10Ϫ3 kJ molϪ1 nmϪ1 eϪ1 ) by less than 1% ͑data not shown͒. Although the convergence parameter ␭ does not influence the final values of ␶ and ⌬Gsolv , it has a strong impact on the convergence rate. For ⑀sϭ78, ␭ϭ0.2 leads to slow convergence, ␭ϭ0.6 to slow convergence and oscillatory evolution of ⌬Gsolv (k) , while the algorithm fails to converge for ␭ϭ0.8 ͑data not shown͒. In practice, it was found that ␭ϭ0.4 is the optimal choice in this case, and is also adequate for ⑀sϭ2 ͑although a somewhat larger values may slightly accelerate convergence͒. This value was adopted for all subsequent calculations. The rates of convergence to the minimum ␶ of ␶(k) appear to be very similar under NPBC or PBC, and for the RF and ST ͑data not shown͒ schemes. The final values of the residual, however, are somewhat lower under NPBC compared to PBC ͑Table II͒. B. Single spherical ion The radial polarization p(r) around around a spherical ion ͓Eqs. ͑22͒ and ͑25͔͒ of charge qIϭ1 e and radius RI ϭ0.4 nm in a solvent of permittivity ⑀sϭ78 is displayed in Fig. 3 for different choices of boundary conditions and treatments of the electrostatic interactions based on a single cutoff RCϭ1.2 nm. The polarization corresponding to the Born model86 ͓CB/NPBC; Eq. ͑26͔͒ or to the lattice-sum case ͑LS/ PBC; computed using the 3D-FFT method64 ͒, and the polarization computed from the 1D-Direct method27,28 ͑ST,RF/ NPBC͒ are also displayed for comparison. For both the ST and RF schemes under NPBC, the agreement between the results of the 1D-Direct and of the present 3D-Direct methods is excellent over the whole range of distances. The only noticeable difference is the more progressive transition of p(r) around RC in the 3D-Direct calculation for the ST scheme, which is due to a significantly lower resolution and to the smoothing of the vacuum field at the ion–solvent cutoff distance ͓function I in Eq. ͑21͔͒. These curves, however, differ significantly from the Born polarization, corresponding to the ideal situation of a spherical ion solvated by a nonperiodic Coulombic continuum of infinite extent. For the ST case, the polarization curve is discontinuous at the ion–solvent cut-off distance RIS . The polarization below RIS is consistently larger than predicted by the Born model, whereas the polarization above is smaller, although always positive. Underpolarization of the solvent above RIS is easily understood since the solvent beyond this distance does not feel directly the electrostatic field of the ion. However, the solvent in this region reacts indirectly to the ionic FIG. 2. Convergence properties of the under-relaxed Jacobi algorithm ͓Eq. ͑16͔͒ used to solve Eq. ͑15͒. The residual ␶(k) ͓Eq. ͑18͔͒ is displayed as a function of the number of iterations k ͑a͒, and the solvation free energy ⌬Gsolv (k) at iteration k ͓Eq. ͑27͔͒ as a function of the corresponding residual ␶(k) ͑b͒. The system consists of a single spherical ion of charge qIϭ1 e and radius RIϭ0.4 nm in a solvent of permittivity ⑀sϭ78, and is either nonperiodic ͑NPBC; spherical domain of radius Sϭ4.0 nm) or periodic ͑PBC; cubic unit cell of edge Lϭ2.6 nm). Electrostatic interactions correspond to the RF scheme, with cut-off radii RISϭRSSϭRCϭ1.2 nm. Three choices of the Jacobi relaxation parameter ␭ are compared ͓for NPBC, only the curve corresponding to ␭ϭ0.4 is displayed in ͑a͔͒. 9135J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Cut-off and periodicity effects in simulations of ions Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp field through interactions with the polarized solvent within the cut-off sphere of the ion, leading to the observed residual polarization. Inside the cut-off sphere of the ion, the solvent is overpolarized because each solvent volume element only interacts with a fraction of the highly polarized solvent within the cut-off sphere of the ion. This partial interaction results in a bias of the solvent polarization towards the ion. For the RF case, the polarization curve is continuous at RIS , although its derivative is not. The polarization is consistently smaller than the corresponding Born curve over the whole range of distance, both below and above RIS . The difference between the curves is largest at distances close to RIS , and becomes progressively smaller at either short or long distances from the ion. It has been shown that in the limit of high solvent permittivities, the Barker–Watts potential represents ͑for the solvent–solvent interactions͒ the cutoff-truncated polynomial of second order ͑terms in rϪ1 to r2 ) that leads to the best agreement between Born and effective polarizations.27,28 A number of additional results related to this comparison are derived in Appendix A, namely that ͑i͒ the RF/NPBC polarization converges to the Born polarization in the limit RIS ,RSS→ϱ; ͑ii͒ in the limit of small distances ͑compared to RC), the RF/NPBC polarization converges towards the Born polarization; ͑iii͒ in the limit of large distances ͑compared to RC), the RF/NPBC polarization becomes proportional to rϪ2 , just as the Born polarization. For both the ST and RF schemes, the polarization curves corresponding to PBC are systematically lower ͑in the range RI to L) compared to the polarization under NPBC. The reason for this is that under PBC, the solvent in the reference TABLE II. Solvation free energy ⌬Gsolv of a single spherical ion ͓Eq. ͑27͔͒ computed using the 3D-Direct method ͑present article͒. The system consists of a single ion of charge qIϭ1 e and radius RI in a solvent of permittivity ⑀s , and is either nonperiodic ͑NPBC; spherical domain of radius Sϭ4 nm) or periodic ͑PBC; cubic unit cell of edge Lϭ2.6 nm). Electrostatic interactions correspond to either the ST or RF schemes, with cut-off radii RISϭRSSϭRC . For the RF scheme, the reaction-field contribution to the solvation free energy ͓␣-dependent term in Eq. ͑27͔͒ is reported between parentheses. The converged value ␶ of the residual is also indicated. The solvation free energies ⌬Gsolv corr corrected by the inclusion of a self-energy term ͓Eq. ͑B3͔͒ are also given. For comparison, the corresponding Born solvation free energies ͓Eq. ͑28͔͒ are Ϫ342.9 (⑀ϭ78, RI ϭ0.2 nm), Ϫ171.4 (⑀ϭ78, RIϭ0.4 nm), Ϫ173.7 (⑀ϭ2, RIϭ0.2 nm), and Ϫ86.8 (⑀ϭ2, RIϭ0.4 nm) kJ molϪ1 . BC Interaction ⑀s RI ͓nm͔ RC ͓nm͔ ␶ ͓kJ molϪ1 nmϪ1 eϪ1 ͔ ⌬Gsolv ͓kJ molϪ1 ͔ ⌬Gsolv ref ͓kJ molϪ1 ͔ ⌬Gsolv corr ͓kJ molϪ1 ͔ NPBC ST 78 0.2 0.8 1.2 Ϫ284.8 Ϫ281.7a Ϫ370.5 NPBC ST 78 0.2 1.2 1.2 Ϫ306.3 Ϫ303.4a Ϫ363.4 NPBC ST 78 0.4 0.8 0.8 Ϫ100.9 Ϫ100.1a Ϫ186.6 NPBC ST 78 0.4 1.2 0.7 Ϫ129.4 Ϫ128.8a Ϫ186.5 PBC ST 78 0.2 0.8 3.0 Ϫ284.1 Ϫ280.3b Ϫ369.8 PBC ST 78 0.2 1.2 2.9 Ϫ301.0 Ϫ297.9b Ϫ355.0 PBC ST 78 0.4 0.8 2.0 Ϫ100.6 Ϫ102.4b Ϫ186.3 PBC ST 78 0.4 1.2 1.8 Ϫ125.3 Ϫ125.7b Ϫ182.4 NPBC RF 78 0.2 0.8 1.1 Ϫ217.0 (37.0) Ϫ213.7 (37.1)a Ϫ345.6 NPBC RF 78 0.2 1.2 1.1 Ϫ259.2 (26.2) Ϫ256.3 (26.4)a Ϫ344.9 NPBC RF 78 0.4 0.8 0.6 Ϫ47.6 (23.7) Ϫ46.8 (23.7)a Ϫ176.2 NPBC RF 78 0.4 1.2 0.6 Ϫ86.4 (22.0) Ϫ85.8 (22.2)a Ϫ172.1 PBC RF 78 0.2 0.8 2.8 Ϫ214.6 (35.8) Ϫ211.0 (35.6)b Ϫ343.2 PBC RF 78 0.2 1.2 2.9 Ϫ253.1 (22.3) Ϫ249.5 (22.5)b Ϫ338.8 PBC RF 78 0.4 0.8 1.6 Ϫ47.5 (22.9) Ϫ48.4 (23.0)b Ϫ176.1 PBC RF 78 0.4 1.2 1.6 Ϫ82.3 (18.7) Ϫ82.3 (18.8)b Ϫ168.0 NPBC ST 2 0.2 0.8 0.0 Ϫ133.4 Ϫ135.2a Ϫ176.8 NPBC ST 2 0.2 1.2 0.0 Ϫ146.7 Ϫ148.3a Ϫ175.6 NPBC ST 2 0.4 0.8 0.0 Ϫ45.6 Ϫ46.4a Ϫ89.0 NPBC ST 2 0.4 1.2 0.0 Ϫ60.3 Ϫ60.8a Ϫ89.2 PBC ST 2 0.2 0.8 0.0 Ϫ133.1 Ϫ135.0b Ϫ176.5 PBC ST 2 0.2 1.2 0.0 Ϫ145.6 Ϫ147.6b Ϫ174.5 PBC ST 2 0.4 0.8 0.0 Ϫ45.6 Ϫ46.3b Ϫ89.0 PBC ST 2 0.4 1.2 0.0 Ϫ59.6 Ϫ60.2b Ϫ88.5 NPBC RF 2 0.2 0.8 0.0 Ϫ99.8 (13.7) Ϫ101.0 (13.8)a Ϫ164.9 NPBC RF 2 0.2 1.2 0.0 Ϫ123.1 (9.8) Ϫ124.4 (9.8)a Ϫ166.5 NPBC RF 2 0.4 0.8 0.0 Ϫ20.4 (9.0) Ϫ20.6 (9.1)a Ϫ85.5 NPBC RF 2 0.4 1.2 0.0 Ϫ39.2 (8.3) Ϫ39.5 (8.4)a Ϫ82.6 PBC RF 2 0.2 0.8 0.0 Ϫ99.5 (13.8) Ϫ100.9 (13.8)b Ϫ164.6 PBC RF 2 0.2 1.2 0.0 Ϫ122.4 (9.5) Ϫ124.0 (9.5)b Ϫ165.8 PBC RF 2 0.4 0.8 0.0 Ϫ20.4 (9.0) Ϫ20.7 (9.1)b Ϫ85.5 PBC RF 2 0.4 1.2 0.0 Ϫ38.9 (8.0) Ϫ39.2 (8.1)b Ϫ82.3 a Solvation free energies ⌬Gsolv ref estimated from the 1D-Direct method ͑Refs. 27 and 28͒ are given for comparison ͑ST,RF/NPBC͒. b Solvation free energies ⌬Gsolv ref estimated from the 3D-FFT method ͑Ref. 60͒ are given for comparison ͑ST,RF/PBC͒. 9136 J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Bergdorf, Peter, and Hu¨nenberger Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp unit cell is perturbed by its interaction with the solvent in adjacent unit cells ͑itself polarized by the periodic copies of the ion͒ and, for rϾLϪRIS , with the periodic copies of the ion themselves. The consequences of these interactions are a depolarization of the solvent in the reference unit cell ͑compared to NPBC͒, and the occurence of negative polarization values for the solvent in the neighboring unit cells. At large distances, p(r) displays an irregular oscillatory behavior with values close to zero at the location of the nearest neighbor ions ͑i.e., L, &L, )L, . . . ; data not shown͒. The depolarization of the solvent within the reference unit cell is slightly more important in the RF case compared to the ST case, which is probably a consequence of the larger magnitude of the residual solvent polarization above RIS observed in the RF/NPBC case. In both cases, however, the solvent depolarization within the reference unit cell remains relatively small because, due to the truncation of ion–solvent interactions at RISϽL/2, dipoles in the reference unit cell do not interact directly with the periodic copies of the ion. Finally, the polarization corresponding to the RF/PBC scheme is seen to agree reasonably well with the LS/PBC curve, the difference being expectedly largest in the neighborhood of the cut-off distance. There is, however, an important difference between the two schemes. When cut-off truncation is applied, the solvation free energy only depends on the polarization in the range RI to RISϽL/2 ͓Eqs. ͑10͒ and ͑20͔͒ and the effect of periodicity on the ionic solvation free energy is expected to be relatively small. If this restriction is removed ͑e.g., when nontruncated LS interactions are considered͒, the effect of artificial periodicity on the ionic solvation free energy becomes dramatically more important.19,55,57 Ionic solvation free energies ⌬Gsolv computed for spherical ions with different parameter combinations, electrostatic schemes, and boundary conditions are listed in Table II. The values ⌬Gsolv ref computed using the 1D-Direct method27,28 ͑ST,RF/NPBC͒ or the 3D-FFT method60 ͑ST,RF/ PBC͒ are also listed for comparison. Note that, while the former values are certainly very accurate, the latter values are probably subject to errors of a similar magnitude as the present method. The agreement between the values computed using different methods is in general very good. The average and maximal relative differences between the present 3D-Direct and the reference values are 1.1% and 1.9%, respectively. Not unexpectedly, these relative differences tend to be somewhat larger for ͑i͒ the smaller ionic radius, ͑ii͒ the larger permittivity value, ͑iii͒ periodic boundary conditions. The following observations can be made: ͑i͒ the solvation free energies are larger in magnitude for the smaller ion and the larger permittivity value, in qualitative agreement with the Born model; ͑ii͒ the solvation free energies are larger in magnitude for the larger cut-off value, a consequence of including a larger amount of polarized solvent within the cut-off sphere of the ion; ͑iii͒ the solvation free energies are more negative for the ST scheme compared to the RF scheme, a consequence of the solvent overpolarization within the cut-off sphere of the ion for the ST scheme ͑Fig. 3͒ and of the inclusion of an additional positive term in the solvation free energy for the RF scheme ͓␣-dependent term in Eq. ͑27͒; reported in Table II between parentheses͔; ͑iv͒ the solvation free energies are larger in magnitude under NPBC compared to PBC, a consequence of the periodicityinduced solvent depolarization within the reference unit cell ͑Fig. 3͒; ͑v͒ the periodicity-induced perturbation of the solvation free energy (NPBC→PBC) is larger for the smaller ion, for the larger cut off, for the higher permittivity value, and for RF compared to ST. The latter effect is a consequence of the larger periodicity-induced solvent depolarization within the reference unit cell for the RF scheme ͑Fig. 3͒. The values reported in Table II are strongly cut-offdependent and compare poorly with the corresponding Born solvation free energies of Ϫ342.9 and Ϫ171.4 kJ molϪ1 (⑀sϭ78, qIϭ1 e, RIϭ0.2 or 0.4 nm͒ or Ϫ173.7 and Ϫ86.8 kJ molϪ1 (⑀sϭ2, qIϭ1 e, RIϭ0.2 or 0.4 nm͒. As discussed in Appendix B, these large discrepancies could be reduced by the inclusion of a charge self-energy term into the total electrostatic energy of the system. It is also suggested that such a self-energy term should be systematically included in the total electrostatic energy during molecular simulations relying on effective cut-off-based electrostatic interaction functions to ensure the obtension of meaningful energies. In this context, a new definition ͓Eq. ͑B4͔͒ is proposed for the electrostatic interaction energy in simulations employing the Barker–Watts reaction-field scheme. The effect of artificial periodicity on the solvation free energy of a spherical ion computed using cut-off-based ͑ST or RF͒ electrostatics is illustrated in Fig. 4 for an ion of charge qIϭ1 e, a solvent of permittivity ⑀sϭ78, and for different values of the ionic radius RI and cut-off radius RC . The relative periodicity-induced perturbation ␥(L) of the FIG. 3. Radial polarization p(r) around a solvated spherical ion ͓Eqs. ͑22͒ or ͑25͔͒. The system consists of a single ion of charge qIϭ1 e and radius RIϭ0.4 nm in a solvent of permittivity ⑀sϭ78, and is either nonperiodic ͑NPBC; spherical domain of radius Sϭ4.0 nm) or periodic ͑PBC; cubic unit cell of edge Lϭ2.6 nm). Electrostatic interactions correspond to either the ST ͑a͒ or RF ͑b͒ schemes, with cut-off radii RISϭRSSϭRCϭ1.2 nm. In addition to the results of the 3D-Direct method ͑present article͒, the analytical polarization corresponding to the Born model ͓CB/NPBC; Eq. ͑26͔͒ and the lattice-sum case ͓LS/PBC; computed using the 3D-FFT method ͑Ref. 64͔͒, and the polarization computed from the 1D-Direct method ͑Refs. 27 and 28͒ for the specific interaction scheme ͑ST,RF/NPBC͒ are also presented for comparison. The cut-off distance as well as the ͑half-͒box edge ͑PBC only͒ are indicated by arrows. 9137J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Cut-off and periodicity effects in simulations of ions Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp ionic solvation free energy ͓Eq. ͑29͔͒ is displayed as a function of L/2 in Fig. 4͑a͒. All curves converge to a limiting value of one when L/2ӷRC , indicating that the solvation free energy under PBC indeed converges to its NPBC value when the computational box becomes large compared to the cut-off radius. For example, the magnitude of ␥(L) for L/2 ϭ4 nm is smaller than 10Ϫ4 for all parameter combinations considered. When L/2 is only moderately larger than RC , artificial periodicity causes a depolarization of the solvent ͑Fig. 3͒ and a decrease in the magnitude of the solvation free energy. As a consequence, ␥(L) becomes negative. When L/2ϭRC , the solvation free energy is reduced by 2%–9% compared to its NPBC value for the parameter combinations considered. In agreement with previous observations ͑Table II͒, the relative periodicity-induced perturbation of the solvation free energy ͑i͒ increases in magnitude with increasing ionic radius; ͑ii͒ increases in magnitude with increasing cutoff radius; ͑iii͒ is larger for the RF scheme compared to the ST scheme. The different curves in Fig. 4͑a͒ can be adequately represented by exponential functions. This is shown in Fig. 4͑b͒, where the quantity log10͓Ϫ(RC /RI)␥(L)͔ is displayed as a function of L/(2RC). The resulting data can be fitted by two straight lines, corresponding to the ST and RF schemes, with linear correlation coefficients ͑over the interval L/(2RC) ϭ0 to 2.5͒ of Ϫ0.9998 and Ϫ0.9985, respectively. Thus, irrespective of RI and RC , the relative periodicity-induced perturbation appears to be approximately of the form ␥͑L͒ϷϪ RI RC 10␮ L/͑2RC͒ ϩ␯ . ͑32͒ Based on all available data for ␧sϭ78, the constants in Eq. ͑32͒ evaluated to ␮ϭϪ2.203 and ␯ϭ1.290 for the ST scheme, and ␮ϭϪ1.468 and ␯ϭ0.715 for the RF scheme. This behavior can be contrasted to the case of nontruncated electrostatic interactions. In this case, the solvation free energy corresponding to CB/NPBC is given by the Born expression86 ͓Eq. ͑28͔͒. A corresponding analytical expression57 has been derived for the LS/PBC case, namely ⌬Gsolv LS/PBC ϭϪ qI 2 8␲⑀o ⑀sϪ1 ⑀s ͫL RI ϩ␰EW ϩ 4␲ 3 ͩRI L ͪ2 Ϫ 16␲2 45 ͩRI L ͪ5 ͬLϪ1 , ͑33͒ with ␰EWϷϪ2.837 297. Thus, it follows from Eq. ͑29͒ that: ␥͑L͒ϭ RI L ͫ␰EWϩ 4␲ 3 ͩRI L ͪ2 Ϫ 16␲2 45 ͩRI L ͪ5 ͬ. ͑34͒ In this case, the evolution of ␥(L) towards zero when L ӷRI is in LϪ1 , i.e., much slower than the exponential distance-dependence observed for cut-off-based schemes ͓Eq. ͑32͔͒. For example, for an ion of radius RIϭ0.4 nm, ␥(L) evaluates to Ϫ0.95 for L/2ϭ0.4 nm, Ϫ0.19 for L/2 ϭ3.0 nm, and is above Ϫ0.1 for L/2Ͼ5.7 nm ͓compare with the smaller magnitude and faster relaxation observed for cut-off-based schemes in Fig. 4͑a͔͒. This shows that the application of a cut-off in the computation of ionic solvation free energies by explicit-solvent simulation dramatically reduces the system-size dependence of the calculated solvation free energies compared to lattice-sum methods.19,55,57 More generally, cut-off truncation ͑with the possible inclusion of a reaction-field correction͒ efficiently reduces the impact of finite-size effects and artificial periodicity on the energies and forces in any molecular dynamics simulation. However, this is at the expense of introducing other ͑potentially more harmful͒ artifacts related with the cut-off truncation itself. C. Interaction between two spherical ions The electrostatic solvation free energy profiles ⌬Gsolv(d) for a pair of monovalent spherical ions ͑same or opposite charges, identical radii of 0.4 nm͒ in a solvent of permittivity ⑀sϭ78 are displayed in Fig. 5 as a function of the interionic distance d for different choices of boundary conditions and treatments of the electrostatic interactions based on a single cut-off RCϭ1.2 nm. The PBC curves correspond to ions aligned along an axis of a cubic unit cell of edge Lϭ6 nm. The corresponding profiles computed using the 3D-FFT method60 ͑ST,RF/PBC͒ are also displayed for comparison. As expected, the curves corresponding to the NPBC case present a minimum ͑maximum͒ at ionic contact for ions of identical ͑opposite͒ charges, and asymptotically converge to a common value for a given scheme. More precisely, in the limit of large interionic distances ͑isolated ions͒, ⌬Gsolv(d) converges towards twice the solvation free energy ⌬Gsolv ion of a single ion (⌬Gsolv ion ϭϪ129.4 or Ϫ86.4 kJ molϪ1 for the ST or RF schemes; see Table II͒. In the limit d→0 ͑superimFIG. 4. Periodicity-induced perturbation of the solvation free energy of a spherical ion. The system consists of a single ion of charge qIϭ1 e and radius RI in a cubic periodic box of edge L filled by a solvent of permittivity ⑀sϭ78. Electrostatic interactions correspond to either the ST or RF schemes, with cut-off radii RISϭRSSϭRC . ͑a͒ Relative periodicity-induced shift ␥(L) in the solvation free energy ͓Eq. ͑29͔͒, displayed as a function of L/2. ͑b͒ Logarithm of minus ␥(L) amplified by RC /RI , displayed as a function of L/2RC . The dashed lines corresponding to a least-squares fit ͓over the interval L/(2RC)ϭ0–2.5] corresponding to either the RF or the ST schemes. 9138 J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Bergdorf, Peter, and Hu¨nenberger Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp posed ions͒, ⌬Gsolv(d) converges towards four times ⌬Gsolv ion ͑zero͒ for ions of identical ͑opposite͒ charges. Under PBC, symmetry and periodicity constraints impose that the profile possesses a stationary point at dϭL/2 and is symmetrical with respect to this point. This induces a difference between the NPBC and PBC curves for d in the interval ͓0;L/2͔. Within this range, the perturbation caused by the introduction of periodicity is attractive ͑repulsive͒ for ions of identical ͑opposite͒ charges. However, the magnitude of the effect is very small. For example, the differences between the values of ⌬Gsolv(L/2) under NPBC and PBC is only about 0.45 kJ molϪ1 in magnitude for all cases considered. Thus, in contrast to the case of lattice-sum methods,57 artificial periodicity has very little influence on the solvation free energy profile for pairs of small monovalents ions in a solvent of high permittivity when cut-off truncation is applied. Finally, the agreement between the present calculations employing the 3D-Direct method under PBC and the 3D-FFT method60 is quite good, especially for the ST scheme. For the RF scheme, the agreement is slightly worse, probably due to small differences in the application of boundary smoothing at the ion surface and ion–solvent cutoff distance.60 The corresponding profiles for the overall electrostatic contribution ⌬Gel(d) to the potential of mean force ͓Eq. ͑30͔͒ are displayed in Fig. 6 as a function of the interionic distance d. At short distances, the curves corresponding to NPBC and PBC are nearly identical. In the ST case, the curves present minima ͑maxima͒ at contact and at the cut-off distance for ions of identical ͑opposite͒ charges. The presence of an extremum at the cut-off distance is clearly an artifact related to the use of truncated electrostatic interactions. These profiles provide an explanation for a number of observations made in explicit-solvent simulations employing the ST scheme: ͑i͒ The tendency for ion pairs with like charges to cluster at distances close to the cut-off distance;9,96 ͑ii͒ the tendency for ion pairs of opposite charges to avoid distances close to the cut-off distance;9,42,43,96 ͑iii͒ the artificially increased38,40,41,57 stability of contact ion pairs for ions of like charges;36,37,39,41,87 ͑iv͒ the artificially decreased stability of contact pairs for ions of opposite charges.42,44,80 In the RF case, the profiles nearly present the expected behavior, namely repulsion ͑attraction͒ for ions of like ͑opposite͒ charges, except for a significant artifact in the neighborhood of the cut-off distance. For ions of like charges, a spurious minimum occurs just below the cut-off distances, while for ions of opposite charges, a spurious maximum occurs just above the cut-off distance. Although these artifacts might affect the populations of contact pairs for ions of like or opposite charges in simulations of ionic solution,88,89 the magnitude of these artifacts is limited compared to the ST scheme, in agreement with previous observations.44,90 At large distances from the ion, and for both the ST and RF schemes, the NPBC profiles tend to be close to the expected Coulombic limit ͓Eq. ͑31͔͒. However, the exact agreement is difficult to assess since positive deviations occur, which are probably related to the limited size of the computational domain. V. CONCLUSION In the present study, continuum electrostatics was used to investigate the nature and magnitude of the perturbations induced by cut-off truncation and artificial periodicity in FIG. 5. Electrostatic solvation free energy profile ⌬Gsolv(d) for a pair of monovalent spherical ions. The system consists of two ions of radii RI ϭ0.4 nm bearing identical ͑a and c͒ or opposite ͑b and d͒ charges, and separated by a distance d in a medium of permittivity ⑀sϭ78. It is either nonperiodic ͑NPBC; spherical domain of radius Sϭ4.0 nm) or periodic ͑PBC; cubic unit cell of edge Lϭ6 nm; ions aligned with an axis of the unit cell͒. Electrostatic interactions correspond to either the ST ͑a and b͒ or RF ͑c and d͒ schemes, with cut-off radii RISϭRSSϭRCϭ1.2 nm. In addition to the results of the 3D-Direct method ͑present article͒, the solvation free energies computed from the 3D-FFT method ͑Refs. 51 and 60͒ for the specific interaction scheme ͑ST,RF/PBC͒ and the same value of L are also presented for comparison. FIG. 6. Electrostatic contribution ⌬Gel(d) to the potential of mean force for a pair of monovalent spherical ions. The system consists of two ions of radii RIϭ0.4 nm bearing identical ͑a and c͒ or opposite ͑b and d͒ charges, and separated by a distance d in a medium of permittivity ⑀sϭ78. It is either nonperiodic ͑NPBC; spherical domain of radius Sϭ4.0 nm) or periodic ͑PBC; cubic unit cell of edge Lϭ6 nm; ions aligned with an axis of the unit cell͒. Electrostatic interactions correspond to either the ST ͑a and b͒ or RF ͑c and d͒ schemes, with cut-off radii RISϭRSSϭRCϭ1.2 nm. The ideal longrange limit ⌬Gel lr (d) is also presented for comparison ͓Eq. ͑31͔͒. It is calculated using ⌬Gsolv ion ϭϪ129.4 (ST) or Ϫ86.4 (RF) kJ molϪ1 ͑Table II͒. 9139J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Cut-off and periodicity effects in simulations of ions Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp explicit-solvent simulations of ions and ion pairs. To this purpose, a new algorithm relying on finite integration was developed to solve the equations of continuum electrostatics based on truncated ͑and possibly reaction-field corrected͒ solute–solvent and solvent–solvent electrostatic interactions, under either nonperiodic ͑NPBC͒ or periodic ͑PBC͒ boundary conditions. This algorithm was tested and validated by comparison with available methods ͑Table I͒ whenever pos- sible. In the context of the solvation of a single spherical ion, the main observations can be summarized as follows: ͑A͒ The application of cut-off truncation ͑under NPBC͒ significantly affects the solvent polarization around a spherical ion ͑compared to the ideal Born result͒. With straight truncation ͑ST͒ of the interactions, the solvent is overpolarized within the cut-off sphere of the ion and underpolarized outside this sphere. When a reactionfield ͑RF͒ correction is applied, the agreement with the Born ͑NPBC͒ or lattice-sum ͑PBC͒ polarization is significantly improved, the deviations being largest in the neighborhood of the cut-off distance. ͑B͒ The introduction of artificial periodicity (NPBC →PBC) when applying cut-off-based electrostatics leads to a depolarization of the solvent around the ion in the reference unit cell. This effect is caused by the indirect ͑solvent-mediated͒ perturbation of the solvent molecules in this reference cell by the periodic copies of the ion. The depolarization is more significant for the RF scheme compared to the ST scheme. ͑C͒ The application of cut-off truncation ͑under NPBC͒ decreases the magnitude of the ionic solvation free energy of a spherical ion ͑compared to the ideal Born result͒. The magnitude of this effect is highly sensitive to the electrostatic scheme ͑ST or RF͒ and to the choice of a cut-off radius. However, as discussed in Appendix B, the problem could be largely ͑though approximately͒ remedied in explicit-solvent simulations by the systematic inclusion of an appropriate self-energy term in the total electrostatic energy of the system. Alternatively, an exact correction term to ionic solvation free energies computed from explicit-solvent simulations can be obtained by the application of the present continuum-electrostatics method under NPBC or of its one-dimensional analog.27,28 ͑D͒ The introduction of artificial periodicity (NPBC →PBC) when applying cut-off-based electrostatics causes a further decrease in the magnitude of the ionic solvation free energy. In contrast to lattice-sum methods,57 where this free-energy shift is important even for relatively large system sizes ͑proportional to LϪ1 , L being the edge length of the cubic unit cell͒, the effect decays rapidly with increasing system sizes ͑proportional to RϪ1 exp(ϪcL/R), R being the cut-off distance͒ in the case of cut-off-based electrostatics. Here also, an approximate correction term was derived ͓Eq. ͑34͔͒ that can be applied to correct ionic solvation free energies evaluated from explicit-solvent simulations under PBC. The relevance of these observations can be appreciated by recalling the five sources of error related to the computation of ionic solvation free energies from explicit-solvent simulations relying on cut-off-based electrostatic interactions: ͑i͒ Incorrect polarization of the solvent around the ion due to truncated ͑and possibly modified͒ electrostatic interactions; ͑ii͒ Cut-off- and system size-dependent perturbation of the solvent polarization due to artificial periodic boundary conditions; ͑iii͒ artifacts at the cut-off distance arising from the finite size of the solvent molecules, and related to the use of either a molecular or an atomic cut-off; ͑iv͒ artificial heating during molecular-dynamics simulations due to the possible presence of discontinuities in the atomic forces; ͑v͒ inaccuracy of the ion–solvent and solvent–solvent interaction functions and parameters ͑force fields͒. Only with the understanding of these five sources of error and the design of appropriate correction schemes will it be possible to obtain accurate ionic solvation free energies from explicit-solvent simulations. The discussion ͑and correction͒ of the two first sources of error has been the focus of the present article. The third problem has been previously discussed by a number of groups.55,72–75,77,78,91 Due to the finite size of solvent molecules, the solvent-generated potential at the ion site ͑and thus the solvation free energy͒ may vary considerably depending on whether cut-off truncation is applied on an atomic or on a molecular basis ͑and in the latter case, on the choice of a molecular center͒. Although the debate is not yet completely settled, a number of arguments suggest that molecular truncation ͑based on the center of van der Waals interactions for a spherical molecule77 ͒ represents the appropriate method for evaluating the solventgenerated potential at the ion site,74,77,91 while a ͑generally sizeable͒ correction term must be applied if atomic truncation ͑or a lattice-sum method͒ is employed instead. The fourth problem, namely the artificial heating of molecules at distances close to the cut-off radius, may be alleviated by the use of an effective truncated electrostatic interaction that is continuously differentiable at the cut-off distance, together with atomic truncation. For example, the Barker–Watts reaction-field interaction14–16 causes very limited heating provided that the permittivity of the solvent considered is high and that an ͑unusual͒ atomic-cut-off implementation is applied.16 Finally, in regard to the fifth problem, it should be stressed that the derivation of force-field parameters for ion– solvent interactions based on experimental ionic solvation free energies makes little sense before the four other problems are solved ͑i.e., methodology-independent ionic solvation free energies can be be obtained from explicit-solvent simulations͒. In the context of the potential of mean force for the interaction between two spherical ions, the main observations can be summarized as follows: ͑A͒ The application of cut-off truncation ͑under NPBC͒ induces serious artifacts in the overall electrostatic contribution to the potential of mean force for the interaction between two spherical ions. As previously observed in explicit-solvent simulations, these lead to spurious features in the radial distribution functions 9140 J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Bergdorf, Peter, and Hu¨nenberger Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp close to the cut-off distance9,42,43 and to artifacts in the relative stability of the contact, solvent-separated and free ion pairs.36–42,44,57,80,87 These effects are reduced ͑although not compleletly eliminated͒ by the application of a reaction-field correction. ͑B͒ The introduction of artificial periodicity (NPBC →PBC) when applying cut-off-based electrostatics appears to cause very small changes in the electrostatic contribution to the ͑minimum-image͒ potentials of mean force for small monovalent ions in a solvent of high permittivity. A rather weak periodicity-induced perturbation was also reported in this case for latticesum methods.57 However, the causes of the limited overall effect are different. In the cut-off case, both the solvation free energy profile and the direct ion–ion interaction are almost unaffected by periodicity. In the lattice-sum case, both contributions are largely affected, but the two perturbations nearly cancel each other. Explicit-solvent simulations of ion solvation and ion– ion interactions are currently in progress to confirm the validity of the above considerations derived from a continuumelectrostatics analysis, and their compatibility with the results of simulations employing lattice-sum methods. ACKNOWLEDGMENTS The Swiss National Science Foundation ͑Project 2100- 061939͒ is gratefully acknowledged for financial support, as well as Peter Gee for his careful reading of the manuscript and helpful suggestions. APPENDIX A: POLARIZATION IN THE BARKER–WATTS REACTION-FIELD SCHEME As shown in Fig. 3, the polarization corresponding to the RF scheme under NPBC is very close to the corresponding Born polarization, the deviation being largest at distances close to the cut-off radius. Here, we derive a number of results related to this comparison, in the more general context of the Barker–Watts ͑BW͒ interaction function ͓Eq. ͑2͔͒, namely that ͑i͒ the BW/NPBC and CB/NPBC ͑Born͒ polarizations become identical in the limit RIS ,RSS→ϱ, irrespective of the value of ␣; ͑ii͒ when RISϭRSS , the BW/NPBC polarization converges towards the Born polarization at short distances from the ion, provided that ␣ϭ2(⑀sϪ1)/(2⑀s ϩ1); ͑iii͒ the BW/NPBC polarization, just as the Born polarization, is proportional to r2 at large distances from the ion, provided that ␣ϭ(⑀sϩ2)/(⑀sϪ1). The latter results obviously remain approximately valid for a solvent of high permittivity (⑀sӷ1) when ␣ is set to one ͑RF͒ or close to one. The derivations are based on continuum-electrostatics results presented in a previous article,27,28 and applied to the specific case of truncated electrostatic interactions corresponding to the Barker–Watts potential ͓Eq. ͑2͔͒. In this case the radial polarization around a solvated spherical ion is a solution of the integral equation p͑r͒ϭg͑r͒ϩ ⑀sϪ1 ⑀s ͵max(RI ,͉rϪRSS͉) rϩRSS K͑r,rЈ͒p͑rЈ͒drЈ for rϾRI , ͑A1͒ together with p(r)ϭ0 for rрRI . The inhomogeneous term g(r) is related to the vacuum field generated by the ion g͑r͒ϭ qI 4␲ ⑀sϪ1 ⑀s H͑RISϪr͒ͩrϪ2 Ϫ ␣r RIS 3 ͪ, ͑A2͒ and the kernel K(r,rЈ) to the form of the solvent–solvent interactions K͑r,rЈ͒ϭϪ͑␣ϩ2͒ r4 ϩ͓RSS 2 Ϫ͑rЈ͒2 ͔2 Ϫ2r2 ͓RSS 2 ϩ͑rЈ͒2 ͔ 16r2 RSS 3 . ͑A3͒ First, it is shown that in the limit RIS ,RSS→ϱ, the polarization defined by Eq. ͑A1͒ converges to the Born polarization ͓Eq. ͑26͔͒. Due to the form of Eq. ͑A2͒, limRIS→ϱ͓g(r)ϪpBorn(r)͔ϭ0. It is thus sufficient to prove that the integral term in Eq. ͑A1͒ vanishes in the limit of an infinite solvent–solvent cut-off radius. In this limit, the lower bound of integration can be set to RSSϪr. For rЈ within the interval RSSϪr to RSSϩr, K(r,rЈ) is positive and possesses a single maximum at r˜Јϭ(RSS 2 ϩr2 )1/2 with K(r,r˜Ј)ϭ(␣ ϩ2)/(4RSS). Assuming that the polarization is positive and finite over the whole range of distance ͑with a maximum value p˜), upper and lower bounds can be given to the integral in Eq. ͑A1͒ 0р ͵RSSϪr RSSϩr K͑r,rЈ͒p͑rЈ͒drЈ р2rp˜K͑r,r˜Ј͒ϭ ␣ϩ2 2 p˜ r RSS , ͑A4͒ which shows that the integral vanishes for any finite r in the limit RSS→ϱ. Thus, the BW/NPBC polarization converges to the Born polarization in the limit of large cut-off radii, irrespective of the value of ␣. This result is in particular valid for the ST27 (␣ϭ0) and RF (␣ϭ1) schemes. As a second result, it is shown that the polarization defined by Eq. ͑A1͒ converges to the Born polarization in the limit of short distances when RISϭRSS , provided that ␣ ϭ2(⑀sϪ1)/(2⑀sϩ1), i.e., when Eq. ͑3͒ is used with ⑀Ј ϭ⑀s ͑adjusted boundary conditions61 ͒. For rрmin͓RIS ,RSS ϪRI͔, the Heaviside function in Eq. ͑A2͒ can be omitted and the lower bound in Eq. ͑A1͒ replaced by RSSϪr. Thus, one looks for a solution of p͑r͒ϭ ⑀sϪ1 ⑀s ͫqI 4␲ ͩrϪ2 Ϫ ␣r RIS 3 ͪ ϩ ͵RSSϪr RSSϩr K͑r,rЈ͒p͑rЈ͒drЈͬ. ͑A5͒ Using the result ͵RSSϪr RSSϩr K͑r,rЈ͒͑rЈ͒Ϫ2 drЈϭ ␣ϩ2 3 r RSS 3 , ͑A6͒ 9141J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Cut-off and periodicity effects in simulations of ions Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp one shows easily that the function satisfying Eq. ͑A5͒ when RISϭRSS and ␣ϭ2(⑀sϪ1)/(2⑀sϩ1) is the Born polarization ͓Eq. ͑26͔͒. Therefore, with this specific value of ␣, the BW/NPBC polarization will converge to the Born polarization at short distances, irrespective of the cut-off values. For a solvent of high permittivity (⑀sӷ1), this result will remain approximately valid when ␣ϭ1 ͑RF͒. As a third result, it is shown that the polarization defined by Eq. ͑A1͒ possesses an rϪ2 dependence in the limit of large distances ͑just as the Born polarization͒, provided that ␣ϭ(⑀sϩ2)/(⑀sϪ1). For rуmax͓RIS ,RSSϩRI͔, Eq. ͑A2͒ implies g(r)ϭ0 and the lower bound in Eq. ͑A1͒ can be replaced by rϪRSS . Thus, one looks for a solution of p͑r͒ϭ ⑀sϪ1 ⑀s ͵rϪRSS rϩRSS K͑r,rЈ͒p͑rЈ͒drЈ. ͑A7͒ Using the result ͵rϪRSS rϩRSS K͑r,rЈ͒͑rЈ͒Ϫ2 drЈϭ ␣ϩ2 3 rϪ2 , ͑A8͒ one shows easily that the function satisfying Eq. ͑A7͒ when ␣ϭ(⑀sϩ2)/(⑀sϪ1) is crϪ2 , where c is a constant. Therefore, with this specific value of ␣, the BW/NPBC polarization will possess a rϪ2 dependence in the limit of large distances. For solvent of high permittivities (⑀sӷ1), this result will remain approximately valid when ␣ϭ1 ͑RF͒ or when ␣ is given by Eq. ͑3͒ with ⑀Јϭ⑀s ͑adjusted boundary condi- tions͒. The above observations are illustrated in Fig. 7 for the case of an ion of radius RIϭ0.4 nm immersed in a solvent of permittivity ⑀sϭ78. In the inset, the polarization p(r) is displayed as a function of rϪ2 for the RF scheme (␣ϭ1), using four different values of the cut-off radius RC . The Born polarization is linear in rϪ2 and corresponds to the diagonal of the graph. Besides a small artifact close to the cut-off distance ͑only visible in the graph for RCϭ1.2 and 1.6 nm͒ and a slight offset ͑only visible in the graph for RCϭ1.2), the curves are nearly indistinguishable from each other and from the Born polarization. In the main graph, the differences between p(r) and pBorn(r) are displayed as a function of rϪ2 for the same values of RC . In the short distance limit ͑right side of the graph͒, convergence towards the Born polarization is evident although slow. However, the differences are of rather small magnitude. For example, the relative difference at the surface of the ion, ͓pBorn(RI)Ϫp(RI)͔/pBorn(RI), evaluates to 0.91, 0.29, 0.12, or 0.06% for RCϭ1.2, 1.6, 2.0 or 2.4 nm. In the long distance limit ͑left side of the graph͒, the approximate rϪ2 evolution of p(r)ϪpBorn(r) can also be observed. The maximal error in the polarization occurs exactly at the cut-off distance. Increasing the cut-off distance rapidly reduces the magnitude of this error, the difference pBorn(RC)Ϫp(RC) evaluating to Ϫ0.021, Ϫ0.009, Ϫ0.004, Ϫ0.003 e nmϪ2 for RCϭ1.2, 1.6, 2.0, or 2.4 nm. These results indicate that for large-enough cut-off distances, the RF/NPBC scheme provides an essentially correct representation of the polarization around a spherical ion, in both the short- and long-distance ranges. APPENDIX B: SELF-ENERGY TERM FOR CUT-OFF-BASED INTERACTION FUNCTIONS Here it is shown that when an effective cut-off-based interaction function is used to handle electrostatic interactions in an explicit-solvent simulation, a charge self-energy term should be included into the total electrostatic energy of the system to ensure a fast convergence of ionic solvation free energies towards the Born result in the limit of large cut-off distances. Generalizing this observation to the case of more complex molecular systems, a new definition ͓Eq. ͑B4͔͒ is proposed for the electrostatic interaction energy in simulations employing the Barker–Watts reaction-field scheme. Consider an effective cut-off-based electrostatic interaction function where the potential generated at r by a unit charge at the origin is given by ␺͑r͒ϭ 1 4␲⑀o H͑RϪr͓͒rϪ1 ϩ␺˜ ͑r͔͒, ͑B1͒ where R is the cut-off radius, chosen smaller than half the smallest dimension of the computational box. It is further assumed that ͑i͒ the interaction function vanishes at rϭR, i.e., ␺(R)ϭ0, ͑ii͒ ␺˜ is a sum of terms of the form RϪlϪ1 rl with lу0, and ͑iii͒ the polarization around a spherical ion converges to the Born polarization ͓Eq. ͑26͔͒ in the limit of an infinite cut-off distance. For example, as shown in Appendix A, the Barker–Watts interaction function ͓Eq. ͑2͔͒ satisfies the three conditions irrespective of the value of ␣. In fact, there is some hint that the form of Eq. ͑B1͒ and the second condition automatically imply the third one.60 FIG. 7. Radial polarization p(r) around a solvated spherical ion ͓Eqs. ͑22͔͒ compared to the Born polarization ͓Eq. ͑26͔͒. The system consists of a single ion of charge qIϭ1 e and radius RIϭ0.4 nm in a solvent of permittivity ⑀sϭ78 under NPBC ͑spherical domain of radius Sϭ10.0 nm). Electrostatic interactions correspond to the RF scheme, with cut-off radii RIS ϭRSSϭRCϭ1.2, 1.6, 2.0, or 2.4 nm. The polarization is computed using the 1D-Direct method ͑Refs. 27 and 28͒ for the specific interaction scheme ͑RF/NPBC͒. In the inset, p(r) is displayed as a function of rϪ2 , while the Born polarization pBorn(r) ͑not displayed͒ is a straight line corresponding to the diagonal of the graph. In the main graph, the difference p(r) ϪpBorn(r) is displayed as a function of rϪ2 . 9142 J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Bergdorf, Peter, and Hu¨nenberger Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp Under these assumptions and for large-enough cut-off distances, the polarization within the cut-off sphere of an ion should be quite close to the Born polarization ͓see Fig. 3͑b͒ for the RF scheme͔. Using the Born polarization as an approximation in this case, one may estimate the corresponding ionic solvation free energy. Combining Eq. ͑10͒ with V(r) ϭϪqIٌ␺(r) and Eq. ͑26͒, one obtains the approximate ex- pression ⌬GsolvϷ qI 2 8␲ ⑀sϪ1 ⑀s ͵RI R dr4␲ r r •ٌ␺͑r͒ ϭ qI 2 2 ⑀sϪ1 ⑀s ͓␺͑R͒Ϫ␺͑RI͔͒ ϭ⌬GBornϪ qI 2 8␲⑀o ⑀sϪ1 ⑀s ␺˜ ͑RI͒, ͑B2͒ where we used ␺(R)ϭ0 and inserted Eq. ͑28͒. In the RF case, estimates based on this equation can be compared to the data in Table II. For example, for ⑀sϭ78, qIϭ1 e, RI ϭ0.2 nm and RISϭRSSϭ1.2 nm, Eq. ͑B2͒ ͓using Eq. ͑2͒ with ␣ϭ1] gives an estimate of Ϫ258.0 kJ molϪ1 ͑including 27.8 kJ molϪ1 for the ␣-dependent contribution͒, to be compared with the numerical value of Ϫ259.2 kJ molϪ1 ͑including 26.2 kJ molϪ1 for the ␣-dependent contribution͒ in Table II. If ␺˜ is a sum of terms of the form RϪlϪ1 rl with l у0, ␺˜ (RI) can be approximated by ␺˜ (0) in the limit of large cut-off radii and small ions. In this case, Eq. ͑B2͒ shows that the ionic solvation free energy computed from an explicitsolvent simulation employing a cut-off-based interaction function will converge significantly faster towards the Born result upon increasing the cut-off distance if a self-energy term of the form ⌬Gselfϭ qI 2 8␲⑀o ⑀sϪ1 ⑀s ␺˜ ͑0͒, ͑B3͒ is included in the electrostatic energy of the system. Note that ⌬Gself converges towards zero in the limit R→ϱ. However, because this term is generally large and converges only as RϪ1 , its inclusion makes a significant difference even for relatively large cut-off radii. Generalizing this observation to more complex molecular systems suggests that a charge self-energy term should be included in explicit-solvent molecular dynamics simulations employing any effective cut-off-based electrostatic interaction function. Intuitively, this term may be interpreted as the reversible work required to individually charge the atoms when they are at infinite separation. This work excludes the ͑infinite͒ Coulombic self-energy, but retains the contribution arising from the non-Coulombic term associated with ␺˜ in Eq. ͑B1͒. In the specific case of the Barker–Watts reaction-field method ͓Eq. ͑2͔͒, a reasonable expression for the total electrostatic energy UBW could be UBWϭ͚i ͚jϾi,j excl(i) qiqj␺BW͑r¯ij͒ ϩ 1 4␲⑀o ͚ͭi ͚jϾi,j෈excl(i) qiqj␺˜ BW͑r¯ij͒ ϩ 1 2 ␺˜ BW͑0͚͒ͫi qi 2 Ϫ⑀s Ϫ1 ͚ͩi qiͪ2 ͬͮ, ͑B4͒ where r¯ij is the minimum-image vector corresponding to rij , excl(i) denotes the exclusion list of atom i ͑the distance between any two excluded atoms is assumed to be smaller than R), and ␺˜ BW͑r͒ϭ ␣r2 2R3 Ϫ ␣ϩ2 2R , ͑B5͒ with ␣ defined by Eq. ͑3͒. Note that current simulation programs ͑e.g., GROMOS 92 and GROMACS 93 ͒ typically restrict the implementation of the Barker–Watts reaction-field method to the first term in Eq. ͑B4͒. The second term is explicitly included here because so-called excluded neighbors ͑usually first and second covalent neighbors͒ should only be excluded from the summation of the Coulombic (rϪ1 ) contribution, but not of the reaction-field (␺˜ BW) contribution. The form of the third term has been chosen for consistency in the context of small molecules ͑compared to the cut-off radius and unitcell size͒. For a small molecule ͑or ion͒ gathered by periodicity around its center, r¯ij can be replaced by rij in Eq. ͑B4͒ and the Heaviside function involved in ␺BW can be omitted. In this case, the reaction-field ͑non-Coulombic͒ contribution contribution to UBW can be written UBW rf ϭ 1 8␲⑀o ͭ␣ R3 ͚i ͚jϾi qiqjrij 2 ϩ ⑀sϪ1 ⑀s ␺˜ BW͑0͚͒ͩi qiͪ2 ͮ. ͑B6͒ For a neutral molecule, one has UBW rf,dip ϭϪ 1 8␲⑀o ␣␮2 R3 , ͑B7͒ where ␮ is the molecular dipole moment, which matches the Onsager expression for a dipole in a spherical cavity94 provided that adjusted boundary conditions ͓⑀Јϭ⑀s in Eq. ͑3͔͒ are applied. For a monoatomic ion, one has UBW rf,ion ϭ⌬Gself , i.e., the self-energy term suggested by Eq. ͑B3͒. Note that the last term in Eq. ͑B4͒ only affects the energy of the system, but does not induce atomic forces. However, it may be essential to include it in free-energy calculations involving alterations of the atomic partial charges. In the specific case of a single ion, the inclusion of such a self-energy term should substantially reduce the error on ionic solvation free energies computed from explicit-solvent simulations with finite cut-off distances. This can be seen from the corresponding corrected values ⌬Gsolv corr ϭ⌬Gsolv ϩ⌬Gself reported in Table II. In the RF case, taking the same example as above (⑀sϭ78, qIϭ1 e, RIϭ0.2 nm and RIS ϭRSSϭ1.2 nm), ⌬Gsolv Born evaluates Ϫ342.9 kJ molϪ1 , to be compared with an estimate ⌬Gsolv corr of Ϫ344.9 kJ molϪ1 . The corresponding estimate for the ST case, Ϫ363.4 kJ molϪ1 , is significantly less accurate. This is probably due to the poorer 9143J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Cut-off and periodicity effects in simulations of ions Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp agreement between the polarization within the cut-off sphere of an ion and the Born polarization in this case ͓see Fig. 3͑a͔͒. 1 M. Karplus and J. A. McCammon, Nat. Struct. Biol. 9, 646 ͑2002͒. 2 X. Daura, A. Gla¨ttli, P. Gee, C. Peter, and W. F. van Gunsteren, Adv. Protein Chem. 62, 341 ͑2002͒. 3 T. E. Cheatham, III and P. A. Kollman, Annu. Rev. Phys. Chem. 51, 435 ͑2000͒. 4 L. Saiz, S. Bandyopadhyay, and M. L. Klein, Biosci Rep. 22, 151 ͑2002͒. 5 S. B. Engelsen, C. Monteiro, C. Herve´ de Penhoat, and S. Pe´rez, Biophys. Chem. 93, 103 ͑2001͒. 6 W. F. van Gunsteren and H. J. C. Berendsen, Angew. Chem., Int. Ed. Engl. 29, 992 ͑1990͒. 7 M. P. Allen and D. J. Tildesley, Computer Simulation of Liquids ͑Oxford University Press, New York, 1987͒. 8 B. R. Brooks, R. E. Bruccoleri, B. D. Olafson, D. J. States, S. Swaminathan, and M. Karplus, J. Comput. Chem. 4, 187 ͑1983͒. 9 C. L. Brooks III, B. M. Pettitt, and M. Karplus, J. Chem. Phys. 83, 5897 ͑1985͒. 10 K. S. Kim, Chem. Phys. Lett. 156, 261 ͑1989͒. 11 R. J. Loncharich and B. R. Brooks, Proteins: Struct., Funct., Genet. 6, 32 ͑1989͒. 12 P. J. Steinbach and B. R. Brooks, J. Comput. Chem. 15, 667 ͑1994͒. 13 J. Norberg and L. Nilsson, Biophys. J. 79, 1537 ͑2000͒. 14 J. A. Barker and R. O. Watts, Mol. Phys. 26, 789 ͑1973͒. 15 I. G. Tironi, R. Sperb, P. E. Smith, and W. F. van Gunsteren, J. Chem. Phys. 102, 5451 ͑1995͒. 16 P. H. Hu¨nenberger and W. F. van Gunsteren, J. Chem. Phys. 108, 6117 ͑1998͒. 17 G. Hummer and D. M. Soumpasis, Phys. Rev. E 49, 591 ͑1994͒. 18 G. Hummer and D. M. Soumpasis, J. Phys.: Condens. Matter 6, A141 ͑1994͒. 19 G. Hummer, L. R. Pratt, and A. E. Garcia, J. Phys. Chem. 100, 1206 ͑1996͒. 20 P. P. Ewald, Ann. Phys. ͑Leipzig͒ 64, 253 ͑1921͒. 21 R. W. Hockney and J. W. Eastwood, Computer Simulation Using Particles ͑Institute of Physics, Bristol, 1981͒. 22 T. Darden, D. York, and L. Pedersen, J. Chem. Phys. 98, 10089 ͑1993͒. 23 U. Essmann, L. Perera, M. L. Berkowitz, T. Darden, H. Lee, and L. G. Pedersen, J. Chem. Phys. 103, 8577 ͑1995͒. 24 M. Neumann, Mol. Phys. 50, 841 ͑1983͒. 25 M. Neumann, O. Steinhauser, and G. S. Pawley, Mol. Phys. 52, 97 ͑1984͒. 26 R. H. Wood, J. Chem. Phys. 103, 6177 ͑1995͒. 27 N. A. Baker, P. H. Hu¨nenberger, and J. A. McCammon, J. Chem. Phys. 110, 10679 ͑1999͒. 28 N. A. Baker, P. H. Hu¨nenberger, and J. A. McCammon, J. Chem. Phys. 113, 2510 ͑1999͒. 29 C. L. Brooks III, J. Phys. Chem. 90, 6680 ͑1986͒. 30 C. L. Brooks III, J. Chem. Phys. 86, 5156 ͑1987͒. 31 J. D. Madura and B. M. Pettitt, Chem. Phys. Lett. 150, 105 ͑1988͒. 32 T. P. Straatsma and H. J. C. Berendsen, J. Chem. Phys. 89, 5876 ͑1988͒. 33 S. G. Kalko, G. Sese, and J. A. Padro, J. Chem. Phys. 104, 9578 ͑1996͒. 34 H. Resat and J. A. McCammon, J. Chem. Phys. 104, 7645 ͑1996͒. 35 H. Resat and J. A. McCammon, J. Chem. Phys. 108, 9617 ͑1998͒. 36 B. M. Pettitt and P. J. Rossky, J. Chem. Phys. 84, 5836 ͑1986͒. 37 L. X. Dang and B. M. Pettitt, J. Phys. Chem. 94, 4303 ͑1990͒. 38 E. Gua`rdia, R. Rey, and J. A. Padro´, J. Chem. Phys. 95, 2823 ͑1991͒. 39 L. X. Dang, B. M. Pettitt, and P. J. Rossky, J. Chem. Phys. 96, 4046 ͑1992͒. 40 G. Hummer, D. M. Soumpasis, and M. Neumann, Mol. Phys. 81, 1155 ͑1994͒. 41 J. S. Bader and D. Chandler, J. Phys. Chem. 96, 6423 ͑1992͒. 42 R. A. Friedman and M. Mezei, J. Chem. Phys. 102, 419 ͑1994͒. 43 G. S. Del Buono, F. E. Figueirido, and R. M. Levy, Chem. Phys. Lett. 263, 521 ͑1996͒. 44 X. Rozanska and C. Chipot, J. Chem. Phys. 112, 9691 ͑2000͒. 45 H. Schreiber and O. Steinhauser, Biochemistry 31, 5856 ͑1992͒. 46 H. Schreiber and O. Steinhauser, Chem. Phys. 168, 75 ͑1992͒. 47 H. Schreiber and O. Steinhauser, J. Mol. Biol. 228, 909 ͑1992͒. 48 P. E. Smith and B. M. Pettitt, J. Chem. Phys. 95, 8430 ͑1991͒. 49 G. S. Del Buono, T. S. Cohen, and P. J. Rossky, J. Mol. Liq. 60, 221 ͑1994͒. 50 L. Perera, U. Essmann, and M. L. Berkowitz, J. Chem. Phys. 102, 450 ͑1994͒. 51 S. Boresch and O. Steinhauser, J. Chem. Phys. 115, 10780 ͑2001͒. 52 P. H. Hu¨nenberger, in Simulation and Theory of Electrostatic Interactions in Solution: Computational Chemistry, Biophysics, and Aqueous Solution, edited by G. Hummer and L. R. Pratt ͑American Institute of Physics, New York, 1999͒, pp. 17–83. 53 M. Deserno and C. Holm, J. Chem. Phys. 109, 7678 ͑1998͒. 54 M. Deserno and C. Holm, J. Chem. Phys. 109, 7694 ͑1998͒. 55 G. Hummer, L. R. Pratt, and A. E. Garcia, J. Phys. Chem. A 102, 7885 ͑1998͒. 56 M. Kastenholz and P. H. Hu¨nenberger, J. Phys. Chem B ͑to be published͒. 57 P. H. Hu¨nenberger and J. A. McCammon, J. Chem. Phys. 110, 1856 ͑1999͒. 58 P. H. Hu¨nenberger and J. A. McCammon, Biophys. Chem. 78, 69 ͑1999͒. 59 W. Weber, P. H. Hu¨nenberger, and J. A. McCammon, J. Phys. Chem. B 104, 3668 ͑2000͒. 60 C. Peter, W. F. van Gunsteren, and P. H. Hu¨nenberger, J. Chem. Phys. ͑to be published͒. 61 S. Boresch and O. Steinhauser, J. Chem. Phys. 111, 8271 ͑1999͒. 62 S. Boresch and O. Steinhauser, J. Chem. Phys. 115, 10793 ͑2001͒. 63 M. Marchi, D. Borgis, N. Levy, and P. Ballone, J. Chem. Phys. 114, 4377 ͑2001͒. 64 C. Peter, W. F. van Gunsteren, and P. H. Hu¨nenberger, J. Chem. Phys. 116, 7434 ͑2002͒. 65 J. D. Madura, M. E. Davis, M. K. Gilson, R. C. Wade, B. A. Luty, and J. A. McCammon, in Reviews in Computational Chemistry, Vol. 4, edited by K. B. Lipkowitz and D. B. Boyd ͑VCH, New York, 1994͒, pp. 229–267. 66 B. Honig, K. Sharp, and A.-S. Yang, J. Phys. Chem. 97, 1101 ͑1993͒. 67 M. K. Gilson, Curr. Opin. Struct. Biol. 5, 216 ͑1995͒. 68 B. Honig and A. Nicholls, Science 268, 1144 ͑1995͒. 69 F. Figueirido, G. S. Del Buono, and R. M. Levy, J. Phys. Chem. B 101, 5622 ͑1997͒. 70 G. Hummer, L. R. Pratt, and A. E. Garcia, J. Chem. Phys. 107, 9275 ͑1997͒. 71 S. Sakane, H. S. Ashbaugh, and R. H. Wood, J. Phys. Chem. B 102, 5673 ͑1998͒. 72 G. Hummer, L. R. Pratt, A. E. Garcia, B. J. Berne, and S. W. Rick, J. Phys. Chem. B 101, 3017 ͑1997͒. 73 H. S. Ashbaugh and R. H. Wood, J. Chem. Phys. 106, 8135 ͑1997͒. 74 J. A˚ qvist and T. Hansson, J. Phys. Chem. B 102, 3837 ͑1998͒. 75 H. S. Ashbaugh and R. H. Wood, J. Phys. Chem. B 102, 3844 ͑1998͒. 76 G. Hummer, L. R. Pratt, A. E. Garcia, S. Garde, B. J. Berne, and S. W. Rick, J. Phys. Chem. B 102, 3841 ͑1998͒. 77 Y. N. Vorobjev and J. Hermans, J. Phys. Chem. 103, 10234 ͑1999͒. 78 T. Darden, D. Pearlman, and L. G. Pedersen, J. Chem. Phys. 109, 10921 ͑1998͒. 79 Y. Y. Sham and A. Warshel, J. Chem. Phys. 109, 7940 ͑1998͒. 80 H. Resat, M. Mezei, and J. A. McCammon, J. Phys. Chem. 100, 1426 ͑1996͒. 81 F. Figueirido, G. S. Del Buono, and R. M. Levy, J. Chem. Phys. 103, 6133 ͑1995͒. 82 L. R. Pratt, G. Hummer, and A. E. Garcia´, Biophys. Chem. 51, 147 ͑1994͒. 83 A. Kovalenko and F. Hirata, J. Chem. Phys. 112, 10403 ͑2000͒. 84 A. Masunov and T. Lazaridis, J. Am. Chem. Soc. 125, 1722 ͑2003͒. 85 M. Neumann and O. Steinhauser, Mol. Phys. 57, 97 ͑1986͒. 86 M. Born, Z. Phys. 1, 45 ͑1920͒. 87 J.-C. Soetens, C. Millot, C. Chipot, G. Jansen, J. G. A´ ngya´n, and B. Maigret, J. Phys. Chem. B 101, 10910 ͑1997͒. 88 L. Degre`ve and F. L. B. da Silva, J. Chem. Phys. 110, 3070 ͑1999͒. 89 D. Sardella Vieira and L. Degre`ve, J. Mol. Struct. 580, 127 ͑2002͒. 90 G. Hummer, D. M. Soumpasis, and M. Neumann, Mol. Phys. 77, 769 ͑1992͒. 91 C. Satheesan Babu, P.-K. Yang, and C. Lim, J. Biol. Phys. 28, 95 ͑2002͒. 92 W. F. van Gunsteren, S. R. Billeter, A. A. Eising, P. H. Hu¨nenberger, P. Kru¨ger, A. E. Mark, W. R. P. Scott, and I. G. Tironi, Biomolecular Simulation: The GROMOS96 Manual and User Guide ͑Verlag der Fachvereine, Zu¨rich, 1996͒. 93 H. J. C. Berendsen, D. van der Spoel, and R. van Drunen, Comput. Phys. Commun. 91, 43 ͑1995͒. 94 L. Onsager, J. Am. Chem. Soc. 58, 1486 ͑1936͒. 95 J. Warwicker and H. C. Watson, J. Mol. Biol. 157, 671 ͑1982͒. 96 P. Auffinger and D. L. Beveridge, Chem. Phys. Lett. 234, 413 ͑1995͒. 9144 J. Chem. Phys., Vol. 119, No. 17, 1 November 2003 Bergdorf, Peter, and Hu¨nenberger Downloaded 13 Jun 2006 to 128.200.197.134. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp