Ab Initio Quantum Mechanics-Based Free Energy Perturbation Method for Calculating Relative Solvation Free Energies

M. RAMI REDDY,1 U. C. SINGH,2 MARK D. ERION1

1

Metabasis Therapeutics, 11119 North Torrey Pines Road, La Jolla, California 92037 2 Azaya Therapeutics, 130 West Rhapsody Drive, San Antonio, Texas 78216

Received 23 January 2006; Revised 16 March 2006; Accepted 15 April 2006 DOI 10.1002/jcc.20510 Published online 21 December 2006 in Wiley InterScience (www.interscience.wiley.com).

Abstract: A free energy perturbation (FEP) method was developed that uses ab initio quantum mechanics (QM) for
treating the solute molecules and molecular mechanics (MM) for treating the surroundings. Like our earlier results using AM1 semi empirical QMs, the ab initio QM/MM-based FEP method was shown to accurately calculate relative solvation free energies for a diverse set of small molecules that differ signicantly in structure, aromaticity, hydrogen bonding potential, and electron density. Accuracy was similar to or better than conventional FEP methods. The QM/ MM-based methods eliminate the need for time-consuming development of MM force eld parameters, which are frequently required for drug-like molecules containing structural motifs not adequately described by MM. Future automation of the method and parallelization of the code for Linux 128/256/512 clusters is expected to enhance the speed and increase its use for drug design and lead optimization. q 2006 Wiley Periodicals, Inc. J Comput Chem 28: 491494, 2007 Key words: free energy perturbation method; quantum mechanics; molecular mechanics; molecular dynamics; relative solvation free energies

Introduction
Free energy perturbation (FEP) methods1 are well recognized as the most accurate and, accordingly, most preferred computational strategy for estimating free energies associated with covalent2 and noncovalent3 interactions between small molecules and solvent, and between small molecules and macromolecules.4 The latter calculations have been used to estimate relative solvation and binding free energies between two drug candidates and therefore aid in drug design.5,6 Despite numerous successes,7 FEP-based methods are rarely used by computational chemists in the pharmaceutical industry. One reason likely relates to the complexity of the calculations coupled with the lack of user-friendly software. Another reason relates to the dependence of the calculations on accurate molecular mechanics force eld parameters, which are available for amino acid residues, nucleotides, and some small molecules8 but not for the majority of structurally unique molecules that represent potential drug candidates. Furthermore, the process used to derive and input the new force eld parameters is time-consuming and often hampered by the lack of relevant experimental data. One strategy that eliminates the need for force eld parameter development and may enable automation of FEP calculations is

to use quantum mechanics (QM) to describe the ligand structure, and molecular mechanics (MM) to describe the surrounding environment (e.g. protein, water). The coupling of QM and MM has been successfully used to determine activation free energies9 and free energy proles of enzymatic reactions.10 QM/MM methods have also been used to calculate polarization and charge transfer effects in aqueous simulations,11 and to study the mechanism and kinetic isotope effects of the nucleophilic substitution reaction in haloalkane dehalogenase.12 Recently, quantum mechanical/molecular mechanical (QM/MM) potentials combined with free energy perturbation techniques were used to calculate pKa and redox potentials in macromolecular systems.13,14 Previously, we described the development and validation of a QM/MM-based free energy methodology, which used AM1 for forces and energies and ab initio HF/6-31G*/ESP for partial atomic charges.15 Herein, we describe the integration of ab initio QM with the FEP calculations and the relative solvation free energies calculated for a series of molecular pairs using this methodology.

Correspondence to: M. R. Reddy; e-mail: reddy@mbasis.com

q 2006 Wiley Periodicals, Inc.

492

Reddy, Singh, and Erion Vol. 28, No. 2 Journal of Computational Chemistry

perturbing system and the surroundings (C) at any value of  is given by El lEAC 1 lEBC (2)

The partial charge on atom i at any value of  is given by eq. 3. qil lqA 1 lqB i i
Figure 1. Molecular threading of phenylalanine (molecule A) and isoleucine (molecule B). Common atoms are the atoms of the amino acid backbone. The noncommon atoms are the side chain atoms, which for l 0 are real atoms for molecule A and dummy atoms for molecule B (dashed structure). At the completion of the transformation (l 1), the side chain atoms for molecules A are dummy atoms and for molecule B are real atoms. Hydrogen atoms are removed for clarity.

(3)

The radius and well depth of atom i at a given value of  is given by ril lriA 1 lriB "il l"A 1 l"B i i (4) (5)

Theory
As in conventional FEP methods,1 relative solvation free energies were calculated by transforming solute A into solute B using the l coupling method. At any l, the system is described as Hl lHA 1 lHB (1)

Ab initio QM forces and energies are calculated for each molecule, A and B, separately and then scaled based on  using eqs. (6) and (7),
i i i fl lfA 1 lfB

(6) (7)

EQM lEQM 1 lEQM B A l

Since A and B are often structurally dissimilar molecules, we used the thread method16 for all the transformations instead of more commonly used atom-to-atom mapping procedures. Accordingly, as exemplied in Figure 1, each transformation required threading together two molecules, referred to herein as molecule A (e.g. phenylalanine) and molecule B (e.g. isoleucine), such that the common atoms (amino acid backbone) were represented by a single topology (see Fig. 1). The noncommon atoms of each molecule, i.e. the portion of the molecules that must be transformed, were represented by two separate topologies each dened using their associated geometries and each capable of interacting only with the environment and not each other. At the start of the simulation, the noncommon atoms of molecule B exist entirely as dummy atoms, which are atoms identical to real atoms except that their Lennard-Jones parameters and atomic charges are set to zero. At intermediate points during the transformation, all atoms in both topologies have fractional Lennard-Jones parameters and charges. At the end of the simulation, the topology that started as dummy atoms (molecule B) is entirely real atoms, whereas the topology that started as real atoms (molecule A) is entirely dummy atoms. The thread method in conventional FEP calculations entails scaling the MM parameters according to l and calculating the corresponding MM energies and forces. In contrast, the QM/MMbased FEP method uses ab initio quantum mechanical methods to calculate the energies and forces for the solute in the system and MM to describe the solvent. To calculate the QM energy and forces, we implemented a procedure17 that separated the threaded molecule into two molecules (A and B) at each dynamic step. The ab initio quantum mechanical energies and forces were then computed and the combined energies and forces were recomputed using the  coupling method. The interaction energy between the

Accordingly, the energy and force will correspond to molecule B when  0 and to molecule A when  1. The total energy for the system is determined using eq. 8 wherein, the term EQM/MM represents the interaction energy involving an atom i in the MM part of the system and an atom j in the QM part of the system. The free energy change (eq. 9) is decomposed into the free energy contribution from the subsystem treated by ab initio QM and the free energy contribution from the surroundings, i.e. the subsystem not treated by QM (non-QM or NQM). Etot EQM EMM
M L XX i1 J1

Eij QM=MM

(8)

Gtot GQM GNQM

(9)

Computational Details
Relative differences in solvation free energies for the transformations shown in Table1 were calculated using the newly implemented ab initio QM/MM based-FEP method with the thermodynamic cycle depicted in Figure 2 and the corresponding eq. 10. The ab initio part of the code originated from the QUEST19 program developed at UCSF. Gsol Gaq Ggas (10)

The FEP calculations in the solvent phase were performed starting with one solute, which was mapped onto the other solute and solvated in a rectangular box of SPC/E water.20,21 All water mol ecules located greater than 13 A from the nearest solute atom or DOI 10.1002/jcc

Journal of Computational Chemistry

Ab Initio QM-Based FEP Method for Calculating Relative Solvation Free Energies

493

Table 1. Relative Solvation Free Energies (kJ/mol).

Transformation CH3OH?EtH AcCH3?AcNH2 C6H6?C6H5OH CH3CCl3?EtH C6H6?C5H5N Ser?Cys Phe?Ile Cyt?Thy Ade?Gua

DDG(ab)a 29.5 6 1.8 25.9 6 2.1 22.6 6 1.7 9.0 6 2.1 13.7 6 1.6 17.7 6 1.7 11.2 6 3.1 24.8 6 2.1 37.3 6 2.5

DDG(AM)b 29.7 6 1.8 27.9 6 2.3 22.2 6 1.7 9.1 6 2.1 13.2 6 1.5 18.8 6 1.8 10.7 6 3.2 25.6 6 2.3 38.1 6 2.6

DDG(FP)c 31.5 6 1.9 28.8 6 2.2 21.6 6 1.7 9.3 6 2.2 12.6 6 1.5 18.5 6 1.8 10.5 6 3.1 24.1 6 2.2 36.3 6 2.6

DDG(E)d 29.0 24.8 23.5 8.4 14.7 16.0 12.1

turned on. The transformation occurred over a total of 51 windows with each window consisting of 1.5 ps of equilibration and 3 ps of data collection. Thus, each mutation required 479 ps of MD. Errors were estimated for each window by dividing the window statistics into four groups and computing the standard deviation.5,6 The reported standard deviation is the root mean square of the window errors.

Results and Discussion

To validate the ab initio QM/MM-based free energy perturbation method, the relative solvation free energy differences for nine molecular pairs were calculated using eq. 10 (Table 1). These test cases were chosen since they constitute relatively large structural changes (e.g. phenylalanine to isoleucine), as well as changes in hydrogen bonding potential (e.g. acetone to acetamide, cytosine to thymine and adenine to guanine), aromaticity (e.g. benzene to pyridine) and electron density (1,1,1-trichloroethane to ethane). The relative solvation free energies, DDG(ab), were calculated using the HF/3-21G* basis set for calculating energies and forces of solute molecules at every step during the molecular dynamics (MD) simulations. Solute partial atomic charges were calculated using a HF/6-31G* basis set with ESP tting at the beginning and maintained throughout the mutation. Results were compared with the relative solvation free energies calculated using AM1 instead of ab initio QM for the forces and energies (DDG(AM)). In addition, results were compared to free energies calculated using the conventional FEP method with HF/6-31G*/ESP partial atomic charges for solute molecules (DDG(FP)). As shown in Table 1, the relative solvation free energies calculated using the current method (DDG(ab)) were as good as or better than the conventional method (DDG(FP)) and accurately reected the experimental results (DDG(E)).18 The CPU time required to complete the methanol to ethane and phenylalanine to isoleucine mutations (479 ps of MD for each system) was 169 and 845 h, respectively, using a P650 IBM single processor computer. The solute/inhibitor molecules with more atoms would require even more CPU time with practical limits with current computer hardware

Ser, serine; Cys, cysteine; Phe, phenylalanine; Ile, isoleucine; Cyt, cytosine; Thy, thymine; Ade, adenine; Gua, guanine. a Calculated using HF/3-21G* for forces and energies and HF/6-31G*/ESP for partial atomic charges. b Calculated using AM1 for forces and energies, and ab initio (HF/631G*)/ESP for partial atomic charges. c Calculated using a conventional FEP method and HF/6-31G*/ESP for partial atomic charges. d Values obtained from experimental data reported in the literature.18

less than 2.5 A from any solute atom were removed. The partial atomic charges for all the solute molecules were calculated using a HF/6-31G* basis set with ESP tting. The van der Waals parameters for all the solute atoms were taken from the Galaxy data base.17 The HF/3-21G* basis set was used for calculating energies and forces of the solute molecules at every step during the molecular dynamics simulations. Energy minimizations and subsequent FEP calculations used molecular mechanics (MM) and molecular dynamics (MD) simulations, which were carried out using periodic boundary conditions in all directions. In the MD simulations, Newtons equations of motion for all the atoms were solved using the Verlet algorithm22 with a 1 fs time step. SHAKE was used to constrain all bond lengths.23 Constant temperature (N, P, T ensemble) was maintained by scaling velocities of all atoms in the system. The non bonded interaction energies were calculated using a 12 A residue based cutoff. Initially, each system was minimized using a conjugate gradient method for 2000 cycles with the solutes frozen and only the water molecules allowed to move. Next, the whole system was energy minimized for an additional 2000 cycles followed by 100 cycles of minimization using the SHAKE option to constrain the bond lengths to their equilibrium values. The system was then equilibrated with 20 ps of MD using a time step of 1 fs at a constant temperature (300 K) and pressure (1 atm). The FEP calculations in the gas phase were performed using a similar procedure but without inclusion of the water molecules. The transformation was achieved using the window method1 as implemented in the Galaxy program17 and a two-stage procedure previously shown to enhance convergence.5,16,24 In the rst stage, the atomic charges of molecule A (the starting topology consisting of real atoms) were slowly turned off while the Lennard-Jones parameters of the atoms of molecule B (the starting topology consisting of dummy atoms) were slowly turned on. During the second stage, the Lennard-Jones parameters of molecule A were turned off while the charges of molecule B were

Figure 2. Thermodynamic cycle for computing solvation free energy differences between two solutes A and B. QM(ab), indicates that the solutes were treated using ab initio quantum mechanics.

Journal of Computational Chemistry

DOI 10.1002/jcc

494

Reddy, Singh, and Erion Vol. 28, No. 2 Journal of Computational Chemistry

likely around 30 atoms. While CPU requirements of this magnitude greatly limit throughput, computer technology continues to advance results in faster and cheaper computers, which may mean that, the higher CPU requirements for this method could be addressed in the future by parallelization of the code and using 128/256/512 Linux clusters. In summary, ab initio QM was integrated into the QM/MM based free energy method and implemented into the Galaxy program.17 Use of ab initio quantum mechanics eliminates time-consuming and often difcult development of ligand-specic molecular mechanics force eld parameters and thereby should enable future automation of FEP-based calculations. Automation will increase usage of FEP-based calculations in the pharmaceutical industry by computational chemists, who at the present time rarely use these methods despite knowledge of their superior accuracy. More accurate results are expected to enhance both drug design and lead optimization and thereby shorten the time to discover new drug candidates.

References
1. (a) Zwanzig, R. J. J Chem Phys 1954, 22, 1420; (b) Tembe, B. L.; McCammon, J. A. Comput Chem 1984, 8, 281; (c) Pearlman, D. A. In Free Energy Calculations in Rational Drug Design; Reddy, M. R.; Erion, M. D., Eds.; Kluwer: Dordrecht, 2001; pp. 935. 2. Erion, M. D.; Reddy, M. R. J Am Chem Soc 1998, 120, 3295. 3. (a) Jorgensen, W. L.; Ravimohan, C. J Chem Phys 1985, 83, 3050; (b) Agarwal, A.; Brown, F. B.; Reddy, M. R. In Free Energy Calculations in Rational Drug Design; Reddy, M. R.; Erion, M. D., Eds.; Kluwer: New York, 2001; pp. 97117. 4. (a) Brooks, C. L.; Fleischman, S. F. J Am Chem Soc 1990, 112, 3307; (b) Bash, P.; Singh, U. C.; Brown, F. K.; Langridge, R.; Kollman, P. A. Science 1987, 235, 564; (c) Reddy, M. R.; Erion, M. D.; Agarwal, A. In Reviews in Computational Chemistry; Lipkowitz, K. B.; Boyd, D. B., Eds.; Wiley-VCH: New York, 2000; pp. 217304. 5. (a) Reddy, M. R.; Viswanadhan, V. N.; Weinstein, J. N. Proc Natl Acad Sci USA 1991, 88, 10287; (b) Reddy, M. R.; Bacquet, R. J.; Zichi, D.; Mathews, D. A.; Welsh, K. M.; Jones, T. R.; Freer, S. J Am Chem Soc 1992, 114, 10117.

6. (a) Erion, M. D.; van Poelje, P. D.; Dang, Q.; Kasibhatla, S. R.; Potter, S. C.; Reddy, M. R.; Reddy, K. R.; Jiang, T.; Lipscomb, W. N. Proc Natl Acad Sci USA 2005, 102, 7970; (b) Reddy, M. R.; Erion, M. D. J Am Chem Soc 2001, 123, 6246; (c) Erion, M. D.; van Poelje, P. D.; Reddy, M. R. J Am Chem Soc 2000, 122, 6114. 7. Reddy, M. R.; Erion, M. D., Eds. Free Energy Calculations in Rational Drug Design; Kluwer: New York, 2001. 8. (a) Bowen, J. P.; Allinger, N. L. In Reviews in Computational Chemistry; Lipkowitz, K. B.; Boyd, D. B., Eds.; VCH Publishers: New York, 1991; pp. 8197; (b) Todebush, P. M.; Bowen, J. P. In Free Energy Calculations in Rational Drug Design; Reddy, M. R.; Erion, M. D., Eds.; Kluwer: New York, 2001; pp. 3759. 9. Singh, U. C.; Kollman, P. A. J Comput Chem 1986, 7, 718. 10. Rosta, E.; Klahn, M.; Warshel, A. J Phys Chem B 2006, 110, 2934. 11. Mo, Y.; Gao, J. J Phys Chem B 2006, 110, 2976. 12. Devi-Kesavan, L. S.; Gao, J. J Am Chem Soc 2003, 125, 1532. 13. Li, G.; Cui, Q. J Phys Chem B 2003, 107, 14521. 14. Riccardi, D.; Schaefer, P.; Yang, Y.; Yu, H.; Ghosh, N.; Prat-Resina, X.; Konig, P.; Li, G.; Xu, D.; Guo, H.; Elstner, M.; Cui, Q. J Phys Chem B 2006, 110, 6458. 15. Reddy, M. R.; Singh, U. C.; Erion, M. D. J Am Chem Soc 2004, 126, 6224. 16. Singh, U. C.; Benkovic, S. J. Proc Natl Acad Sci USA 1988, 85, 9518. 17. Galaxy Molecular Modeling Software and AM2000 Macromolecular Simulation Package; AM Technologies: San Antonio, Texas, 1995. 18. (a) Hine, J.; Mookerjee, P. K. J Org Chem 1975, 40, 292; (b) Cabani, S.; Gianni, P.; Mollica, V.; Lepori, L. J Solution Chem 1981, 10, 563. 19. (a) Singh, U. C.; Kollman, P. A. J Comput Chem 1984, 5, 129; (b) Singh, U. C.; Kollman, P. A. QUEST (Version 1.1), University of California, San Francisco, 1986. 20. Berendsen, H. J. C.; Grigera, J. R.; Straatsma, T. P. J Phys Chem 1987, 91, 6269. 21. Reddy, M. R.; Berkowitz, M. Chem Phys Lett 1989, 155, 173. 22. Verlet, L. Phys Rev 1967, 159, 98. 23. Ryckaert, J. P.; Ciccotti, G.; Berendsen, H. J. C. J Comput Phys 1997, 23, 327. 24. (a) Singh, U. C. Proc Natl Acad Sci USA 1988, 85, 4280; (b) Reddy, M. R.; Varney, M. D.; Kalish, V.; Viswanadhan, V. N.; Appelt, K. J Med Chem 1994, 114, 10117.

Journal of Computational Chemistry

DOI 10.1002/jcc