Long-range electrostatic corrections in multipolar/polarizable QM/MM simulations

Abstract

Taking long-range electrostatic effects into account in classical and hybrid quantum mechanics–molecular mechanics (QM/MM) simulations is necessary for an accurate description of the system under study. We have recently developed a method, termed long-range electrostatic corrections (LREC), for monopolar QM/MM calculations. Here, we present an extension of LREC for multipolar/polarizable QM/MM simulations within the LICHEM software package. Reaction barriers and QM–MM interaction energies converge with a LREC cutoff between 20 and 25 Å, in agreement with our previous results. Additionally, the LREC approach for the QM–MM interactions can be smoothly combined with standard shifting or Ewald summation methods in the MM calculations. We recommend the use of QM(LREC)/MM(PME), where the QM region is treated with LREC and the MM region is treated with particle mesh Ewald (PME) summation. This combination is an excellent compromise between simplicity, speed, and accuracy for large QM/MM simulations.

1 Introduction

Long-range electrostatic (LRE) interactions are important for the accurate determination of enzyme structures and reaction paths [7, 8, 38, 51]. Unfortunately, directly incorporating LRE interactions in quantum mechanical–molecular mechanical (QM/MM) simulations can be challenging due to the necessity of modifying the Fock matrix elements.

The most straightforward approach for the treatment of electrostatic interactions in QM/MM simulations is to simulate only a small portion of the solvent around a protein or enzyme active site [5] (i.e., embedding a protein in a water droplet). The droplet’s liquid-like environment can be maintained by either freezing the positions of the molecules in the outer edge or applying a stochastic boundary potential [5]. Droplet simulations are efficient due to the small number of atoms, although, the LRE interactions are neglected. However, some of the missing LRE effects can be implicitly included by embedding the droplet in a classical continuum [2–4, 11, 23, 41].

An alternative approach involves using Mulliken charges to approximate the QM charge distribution in periodic images, so that the QM/MM calculation can be carried out using the Ewald summation or smooth particle mesh Ewald (PME) methods [18, 27, 34, 47]. QM/MM calculations with Ewald/PME methods are effective and can readily be improved by replacing Mulliken charges with charges fit from the electrostatic potential. Similar approaches have been developed to extend the Ewald, Wolf, and isotropic periodic sum methods to calculations with semi-empirical, density functional tight binding, Hartree–Fock, and density functional theory [22, 35, 36, 40, 50]. Very recently, Giese and York have reported the ambient-potential composite Ewald method, where no approximations are employed to represent the QM charge density in reciprocal space [20]. Most of these QM/MM LRE methods require modifications to the self-consistent-field (SCF) matrix elements or post-SCF corrections. While these approaches can be quite accurate, modifying SCF routines can limit which QM packages can be used for the QM/MM calculations.

Recently, Fang et al. [14] introduced a cubic smoothing function to include long-range electrostatic corrections (LREC) in QM/MM simulations. The LREC approach uses a combination of a smoothing function and the minimum image convention to scale the electrostatic interactions such that the potential and forces smoothly decrease to zero at a finite cutoff radius. While energy and force shifting approaches also smoothly truncate the potential at a finite radius [16, 28, 29, 35, 36, 44, 49], LREC has an exceptionally simple implementation in the QM/MM Hamiltonian. In the LREC approach, the external MM monopoles are scaled based on their distance from the QM region, which does not require modifications to the matrix elements or post-SCF corrections. The LREC method has been shown to calculate energies and forces that are within 0.2 % of the PME results when using cutoffs of 20–25 Å [14].

The complications of treating LRE interactions of monopoles also affect potentials that include multipole moments and explicit polarization. Truncation, shifting, and smoothing approaches are appealing for multipolar/polarizable QM/MM simulations due to the relative simplicity of introducing these methods into the Hamiltonian [14]. However, each type of shifted multipole–multipole interaction requires a different correction term. As will be shown below, the LREC approach produces a single scale factor for each MM atom, and hence, does not require modifications to the underlying QM software.

In this paper, we extend the LREC approach to atomic multipole moments within the LICHEM software package and implement QM(LREC)/MM(LRE) calculations, where the QM and MM regions are treated with a different LRE approaches. In Sect. 2, we review multipolar/polarizable QM/MM simulations, followed by a discussion of LRE methods in Sect. 3. Our extension of the LREC method to multipolar/polarizable QM/MM simulations is presented in Sect. 4, and we conclude our discussion by examining the performance of multipolar LREC.

2 Multipolar/polarizable QM/MM

2.1 Multipole moments

The electron density and nuclei in a molecule create a continuous, and often, anisotropic electrostatic potential. In classical simulations, interactions due to the diffuse electron density are generally approximated by a set of pair-wise electrostatic potentials. One of the simplest methods for modeling the molecular electrostatic field is to place point-charges (monopoles) on the atomic centers. Atomic monopoles are often inadequate [8, 25], and some models augment the field by adding massless charged dummy atoms to the molecule [24, 33]. However, the dummy atoms can complicate the potential/dynamics during the simulations.

An alternative approach is to add higher-order moments (multipoles) of the electrostatic potential to the model [8, 30, 43, 45]. A multipole expansion is a Taylor series representation of the continuous electrostatic potential (V_elst) [45],

$$V_{\text{elst}}(r)\approx {\displaystyle \sum _n^{\infty }\frac{\mathrm{d}}{n!\mathrm{d}{r}^n}{V}_{\text{elst}}(r)r^n},$$ (1)

where r is the distance from the point of interest (e.g., nucleus, center of mass, etc) and n is the order of the moment. Typically, MM force fields truncate the Taylor series at the second order [8, 30, 43], and the moments can be determined from an analysis of the electron density or electrostatic potential.

The zeroth moment of the electrostatic potential is the monopole, which is equivalent to a point-charge. The first and second moments are referred to as the dipole and quadrupole, respectively. Multipole moments can produce an anisotropic electrostatic potential around individual atoms, as opposed to the spherical potential wells produced by monopoles. Thus, a set of atomic multipole moments, through the quadrupole moment, can reproduce the molecular electrostatic potential with a higher degree of accuracy than monopole moments alone [8, 45]. Atom centered multipolar models provide a reasonable compromise between computational cost and accuracy in the reproduction of the electrostatic potential at medium and long-range.

2.2 QM/MM Hamiltonian

The polarizable QM/MM Hamiltonian may be expressed as

$$H=H_{\text{qm}}+V_{\text{mm}}+V_{\text{qmmm}}+V_{\text{po}l},$$ (2)

where H_qm is the unperturbed Hamiltonian for the QM subsystem, V_mm is the potential for the MM–MM interactions, V_qmmm is the potential for the QM–MM interactions, and V_pol is the potential due to the polarization of the QM/MM system.

In practice [26], Eq. 2 can be expressed as

$$H=H_{\text{qm}}^{\prime }+V_{\text{mm}}^{\prime }+V_{\text{mm},\text{pol}},$$ (3)

where

$$H_{\text{qm}}^{\prime }=H_{\text{qm}}+V_{\text{qmmm},\text{mtp}},$$ (4)

and

$$V_{\text{mm}}^{\prime }=V_{\text{mm}}+V_{\text{qmmm},\text{bnd}}+V_{\text{qmmm},\text{vdw}}.$$ (5)

Here $H_{\text{qm}}^{\prime }$ adds the MM multipole moments (V_qmmm,mtp) to the QM Hamiltonian, $V_{\text{mm}}^{\prime }$ adds the QM–MM bonded (V_qmmm,bnd) and van der Waals (V_qmmm,vdw) interactions to the MM potential, and V_mm,pol is the many-body MM polarization energy. Thus, Eq. 3 divides the QM–MM interactions and polarization between the QM and MM calculations.

The QM/MM Hamiltonian becomes more complicated when the system is periodic. Since the Coulomb interactions have a relatively long range, particles 20 Å or more from the QM region can have non-negligible contributions to the energy and forces [6, 14, 44, 48]. One would expect that monopole–monopole interactions in a neutral system would begin to cancel with interactions of the opposite sign. In fact, Wolf demonstrated [48] that the r⁻¹ dependence of isolated charged particles behaves as a r⁻⁵ potential in a homogeneous neutral environment. However, since the pair-wise terms in the electrostatic potential often have relatively large magnitudes beyond 10 Å, the potential cannot simply be truncated [1, 6, 8, 15, 16, 37, 38, 42, 44, 48, 51]. As will be discussed below, many LRE methods can be integrated into the periodic QM/MM Hamiltonian.

2.3 Polarization

Polarizable QM/MM simulations combine quantum polarization (due to many-body electrostatic, exchange, and dispersion interactions) with classical polarization (induced dipoles, Drude oscillators, etc). Within LICHEM, the electron density is polarized by both the QM region and the static MM multipole moments [26]. A separate MM calculation is then performed to determine the response of the MM polarizable sites to the QM and MM electrostatic fields. In this manner, only one QM and two MM calculations ( $V_{\text{mm}}^{\prime }$ and V_mm,pol) are required to calculate the relaxed QM/MM total energy and the underlaying QM and MM packages do not need to be modified. Additionally, this approach can easily be implemented for a variety of classical polarizable models.

3 Long-range electrostatics

In this section, some common approaches to calculate long-range electrostatic interactions will be reviewed. While the higher-order moments are neglected in most of this discussion, these long-range methods can readily be extended to include the full multipole–multipole and polarizable interactions [28, 29].

3.1 Shifting

The Coulomb potential produces relatively long-range interactions due to the r⁻¹ distance dependence. The potential is given by

$$V_{ij,\text{cl}}(r_{\mathit{ij}})=\frac{q_i{q}_j}{r_{\mathit{ij}}},$$ (6)

where V_ij,cl is Coulomb’s law in atomic units, q_i is the monopole on atom i, and r_ij is the distance between atoms i and j. For periodic systems, the Coulomb potential is often only calculated up to a distance of 9–20 Å to reduce the computational cost [8, 31]. However, due to the length scale of the potential, the residual energy is often non-negligible at the cutoff radius (R_c) [6, 42, 44]. To mitigate some of the truncation artifacts, the potential can be shifted by a constant; such that it is equal to zero at the cutoff [44]. The energy shifted potential, V_ij,es, is given by

$$V_{ij,\text{es}}(r_{\mathit{ij}})=\frac{q_i{q}_j}{r_{\mathit{ij}}}-\frac{q_i{q}_j}{R_{\mathrm{c}}}.$$ (7)

The force, F_ij, due to a general potential, V_ij, is given by

$$F_{\mathit{ij}}(r_{\mathit{ij}})=-\nabla V_{\mathit{ij}}(r_{\mathit{ij}}),$$ (8)

where ∇ is the gradient operator, and thus,

$$F_{ij,\mathrm{c}l}(r_{\mathit{ij}})=\frac{q_i{q}_j}{r_{\mathit{ij}}^2},$$ (9)

and

$$F_{ij,\text{es}}(r_{\mathit{ij}})=\frac{q_i{q}_j}{r_{\mathit{ij}}^2}.$$ (10)

Since F_ij,cl = F_ij,es, the use of energy-shifted potential still results in finite forces at the cutoff radius [44]. Thus, energy-shifted potentials produce artifacts in calculations which require forces, e.g., MD simulations or geometry optimizations.

The truncated potential can be further improved by shifting the force in a manner analogous to the shifted potential. The shifted force, F_ij,fs, is given by [44]

$$F_{ij,\text{fs}}(r_{\mathit{ij}})=\frac{q_i{q}_j}{r_{\mathit{ij}}^2}-\frac{q_i{q}_j}{R_{\mathrm{c}}^2},$$ (11)

and the force-shifted potential can be derived by integrating the shifted force [44],

$$V_{ij,\text{fs}}(r_{\mathit{ij}})=-{\displaystyle \int F_{ij,\text{fs}}(r_{\mathit{ij}})\mathrm{d}{r}_{\mathit{ij}}}+\mathrm{\Delta }{V}_{\mathit{ij},\text{fs}},$$ (12)

where

$$\mathrm{\Delta }{V}_{ij,\text{fs}}=-\frac{2q_i{q}_j}{R_{\mathrm{c}}}.$$ (13)

The final force-shifted potential is given by

$$V_{ij,\text{fs}}(r_{\mathit{ij}})=\frac{q_i{q}_j}{r_{\mathit{ij}}}+\frac{q_i{q}_j{r}_{\mathit{ij}}}{R_{\mathrm{c}}^2}-\frac{2q_i{q}_j}{R_{\mathrm{c}}}$$ (14)

where both the energy and force are now zero at the cutoff.

Truncation and shifting schemes make the calculation of the Coulomb potential tractable for large systems, but they often produce artifacts due to the artificial nature of the finite cutoff. In principle, the shifting procedure could be continued for higher-order derivatives. However, as will be shown in Sects. 3.3 and 4.1, this approach can drastically alter the electrostatic potential. The energy and force shifting derivations given thus far are by no means the only shifting approaches. Polynomial or exponential damping functions are often applied to the Coulomb potential to shift the energy and accelerate the convergence of the LRE interactions. In general, any shifting approach can be rewritten as

$$V_{ij,\text{sm}}=S(r_{\mathit{ij}})V_{ij,\text{cl}},$$ (15)

where V_ij,sm is the smoothed Coulomb potential and S(r_ij) is a damping function.

3.2 Ewald and PME

The calculation of the total electrostatic potential for an infinite periodic system involves a conditionally convergent sum [48]. However, this issue can be overcome by separating this sum into two absolutely convergent sums, which is the basis of the Ewald method [10]. Ewald summation methods divide the electrostatic potential into a short-range damped Coulomb potential and a long-range Fourier transformed Coulomb potential. The total Coulomb potential, V_tot,cl, may be calculated by [10]

$$V_{\text{tot,cl}}=V_{\text{sr}}+{\stackrel{\sim }{V}}_{\text{lr}}-V_{\text{self}}+{\epsilon }_b,$$ (16)

where V_sr is the sum of the damped short-range Coulomb potential, ${\stackrel{\sim }{V}}_{\text{lr}}$ is the Fourier transformed sum of the long-range Coulomb potential, V_self is the self-interaction potential of the multipoles, and ε_b is a correction term due to the boundary conditions that arises when the unit cell has a finite charge and/or dipole moment [10, 32].

The efficiency of the Ewald method can be improved by employing numerical Fourier transforms. One possibility is by using B-spline functions to interpolate the periodic potential onto a mesh of grid points [9]. This approach, termed (smooth) particle mesh Ewald summation, can be used in conjunction with a fast Fourier transform (FFT) algorithm [13]. The FFT can significantly speed up the calculation of the long-range Coulomb potential and reduces the theoretical computational cost from $\mathcal{O}(N^2)$ to $\mathcal{O}(N\log N)$.

3.3 LREC

Our approach for treating LRE interactions is to scale the electrostatic potential based on the distance between the atoms [14]. This approach smooths the electrostatic potential and forces, which are zero at the cutoff radius. Conceptually, this is equivalent to the procedure used in shifted or Wolf approaches, except that the LREC smoothing function has been designed to produce energies and forces in good agreement with PME.

The smoothed Coulomb potential used by the LREC method, V_ij,_lrec, is given by

$$V_{ij,\text{lrec}}(r_{\mathit{ij}})=f(r_{\mathit{ij}}^{\prime },2)\frac{q_i{q}_j}{r_{\mathit{ij}}},$$ (17)

where

$$f(r_{\mathit{ij}}^{\prime },s)=\left[1-{\left(2r_{\mathit{ij}}^{\prime 3}-3r_{\mathit{ij}}^{\prime 2}+1\right)}^s\right],$$ (18)

and

$$r_{\mathit{ij}}^{\prime }=\left(1-\frac{r_{\mathit{ij}}}{R_{\mathrm{c}}}\right).$$ (19)

Here f is the LREC cubic smoothing function and s is an adjustable integer exponent (see Fig. 1). Unlike most LRE approaches where the damped forces are derived analytically from the damped energy, the LREC force calculations simply use a different exponent in the smoothing function,

$$F_{ij,\text{lrec}}(r_{\mathit{ij}})=-f\left(r_{\mathit{ij}}^{\prime },3\right)\frac{q_i{q}_j}{r_{\mathit{ij}}^2}.$$ (20)

The use of different exponents is related to the length scale of the potential and the gradient. Since the forces act over a shorter range than the potential, it is beneficial to move the inflection point of the smoothing function closer to $\frac{r_{\mathit{ij}}}{R_{\mathrm{c}}}=1$ by increasing the exponent (see Figs. 1, 2). Increasing the exponent reduces the damping at short ranges, while increasing the damping in the regions where the forces are small. In general, the LREC exponent and cutoff can be adjusted to control the convergence, cost, and accuracy of the method, similar to adjusting the short-range cutoff and damping parameter in the Ewald, PME, and Wolf methods. As shown by Fang et al. [14], the LREC approach produces energies (s = 2) and forces (s = 3) in good agreement with PME when the cutoff radius is larger than 20 Å.

Fig. 1.

LREC damping function (Eq. 18) with s = 2 and s = 3. Increasing the exponent moves the inflection point of the LREC function closer to the cutoff radius

Fig. 2.

Energy and force for the Coulomb potential with different damping functions, S(r_ij). At short-range, the LREC smoothing function produces energies and forces nearly identical to those from the true potential, while heavily damping the interactions at longer length scales

4 LREC in QM/MM

4.1 Effective multipole moments

To easily implement any shifting or smoothing function in QM/MM simulations, the MM monopoles can be scaled to a value consistent with the smoothed Coulomb potential,

$$q_j^{\ast }(r_j)=S(r_j)q_j,$$ (21)

where $q_j^{\ast }$ is the scaled monopole in the MM region, S is a generalized damping function, and r_j is the distance between the MM atom and the QM center of mass. Scaling monopoles based on their distance from the QM region is a simple approach which avoids the need to modify the QM integrals or correct the energies. Since the QM external field contains the damped multipole moments, forces on the QM atoms can be calculated using the QM gradients without further corrections. Additionally, there is no need to modify the forces when $r_j^{\prime }$ is slowly varying, which is the case for iterative QM/MM geometry optimizations [52]. However, one needs to be cautious when performing QM/MM–MD simulations, when there is a small number of QM atoms, or when the cutoff radius is small; since $r_j^{\prime }$ will no longer be slowly varying. Figure 3 reports scale factors calculated for a selection of smoothing and shifting methods. It is clear that the LREC approach faithfully represents the electrostatic potential over a longer length scale than other shifting approaches. At $r_j=\frac{R_{\mathrm{c}}}{2}$, the LREC approach retains 75 % of the original Coulomb energy, while the force shift potential has been reduced to 25 %.

Fig. 3.

Monopole scaling factors (Eq. 21) from different electrostatic potentials (CL Coulomb’s law, ES energy shifted, FS force shifted, LREC $f(r_j^{\prime },2)$). LREC faithfully represents the electrostatic field over a larger range than the two shifting approaches

Higher-order multipole moments can be scaled in the same manner as the monopole moments:

$${\mu }_j^{\ast }(r_j)=S_{\text{dp}l}(r_j){\mu }_j,$$ (22)

and

$${\Theta }_j^{\ast }(r_j)=S_{\text{qpl}}(r_j){\Theta }_j,$$ (23)

where S_dpl and S_qpl are the dipole and quadrupole damping functions, respectively. This approach can easily be implemented to produce smoothed multipole expansions. In the LREC approach, Eqs. 22 and 23 reduce to

$${\mu }_j^{\ast }(r_j)=f(r_j^{\prime },s){\mu }_j,$$ (24)

and

$${\Theta }_j^{\ast }(r_j)=f(r_j^{\prime },s){\Theta }_j,$$ (25)

where f and r′ have the same form as in Eqs. 18 and 19. The work Fang et al. [14] and our exploratory calculations (see supporting information) have shown that setting s = 2 is sufficient for QM/MM calculations with only monopoles, while setting s = 3 significantly improves LREC performance for multipolar QM/MM calculations. Within LICHEM, the scaled monopole, dipole, and quadrupole moments can be further approximated as a set of 6 point-charges in an octahedral arrangement [12, 19]. As was shown previously [26], the point-charge approximation is a simple and accurate method for including multipole moments in QM software that can only employ point-charges for the external field.

4.2 Mixed LRE methods

Since QM/MM simulations separate the QM–QM, MM–MM, and QM–MM electrostatic interactions, the long-range electrostatics can be treated with different approximations in each part of Eq. 3. For example, the QM–MM interactions may be treated with the LREC approach while the MM–MM interactions can be calculated with PME. The resulting QM(LREC)/MM(PME) calculations take advantage of the simplicity of the LREC smoothing function in quantum calculations and the speed/accuracy of PME [8, 9, 21] for classical models. It is, perhaps, not intuitively clear why two different LRE approaches can be combined in QM/MM simulations. Separability is key to the validity of smoothly mixing two LRE methods. Two distinct types of separability appear in the QM/MM calculations.

1. Separation of the optimizations: LICHEM employs the iterative QM/MM optimization algorithm [26, 52] shown in Fig. 4. Since the QM and MM degrees of freedom are optimized independently, only a single LRE method is used to calculate the forces during each optimization procedure. Thus, two different LRE approaches are only used simultaneously while calculating the total energy.

2. Separability of the energies: The QM/MM total electrostatic energy, ε_tot, can be written as

   $${\epsilon }_{\text{tot}}={\epsilon }_{\text{qm}}^{\prime }+{\epsilon }_{\text{mm}}^{\prime }+{\epsilon }_{\text{mm,pol}}^{\prime },$$ (26)

   where

   $$\text{QM}(\text{LREC})\to {\epsilon }_{\text{qm}}^{\prime },$$ (27)

   and

   $$\text{MM}(\text{PME})\to {\epsilon }_{\text{mm}}^{\prime },{\epsilon }_{\text{mm,pol}}.$$ (28)

   Here ε represents the electrostatic component of the energies calculated with the QM/MM Hamiltonian (Eq. 3). The QM–MM interaction energies in Eq. 27 are calculated using Gaussian integrals in the QM software, while the electrostatic interactions in Eq. 28 are calculated using the MM force field. Since the QM–MM, MM–MM, and polarization portions of Eq. 26 are already calculated with different algorithms (i.e., Gaussian integrals and force fields), using two sufficiently accurate LRE methods does not significantly affect the total energy. Test calculations with MM(FS)/MM(PME), where the QM region was treated with a force-shifted MM potential, confirmed that there are negligible errors (not reported) compared to full PME calculations.

Fig. 4.

QM/MM optimization procedure (Refs. [26, 52]) implemented in LICHEM. The LRE methods used in each step are shown for QM(LREC)/MM(PME). Based on the change in the atomic positions after optimization; skipped on the first step

QM and polarization calculations are inherently many-body and must include enough of the multipole moments for the calculations to converge. On the other hand, the charges on the QM atoms can simply be neglected in the MM–MM calculations. Since the MM potential is pair-wise additive, calculating the electrostatic interactions without the presence of QM charges generally does not affect the MM–MM interactions.

5 Results

5.1 Computational details

The QM calculations were performed using the 6–31++G(d,p) basis set, and each QM/MM system was solvated using a cubic box of liquid water (32,000 molecules, box length: 98.646 Å). We have tested the LREC method with a variety of cutoff radii up to the maximum cutoff (R_c = 49.323 Å) allowed by the minimum image convention. The QM–MM and MM–MM vdW interactions, on the other hand, were calculated using the default settings of the TINKER software package [39] (polynomial smoothing with R_c = 9 Å). Additionally, we have tested LICHEM’s QM(LREC)/MM(PME) implementation using both the non-polarizable TIP3P [24] and the polarizable AMOEBA [43] potentials.

5.2 QM–MM interactions

To examine the convergence of multipolar LREC calculations, QM/MM calculations were performed on the box of liquid water using the NWChem [46]–TINKER interface in LICHEM. A collection of 5 water molecules in the center of the box were taken as the QM subsystem, and all remaining water molecules were designated as the MM region. The QM–MM interaction energy, E_int, is given by

$$E_{\mathrm{int}}=E_{\text{tot}}-E_{\text{qm}}-E_{\text{mm}},$$ (29)

where E_tot is the QM/MM total energy, E_qm is the energy of the QM subsystem in the gas phase, and E_mm is the MM energy of the system with the QM atoms removed.

Figure 5 reports the convergence of the QM–MM interaction energy as the electrostatic cutoff is increased. The interaction energies from the long-range corrected QM/MM calculations are in good agreement with the results taken from the AMOEBA force field and, as was shown previously [14], the QM/MM energies converge with a cutoff between 20 and 25 Å. The computational cost of LRE methods is almost entirely determined by the cost of the QM calculations and the LRE cutoff radius. As can be seen in Fig. 6, the computational cost of QM/MM simulations can increase rapidly as the cutoff radius is extended. Since multipolar QM/MM simulations are both more expensive and slower to converge than monopolar QM/MM simulations, it is beneficial to use a higher exponent to accelerate the convergence (see supporting information).

Fig. 5.

Convergence of the QM–MM interaction energy (Eq. 29) as the LREC cutoff is increased. For comparison, MM(PME) interaction energies are shown for TIP3P and AMOEBA

Fig. 6.

Computational cost of PME0(LREC)/MM(PME) relative to the cost of PBE0(LREC)/TIP3P(PME) with R_c = 0 Å (i.e., the QM region is in the gas phase)

Additionally, PBE0(LREC)/TIP3P(LREC) calculations, using the in-house modified versions of Gaussian and TINKER, confirmed that the QM/MM total energy was within 0.01 % of the PBE0(LREC)/TIP3P(PME) energy. The excellent agreement between the QM(LREC)/MM(LREC), QM(LREC)/MM(PME), and MM(PME) calculations confirms that the LREC method can be smoothly combined with other LRE methods.

5.3 Reaction barriers

Generally, the determination of reaction barriers is one of the main objectives in QM/MM simulations. To test the QM(LREC)/MM(PME) method on chemical reactions, we have chosen to use the aspartic acid dimer double proton transfer reaction from our previous work [14]. The acid dimer reactant, transition state, and product structures were solvated using the water box from Sect. 5.1 and re-optimized at the ωB97xD/AMOEBA (R_c = 25 Å level of theory.

The reaction barriers were calculated using LICHEM’s Gaussian [17]–TINKER interface for single-point energy calculations with R_c = {2, 5, 10,…, 45, 50} Å. The proton transfer barriers reported in Fig. 7 show large fluctuations with small cutoffs before stabilizing with R_c ≥ 20 Å, around 6.06–6.18 kcal/mol. While the barrier continues to increase beyond R_c = 20 Å, the total change induced by further increasing the cutoff from 20 to 50 Å is ∼0.1 kcal/mol. Furthermore, extrapolating to R_c = ∞ yields a barrier of 6.30 kcal/mol. Thus, the approximate error at R_c = 20 Å is only 0.25 kcal/mol, which is well below the so-called chemical accuracy (1 kcal/mol).

Fig. 7.

Aspartic acid double proton transfer reaction barriers calculated with s = 3 and different values of the LREC cutoff. Between R_c = 20 and 50 Å, the total change in the barrier height is less than ∼ 0.1 kcal/mol, which is within chemical accuracy

6 Conclusion

We have demonstrated that the LREC approach can be extended to multipolar/polarizable models in a straightforward manner. The interaction energies and reaction barriers from the LREC approach converge as the cutoff radius is increased and, in agreement with our previous results, R_c = 20–25 Å is sufficient to obtain converged QM/MM energies. Using multipolar LREC for QM/MM simulations is an easy-to-implement approach for calculating long-range electrostatic interactions, which does not require modifications to the underlying QM software. Furthermore, the QM(LREC) calculations can readily be combined with efficient and accurate long-range electrostatics methods, such as PME, for the MM calculations.