A comparison of different initialization protocols to obtain statistically independent molecular dynamics simulations

We study how the results of molecular dynamics (MD) simulations are affected by various choices during the setup, e.g., the starting velocities, the solvation, the location of protons, the conformation of His, Asn, and Gln residues, the protonation and titration of His residues, and the treatment of alternative conformations. We estimate the binding affinity of ligands to four proteins calculated with the MM/GBSA method (molecular mechanics combined with a generalized Born and surface area solvation energy). For avidin and T4 lysozyme, all variations gave similar results within 2 kJ/mol. For factor Xa, differences in the solvation or in the selection of alternative conformations gave results that are significantly different from those of the other approaches by 4-6 kJ/mol, whereas for galectin-3, changes in the conformations, rotations, and protonation gave results that differed by 10 kJ/mol, but only if residues close to the binding site were modified. This shows that the results of MM/GBSA calculations are reasonably reproducible even if the MD simulations are set up with different software. Moreover, we show that the sampling of phase space can be enhanced by solvating the systems with different equilibrated water boxes, in addition to the common use of different starting velocities. If different conformations are available in the crystal structure, they should also be employed to enhance the sampling. Protonation, ionization, and conformations of Asn, Gln, and His may also be used to enhance sampling, but great effort should be spent to obtain as reliable predictions as possible close to the active site.     

Can MM/GBSA calculations be sped up by system truncation?

We have studied whether calculations of the binding free energy of small ligands to a protein by the MM/GBSA approach (molecular mechanics combined with generalized Born and surface area solvation) can be sped up by including only a restricted number of atoms close to the ligand. If the protein is truncated before the molecular dynamics (MD) simulations, quite large changes are observed for the calculated binding energies, for example, 4 kJ/mol average difference for a radius of 19 Å for the binding of nine phenol derivatives to ferritin. The results are improved if no atoms are fixed in the simulations, with average and maximum errors of 2 and 3 kJ/mol at 19 Å and 3 and 6 kJ/mol at 7 Å. Similar results are obtained for two additional proteins, p38α MAP kinase and factor Xa. On the other hand, if energies are calculated on snapshots that are truncated after the MD simulation, all residues more than 8.5 Å from the ligand can be omitted without changing the energies by more than 1 kJ/mol on average (maximum error 1.4 kJ/mol). At the molecular mechanics level, the gain in computer time for such an approach is small. However, it shows what size of system should be used if the energies instead are calculated with a more demanding method, for example, quantum‐mechanics. © 2017 Wiley Periodicals, Inc.     

A semiempirical approach to ligand‐binding affinities: Dependence on the Hamiltonian and corrections

We present a combination of semiempirical quantum‐mechanical (SQM) calculations in the conductor‐like screening model with the MM/GBSA (molecular‐mechanics with generalized Born and surface‐area solvation) method for ligand‐binding affinity calculations. We test three SQM Hamiltonians, AM1, RM1, and PM6, as well as hydrogen‐bond corrections and two different dispersion corrections. As test cases, we use the binding of seven biotin analogues to avidin, nine inhibitors to factor Xa, and nine phenol‐derivatives to ferritin. The results vary somewhat for the three test cases, but a dispersion correction is mandatory to reproduce experimental estimates. On average, AM1 with the DH2 hydrogen‐bond and dispersion corrections gives the best results, which are similar to those of standard MM/GBSA calculations for the same systems. The total time consumption is only 1.3–1.6 times larger than for MM/GBSA. © 2012 Wiley Periodicals, Inc.     

Comparison of radii sets,entropy, QM methods,and sampling on MM‐PBSA,MM‐GBSA,and QM/MM‐GBSA ligand binding energies of F. tularensis enoyl‐ACP reductase (FabI)

To validate a method for predicting the binding affinities of FabI inhibitors, three implicit solvent methods, MM‐PBSA, MM‐GBSA, and QM/MM‐GBSA were carefully compared using 16 benzimidazole inhibitors in complex with Francisella tularensis FabI. The data suggests that the prediction results are sensitive to radii sets, GB methods, QM Hamiltonians, sampling protocols, and simulation length, if only one simulation trajectory is used for each ligand. In this case, QM/MM‐GBSA using 6 ns MD simulation trajectories together with GB^neck2, PM3, and the mbondi2 radii set, generate the closest agreement with experimental values (r² = 0.88). However, if the three implicit solvent methods are averaged from six 1 ns MD simulations for each ligand (called “multiple independent sampling”), the prediction results are relatively insensitive to all the tested parameters. Moreover, MM/GBSA together with GB^HCT and mbondi, using 600 frames extracted evenly from six 0.25 ns MD simulations, can also provide accurate prediction to experimental values (r² = 0.84). Therefore, the multiple independent sampling method can be more efficient than a single, long simulation method. Since future scaffold expansions may significantly change the benzimidazole's physiochemical properties (charges, etc.) and possibly binding modes, which may affect the sensitivities of various parameters, the relatively insensitive “multiple independent sampling method” may avoid the need of an entirely new validation study. Moreover, due to large fluctuating entropy values, (QM/)MM‐P(G)BSA were limited to inhibitors’ relative affinity prediction, but not the absolute affinity. The developed protocol will support an ongoing benzimidazole lead optimization program. © 2015 Wiley Periodicals, Inc.     

Binding affinities of factor Xa inhibitors estimated by thermodynamic integration and MM/GBSA

We present free energy estimates of nine 3-amidinobenzyl-1H-indole-2-carboxamide inhibitors of factor Xa. Using alchemical thermodynamic integration (TI) calculations, we estimate the difference in binding free energies with high accuracy and precision, except for mutations involving one of the amidinobenzyl rings. Crystal studies show that the inhibitors may bind in two distinct conformations, and using TI, we show that the two conformations give a similar binding affinity. Furthermore, we show that we can reduce the computational demand, while still retaining a high accuracy and precision, by using fewer integration points and shorter protein-ligand simulations. Finally, we have compared the TI results to those obtained with the simpler MM/GBSA method (molecular-mechanics with generalized Born surface-area solvation). MM/GBSA gives better results for the mutations that involve a change of net charge, but if a precision similar to that of the TI method is required, the MM/GBSA method is actually slightly more expensive. Thus, we have shown that TI could be a valuable tool in drug design.     

Assessing the performance of the MM/PBSA and MM/GBSA methods. 1. The accuracy of binding free energy calculations based on molecular dynamics simulations

The Molecular Mechanics/Poisson-Boltzmann Surface Area (MM/PBSA) and the Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) methods calculate binding free energies for macromolecules by combining molecular mechanics calculations and continuum solvation models. To systematically evaluate the performance of these methods, we report here an extensive study of 59 ligands interacting with six different proteins. First, we explored the effects of the length of the molecular dynamics (MD) simulation, ranging from 400 to 4800 ps, and the solute dielectric constant (1, 2, or 4) on the binding free energies predicted by MM/PBSA. The following three important conclusions could be observed: (1) MD simulation length has an obvious impact on the predictions, and longer MD simulation is not always necessary to achieve better predictions. (2) The predictions are quite sensitive to the solute dielectric constant, and this parameter should be carefully determined according to the characteristics of the protein/ligand binding interface. (3) Conformational entropy often show large fluctuations in MD trajectories, and a large number of snapshots are necessary to achieve stable predictions. Next, we evaluated the accuracy of the binding free energies calculated by three Generalized Born (GB) models. We found that the GB model developed by Onufriev and Case was the most successful model in ranking the binding affinities of the studied inhibitors. Finally, we evaluated the performance of MM/GBSA and MM/PBSA in predicting binding free energies. Our results showed that MM/PBSA performed better in calculating absolute, but not necessarily relative, binding free energies than MM/GBSA. Considering its computational efficiency, MM/GBSA can serve as a powerful tool in drug design, where correct ranking of inhibitors is often emphasized.     

The Normal-Mode Entropy in the MM/GBSA Method: Effect of System Truncation, Buffer Region, and Dielectric Constant

We have performed a systematic study of the entropy term in the MM/GBSA (molecular mechanics combined with generalized Born and surface-area solvation) approach to calculate ligand-binding affinities. The entropies are calculated by a normal-mode analysis of harmonic frequencies from minimized snapshots of molecular dynamics simulations. For computational reasons, these calculations have normally been performed on truncated systems. We have studied the binding of eight inhibitors of blood clotting factor Xa, nine ligands of ferritin, and two ligands of HIV-1 protease and show that removing protein residues with distances larger than 8-16 ? to the ligand, including a 4 ? shell of fixed protein residues and water molecules, change the absolute entropies by 1-5 kJ/mol on average. However, the change is systematic, so relative entropies for different ligands change by only 0.7-1.6 kJ/mol on average. Consequently, entropies from truncated systems give relative binding affinities that are identical to those obtained for the whole protein within statistical uncertainty (1-2 kJ/mol). We have also tested to use a distance-dependent dielectric constant in the minimization and frequency calculation (ε = 4r), but it typically gives slightly different entropies and poorer binding affinities. Therefore, we recommend entropies calculated with the smallest truncation radius (8 ?) and ε =1. Such an approach also gives an improved precision for the calculated binding free energies.     

Assessing the performance of the molecular mechanics/Poisson Boltzmann surface area and molecular mechanics/generalized Born surface area methods. II. The accuracy of ranking poses generated from docking

In molecular docking, it is challenging to develop a scoring function that is accurate to conduct high-throughput screenings. Most scoring functions implemented in popular docking software packages were developed with many approximations for computational efficiency, which sacrifices the accuracy of prediction. With advanced technology and powerful computational hardware nowadays, it is feasible to use rigorous scoring functions, such as molecular mechanics/Poisson Boltzmann surface area (MM/PBSA) and molecular mechanics/generalized Born surface area (MM/GBSA) in molecular docking studies. Here, we systematically investigated the performance of MM/PBSA and MM/GBSA to identify the correct binding conformations and predict the binding free energies for 98 protein-ligand complexes. Comparison studies showed that MM/GBSA (69.4%) outperformed MM/PBSA (45.5%) and many popular scoring functions to identify the correct binding conformations. Moreover, we found that molecular dynamics simulations are necessary for some systems to identify the correct binding conformations. Based on our results, we proposed the guideline for MM/GBSA to predict the binding conformations. We then tested the performance of MM/GBSA and MM/PBSA to reproduce the binding free energies of the 98 protein-ligand complexes. The best prediction of MM/GBSA model with internal dielectric constant 2.0, produced a Spearman's correlation coefficient of 0.66, which is better than MM/PBSA (0.49) and almost all scoring functions used in molecular docking. In summary, MM/GBSA performs well for both binding pose predictions and binding free-energy estimations and is efficient to re-score the top-hit poses produced by other less-accurate scoring functions.     

MM/GBSA and LIE estimates of host–guest affinities: dependence on charges and solvation model

The affinities of two sets of guest–host systems were estimated using the popular end-point methods MM/GBSA (molecular-mechanics with generalised Born and surface-area solvation) and LIE (linear interaction energy). A set of six primary alcohols that bind to α-cyclodextrin (α-CD) and a set of eight guest molecules to cucurbit[8]uril (CB8) were considered. Three different charge schemes were used to obtain charges for the host and guest molecules, viz., AM1-BCC, RESP, and the recently suggested xAvESP (which average ESP charges over a number of molecular dynamics snapshots). Furthermore, both the generalised Born and Poisson–Boltzmann solvation models were used in the MM/GBSA calculations. The two solvation models perform equally well in predicting relative affinities, and hence there is no point in using the more expensive Poisson–Boltzmann model for these systems. Both the LIE and MM/GBSA estimates are shown to be robust with respect to the charge model, and therefore it is recommended to use the cheapest AM1-BCC charges. Using AM1-BCC charges, the MM/GBSA method gave a MADtr (mean absolute deviation after removal of systematic error) of 17 kJ/mol and a correlation coefficient (r ²) of 0.67 for the CB8 complexes, and a MADtr of 10 kJ/mol and an r ² of 0.96 for the α-CD complexes. The LIE method gave a MADtr of 20 kJ/mol and an r ² of 0.10 for the CB8 complexes, after optimisation of the non-polar scaling parameter. For the α-CD complexes, no optimisation was necessary and the method gave a MADtr of 2 kJ/mol and a r ² of 0.96. These results indicate that both MM/GBSA and LIE are able to estimate host–guest affinities accurately.     

新型二氟甲基磷酸类酪氨酸蛋白磷酸酯酶1B抑制剂的分子动力学模拟和结合自由能计算

通过分子对接建立了一系列含二氟甲基磷酸基团(DFMP)或二氟甲基硫酸基团(DFMS)的抑制剂与酪氨酸蛋白磷酸酯酶1B(PTP1B)的相互作用模式, 并通过1 ns的分子动力学模拟和molecular mechanics/generalized Born surface area (MM/GBSA)方法计算了其结合自由能. 计算获得的结合自由能排序和抑制剂与靶酶间结合能力排序一致; 通过基于主方程的自由能计算方法, 获得了抑制剂与靶酶残基间相互作用的信息, 这些信息显示DFMP/DFMS基团的负电荷中心与PTP1B的221位精氨酸正电荷中心之间的静电相互作用强弱决定了此类抑制剂的活性, 进一步的分析还显示位于DFMP/DFMS基团中的氟原子或其他具有适当原子半径的氢键供体原子会增进此类抑制剂与PTP1B活性位点的结合能力.     

Binding free energies in the SAMPL6 octa-acid host–guest challenge calculated with MM and QM methods

We have estimated free energies for the binding of eight carboxylate ligands to two variants of the octa-acid deep-cavity host in the SAMPL6 blind-test challenge (with or without endo methyl groups on the four upper-rim benzoate groups, OAM and OAH, respectively). We employed free-energy perturbation (FEP) for relative binding energies at the molecular mechanics (MM) and the combined quantum mechanical (QM) and MM (QM/MM) levels, the latter obtained with the reference-potential approach with QM/MM sampling for the MM → QM/MM FEP. The semiempirical QM method PM6-DH+ was employed for the ligand in the latter calculations. Moreover, binding free energies were also estimated from QM/MM optimised structures, combined with COSMO-RS estimates of the solvation energy and thermostatistical corrections from MM frequencies. They were performed at the PM6-DH+ level of theory with the full host and guest molecule in the QM system (and also four water molecules in the geometry optimisations) for 10–20 snapshots from molecular dynamics simulations of the complex. Finally, the structure with the lowest free energy was recalculated using the dispersion-corrected density-functional theory method TPSS-D3, for both the structure and the energy. The two FEP approaches gave similar results (PM6-DH+/MM slightly better for OAM), which were among the five submissions with the best performance in the challenge and gave the best results without any fit to data from the SAMPL5 challenge, with mean absolute deviations (MAD) of 2.4–5.2 kJ/mol and a correlation coefficient (R²) of 0.77–0.93. This is the first time QM/MM approaches give binding free energies that are competitive to those obtained with MM for the octa-acid host. The QM/MM-optimised structures gave somewhat worse performance (MAD?=?3–8 kJ/mol and R²?=?0.1–0.9), but the results were improved compared to previous studies of this system with similar methods.     

Converging ligand‐binding free energies obtained with free‐energy perturbations at the quantum mechanical level

In this article, the convergence of quantum mechanical (QM) free‐energy simulations based on molecular dynamics simulations at the molecular mechanics (MM) level has been investigated. We have estimated relative free energies for the binding of nine cyclic carboxylate ligands to the octa‐acid deep‐cavity host, including the host, the ligand, and all water molecules within 4.5 Å of the ligand in the QM calculations (158–224 atoms). We use single‐step exponential averaging (ssEA) and the non‐Boltzmann Bennett acceptance ratio (NBB) methods to estimate QM/MM free energy with the semi‐empirical PM6‐DH2X method, both based on interaction energies. We show that ssEA with cumulant expansion gives a better convergence and uses half as many QM calculations as NBB, although the two methods give consistent results. With 720,000 QM calculations per transformation, QM/MM free‐energy estimates with a precision of 1 kJ/mol can be obtained for all eight relative energies with ssEA, showing that this approach can be used to calculate converged QM/MM binding free energies for realistic systems and large QM partitions. © 2016 The Authors. Journal of Computational Chemistry Published by Wiley Periodicals, Inc.     

Binding affinities in the SAMPL3 trypsin and host–guest blind tests estimated with the MM/PBSA and LIE methods

We have estimated affinities for the binding of 34 ligands to trypsin and nine guest molecules to three different hosts in the SAMPL3 blind challenge, using the MM/PBSA, MM/GBSA, LIE, continuum LIE, and Glide score methods. For the trypsin challenge, none of the methods were able to accurately predict the experimental results. For the MM/GB(PB)SA and LIE methods, the rankings were essentially random and the mean absolute deviations were much worse than a null hypothesis giving the same affinity to all ligand. Glide scoring gave a Kendall's τ index better than random, but the ranking is still only mediocre, τ = 0.2. However, the range of affinities is small and most of the pairs of ligands have an experimental affinity difference that is not statistically significant. Removing those pairs improves the ranking metric to 0.4-1.0 for all methods except CLIE. Half of the trypsin ligands were non-binders according to the binding assay. The LIE methods could not separate the inactive ligands from the active ones better than a random guess, whereas MM/GBSA and MM/PBSA were slightly better than random (area under the receiver-operating-characteristic curve, AUC = 0.65-0.68), and Glide scoring was even better (AUC = 0.79). For the first host, MM/GBSA and MM/PBSA reproduce the experimental ranking fairly good, with τ = 0.6 and 0.5, respectively, whereas the Glide scoring was considerably worse, with a τ = 0.4, highlighting that the success of the methods is system-dependent.     

Solvent interaction energy calculations on molecular dynamics trajectories: increasing the efficiency using systematic frame selection

End-point methods such as linear interaction energy (LIE) analysis, molecular mechanics generalized Born solvent-accessible surface (MM/GBSA), and solvent interaction energy (SIE) analysis have become popular techniques to calculate the free energy associated with protein-ligand binding. Such methods typically use molecular dynamics (MD) simulations to generate an ensemble of protein structures that encompasses the bound and unbound states. The energy evaluation method (LIE, MM/GBSA, or SIE) is subsequently used to calculate the energy of each member of the ensemble, thus providing an estimate of the average free energy difference between the bound and unbound states. The workflow requiring both MD simulation and energy calculation for each frame and each trajectory proves to be computationally expensive. In an attempt to reduce the high computational cost associated with end-point methods, we study several methods by which frames may be intelligently selected from the MD simulation including clustering and address the question of how the number of selected frames influences the accuracy of the SIE calculations.     

Stereoselective inclusion mechanism of ketoprofen into β-cyclodextrin: insights from molecular dynamics simulations and free energy calculations

The host–guest inclusion mechanism formed between β-cyclodextrin and those poorly water-soluble drug molecules has important applications in supramolecular chemistry, biology and pharmacy. In this work, the chiral recognition ability of β-cyclodextrin to one of nonsteroidal anti-inflammatory drugs, ketoprofen, has been systematically investigated using molecular dynamics and free energy simulation methods. The R- and S-enantiomers of ketoprofen were explicitly bound within the cyclodextrin cavity in our simulations, respectively. In consistent with experimental observations, tiny structural difference between two isomers could be observed. Calculated absolute binding free energies using adapted biasing force (ABF) method and MM/GBSA approach for both isomers are comparable to experimental values. Significant binding fluctuations along the MD trajectory have been observed. The free energy profiles calculated using two different approaches reveal that the ketoprofen prefers binding in the cavity with the carboxylate group facing the wider edge of β-cyclodextrin. Similar free energy profiles for two enantiomers obtained using ABF calculations indicate that it is very hard to separate and identify the chiral conjugates within the framework of the natural β-cyclodextrin.     

Crystal molecular dynamics simulations to speed up MM/PB(GB)SA evaluation of binding free energies of di-mannose deoxy analogs with P51G-m4-Cyanovirin-N

Complexes of two Cyanovirin-N (CVN) mutants, m4-CVN and P51G-m4-CVN, with deoxy di-mannose analogs were employed as models to generate conformational ensembles using explicit water Molecular Dynamics (MD) simulations in solution and in crystal environment. The results were utilized for evaluation of binding free energies with the molecular mechanics Poisson-Boltzmann (or Generalized Born) surface area, MM/PB(GB)SA, methods. The calculations provided the ranking of deoxy di-mannose ligands affinity in agreement with available qualitative experimental evidences. This confirms the importance of the hydrogen-bond network between di-mannose 3'- and 4'-hydroxyl groups and the protein binding site B(M) as a basis of the CVN activity as an effective HIV fusion inhibitor. Comparison of binding free energies averaged over snapshots from the solution and crystal simulations showed high promises in the use of the crystal matrix for acceleration of the conformational ensemble generation, the most time consuming step in MM/PB(GB)SA approach. Correlation between energy values based on solution versus crystal ensembles is 0.95 for both MM/PBSA and MM/GBSA methods.     

Accurate predictions of nonpolar solvation free energies require explicit consideration of binding-site hydration

Continuum solvation methods are frequently used to increase the efficiency of computational methods to estimate free energies. In this paper, we have evaluated how well such methods estimate the nonpolar solvation free-energy change when a ligand binds to a protein. Three different continuum methods at various levels of approximation were considered, viz., the polarized continuum model (PCM), a method based on cavity and dispersion terms (CD), and a method based on a linear relation to the solvent-accessible surface area (SASA). Formally rigorous double-decoupling thermodynamic integration was used as a benchmark for the continuum methods. We have studied four protein-ligand complexes with binding sites of varying solvent exposure, namely the binding of phenol to ferritin, a biotin analogue to avidin, 2-aminobenzimidazole to trypsin, and a substituted galactoside to galectin-3. For ferritin and avidin, which have relatively hidden binding sites, rather accurate nonpolar solvation free energies could be obtained with the continuum methods if the binding site is prohibited to be filled by continuum water in the unbound state, even though the simulations and experiments show that the ligand replaces several water molecules upon binding. For the more solvent exposed binding sites of trypsin and galectin-3, no accurate continuum estimates could be obtained, even if the binding site was allowed or prohibited to be filled by continuum water. This shows that continuum methods fail to give accurate free energies on a wide range of systems with varying solvent exposure because they lack a microscopic picture of binding-site hydration as well as information about the entropy of water molecules that are in the binding site before the ligand binds. Consequently, binding affinity estimates based upon continuum solvation methods will give absolute binding energies that may differ by up to 200 kJ/mol depending on the method used. Moreover, even relative energies between ligands with the same scaffold may differ by up to 75 kJ/mol. We have tried to improve the continuum solvation methods by adding information about the solvent exposure of the binding site or the hydration of the binding site, and the results are promising at least for this small set of complexes.     

Elucidating the Glucokinase Activating Potentials of Naturally Occurring Prenylated Flavonoids: An Explicit Computational Approach

Glucokinase activators are considered as new therapeutic arsenals that bind to the allosteric activator sites of glucokinase enzymes, thereby maximizing its catalytic rate and increasing its affinity to glucose. This study was designed to identify potent glucokinase activators from prenylated flavonoids isolated from medicinal plants using molecular docking, molecular dynamics simulation, density functional theory, and ADMET analysis. Virtual screening was carried out on glucokinase enzymes using 221 naturally occurring prenylated flavonoids, followed by molecular dynamics simulation (100 ns), density functional theory (B3LYP model), and ADMET (admeSar 2 online server) studies. The result obtained from the virtual screening with the glucokinase revealed arcommunol B (−10.1 kcal/mol), kuwanon S (−9.6 kcal/mol), manuifolin H (−9.5 kcal/mol), and kuwanon F (−9.4 kcal/mol) as the top-ranked molecules. Additionally, the molecular dynamics simulation and MM/GBSA calculations showed that the hit molecules were stable at the active site of the glucokinase enzyme. Furthermore, the DFT and ADMET studies revealed the hit molecules as potential glucokinase activators and drug-like candidates. Our findings suggested further evaluation of the top-ranked prenylated flavonoids for their in vitro and in vivo glucokinase activating potentials.     

QM/MM-PBSA method to estimate free energies for reactions in proteins

We have developed a method to estimate free energies of reactions in proteins, called QM/MM-PBSA. It estimates the internal energy of the reactive site by quantum mechanical (QM) calculations, whereas bonded, electrostatic, and van der Waals interactions with the surrounding protein are calculated at the molecular mechanics (MM) level. The electrostatic part of the solvation energy of the reactant and the product is estimated by solving the Poisson-Boltzmann (PB) equation, and the nonpolar part of the solvation energy is estimated from the change in solvent-accessible surface area (SA). Finally, the change in entropy is estimated from the vibrational frequencies. We test this method for five proton-transfer reactions in the active sites of [Ni,Fe] hydrogenase and copper nitrite reductase. We show that QM/MM-PBSA reproduces the results of a strict QM/MM free-energy perturbation method with a mean absolute deviation (MAD) of 8-10 kJ/mol if snapshots from molecular dynamics simulations are used and 4-14 kJ/mol if a single QM/MM structure is used. This is appreciably better than the original QM/MM results or if the QM energies are supplemented with a point-charge model, a self-consistent reaction field, or a PB model of the protein and the solvent, which give MADs of 22-36 kJ/mol for the same test set.