Sie sind auf Seite 1von 10

THE JOURNAL OF CHEMICAL PHYSICS 130, 104108 2009

Bounding the electrostatic free energies associated with linear continuum models of molecular solvation
Jaydeep P. Bardhan,1,2,a Matthew G. Knepley,2,3 and Mihai Anitescu3
1 2

Biosciences Division, Argonne National Laboratory, Argonne, Illinois 60439, USA Department of Molecular Biophysics and Physiology, Rush University, Chicago, Illinois 60612, USA 3 Mathematics and Computer Science Division, Argonne National Laboratory, Argonne, Illinois 60439, USA

Received 20 December 2008; accepted 23 January 2009; published online 11 March 2009 The importance of electrostatic interactions in molecular biology has driven extensive research toward the development of accurate and efcient theoretical and computational models. Linear continuum electrostatic theory has been surprisingly successful, but the computational costs associated with solving the associated partial differential equations PDEs preclude the theorys use in most dynamical simulations. Modern generalized-Born models for electrostatics can reproduce PDE-based calculations to within a few percent and are extremely computationally efcient but do not always faithfully reproduce interactions between chemical groups. Recent work has shown that a boundary-integral-equation formulation of the PDE problem leads naturally to a new approach called boundary-integral-based electrostatics estimation BIBEE to approximate electrostatic interactions. In the present paper, we prove that the BIBEE method can be used to rigorously bound the actual continuum-theory electrostatic free energy. The bounds are validated using a set of more than 600 proteins. Detailed numerical results are presented for structures of the peptide met-enkephalin taken from a molecular-dynamics simulation. These bounds, in combination with our demonstration that the BIBEE methods accurately reproduce pairwise interactions, suggest a new approach toward building a highly accurate yet computationally tractable electrostatic model. 2009 American Institute of Physics. DOI: 10.1063/1.3081148
I. INTRODUCTION

Electrostatic interactions between biological molecules and the surrounding aqueous solvent play key roles in determining the afnity and specicity with which biomolecules bind one another.13 Modeling these solvation interactions remains a theoretical and computational challenge in part due to the long-range character of electrostatic forces and the need to average over many solvent degrees of freedom.4,5 Implicit-solvent models based on linear, continuum electrostatic theory have been proven highly successful at reproducing a range of experimental phenomenon and explicit-solvent simulations.4,6,7 It has been proven difcult, however, to develop rigorous numerical simulation techniques for the associated system of partial differential equations PDEs that are fast enough to be used in dynamical simulations although notable exceptions do exist8. Generalized-Born GB models911 represent one of the most popular classes of approaches for estimating electrostatic interactions for recent reviews, see Refs. 1215. These methods have in common two key features. First, every solute atom is associated with a parameter called an effective Born radius that measures the magnitude of solvent screening seen by the atom. Second, the interaction energy between atoms is calculated using an interpolation formula, most often the Still equation,9 that recovers the Born expression exactly for a charges self-energy and asymptotically approaches a Coulomb interaction for large interatomic disa

Electronic mail: jbardhan@alum.mit.edu.

tances. Modern GB methods can reproduce more expensive PDE simulations to within 1% or 2% accuracy1623 and are fast enough to be employed in dynamical simulations.19,24,25 Despite these successes, however, numerous studies suggest that GB methods sometimes fail to reproduce relevant features of biomolecular energy landscapes or fail to generate electrostatic forces in agreement with forces from continuum-model calculations.16,2629 The Coulomb-eld approximation CFA plays an important role in many GB models. In the CFA, one assumes that an atoms self-energy contribution to the electrostatic solvation free energy arises from the bare Coulomb eld induced by the atom in the solute volume.10 This assumption gives rise to a volume integral, which Ghosh et al. converted to a surface integral using the divergence theorem. They noted the close relationship between the resulting integral and the apparent-surface-charge ASC integral 3037 The recently introduced boundary-integralequation. based electrostatics estimation BIBEE model for electrostatics takes this relationship as a starting point and exploits it fully to eliminate the need for the nonphysical interpolation inherent to GB models.38 Removing these interpolations allows BIBEE to capture solvent-screened pairwise interactions much more accurately. In earlier work this improved delity was demonstrated by comparing the electrostatic reaction-potential matrices generated by BIBEE, GB models, and high-accuracy boundary-element method BEM simulations of the ASC integral equation. The reaction-potential matrix, which maps the explicitly described charge distribu 2009 American Institute of Physics

0021-9606/2009/13010/104108/10/$25.00

130, 104108-1

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

104108-2

Bardhan, Knepley, and Anitescu

J. Chem. Phys. 130, 104108 2009

tions to the induced reaction potentials, offers a more detailed view of methods failings and strengths than do the overall electrostatic solvation free energies. We demonstrate in the present paper that BIBEE variants can provide mathematically rigorous upper and lower bounds LBs to the true continuum-model free energy that is, the free energy that would be calculated if the mixed-dielectric problem were solved exactly. The BIBEE/CFA method provides a rigorous upper bound to the actual free energy of the continuum model, and a new diagonal approximation BIBEE/LB introduced in this work generates a rigorous LB. In addition, we show empirically that the BIBEE by preconditioning BIBEE/P method described in earlier work38 generates an effective but not yet proven LB. In Sec. II we specify the precise mathematical model of interest and the ASC boundary-integral-equation BIE formulation of the problem. We also describe the BIBEE approach to nding approximate solutions to the BIE. Section III contains proofs that BIBEE approximations can provide upper and LBs for the electrostatic free energy of a given system. Computational results in Sec. IV illustrate the bounds effectiveness. The paper concludes with a brief discussion.

1 r d 2r n r 4 r r
= lim
0

1 r d 2r . r,rr n r 4 r r

Denoting the integral operator in Eq. 1 as D and the identity operator by I, the integral equation can be written as D = qk I +
k=1 nc

1 . n r 4 r r k

Given the surface charge distribution r, the reaction potential induced at a point r in the solute by r is

R r =

r d 2r . 4 1 r r

In the linear response model, the electrostatic component of the solvation free energy can be written as
1 T R q , Ges = 2 1 T =2 q Lq ,

6 7

II. THEORY A. Linear continuum electrostatic model and boundary-integral-equation formulation

We consider a single solute molecule in an innite solvent region. The solute is treated a homogeneous dielectric medium with low dielectric constant 1, with nc discrete point charges located inside the solute-solvent boundary ; the ith charge has value qi and is located at ri. Inside , the electrostatic potential obeys a Poisson equation. The solvent region is modeled as a homogeneous dielectric with dielectric constant 2 1, and in this region the Laplace equation governs the potential. At the boundary, both the potential and the normal displacement eld are continuous, and regularity conditions at innity complete the needed boundary conditions.39,40 This coupled PDE problem can be solved using the BIE r +

where q is the nc-length vector of charge values, R denotes the vector of reaction potentials at the charge locations, and the reaction-potential matrix L maps the charge values to the reaction potentials at the charge locations, where the matrix element Li, j represents the electrostatic potential induced by solvent polarization at ri by a unit charge at r j. Accordingly, the ith column of the reaction-potential matrix represents the vector of nc reaction potentials induced at the charge locations by a unit charge at ri. Using the BIE formalism and the particular set of nc charge locations, L is the product of three operators L = CA1B , 8

1 r d 2r n r 4 r r 1 , n r 4 r r k
1

qk =
k=1

nc

where the unknown surface charge distribution r captures the effects of solvent polarization =

1 2 1 2 1 + 2

where B maps the nc charges to the normal displacement eld at the boundary that is, the right-hand side in Eq. 4 can be written as Bq, A is the operator acting on on the left-hand side of Eq. 4, so A1 determines the induced charge at the boundary, and C represents the integral operator in Eq. 5, which maps the induced charge distribution r to R. The analysis of the present paper focuses on approximations to A. Two assumptions are made about the boundary . First, it is assumed to be Liapunov smooth;45 that is, at every point r there exists a tangent plane, although not necessarily a curvature.46 The Liapunov-smoothness condition is sufcient to ensure that the operator D is compact.45 Also, the boundary is assumed to be connected i.e., there exists only one solute, and there are no water-lled cavities inside the solute to be modeled as regions of high dielectric47,48.
B. Diagonal approximations to the boundary-integral equation

and nr represents the normal direction at r, which points outward from the low-dielectric solute into the solvent.30,33,41,42 Note that 1 2 0. The notation W denotes the Cauchy principal value integral43,44

The BIBEE method38 approximates electrostatic solvation free energies using diagonal approximations to the operator A; different diagonal approximations lead to different

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

104108-3

Bounding electrostatic free energies

J. Chem. Phys. 130, 104108 2009

estimated free energies. The BIBEE/CFA method approximates r as the direct Coulomb eld due to the charges

GD = DG ,

13

1 CFA = Bq . 2

As described in earlier work, the BIBEE/CFA method is closely related to the CFA employed by GB models for electrostatics.38 In particular, the CFA assumes that the eld induced by r at r is zero. This condition holds when the solute charge distribution produces a uniform normal eld at the boundary because the set of uniform elds is in the or1 I + D.38,46,49 thogonal range of the operator 2 The BIBEE/P approximation calculates r as

as can be shown using the Calderon projector.44 Equation 13 and the properties of G demonstrate that D is quasi-Hermitian,5658 and therefore we may dene a Hermitian operator H that is similar to D, H = G1/2DG1/2 . 14

P = Bq .

10

This approximation takes into account not only the eld induced by the solute charge distribution but also the strong normal eld induced at r by an innitesimally small disk of r immediately surrounding r.38 The word preconditioning refers to an approach for accelerating the convergence of Krylov-subspace iterative methods such as GMRES Ref. 50 for solving BIE numerically using the BEM.33,37,40,48,5155 A matrix P is said to be a preconditioner for the BEM system Ax = b if the preconditioned system PAx = Pb can be solved using fewer matrix-vector products more precisely, such a matrix P is a left preconditioner. Finally, we introduce in this paper and dene the BIBEE/LB approximation

The proof proceeds in two stages. In the rst, the operator CT which maps the charge distribution to the direct Coulomb potential at the surface is rewritten using basic tools of potential theory and BIE.46,59 We wish to write this potential not as CTq but in terms of Bq to make the energy expression more symmetric. In the second stage, the resulting expression for the electrostatic solvation free energy is then simplied by exploiting the quasi-Hermiticity of the integral operators. The bounds may then be derived easily. To obtain an equivalent expression for CTq, we study a system with the same charge distribution q situated inside the same boundary , where the dielectric constant is uniform throughout space that is, 1 = 2. Under the assumptions for the connectedness of and the support of the charge distribution, the potential directr = CTq is harmonic in an open set that includes and extends out to innity. From the denition of B, the normal derivative of the potential at 1Bq. The potential at the boundary can can be written as then be determined by solving an exterior Neumann problem,46 the BIE for which is 1Bq , I 2 D = 2 15

1 + LB = Bq . 2

11

where r is the unknown on . The single-layer potential at is then CTq = directr = G , 16 17

III. BOUNDS ON THE ELECTROSTATIC FREE ENERGY

1GI 2D1Bq . =2

We now prove that the BIBEE/CFA and BIBEE/LB approximations bound the electrostatic free energy. In order to analyze the properties of the estimated electrostatic solvation free energy as we introduce different approximations to the operator A, it is valuable to represent A using a decomposition compatible with the approximations. For the space of allowable charge distributions, we use the space of continuous functions whose support is strictly contained within . We dene the single-layer potential operator G such that Gr =

Concluding the rst stage of the proof, the energy expression may be written using Eqs. 17, 1, and 8 as
1 T 1 2 q CA Bq

1qTBTI 2DTGI + D1Bq =

18

after making use of the symmetry of G. For the second stage of the proof, we rewrite the integral operators as I 2D = G1/2I 2HG1/2 , D = G1/2I + H G 1/2 . I + 19 20

r d 2r , 4 r r

12

and it is known that G is self-adjoint and positive denite.56,57 In turn, this implies that it has a square root; that is, there exists a self-adjoint, positive denite operator G1/2, such that G 1/2G 1/2 = G . For almost boundaries, single-layer operator D, all suitably smooth e.g., Liapunov smooth the operators G and D are compact.46 The potential operator G, the double-layer potential and its transpose D are related by

Substituting Eqs. 19 and 20 into Eq. 18 and using the identity G = G1/2G1/2 then give 1qTBTG1/2I 2H1G1/2G1/2G1/2G1/2 H1G1/2Bq , I + so that simplifying we have
1 T 1 2 q CA Bq

21

1qTBTG1/2I 2H1I + H1G1/2Bq . = 22

The two inverse terms commute, and writing

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

104108-4

Bardhan, Knepley, and Anitescu

J. Chem. Phys. 130, 104108 2009

T H = V H DV H ,

23

where D is the eigenvalue matrix of the operator D, we nally have


1 T 1 2 q CA Bq

1qTBTG1/2VHI 2D1 =
T 1/2 D1VH I + G Bq .

24

Finally, we note that our results are proved for the operators acting over the innite dimensional function spaces and not for their approximations obtained by discretization. Nonetheless, consistency of the approximation indicates that the bounds we obtained will be satised for sufciently accurate discretizations.
IV. RESULTS

Because 1 / 2 , + 1 / 2,

1 2 and 43,46

the eigenvalues of D lie in the interval the product


1

i 1 2 i 1 +
1

25

is positive whenever 1 is nite i.e., when the integral equation is well posed49,60. As expected on physical grounds, the symmetric quadratic expression in Eq. 24 is negative when . 1 2 and positive when 2 1 due to the the sign of To obtain the BIBEE bounds, dene the vector
T 1/2 G Bq x = VH

26

Computational results in this section validate the bounds presented in Sec. III. Furthermore, examples demonstrate that the BIBEE/P method provides an effective LB for the free energies when simulating typical molecular surfaces and charge distributions, even when one does not know a priori whether D has positive eigenvalues. Planar-element discretizations of surfaces were generated using previously described techniques.48,63 For all simulations, the program MSMS Ref. 64 was used to generate solvent-excluded surfaces with a probe radius of 1.4 .
A. Met-enkephalin

with components xi, i = 1 , 2 , . . .. This enables Eq. 24 to be written as a sum over the modes 1 T 1 1 1 2 11 2D D q CA Bq = i 1 + i xi . 2 i 27

The BIBEE approximations to A generate the approximate free energies


1 T 1 2 q CCFABq

1 1 1 2D = 1 i i

x2 i,

28

1 T 1 1 1 2 1 1 2D q C P Bq = i 1 xi , 2 i 1 T 1 1 1 1 2D 1+ q CLBBq = i 2 2 i

29
1

x2 i.

30

0, every term in Eq. 27 is negative and Because

D 1 + i i. 2

31

Therefore, applying Eq. 31 to the BIBEE/CFA approximation of Eq. 28 gives an approximate free energy that is an upper bound to the true free energy in the continuum model. Similarly, because


1+

D 1 + i i, 2

32

we can apply Eq. 32 to the BIBEE/LB approximation in Eq. 30 to obtain that BIBEE/LB gives a LB for the true free energy. For surfaces such as spheres and prolate spheroids, which possess operators D with only nonpositive eigenvalues,61,62 it is straightforward to show using Eq. 29 that the BIBEE/P approximation provides a rigorous LB. We did not prove that BIBEE/P bounds the free energy in the general case. However, as we demonstrate in Sec. IV, it seems to offer an effective bound much closer to the actual free energy than the BIBEE/LB.

The peptide met-enkephalin AcYGGFMNH2 was constructed in an extended conformation using CHARMM Ref. 65 and the CHARMM22 force eld66 with the CMAP correction.67,68 The VMD software package69 was used to solvate the peptide such that no peptide atom was within 8 of the edge of the solvent box, and sodium and chloride atoms were added to achieve a salt concentration of 0.145 M. A molecular dynamics MD simulation of 500 ps in NAMD Ref. 70 was then performed following 1000 steps of energy minimization and 100 ps of equilibration at 300 K. Dynamics were performed in the NPT ensemble at 1 atm pressure using Langevin dynamics, particle-mesh Ewald electrostatics, and 2 fs time steps with snapshots of the trajectory saved every 2500 time steps. Electrostatic solvation free energies of the 50 peptide structures were then estimated using Generalized Born with Molecular Volume GBMV,19 BEM, the BIBEE methods, and surface-GB SGB/CFA the SGB method of Ghosh et al. but lacking the empirical correction terms presented in their work71. For these simulations, 1 = 1, 2 = 80, and atomic radii and charges were taken from the 66 CHARMM22 parameter set. The surface discretizations were computed in MSMS such that the density of vertices was 5.0 vertices / 2; doubling the vertex density to 10 vertices / 2 did not materially affect the results. Figures 1a and 1b are time-series plots of the computed electrostatic free energies and Fig. 2 is a scatter plot of the BIBEE/LB, BIBEE/CFA, BIBEE/P, SGB/CFA, and GBMV energies against the BEM simulations. Owing to the difference in scale between the BIBEE/LB estimates and the estimates from the other methods, the BIBEE/LB results are shown on a separate plot. Comparing Figs. 1a and 1b, it is immediately evident that the BIBEE/LB method provides a very loose LB to the true free energy. This bound becomes worse as the ratio 2 / 1 increases and approaches the BIBEE/P bound as the dielectric ratio approaches one data not shown. One of the peptide structures was analyzed in more detail with the reaction-potential matrices calculated using

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

104108-5
0 1000 Electrostatic Free Energy (kcal/mol) 2000 3000 4000 5000 6000 7000 100

Bounding electrostatic free energies


20 40 Electrostatic Free Energy (kcal/mol) 60 80 100 120 140 BEM BIBEE/LB 200 300 Time (ps) 400 500 600 160 100 BEM SGB/CFA GBMV BIBEE/CFA BIBEE/P 200 300

J. Chem. Phys. 130, 104108 2009

Time (ps)

400

500

600

(a)

(b)

FIG. 1. Color online Comparison of electrostatic solvation free energies using met-enkephalin structures taken from a 500-ps MD simulation plotted as time series; snapshots have been taken at 10-ps intervals. Energies are in kcal/mol. a All estimates are plotted. b BIBEE/LB has been omitted for clarity.

BEM, GBMV, SGB/CFA, BIBEE/CFA, and BIBEE/P. Figure 3a is a plot of the eigenvalues of the calculated matrices. The SGB/CFA eigenvalues are slightly more accurate than the BIBEE/CFA eigenvalues for the dominant eigenmodes but less accurate for the smaller eigenmodes. Of the four electrostatic approximations, the GBMV method appears to provide the most accurate eigenvalue estimates. As noted previously, BIBEE/CFA is most accurate for the largestmagnitude eigenvalues and the BIBEE/P method offers the best delity to BEM for the smallest eigenvalues.38 It is important that methods for estimating electrostatic interactions calculate not only an accurate total free energy but also preserve the energetics of interaction between chemical groups. To analyze how different methods preserve pairwise interactions with respect to the BEM calculations, we project the eigenvectors of the approximate reactionpotential matrices onto the eigenvectors of the reactionpotential matrix from BEM. For example, the i , j entry of the matrix
0 Estimated Electrostatic Free Energy (kcal/mol) 10 20 30 y=x SGB/CFA GBMV BIBEE/CFA BIBEE/P

T VSGB/CFA VBEM

33

represents the projection of the jth eigenvector of the SGB/ CFA reaction-potential matrix onto the jth eigenvector of the BEM reaction-potential matrix. In this projection framework, perfect preservation of the pairwise interactions would give rise to a diagonal matrix with diagonal entries of unit magnitude. Conversely, if a method imperfectly reproduces pairwise interactions, the off-diagonal entries are nonzero. The degree to which the approximate-method eigenvectors align with the actual eigenvectors can then be assessed visually using a heat map; Figs. 46 are plots of the projections of the SGB/CFA, GBMV, and BIBEE/P eigenvectors onto the eigenvectors of the BEM reaction-potential matrix.38 We believe that discrepancies between the BIBEE and BEM eigenvectors may be a result of discretizing the integral equation for simulation using BEM.72 However, the various BIBEE methods give rise to essentially identical eigenvectors data not shown, which is expected given that the diagonal ap10
4

10

BEM SGB/CFA GBMV BIBEE/CFA BIBEE/P

Eigenvalue Magnitude
85 80 75 70 65 Electrostatic Free Energy from BEM (kcal/mol) 60

40 50 60 70 80 90 100 90

10

10

10

10

10

10

20

30 40 50 Eigenvalue Index

60

70

80

FIG. 2. Color online Comparison of estimated electrostatic solvation free energies using met-enkephalin structures taken from a 500-ps MD simulation plotted as a scatter plot against energies calculated using BEM. Energies are in kcal/mol.

FIG. 3. Color Eigenvalues of the reaction-potential matrices computed from the nal met-enkephalin structure using BEM, BIBEE/CFA, BIBEE/P, GBMV, and SGB/CFA methods.

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

104108-6
80 70 60 BEM Eigenvalue Index 50 40 30 20 10

Bardhan, Knepley, and Anitescu


80 70 60 50 40 30 20 10
10 20 30 40 50 60 SGB/CFA Eigenvalue Index 70 80

J. Chem. Phys. 130, 104108 2009

10

20

30

40

50

60

70

80

FIG. 4. Color Comparison of the eigenvectors of the SGB/CFA reactionpotential matrix with the eigenvectors from BEM computed from the nal met-enkephalin structure. The color of cell i, j represents the projection of the jth eigenvector of the SGB/CFA reaction-potential matrix onto the ith eigenvector of the BEM reaction-potential matrix.

FIG. 6. Color Comparison of the eigenvectors of the BIBEE/P reactionpotential matrix with the eigenvectors from BEM computed from the nal met-enkephalin structure.

proximations to the integral operator differ only by a multiple of the identity matrix. Compared to the eigenvectors from SGB/CFA, the GBMV eigenvectors are clearly better aligned with the BEM eigenvectors, especially for modes 35-50. However, both GB methods exhibit poor alignment for smaller-magnitude eigenvalues. The BIBEE methods, although not always as accurate for estimating total energies, show very good alignment throughout the spectrum. It is interesting that all three methods suffer from some loss of accuracy for modes 20-35, and we are investigating possible reasons for these systematic deviations. The discrepancies between the GBMV and BEM eigenvectors stand in contrast to the good agreement between the GBMV and BEM eigenvalues Fig. 3 and the high accuracy of modern GB methods in reproducing PoissonBoltzmann PB calculations.15 To reconcile these results, it is necessary
80 70 60 BEM Eigenvalue Index 50 40 30 20 10

to understand the relationship between the eigenvectors of the reaction-potential matrix and the charge vector q. Figure 7 is a superposition of two plots, the projection of q onto the BEM eigenvectors, and the cumulative electrostatic free energy obtained as one sums the individual mode contributions beginning with the largest-magnitude eigenvalues. It appears that GB methods based on the CFA achieve good accuracy by nding good approximations to the dominant eigenvalues and eigenvectors. These modes contribute more to the total electrostatic free energy of solvation than the modes associated with smaller eigenvalues, which are not approximated as well. Not all of the modes important to the solvation free energy are captured, however. The SGB/CFA reactionpotential matrix captures the rst 30 eigenmodes slightly more accurately than the BIBEE approximations, and the remaining modes are modeled signicantly more accurately using BIBEE. Approximately one third of the total electro2
0

50

0 0

10

20

30

40 50 60 Eigenvalue Index

70

80

90

100

10

20

30 40 50 GBMV Eigenvalue Index

60

70

80

FIG. 5. Color Comparison of the eigenvectors of the GBMV reactionpotential matrix with the eigenvectors from BEM computed from the nal met-enkephalin structure.

FIG. 7. Color online The contribution of different eigenmodes to the total electrostatic solvation free energy computed from the nal met-enkephalin structure. Left axis dashed-dotted line: magnitude of the projection of the solute charge vector q onto the corresponding eigenmode. Right axis solid line: the cumulative electrostatic solvation free energy beginning with the contributions from the dominant eigenmodes. Energies are in kcal/mol.

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

Cumulative Electrostatic Free Energy (kcal/mol)

Magnitude of Projection Onto Vpb

104108-7
0 Electrostatic Free Energy from BIBEE (kcal/mol) 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5000

Bounding electrostatic free energies

J. Chem. Phys. 130, 104108 2009

y=x BIBEE/CFA BIBEE/P

the calculated BEM free energies. In these simulations, atomic radii and charges were taken from the PARSE parameter set,73 and the dielectric constants were 1 = 4 and 2 = 80. For every BEM calculation, there exists one blue point above the y = x line, from the BIBEE/CFA method, and a corresponding red point from the BIBEE/P approximation directly beneath. The results illustrate that the BIBEE/CFA bound holds as proven and that the BIBEE/P method seems to provide a LB in practice, even though its bounding properties are unproven.
C. Interpolating between the BIBEE/CFA and BIBEE/P estimates

4000 3000 2000 1000 Electrostatic Free Energy from BEM (kcal/mol)

FIG. 8. Color online Electrostatic solvation free energies calculated using BEM and the free energy bounds estimated using BIBEE/CFA and BIBEE/P using protein geometries employed by Feig and Brooks Ref. 15. Energies are in kcal/mol.

static solvation free energy arises from these modes that correspond to smaller eigenvalues. The GBMV method captures the rst 50 modes reasonably well, and this is reected in the improved accuracy of GBMV with respect to the BEM calculations.
B. Protein test suite

Feig and Brooks,15 in their analysis of modern GB electrostatic models and the models accuracy relative to solutions of the PB equation, assembled and prepared a test set of over 600 protein geometries and have made the data available for download. To demonstrate the performance of the BIBEE/CFA and BIBEE/P bounds over a broad range of molecular sizes and shapes, we have estimated the electrostatic free energy of solvation for the proteins in this data set. Figure 8 is a scatter plot of the free energy bounds against
20

The existence of both upper and LBs for the electrostatic free energy immediately suggests the possibility that one might nd an inexpensive interpolation between the bounds to provide a close estimate for the actual free energy, with an ideal scheme possessing strong mathematical and physical foundations. Unfortunately, we have not yet been able to devise such a theoretically satisfactory solution. In this section, we present instead the application of one of the simplest possible schemes, in which weights CFA and P are found empirically such that the weighted sum CFAECFA + PE p is approximately equal to the BEM-calculated free energy EBEM. Because ECFA and E P bound EBEM it is clear that 0 CFA, P 1. For calculations in which the free energies of different conformations of the same molecule are of interest, weights could be determined from an initial structure or from the mean of a set of representative structures. The ensemble of met-enkephalin snapshots provides an illustration. Figures 9a and 9b plot the interpolated BIBEE free energy calculated using weights determined by forcing the interpolated free energy of the rst snapshot to match the free energy from BEM simulation of the rst snapshot. The correlation between the BEM and interpolated BIBEE free energies is remarkable. We wish to emphasize, however, the purely em0 Estimated Electrostatic Free Energy (kcal/mol) 10 20 30 40 50 60 70 80 90 100 90 85 80 75 70 65 Electrostatic Free Energy from BEM (kcal/mol) 60 y=x SGB/CFA GBMV BIBEE/CFA BIBEE/P Interpolated BIBEE

40 Electrostatic Free Energy (kcal/mol)

60

80

100

120

140

BEM SGB/CFA GBMV BIBEE/CFA BIBEE/P Interpolated BIBEE 200 300 Time (ps) 400 500 600

160 100

(a)

(b)

FIG. 9. Color online Interpolation between BIBEE/CFA and BIBEE/P to estimate electrostatic solvation free energies of met-enkephalin structures taken from a 500-ps molecular-dynamics simulation; snapshots have been taken at 10-ps intervals. The weighting has been selected so that interpolation applied to the the rst snapshot gives the BEM-calculated electrostatic free energy. Energies are in kcal/mol. a Energies plotted as time series. b Energies plotted as a scatter plot against the BEM calculations.

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

104108-8
0 Electrostatic Free Energy from BIBEE (kcal/mol) 100 200 300 400 500 600 700 800 900 1000 1000

Bardhan, Knepley, and Anitescu


0.8 y=x BIBEE/CFA BIBEE/P Interpolated BIBEE (Met weight) Interpolated BIBEE (mean Feig weight) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 3000

J. Chem. Phys. 130, 104108 2009

800 600 400 200 Electrostatic Free Energy from BEM (kcal/mol)

(EBEM ECFA) / (EP ECFA)

2500 2000 1500 1000 500 Electrostatic Free Energy from BEM (kcal/mol)

(a)

(b)

FIG. 10. Color online Interpolation between BIBEE/CFA and BIBEE/P to estimate the electrostatic solvation free energies of the protein geometries of Feig and Brooks Ref. 15. All energies are in kcal/mol. a Free energies interpolated using either the exact weighting for the rst met-enkephalin snapshot such that the interpolated free energy matches the BEM calculation exactly or the mean of the exact weights over all the proteins in the test set of Feig. b The exact weights for the BIBEE/P approximation plotted against the electrostatic free energies calculated using BEM.

pirical nature of the interpolation. The fact that the metenkephalin weights used to generate the results in Fig. 9 are not necessarily suitable for all proteins becomes clear upon using the met-enkephalin weights to interpolate the BIBEE/ CFA and BIBEE/P free energies for the proteins of the test set of Feig and Brooks.15 In Fig. 10a are plotted BIBEEestimated free energies that have been interpolated using two different sets of weights. The rst set of weights consists of the exact met-enkephalin weights. It is clear that interpolated free energies calculated using these weights deviate systematically from the BEM-calculated free energies. The second set of weights used to calculate interpolatedBIBEE energies in Fig. 10a is determined by calculating for each protein separately the weights that would cause the interpolated BIBEE free energy to match the BEM calculation. The mean of these weights can then be used generally for proteins. It is not surprising that this set of interpolated free energies correlates better with BEM calculations than do the free energies from the met-enkephalin weights. The range of the exact weights in this test set suggests that it may be difcult to develop and justify empirical approaches to interpolating between the BIBEE approximations; Fig. 10b is a scatter plot of the exact BIBEE/P weights against the electrostatic free energy calculated using BEM.
V. DISCUSSION

In the present paper, we have described how the BIBEE approach to approximating electrostatic interactions can be used to rigorously bound the true electrostatic free energy of interaction between a molecular solute and the surrounding solvent. The BIBEE bounds are derived from diagonal approximations to a BIE formulation of the mixed-dielectric Poisson problem using quasi-Hermitian operator theory58 to characterize the relationship between the approximate free energies and the approximation to the boundary-integral operator. The bounds appear to hold when the integral equation

is discretized and approximated numerically using standard BEM techniques, although more detailed numerical analysis is warranted. Structures of met-enkephalin taken from a MD simulation have been used to demonstrate the bound performance relative to high-resolution BEM simulations and to GB calculations. We have also employed a set of several hundred protein structures15 to illustrate that the effective but unproven bound from the BIBEE/P approximation holds in practice over a variety of macromolecular geometries. Clearly, in order for the BIBEE bounds to be valuable in practice, it is necessary that they be computable in much less time than is required for full solution of the mixed-dielectric Poisson problem. The computational complexity of BIBEE calculations therefore merits discussion. Both the BEM and BIBEE calculations scale essentially linear with the surface area of the macromolecule if one uses appropriate computational techniques e.g., fast multipole53,55 or FFTSVD Ref. 48. In previous work38 it was shown that the BIBEE methods are typically 23 times slower than SGB/CFA methods if one uses the same surface discretization to evaluate the reaction-potential matrices explicitly and approximately ten times faster than BEM simulations based on compressing the dense matrices that represent the discretized integral operators.48 This order-of-magnitude improvement is likely only a weak LB because the BIBEE implementation has not been optimized for performance. In particular, we expect signicant further improvements to arise in applications similar to the met-enkephalin calculations in the present work, in which the electrostatic free energies of a large number of very similar geometries are to be estimated. Lower and upper electrostatic free-energy bounds can be proven rigorously using BIBEE; however, the provable LB from BIBEE/LB is so loose that it is likely to be of minimal practical value. In contrast, the effective bound generated by BIBEE/P might well be valuable, but its bounding properties remain unproven. Current work therefore focuses on identi-

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

104108-9

Bounding electrostatic free energies

J. Chem. Phys. 130, 104108 2009

fying why the BIBEE/P method works in practice to generate LBs for electrostatic free energies when considering molecular charge distributions and surfaces. Because it is known that some simple surfaces do in fact have positive eigenvalues,61,62 charge distributions that excite these eigenstates would have their responses mispredicted by the BIBEE/P method. However, we emphasize that bounding the electrostatic free energy does not necessarily require bounding all the eigenvalues, but only the ones corresponding to eigenvectors that realistic charge distributions can excite. It is possible that point-charge-based charge distributions, which rarely have charges within an atomic radius of the dielectric boundary, do not excite the modes that BIBEE/P would mispredict. It is also possible that the combination of BIEs and quasi-Hermitian theory may yield further insights into BIBEE/P or the precise relationship between the eigenvectors of the true and BIBEE-approximated reaction-potential matrices. The derived bounds do not demonstrate that the actual reaction-potential-matrix eigenvalues are bounded by the eigenvalues from the BIBEE reaction-potential matrices. Such a proof would require that the BIBEE methods give rise to reaction-potential matrices with the same eigenvectors as the true reaction-potential matrix. Instead, the present work demonstrates that for a given charge distribution, the BIBEE methods generate upper and LBs for the true electrostatic solvation free energy. The analysis of the BIBEE approximations rests on the compactness of the normal electric-eld operator. Compactness depends, in part, on smoothness properties of the solutesolvent boundary. Many of the most popular denitions, including the van der Waals surface, the solvent-accessible surface, and the solvent-excluded also known as molecular surface,74,75 can violate the Liapunov-smoothness criterion in many places on the boundary. Although the bounds appear to still hold for the molecular surfaces employed here, one subject for future work on BIBEE is incorporating algorithms that generate suitably smooth surfaces.7679 It is possible that algorithms introduced recently to dene excluded volumes in the GB model80 will be preferable to the standard surfaces. The rigorous alpha-shapes method developed by Liang et al.81,82 may also be a suitable starting point. It is clear that methods for estimating electrostatic interactions must reproduce both the eigenvalues and the eigenvectors of the reaction-potential matrices of high-resolution simulations of the linear-continuum model if such methods are to be described as accurate. As shown in earlier work38 and in Sec. IV of this paper, BIBEE methods reproduce with high delity component-wise interactions, as indicated by the excellent alignment between the eigenvectors of the BEM and BIBEE reaction-potential matrices. The BIBEE approach thus satises one of the necessary conditions for a satisfactory implicit-solvent model. The bounds derived in the present work suggest that it is possible to nd accurate eigenvalue estimates using BIBEE. Current work is directed toward developing methods for inexpensively estimating these eigenvalues, and we acknowledge that the presented BIBEE methods do not provide eigenvalue approximations of uniformly high quality as does the GBMV approach.

However, it should be emphasized that BIBEE achieves its level of accuracy without any empirical tting terms and without any parametrization beyond what is required for the continuum model itself. It is possible that the BIBEE method may be useful for the development of further renements of GB methods, or that the corrections developed in GB could be used to improve BIBEE eigenvalue estimates. Explicitsolvent MD simulations represent another model against which GB calculations can be assessed,18,8387 and many GB models have been presented in combination with physical interpretations drawn from the continuum model or using parameters t against continuum calculations.911,20,21,23,71,88 The comparison of BIBEE calculations to explicit-solvent simulations is an interesting question for future work. It is not clear whether BIBEE can be adapted to provide forces for dynamics, as GB has been adapted.10,89 However, BIBEE in its present form may still be useful for calculations in which continuity of forces is not a requirement, similar to grid-based GB methods. The BIBEE bounds themselves may be useful for Monte Carlo simulations of macromolecules. In particular, the availability of rapidly calculable bounds makes it possible to accelerate sampling by avoiding expensive electrostatics simulations.90 The existence of reasonable upper and LBs has implications for molecular design as well, particularly branch-and-bound pruning algorithms and hierarchical search techniques. Our results may also have implications for optimization of electrostatic interactions between molecules.91,92 The desolvation matrix associated with rigid molecular binding is a difference of reaction-potential matrices, which raises the possibility that the BIBEE bounds presented here may be used to bound the desolvation-matrix eigenvalues. Such bounds, if obtained, could be useful for screening candidate ligand geometries in molecular design. Finally, extending the proof for the BIBEE bounds to disconnected dielectric boundaries93,94 would allow new approaches to simulating ionic or protein solutions with many low-dielectric bodies.53
ACKNOWLEDGMENTS

The authors thank B. Roux for the use of CHARMM. J.P.B. gratefully acknowledges support from a Wilkinson Fellowship in Scientic Computing. This work was supported by the Mathematical, Information, and Computational Sciences Division Subprogram of the Ofce of Advanced Scientic Computing Research, Ofce of Science, U.S. Department of Energy under Contract No. DE-AC0206CH11357.
J. G. Kirkwood, J. Chem. Phys. 2, 351 1934. C. Tanford and J. G. Kirkwood, J. Am. Chem. Soc. 79, 5333 1957. 3 A. Warshel and S. T. Russell, Q. Rev. Biophys. 17, 283 1984. 4 K. A. Sharp and B. Honig, Annu. Rev. Biophys. Biophys. Chem. 19, 301 1990. 5 B. Honig and A. Nicholls, Science 268, 1144 1995. 6 J. Warwicker and H. C. Watson, J. Mol. Biol. 157, 671 1982. 7 M. K. Gilson, A. Rashin, R. Fine, and B. Honig, J. Mol. Biol. 184, 503 1985. 8 R. Luo, L. David, and M. K. Gilson, J. Comput. Chem. 23, 1244 2002. 9 W. C. Still, A. Tempczyk, R. C. Hawley, and T. F. Hendrickson, J. Am. Chem. Soc. 112, 6127 1990. 10 D. Qiu, P. S. Shenkin, F. P. Hollinger, and W. C. Still, J. Phys. Chem. A
1 2

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

104108-10

Bardhan, Knepley, and Anitescu

J. Chem. Phys. 130, 104108 2009 3 1996. F. G. Scholtz, H. B. Geyer, and F. J. W. Hahne, Ann. Phys. 213, 74 1992. 59 O. D. Kellogg, Foundations of Potential Theory Dover, New York, 1953. 60 M. A. Jaswon and G. T. Symm, Integral Equation Methods in Potential Theory and Elastostatics Academic, New York, 1977. 61 J. F. Ahner, J. Math. Anal. Appl. 181, 328 1994. 62 J. F. Ahner and R. F. Arenstorf, J. Math. Anal. Appl. 117, 187 1986. 63 J. P. Bardhan, M. D. Altman, J. K. White, and B. Tidor, J. Chem. Phys. 127, 014701 2007. 64 M. F. Sanner, http://www.scripps.edu/sanner/html/msms_home.html, 1996. 65 B. R. Brooks, R. E. Bruccoleri, B. D. Olafson, D. J. States, S. Swaminathan, and M. Karplus, J. Comput. Chem. 4, 187 1983. 66 A. D. MacKerell, Jr., D. Bashford, M. Bellott, R. L. Dunbrack, Jr., J. D. Evanseck, M. J. Field, S. Fischer, J. Gao, H. Guo, S. Ha, D. Joseph McCarthy, L. Kuchnir, K. Kuczera, F. T. K. Lau, C. Mattos, S. Michnick, T. Ngo, D. T. Nguyen, B. Prodhom, W. E. Reiher, III, B. Roux, M. Schlenkrich, J. C. Smith, R. Stote, J. Straub, M. Watanabe, J. WiorkiewiczKuczera, D. Yin, and M. Karplus, J. Phys. Chem. B 102, 3586 1998. 67 A. D. MacKerell, Jr., M. Feig, and C. L. Brooks III, J. Am. Chem. Soc. 126, 698 2004. 68 A. D. MacKerell, Jr., M. Feig, and C. L. Brooks III, J. Comput. Chem. 25, 1400 2004. 69 W. Humphrey, A. Dalke, and K. Schulten, J. Mol. Graphics 14, 33 1996. 70 J. C. Phillips, R. Braun, W. Wang, J. Gumbart, E. Tajkhorshid, E. Villa, C. Chipot, R. D. Skeel, L. Kale, and K. Schulten, J. Comput. Chem. 26, 1781 2005. 71 A. Ghosh, C. S. Rapp, and R. A. Friesner, J. Phys. Chem. B 102, 10983 1998. 72 M. D. Altman, J. P. Bardhan, J. K. White, and B. Tidor, Engineering in Medicine and Biology Conference, 2005 unpublished. 73 D. Sitkoff, K. A. Sharp, and B. Honig, J. Phys. Chem. B 98, 1978 1994. 74 F. M. Richards, Annu. Rev. Biophys. Bioeng. 6, 151 1977. 75 M. L. Connolly, J. Appl. Crystallogr. 16, 548 1983. 76 R. J. Zauhar, J. Comput.-Aided Mol. Des. 9, 149 1995. 77 Y. N. Vorobjev and H. A. Scheraga, J. Comput. Chem. 18, 569 1997. 78 S. Bhat and E. O. Purisima, Proteins: Struct., Funct., Bioinf. 62, 244 2006. 79 P. W. Bates, G. W. Wei, and S. Zhao, J. Comput. Chem. 29, 380 2008. 80 M. S. Lee, M. Feig, F. R. Salsbury, and C. L. Brooks III, J. Comput. Chem. 24, 1348 2003. 81 J. Liang, H. Edelsbrunner, P. Fu, P. V. Sudhakar, and S. Subramaniam, Proteins: Struct., Funct., Bioinf. 33, 1 1998. 82 J. Liang, H. Edelsbrunner, P. Fu, P. V. Sudhakar, and S. Subramaniam, Proteins: Struct., Funct., Bioinf. 33, 18 1998. 83 B. D. Bursulaya and C. L. Brooks III, J. Phys. Chem. B 104, 12378 2000. 84 B. Xia, V. Tsui, D. A. Case, H. Jane Dyson, and P. E. Wright, J. Biomol. NMR 22, 317 2002. 85 H. Nymeyer and A. E. Garca, Proc. Natl. Acad. Sci. U.S.A. 100, 13934 2003. 86 Z. A. Sands and C. A. Laughton, J. Phys. Chem. B 108, 10113 2004. 87 M. S. Formaneck and Q. Cui, J. Comput. Chem. 27, 1923 2006. 88 T. Grycuk, J. Chem. Phys. 119, 4817 2003. 89 G. D. Hawkins, C. J. Cramer, and D. G. Truhlar, J. Phys. Chem. A 100, 19824 1996. 90 L. David, R. Luo, and M. K. Gilson, J. Comput.-Aided Mol. Des. 15, 157 2001. 91 E. Kangas and B. Tidor, J. Chem. Phys. 112, 9120 2000. 92 L.-P. Lee and B. Tidor, J. Chem. Phys. 106, 8681 1997. 93 A. Greenbaum, L. Greengard, and G. B. McFadden, J. Comput. Phys. 105, 267 1993. 94 J. Helsing and E. Wadbro, J. Comput. Phys. 202, 391 2005.
58

101, 3005 1997. M. Schaefer and M. Karplus, J. Phys. Chem. 100, 1578 1996. 12 D. Bashford and D. A. Case, Annu. Rev. Phys. Chem. 51, 129 2000. 13 N. A. Baker, Curr. Opin. Struct. Biol. 15, 137 2005. 14 M. Feig, J. Chocholousov, and S. Tanizaki, Theor. Chem. Acc. 116, 194 2006. 15 M. Feig and C. L. Brooks III, Curr. Opin. Struct. Biol. 14, 217 2004. 16 S. R. Edinger, C. Cortis, P. S. Shenkin, and R. A. Friesner, J. Phys. Chem. B 101, 1190 1997. 17 L. David, R. Luo, and M. K. Gilson, J. Comput. Chem. 21, 295 2000. 18 L. Y. Zhang, E. Gallicchio, R. A. Friesner, and R. M. Levy, J. Comput. Chem. 22, 591 2001. 19 M. S. Lee, F. R. Salsbury, and C. L. Brooks III, J. Chem. Phys. 116, 10606 2002. 20 A. Onufriev, D. A. Case, and D. Bashford, J. Comput. Chem. 23, 1297 2002. 21 M. Wojciechowski and B. Lesyng, J. Phys. Chem. B 108, 18368 2004. 22 J. Zhu, E. Alexov, and B. Honig, J. Phys. Chem. B 109, 3008 2005. 23 J. Mongan, C. Simmerling, J. A. McCammon, D. A. Case, and A. Onufriev, J. Chem. Theory Comput. 3, 156 2007. 24 B. N. Dominy and C. L. Brooks III, J. Phys. Chem. B 103, 3765 1999. 25 W. Im, M. S. Lee, and C. L. Brooks III, J. Comput. Chem. 24, 1691 2003. 26 R. Zhou and B. J. Berne, Proc. Natl. Acad. Sci. U.S.A. 99, 12777 2002. 27 C. M. Stultz, J. Phys. Chem. B 108, 16525 2004. 28 S. Jang, E. Kim, and Y. Pak, Proteins: Struct., Funct., Genet. 62, 663 2006. 29 R. Geney, M. Layten, R. Gomperts, V. Hornak, and C. Simmerling, J. Chem. Theory Comput. 2, 115 2006. 30 S. Miert, E. Scrocco, and J. Tomasi, Chem. Phys. 55, 117 1981. 31 D. G. Levitt, Biophys. J. 22, 209 1978. 32 P. C. Jordan, Biophys. J. 39, 157 1982. 33 P. B. Shaw, Phys. Rev. A 32, 2476 1985. 34 R. J. Zauhar and R. S. Morgan, J. Mol. Biol. 186, 815 1985. 35 R. J. Zauhar and R. S. Morgan, J. Comput. Chem. 9, 171 1988. 36 R. Cammi and J. Tomasi, J. Comput. Chem. 16, 1449 1995. 37 E. O. Purisima and S. H. Nilar, J. Comput. Chem. 16, 681 1995. 38 J. P. Bardhan, J. Chem. Phys. 129, 144105 2008. 39 J. D. Jackson, Classical Electrodynamics, 3rd ed. Wiley, New York, 1998. 40 A. H. Juffer, E. F. F. Botta, B. A. M. van Keulen, A. van der Ploeg, and H. J. C. Berendsen, J. Comput. Phys. 97, 144 1991. 41 S. M. Rao, T. K. Sarkar and R. F. Harrington, IEEE Trans. Microwave Theory Tech. 32, 1441 1984. 42 D. Boda, M. Volisk, B. Eisenberg, W. Nonner, D. Henderson, and D. Gillespie, J. Chem. Phys. 125, 034901 2006. 43 K. E. Atkinson, The Numerical Solution of Integral Equations of the Second Kind Cambridge University Press, Cambridge, England, 1997. 44 G. C. Hsaio and W. L. Wendland, Boundary Element Methods: Foundations and Error Analysis Wiley, New York, 2004. 45 W. Pogorzelski, Integral Equations and Their Applications Pergamon, New York, 1966. 46 R. Kress, Linear Integral Equations Springer-Verlag, Berlin, 1989. 47 S. Hnger and T. Simonson, J. Comput. Chem. 22, 290 2001. 48 M. D. Altman, J. P. Bardhan, J. K. White, and B. Tidor, J. Comput. Chem. 30, 132 2009. 49 J. Tausch, J. Wang, and J. White, IEEE Trans. Comput.-Aided Des. 20, 1398 2001. 50 Y. Saad and M. Schultz, SIAM Soc. Ind. Appl. Math. J. Sci. Stat. Comput. 7, 856 1986. 51 H. X. Zhou, Biophys. J. 65, 955 1993. 52 J. Liang and S. Subramaniam, Biophys. J. 73, 1830 1997. 53 B. Z. Lu, D. Q. Zhang, and J. A. McCammon, J. Chem. Phys. 122, 214102 2005. 54 B. J. Yoon and A. M. Lenhoff, J. Comput. Chem. 11, 1080 1990. 55 M. O. Fenley, W. K. Olson, K. Chua, and A. H. Boschitsch, J. Comput. Chem. 17, 976 1996. 56 V. Ya. Raevskii, Theor. Math. Phys. 100, 1040 1994. 57 J. F. Ahner, V. V. Dyakin, V. Ya. Raevskii, and St. Ritter, Math. Notes 59,
11

Downloaded 11 Mar 2009 to 18.51.1.222. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/jcp/copyright.jsp

Das könnte Ihnen auch gefallen