SB Drug Design

7/30/2019 SB Drug Design

1/20

Annu. Rev. Pharmacol. Toxicol. 1997. 37:7190Copyright c 1997 by Annual Reviews Inc. All rights reserved

STRUCTURE-BASED DRUG DESIGN:Computational Advances

Tami J. Marrone, James M. Briggs, and J. Andrew McCammonDepartment of Chemistry and Biochemistry and Department of Pharmacology,University of California San Diego, La Jolla, California 92093-0365

KEY WORDS: structure-based drug design, free energy perturbation, docking, computationalchemistry

ABSTRACTStructure-based computational methods continue to enhance progress in the dis-covery and renement of therapeutic agents. Several such methods and theirapplications are described. These include molecular visualization and molecularmodeling, docking, fragment methods, 3-D database techniques, and free-energyperturbation. Related issues that are discussed include the use of simplied po-tential energy functions and the determination of the positions of tightly boundwaters. Strengths and weaknesses of the various methods are described.

INTRODUCTIONComputing is used in various ways in drug discovery. Important examplesinclude QSAR, articial intelligence, and structure-based methods. Here, wefocus on the structure-based methods. These show increasing utility for the dis-covery of lead compounds, and especially for the renement of lead compoundsand for the re-engineering of drugs to overcome certain types of resistance. The

structure-based methods are becoming increasingly important due in part to therapidgrowth in structural data (1004 structures released in the Protein DataBank in 1995 alone) (1), and the particularly high speed with which structures can bedetermined as part of a focused drug-discovery effort with a well-characterizedtarget.

This review is divided into two major sections: the discovery of leads, andthe renement and re-engineering of leads. The methods and topics describedhere include molecular visualization and modeling, docking from structuraldatabases, assembly of leads from fragments, 3-D database methods, simplied

potential energy functions, detailed calculations of equilibrium constants, and

710362-1642/97/0415-0071$08.00


2/20

72 MARRONE, BRIGGS & McCAMMON

NEW LEAD REFINED LEAD

Structure of Compound-Target

DockingComputational AlchemyMolecular Visualization

Docking from databasesAssembly of Leads from

Fragments3D Database MethodsMolecular Visualization

Figure 1 Schematic diagram of the structure-based drug design cycle.

methods to allow for water-binding sites. Since many of the methods usedto discover new classes of lead compounds can also be used to rene them,Figure 1 shows schematically where each of these methods can be used in thestructure-based drug design process.

DISCOVERY OF LEAD COMPOUNDS

Basic Molecular Visualization and ModelingAdvances in the ability to visualize molecular structure and properties haveled to a revolution in computer modeling. Many commercial (InsightII (2),Quanta (2), Cerius2 (2), Sybyl (3), CAChe (4), etc) and academic (Macro-Model (5), Grasp (6), etc) programs are available for this task, and many provideinterfaces to computational codes. Molecular structure and property visualiza-

tion are important in all phases of the molecular design process, although theyare used more heavily in some phases than in others. Some determinations of the goodness of a t of a model compound to a target binding site are made viavisual inspections of the docked structures, with interactive feedback of inter-atomic distances and energy components. Since most modeling efforts utilizevisualization, there are countless applications in the literature. However, onlya few of these applications are described here as examples of the importance of molecular visualization and modeling in structure-based drug design.

In a recent study directed at the rational design of nonpeptide-based inhibitors

of the HIV protease, molecular visualization and modeling played a signicant


3/20

STRUCTURE-BASED DRUG DESIGN 73

role (7). The work began with the knowledge of X-ray structures of HIVprotease/inhibitor complexes. It was clear from those structures that a tetraco-ordinated water molecule was present in the active site of the enzyme where itwas hydrogen bonded to the backbone amide hydrogens of two residues (Ile50

and Ile50 ) and donated two hydrogen bonds to the inhibitor. A goal of thedesign process was to incorporate the function of this water molecule into anew inhibitor. The design cycle started with the development of a pharma-cophore model based on the available X-ray data. The initial models includedtwo symmetric hydrophobic groups to t into the S1 and S1 hydrophobicpockets in the enzyme and one hydrogen bonding site that would bind to thecatalytic aspartates. The pharmacophore model was used to search a databaseof 3-D molecular structures and resulted in a molecule with a phenyl groupthat included an oxygen to take the place of the structural water molecule. Vi-

sual analysis of the structure of the HIV protease with the phenyl-containingmolecule suggested that a benzene ring might not be able to place all of thegroups where they needed to be simultaneously. A cyclohexanone-containingmolecule was next suggested and then subsequently modied to a syntheti-cally more accessible seven-membered-ring cyclic urea, with the added benetthat the nitrogens could be easily functionalized. Two hydroxyl groups wereused to hydrogen bond to the two aspartates in the active site of the protease.Further modeling suggested appropriate stereochemistries for all of the chiralsites and predicted that the cyclic urea should bind quite well. Later crystal-

lographic work conrmed the conclusions of the modeling work. Subsequentuse of medicinal chemistry to enhance the oral bioavailability resulted in asubnanomolar clinical trial candidate with good pharmacokinetics.

Another picomolar inhibitor of the HIV-1 protease was discovered by usingenergy minimization of new molecular structures within the active site of theenzyme (8). Only atoms in the inhibitors were allowed to move. Primarilythrough a visual inspection of the docked complexes, it was noticed that therewas unoccupied volume near Asp 29/30 and Asp 129/130 and that the backboneamide hydrogens could hydrogen-bond with appropriate acceptors on groups

in those pockets. Hydroxyl groups were placed on the phenyl rings in thosepockets, which resulted in an increase in binding by three orders of magnitudein the best case.

Visualization was used recently to help develop a simplied potential functionto quickly estimate the relative free energies of binding for inhibitors to theHIV-1 protease (9). The function consists of a molecular-mechanics-basedenthalpy term and a hydrophobic term. New potential inhibitors were dockedinto the active site by overlaying them with positions of known inhibitors fromX-ray structures and then relaxing the whole system using the AMBER (10)

program. An atomic hydrophobic interaction energy based on the molecular


4/20


surface was used; the hydrophobicity parameters were derived from a largedatabase of 1-octanol-water partition coefcients for a variety of molecules.Molecular hydrophobicity maps were visualized in efforts to understand thecritical interactions needed between inhibitors and the active-site hydrophobic

pockets; the maps helped in the rationalization of binding differences. Trends,but not magnitudes, in the calculated relative free energies of binding matchthose from experiment and from free-energy-perturbation calculations.

Docking from Structural DatabasesDockingcompounds from databases to targets of known structurecanbe utilizedto discover or rene new leads. Structures from a database of small-moleculecompounds are t into the target structure using a docking program. Theenergies of the resulting complexes are evaluated and those that show the most

promise can be experimentally tested as possible lead compounds. Dockinghas been reviewed extensively (1115), so this section only briey describesa few aspects and applications of docking as a method for lead generation instructure-based drug design.

The scoring method is critical in ranking the docked structures. Energy eval-uations can be computationally expensive, so simplied energy functions areused to rank the structures. Some of these energy functions are described inmore detail below. Besides using a simplied energy function, methods areemployed to make the energy evaluation more efcient. An example is the use

of grid-based methods (1619) for docking of exible or rigid compounds to arigid target. In these, the interaction energy of the target with appropriate probesis precomputed and mapped out onto a three-dimensional grid.

Initial docking algorithms (20, 21) involved tting a rigid compound witha rigid target. However, conformational exibility has been incorporated intoseveral docking algorithms (16, 17, 2228) at additional computational cost.Including conformational exibility increases the chance of nding the lowest-energy structure of the complex, because compounds and targets are capableof altering their structures upon binding. However, even where using confor-

mational exibility is not cost efcient for screening compounds in a database,it could be useful in further renement of leads.Several new lead compounds have been generated using docking from data-

bases (2931). One good example, which demonstrates how docking fromstructural databases can be used to generate and rene leads, is the screening of the Fine Chemicals Directory for possible new inhibitors of thymidylate syn-thase (31). Using a crystal structure of thymidylate synthase as the target, anaverage of 10 4 orientations of each compound were docked to the target usinga steric t. The electrostatic energy was used to score the complexes, and a

solvation correction was applied to the top-scoring compounds to further screen


5/20


the candidates. Based on these scores and lack of similarity to the thymidylatesynthase substrate, several possible inhibitors were proposed and screened forinhibition. The crystal structure of the complex of one of the leads, sulisoben-zone, with the enzyme showed another possible site that could be exploited.

Subsequent docking simulations with molecules sterically similar to sulisoben-zone identied phenolphthalein analogs that were found to have inhibitionconstants in the 130 M range. The crystal structure of phenolpthalein withthymidylate synthase showed that phenolpthalein bound in the alternate bind-ing site. This application nicely demonstrates one approach to structure-baseddrug design.

Faster computers and more efcient algorithms will permit the use of con-formational exibility during the screening process to identify leads that maybe missed within the context of rigid docking. More detailed and accurate po-

tential functions could also be used in the scoring phase of the docking process.Databases of commercially available compounds are sometimes preferred fordocking applications to avoid possible difculty in the syntheses of the leadcompounds generated by the docking simulations. The number of chemicallibraries has exploded with the advent of combinatorial synthesis, and per-haps databases of these libraries will offer a larger selection of commerciallyavailable compounds to investigate as leads in docking simulations.

Assembly of Leads from Fragments

When the structure of the target is known, novel leads can be generated us-ing fragment methods. The interactions of several chemical groups with thetarget are calculated, and possible binding sites are identied. Databases canbe searched to match small molecules or chemical groups with the possiblebinding site. These candidates can then be assembled into one compound usinglinking groups.

One of the earliest attempts to model the interaction of chemical groups witha protein was made by Goodford using GRID (19). This method was developedto quickly determine possible binding sites of a protein, and it is the basis for

several of the de novo design methods. The interaction energies of variousprobes representing chemical groups are mapped onto a grid. The interactionenergy was originally described using a coulombic term, a Lennard-Jones term,and a hydrogen-bonding term. The hydrogen-bonding term was subsequentlyrened by tting to experimental data (3234). Possible binding sites for aparticular chemical group are identied as the sites with the most favorableinteraction energies for a chemical group.

MCSS/HOOK The MCSS (Multiple Copy Simultaneous Search) (35) program

is used to nd energetically favorable binding sites and orientations for peptide


6/20


7/20


fragment library is constructed with the use of this scoring function, after whichthe fragments are arranged according to their score (including the solvationenergy). Fragments from the previously generated library of peptide fragmentsare then fused to the appropriate end of the seed and the energy of the complex

is evaluated (scored). Peptides with favorable scores are carried through theprocess for a second round of fusions, and so on. Once a collection of peptidesof desired length is obtained, a subset is subjected to energy minimization in thebinding site (the peptides and a selected portion of the binding site are permittedto move). The peptides are also energy minimized outside of the binding siteto estimate whether the conformational change that is required for binding willcause the binding energy to become unfavorable. This procedure was carriedout for two aspartyl proteases, rhizopuspepsin, and a model of the renin activesite. In these cases, after relaxation of the most promising leads in the active

sites of the X-ray complexes, structures were obtained that closely matched thecrystal structures. As with most computational methods, graphical analyses areused heavily in the rst and nal stages.

MCDNLG The Monte Carlo De Novo Ligand Generator (MCDNLG) (40) usesa binding cavity from an X-ray or NMR structure that is then lled with adensely packed array of atoms of random type. Each atom has three proper-ties: element type, hybridization, and the number of implicit hydrogens. Thepseudo-atoms can have a maximum of 12 bonds simultaneously. The energy

function is composed of intra- and intermolecular components that control lig-and geometry and ultimate shape and chemical complementarity to the bindingsite. Some intramolecular terms include bond energy, angle and dihedral anglestrain, and valence strain (this allows for bond making and breaking). The inter-molecular components include dispersion, repulsion, hydrogen bonding, and adesolvation penalty for polar atoms. A Monte Carlo procedure is employed torandomly select atoms to appear or disappear or change type or change bondingarrangements or rotate or translate. A novel ligand can be created in 300,000Monte Carlo steps. Quite diverse ligands result from different initial conditions

(random number seeds). As a test, the method was applied to the identica-tion of inhibitors of dihydrofolate reductase, thymidylate synthase, and HIV-1protease. Sets of compounds were generated, which fell into families that com-pared well with molecules known to bind. It should be noted that this methodis most useful for coming up with novel leads, not for nding either the nalproduct or molecules that are synthetically accessible.

SPROUT The SPROUT program (41) requires information about the bindingsite from either X-ray or NMR data, or from a pharmacophore model from the

overlay of known inhibitors in their bioactive conformations. Target sites in the


8/20


binding pocket are identied and labeled by type (primarily hydrogen-bonding,electrostatic, and hydrophobic). The binding site dimensions are conferred bya xed boundary. A fragment library is used that has been presorted accordingto atomic and molecular properties (shape, hydrogen-bond, etc). Fragments of

appropriate shape are selected and overlaid on a target site. The fragment canbe rotated around a pivot point to nd an ideal docking arrangement (i.e. onethat satises the steric requirements but does not violate the boundary). Oncefragments have been docked into all of the target sites, fragments are joined ina manner dependent on the identity of the fragments (acyclic, cyclic, fusion,bridging, and spiro). In the second phase of the sprouting, atom types are inter-changed with others of the same hydridization in order to nd a combinationwith optimal interactions with the binding site. The scoring function is com-posed of terms accounting for the loss of translational, rotational, and torsional

degrees of freedom on binding, hydrophobic, van der Waal, and solvation freeenergies, and an enthalpy term to account for the energy required to obtainthe bioactive conformation. The methods were applied to trypsin and HIV-1protease and generated ligands that resembled known inhibitors.

3-D Database MethodsCLIX The CLIX program (42) uses a combination of the GRID (19) pro-gram and a search and docking from the Cambridge Crystallographic Database(CCDB) in order to identify small molecules that have both the steric and

chemical likelihood of binding. The GRID program is used to nd discretebinding sites for a large variety of small representative groups. Once identied,a database search is undertaken of the CCDB for ligands that contain pairs of interaction sites identied by the GRID search. An iterative search is performedto ensure that selected molecules from the CCDB contain groups representativeof the GRID interaction sites. Each molecule is then placed into the bindingsite such that two of the matching groups in the molecule are overlaid on theirrespective GRID interaction sites. The molecule is rigidly rotated about the linethat connects the two interaction sites to see whether an overlap can be found

that overlays other functionalities in the molecule with other GRID interactionsites but does not cause steric interactions with the binding site. The GRIDenergy-scoring method is used. Further analyses with the GRID program areused to identify places in the binding site that are not satised with a particularsmall molecule. Corresponding portions in the small molecule are changed tomore appropriate groups (i.e. those that correspond to the GRID probes thathave high scores in those regions). Users of CLIX try to ensure that chemi-cally reasonable choices are made by considering adjacent functionalities. Atest of the procedure was made with hemagglutinin for the identication of the

binding mode of sialic acid. Of the seven resulting structures with the lowest


9/20


energy, three were very close to the crystal structure (0.420.52 A), while theremaining four were far off (46 A).

CAVEAT 3-D DATABASE SEARCHING The CAVEAT program (39) represents a

novel approach to 3-D database searching (43). Most 3-D database searchingmethods involve the overlaying of molecules in a database with a known in-hibitor or a pharmacophore model. The CAVEAT program operates under theassumption that it is enough, at least initially, to focus on the orientation of bonds and not on the placement of atoms. In this way, a 3-D database can beprescreened for pairs of bonds and their relative orientations. User-speciedbond types are used to construct a sorted CAVEAT database of bond unit vectorpairs. A major goal of the CAVEAT work was to generate a method that couldgive results in an interactive fashion (i.e. fast). This is accomplished by do-

ing most of the processing work during the construction of the CAVEAT bondvector database.Queries are constructed via the consideration of pairs of bond vectors. One

seeks to address the smallest set of relevant pairs. Tolerances are assigned toeach query pair so that close matches can be retained. The nal phase of thequery is to group the results into families. Graph-theoretical methods are used,among others, for the clustering of similar/related structures. A CAVEATdatabase consisting of 50,000,000 bond pairs can be searched to generatehundreds to thousands of matches in 0.52.0 minutes.

Simplied Potential Energy FunctionsSimplied empirical potential energy functions are used to rank docked struc-tures or predict binding afnities. This section briey discusses molecular me-chanics potential functions. Some methods used to include solvation implicitlyin molecular mechanics calculations are discussed in detail along with theirstrengths and weaknesses. Applications of these methods are also described.

Traditional energy renement, molecular dynamics, and Monte Carlo sim-ulations of molecules often use molecular-mechanics-based potential func-

tions. These functions consist of bonded interactions (bonds, bond angles,dihedrals) and nonbonded interactions (coulombic, Lennard-Jones). The pa-rameters and functional forms used to describe the energetics of the system areoften called a force eld. Several force elds have been developed for pro-teins (4449) and nucleic acids (4446, 48, 50). The number of nonbondedinteractions among the atoms usually increases on the order of n2, where n isthe number of atoms in the system. Energy evaluations become more com-putationally expensive as the size of the compound-target complex increasesor if explicit waters are added to the system to describe solvation. Quantita-

tive calculation of the relative free energies of binding of two compounds to


10/20


the same target is quite difcult using molecular-mechanics potential functionsand including explicit solvation. However some success has been obtainedwith this approach, as discussed in detail in the section below on computationalalchemy. To obtain qualitative relative free energies of binding or to rank several

structures of a compound-target complex, simpler energy functions have beendeveloped to predict trends without the cost of a detailed simulation. Thesefunctions describe the free energies of binding or the energy of a conforma-tion in simplied, empirical terms for rapid screening. Several of these energyfunctions have been described in recent reviews (11) on docking and usuallyinclude an electrostatic term and a hydrophobic term that can be based on thesurface area of the system.

Some energy functions use molecular-mechanics-based force elds and sim-ple models of solvation. One common empirical method makes use of the

solvent-accessible surface area of each amino acid or other group (5154).Atomic solvation parameters for these groups are derived from empirical vapor-to-water or water-to-octanol transfer free energies for appropriate analogues. Ahydrophilic group has a favorable solvation energy, while a hydrophobic grouphas an unfavorable solvation energy. Derivatives of the accessible surface areawith respect to the atomic positions can be used in molecular-mechanics andmolecular-dynamics simulations. A similar solvation term was developed (55)where the interaction between a protein and solvent is described as a functionof atom occupancies. Here the occupancy of an atom is the volume around the

atom that is available for water to occupy. The overlap of the solute atoms andthe occupancy is described as a Gaussian function that is easily differentiated.This allows the method to be used in dynamics (55, 56) and mechanics. Others(57) have also parameterizered atomic solvation parameters to be compatiblewith molecular mechanics calculations. Such solvation models have been com-pared in the literature and shown to be very sensitive to the number and choiceof atomic solvation parameters (58, 59). These methods have a major drawback in that they essentially include only the rst hydration shells of water and do nottake into account any longer-range forces such as electrostatics (59). Another

way to describe solvation is through the use of continuum models, where thesolute is treated as a set of point charges in a dielectric continuum. Solvationenergies are calculated by solving the Poisson-Boltzmann (PB) equation, andthe derivatives of the associated electrostatic forces can be used in stochasticdynamics simulations (60). The PB approach only calculates the electrostaticinteractions of the solutes and the solvent. Lennard-Jones interactions as wellas the energy to form the solute cavity in the continuum must be added-inwhere necessary. However, in comparing a series of similar compounds, theLennard-Jones and cavity terms would be similar and could be neglected. This

approach was utilized (61) to show that PB electrostatics calculations better


11/20


estimated the experimental relative free energies of solvation of organophos-phorous molecules than explicit free energy calculations. In the case that elec-trostatics dominate the binding of a compound to a target, the relative freeenergies of binding can be calculated for a series of compounds using the PB

approach (6264). The PB approach has been successful in predicting rateconstants of diffusion-controlled enzymatic reactions from Brownian dynam-ics simulations (65). These calculations have also shown success in estimatingpKa s and ionization states of ionizable groups in proteins (66).

To account for the missing cavitation and Lennard-Jones terms in continuummodels, hydrophobic solvation can be described as an apolar surface areadependent term. It has been shown that the vapor-to-water and water-to-octanoltransfer free energies of alkanes are approximately linear with respect to sur-face area, although there is a debate on how to extract the magnitude of the

hydrophobic interaction with water. A new parameter (67) set has been de-rived that includes a surface areadependent term for use in continuum modelsto predict the free energies of solvation of organic molecules. The stabilitiesof loops (68), helices (69), and beta sheets (70) have been examined with thecontinuum-plus-apolar model. Structures have been successfully ranked fromdocking procedures (71), and the effects of mutation in the repressor-operatorcomplex have been studied (72) using this model.

The PB-plus-apolar model has several advantages. The solvent is a contin-uum, so computational time is saved by avoiding sampling of solvent congu-

rations (73). Also, long-range electrostatics and ionic strength are included inthe calculation via the PB equation. Polarization can be included by varyingthe dielectric coefcients of the solute. However, this method may not be wellsuited forelectrically neutral solutes (i.e. no partial charges on the solute model)because the hydrophobic effect is taken into account in a very approximate waythrough the surface term.

REFINEMENT AND RE-ENGINEERING

Computational AlchemyThis section briey reviews the thermodynamic cycle perturbation method,which in principle allows exact calculations of the relative binding afnities of compounds for a given target. A general description of the method is given,followed by examples of its use in drug discovery efforts. The concluding dis-cussion describes some of the limitations, recent advances, and future prospectsof this method.

The binding constant of a compound to a target is related of course to thefree energy of binding. Many of the computational tools described above work

by providing rapid estimates of these free-energy changes. The exact relation


12/20


between binding contants and free-energy changes is given by the familiarexpression

K = exp( G / RT ),

where K is the thermodynamic equilibrium constant, G

is the standardchange in free energy, R is the gas constant, and T is the absolute temperature.In principle, it would be possible to determine which of two compounds, C 1 andC 2, would bind more strongly to a target M by simulating the two binding pro-cesses in the appropriate solvent and calculating the corresponding changes infree energy, G 1 and G

2. Such calculations are in fact possible for very sim-

ple solutes, using either molecular dynamics or Monte Carlo simulations (74).For solutes of the complexity of pharmacologic compounds and targets, di-

rect calculations of the free energies of binding are seldom possible. This is due

to the signicant structural changes involved, which may include desolvation of charged groups on the compound and of recessed binding sites in the target, andchanges in the conformation and titration states of either molecule. The poten-tial energy functions used in current simulation studies generally do not allowfor automatic adjustment of titration states. But a more general problem en-countered in efforts to simulate binding is that the solvation and conformationalchanges can not be simply accounted for because they occur on time scales thatarelongcompared to typicalsimulations. With current supercomputers, simula-tions of an enzyme-inhibitor complex in water can be extended to perhaps 10 ns.

Actual binding processes may require many orders of magnitude more time, andattempts to simulate them on shorter time scales can lead to artifactual results.To overcome this problem, an indirect approach to calculating relative free

energies of binding was introduced in 1984 (75). This makes use of the ther-modynamic cycle shown in Figure 2. Because free energy is a state function,the relative free energy of binding, G = G 2 G

1 = G

4 G

3. The

determination of G 3 and G4 involves simulations in which the molecule C 1

is gradually transformed into the molecule C 2. In the rst biological applicationof the method, for example, benzamidine was transformed into parauoroben-

zamidine, both in water, to provide G3, and in the substrate recognition pocketof trypsin, to provide G 4 (76). Because these nonphysical transformations of-

ten involve transmutation of one type of atom into another, such calculationshave been colorfully described as computational alchemy (74). Comparisonof the magnitudes of G 3 and G

4 provides insight into the relative contri-

butions of solvation and compound-target interactions in the analysis of theorigins of the binding selectivity.

To carry out an alchemical simulation, one must gradually change the po-tential energy function from that for C 1 and its surroundings to that for C 2 and

its surroundings, during the course of a molecular-dynamics or Monte Carlo


13/20


M + C 2 C 2 M

C1

MM + C 11

2

3 4

Figure 2 Schematic diagram of a thermodynamic cycle.

simulation. Two general approaches have been widely used for calculating thefree-energy changes. In the perturbation method, a simulation is run for C 1 inits surroundings, and then the changes in potential energy upon making a nitebut small change in C 1 (e.g. changing the C - H bondlength 25% of the way to-ward the C -F bondlength, in the benzamidine example) are calculated for eachof a sequence of snapshots from the simulation. Properly averaged, these en-

ergy changes provide an estimate of the free-energy change for the small changein C 1. Then a simulation is run for the partly changed molecule in its surround-ings and analyzed as above to estimate the next increment in the free-energychange. This procedure is continued until one has a set of nite differences thatcan be added to provide an estimate of the total free-energy change, G 3 or

G 4. The other widely used approach for calculating alchemical free-energychanges is the thermodynamic integration method. This method providesestimates of derivatives of the free energy with respect to a parameter thatdenes the extent of the change in C 1, again using a sequence of simulations

corresponding to different values of in a standard implementation. A simpleintegral of these derivatives from = 0 to = 1 then provides the estimateof G 3 or G

4. An excellent account of the technical details of alchemical

simulations has been published very recently (77).Computational alchemy methods have been applied to a number of systems

of pharmaceutical interest in recent years, primarily to explore ways in whichcompounds might be modied to increase their binding to soluble moleculartargets. The early availability of crystallographic structures of the HIV pro-tease complexed with a variety of inhibitors has stimulated many such studies,

several of which have been described in the literature. A good example, which


14/20


includes references to earlier studies, is a very recent analysis of modicationsof a parent N , N -disubstituted benzene sulfonamide (78). The parent com-pound, which was arrived at by structure-based drug design, was modied inways intended to increase its solubility (and bioavailability) while maintain-

ing its charge, molecular weight, and inhibitory potency. The modicationsinvolved the replacement of an aromatic H by OH and NH 2. The calculatedchanges in the free energies of hydration and of binding were found to be ingood agreement with experimental results, although the experimental hydra-tion data were estimated from the solubilities of the corresponding solids. Inanother good example, signicantly larger changes were made to evaluate theprospective afnities of a series of inhibitors based on a hydroxyethylene back-bone (79). This work was actually used to guide the synthesis of compounds fortesting, and the predicted afnities were found to be in general agreement with

the subsequent experiments. This study also emphasized the value of the com-putational alchemy approach in accounting for differences in the desolvationcontributions to the binding constants.

Other systems of pharmacologic interest that have been studied by computa-tional alchemy methods include the following enzymes and sets of related in-hibitors: thymidylate synthase (80), acetylcholinesterase (81), adenosine deam-inase (82), and elastase (83).

Experience obtained from the simulations described above, and from appli-cations of computational alchemy to other systems (74, 84), point to a number

of strengths and limitations of the method. The primary strength, of course,is the rigorous foundation of the method. As the speed of computers and thequality of the potential functions increase, the calculations will become in-creasingly accurate. Indeed, for small- or medium-sized solutes (e.g. synthetichost-guest systems), discrepancies between theoretical and experimental mea-sures of molecular recognition have been resolved by corrections of the latterin some cases (85). Such accuracy is not yet routinely possible in problems of pharmacological interest, as discussed below.

Perhaps the greatest challenge in attempting to calculate the relative free

energy of binding of two or more compounds to a target is adequately sam-pling the relevant congurations of the atoms in the system (86). Even wherea high-resolution experimental structure of an initial compound-target com-plex is available, this does not describe the large number of slightly differentconformations that are reected in the thermodynamics of the system (8789).This distribution of congurations will in general change somewhat when onecompound is replaced by another in the target complex. Owing to the large sizeand inherent exibility of biopolymers, it is generally not possible to samplea representative set of the congurations involved (90). In fact, the sampling

problem is seen to be even more challenging when one recognizes that target


15/20


proteins, for example, contain many titratable side chains, and the protonationstates of these may change when different compounds are bound.

Another challenge facing the practical application of computational alchemyto drug discovery is insuring the adequacy of the underlying potential energy

function. Although existing potential functions that assume pairwise additivityof nonbonded interatomic interactions appear to be sufcient in many cases(91), the description of certain interactions may require explicit treatment of electronic polarizability (92) or special techniques to avoid the usual truncationof long-ranged electrostatic interactions (93). Also, existing potential func-tions are largely limited to amino acid residues, nucleic acid bases, and othercommonly occurring groups. The development of potential functions that areof the same quality for prosthetic groups or model compounds is a continuingchallenge (94).

The development of new methods to meet challenges such as those listedabove is an increasingly active area of research. Some notable lines of progressinclude the development of parallel computing tools that will increase the speedof the calculations (95); new approaches for choosing reference states in al-chemical simulations that will speed convergence of the free energy calcula-tions (96, 97); improved methods for assigning protonation states (98, 99);and improvements in the potential energy functions, both for specic chemicalgroups (92, 94) and for the treatment of long-range electrostatic interactions(93, 100).

Computational alchemy is already contributing to drug discovery efforts,particularly in the late stages of structure-based design projects and where thesampling and potential-function difculties are not too large (78, 79). Theutility of these methods will increase as the basic research described aboveyields faster, more generally reliable tools.

Structural Water Rening and re-engineering compounds can include designing the compoundto either mimic tightly bound waters or to utilize waters as bridging groups.

These design methods are currently being explored by computational methods.Several computational methods, such as grand canonical ensemble simulationsand neural networks, are used to determine water binding sites in targets. Grandcanonical methods allow water to be inserted/deleted during a Monte Carlo ormolecular-dynamics simulation. The probability of a water molecules occu-pying a position in the target can be calculated and used to predict the positionsof tightly bound waters. In particular, grand canonical Monte Carlo has beenused successfully to predict water binding sites in hyaluronic acid (101) andDNA (102) crystal structures. In these simulations, the target was held xed in

order to compare with the experimentally determined crystal waters. However,


16/20


17/20


9. Viswanadhan VN, Reddy MR, WlodawerA, Varney MD, Weinstein JN. 1996. Anapproach to rapid estimation of relativebinding afnities of enzyme inhibitors:application to peptidomimetic inhibitorsof the human immunodeciency virus

type 1 protease. J. Med. Chem. 39:7051210. Weiner SJ, Kollman PA, Case DA, Singh

UC, Ghio C, et al. 1984. A new force eldfor molecular mechanical simulation of nucleic acids and proteins. J. Am. Chem.Soc. 106:76584

11. Rosenfeld R, Vajda S, DeLisi C. 1995.Flexible docking and design. Annu. Rev. Biophys. Biomol. Struct. 24:677700

12. Lybrand TP. 1995. Ligand-protein dock-ing and rational drug design. Curr. Opin.Struct. Biol. 5:22428

13. Jones G, Willett P. 1995. Docking small-molecule ligands into active sites. Curr.Opin. Biotechnol. 6:65256

14. Kuntz ID, Meng EC, Shoichet BK. 1994.Structure-based molecular design. Acc.Chem. Res. 27:11723

15. Schoichet BK, Kuntz ID. 1996. Predict-ing the structure of protein complexes: astep in the right direction. Chem. & Biol.3:15156

16. Goodsell DS,Olsen AJ.1990. Automateddocking of substrates to proteins by sim-ulated annealing. Proteins: Struct. Funct.Genet. 8:195202

17. Luty BA, Wasserman ZR, Stouten PFW,Hodge CN, Zacharias M, McCammonJA. 1995. A molecular-mechanics/gridmethod for evaluation of ligand receptorinteractions. J. Comput. Chem. 16:45464

18. Meng EC, Shoichet BK, Kuntz ID.1992. Automated docking with grid-based energy evaluation. J. Comput.Chem. 13:50524

19. Goodford PJ. 1985. A computational pro-cedure for determining energetically fa-vorable binding sites on biologically im-portantmolecules. J.Med. Chem. 28:84957

20. Wodak SJ, Janin J. 1978. Computer anal-ysis of protein-protein interaction. J. Mol. Biol. 124:32342

21. Kuntz ID, Blaney JM, Oatley SJ, Lan-gridge R, Ferrin TE. 1982. A geometricapproach to macromolecule-ligand inter-actions. J. Mol. Biol. 161:26988

22. Judson RS, Jaeger EP, Treasurywala AM.1994. A genetic algorithm based methodfor docking exible molecules. J. Mol.Struct. Theochem. 308:191206

23. JonesG, Willett P, Geln RC.1995. Molec-

ular recognition of receptor sites using a

genetic algorithmwitha descriptionofde-solvation. J. Mol. Biol. 245:4353

24. Hart TN, Read RJ. 1992. A multiple-startMonte Carlo docking method. Proteins:Struct. Funct. Genet. 13:20622

25. Leach AR. 1994. Ligand docking to pro-

teins with discrete side-chain exibility. J. Mol. Biol. 243:3102626. Gulukota K, Vadja S, Delisi C. 1996.

Peptide docking using dynamic program-ming. J. Comput. Chem. 17:41828

27. DesJarlais RL, Sheridan RP, Dixon JS,Kuntz ID, Venkataraghaven R. 1986.Docking exible ligands to macromolec-ular receptorsby molecular shape. J. Med.Chem. 29:214953

28. Clark KP, Ajay. 1995. Flexible liganddocking without parameter adjustmentacross four ligand-receptor complexes. J.Comput. Chem. 16:121026

29. DesJarlais RL, Seibel GL, Kuntz ID,Furth PS, Alvarez JC, et al. 1990.Structure-based design of nonpeptide in-hibitors specic for the human immun-odeciency virus 1 protease. Proc. Natl. Acad. Sci. USA 87:664448

30. Ring CS, Sun E, McKerrow JH, LeeGK, Rosenthal PJ, et al. 1993. Structure-based inhibitor design by using proteinmodels for the development of antipara-sitic agents. Proc. Natl. Acad. Sci. USA90:358387

31. Schoichet BK, Stroud RM, Santi DV,

Kuntz ID, Perry KM. 1993. Structure-based discovery of inhibitors of thymidy-late synthase. Science 259:144550

32. Boobbyer DNA, Goodford PJ, McWhin-nnie PM,WadeRC. 1989. New hydrogen-bond potentials for use in determiningenergetically favorable binding sites onmolecules of known structure. J. Med.Chem. 32:108394

33. Wade RC, Clark KJ, Goodford PJ. 1993.Further development of hydrogen bondfunctions for use in determining energeti-cally favorable binding sites on moleculesof knownstructure. 1. ligandprobegroupswith the ability to form two hydrogenbonds. J. Med. Chem. 36:14047

34. Wade RC, Goodford PJ. 1993. Furtherdevelopment of hydrogen bond functionsfor use in determining energetically favor-able binding sites on molecules of knownstructure. 2. ligand probe groups with theability to form more than two hydrogenbonds. J. Med. Chem. 36:14856

35. Caish A, Miranker A, Karplus M. 1993.Multiple copy simultaneous search andconstruction of ligands in binding sites:application to inhibitors of HIV-1 aspar-

tic proteinase. J. Med. Chem. 36:214267


18/20


36. Eisen MB, Wiley DC, Karplus M, Hub-bard RE. 1994. HOOK: A program fornding novel molecular architectures thatsatisfy the chemical and steric require-ments of a macromolecule binding site.Proteins: Struct. Funct. Genet. 19:199

22137. Moon JB, Howe WJ. 1991. Computer de-sign of bioactive molecules: a methodfor receptor-based de novo ligand design.Proteins: Struct. Funct. Genet. 11:31428

38. Bohm H-J. 1992. The computer programLUDI: a new method for the de novo de-sign of enzyme inhibitors. J. Comput.- Aided Mol. Design 6:6178

39. Lauri G, Bartlett PA. 1994. CAVEAT: aprogram to facilitate thedesign of organicmolecules. J. Comput.-Aided Mol. Design8:5166

40. Gehlhaar DK, Moerder KE, Zichi D,Sherman CJ, Ogden RC, Freer ST. 1995.De novo design of enzymes inhibitors byMonte Carlo ligand generation. J. Med.Chem. 38:46672

41. Gillet V, Johnson AP, Mata P, SikeS, Williams P. 1993. SPROUT: a pro-gramfor structuregeneration. J. Comput.- Aided Mol. Design 7:12753

42. Lawrence MC, Davis PC. 1992. CLIX:a search algorithm for nding novelligands capable of binding proteins of known three-dimensional structure. Pro-

teins: Struct. Funct. Genet. 12:314143. Martin YC. 1992. 3-D database searchingin drug design. J. Med. Chem. 35:214554

44. Brooks BR, Bruccolari RE, Olafson BD,States DJ, Swaminathan S, Karplus M.1983. CHARMM: a program for macro-molecular energy minimizations and dy-namics calculations. J. Comput. Chem.4:187217

45. Weiner SJ, Kollman PA, Case DA, SinghUC, Ghio C, et al. 1984. A new force eldfor molecular mechanical simulation of nucleicacids. J. Am. Chem. Soc. 106:76584

46. Weiner SJ, Kollman PA, Nguyen DT,Case DA.1986. Anall atomforce eld forsimulation of proteins and nucleic acids. J. Comput. Chem. 7:23052

47. Jorgensen WL, Tirado-Rives J. 1988.The OPLS potential functions for pro-teins. Energy minimizations for crystalsof cyclic peptides and crambin. J. Am.Chem. Soc. 110:165766

48. Cornell WD, Cieplak P, Bayly CI, GouldIR, Merz KM Jr., et al. 1995. A secondgeneration force eld for the simulation

of proteins, nucleic acids, and organic

molecules. J. Am. Chem. Soc. 117:517997

49. van GunsterenWF, BerendsenHJC.1988.Groningen molecular simulation (GRO-MOS) library manual. BIOMOS B. V.,Nijenborgh 16, 9747 AG Groningen, The

Netherlands50. Mackerell AD, Wiorkiewicz-Kuczera J,Karplus M. 1995. An all-atom empiricalenergy function for the simulation of nu-cleic acids. J. Am. Chem. Soc. 117:1194675

51. Eisenberg D, McLachlan AD. 1986. Sol-vation energy in protein folding and bind-ing. Nature 319:199203

52. Wesson L, Eisenberg D. 1992. Atomicsolvation parameters applied to moleculardynamics of proteins in solution. ProteinScience 1:22735

53. Ooi T, Oobatake M, Nemethy G, Scher-agaHA.1987. Accessible surface areasasa measure of the thermodynamic param-eters of hydration of peptides. Proc. Natl. Acad. Sci. USA 84:308690

54. Vila J, Williams RL, Vasquez M, Scher-aga HA. 1991. Empirical solvation mod-els can be used to differentiate na-tive from near-native conformations of bovine pancreatic-trypsin inhibitors. Pro-teins: Struct. Funct. Genet. 10:199218

55. Stouten PFW, Frommel C, Nakamura H,Sander C. 1993. An effective solvationterm based on atomic occupancies for use

in protein simulations. Mol. Simul. 10:9712056. Wasserman ZR, Hodge CN. 1996. Fit-

ting an inhibitor into the active site of thermolysin: a molecular dynamics study.Proteins: Struct. Funct. Genet. 4:22737

57. Schiffer CA, Caldwell JW, Kollman PA,Stroud RM. 1993. Protein structure pre-diction with a combined solvation freeenergy-molecular mechanics force eld. Mol. Simul. 10:12149

58. von Freyburg B, Richmond TJ, Braun W.1993. Surface area included in energy re-nementof proteins.A comparative studyon atomic solvation parameters. J. Mol. Biol. 233:27592

59. Juffer A, Eisenhaber F, Hubbard SJ,Walther D, Argos P. 1995. Comparison of atomic solvation parametric sets: appli-cability and limitations in protein foldingand binding. Protein Science 4:2499509

60. Gilson MK, McCammon JA, Madura JD.1995. Molecular dynamics simulationswith a continuum electrostatic model of solvent. J. Comput. Chem. 16:108195

61. Ewing TJA, LybrandTP. 1994. A compar-isonof perturbationmethodsandPoisson-

Boltzmann electrostatics calculations for


19/20


estimation of relative solvation free ener-gies. J. Phys. Chem. 98:174852

62. Shen J, Quiocho FA. 1995. Calcu-lation of binding energy differencesfor receptor-ligand systems using thePoisson-Boltzmann method. J. Comput.

Chem. 16:4454863. Shen J, Wendoloski J. 1995. Bind-ing of phosphorus-containing inhibitorsto thermolysin studied by the Poisson-Boltzmann method. Protein Science4:37381

64. Shen J, Wendoloski J. 1996. Electro-static binding energy calculation usingthe nite difference solution to the lin-earized Poisson-Boltzmann equation: as-sessment of its accuracy. J. Comput.Chem. 17:35057

65. Antosiewicz J, McCammon JA, Wlodek S, Gilson MK. 1995. Simulation of ch-arge-mutant acetylcholinesterases. Bio-chemistry 34:421119

66. Antosiewicz J, McCammon JA, GilsonMK. 1994. Prediction of pH-dependentproperties of proteins. J. Mol. Biol.238:41536

67. Sitkoff D, Sharp KA, Honig B. 1994. Ac-curate calculation of hydration free ener-giesusing macroscopic solvent models. J.Phys. Chem. 98:197888

68. Smith KC, Honig B. 1994. Evaluationof the conformational free energies of loops in proteins. Proteins: Struct. Funct.

Genet. 18:1193269. Yang AS, Honig B. 1995. Free energydeterminants of secondary structure for-mations. 1. Alpha-helices. J. Mol. Biol.252:35165

70. Yang AS, Honig B. 1995. Free energy de-terminants of secondary structure forma-tions. 2. Antiparallel beta-sheets. J. Mol. Biol. 252:36676

71. Jackson R, Sternberg MJE. 1995. A con-tinuum model for protein-protein interac-tions: application to thedocking problem. J. Mol. Biol. 250:25875

72. Zacharias M, Luty BA, Davis ME,McCammon JA. 1994. Combined con-formational search and nite-differencePoisson-Boltzmann approach for exibledocking: application to an operator muta-tion in the repressor-operator complex. J. Mol. Biol. 238:45565

73. Marrone TJ,Gilson MK,McCammon JA.1996. Comparison of continuum and ex-plicit models of solvation-potentials of mean force for alanine dipeptide. J. Phys.Chem. 100:143941

74. Straatsma TP, McCammon JA. 1992.Computationalalchemy. Annu. Rev. Phys.Chem. 43:40735

75. Tembe BL, McCammon JA. 1984.Ligand-receptor interactions. Comput.Chem. 8:28183

76. WongCF, McCammonJA. 1986. Dynam-ics and design of enzymes and inhibitors. J. Am. Chem. Soc. 108:383032

77. Straatsma TP. 1996. Free energy bymolecular simulation. In Reviews in Com- putational Chemistry , ed. KB Lipkowitz,DB Boyd, 9:81127. New York: VCHPublishers

78. Rao BG, Kim EE, Murcko MA. 1996.Calculation of solvation and binding freeenergy differences between VX-478 andits analogs by free energy perturbationand amsol methods. J. Comput.-Aided Mol. Design 10:2330

79. Rami Reddy M, Varney MD, Kalish V,Viswanadham VN, Appelt K. 1994. Cal-

culationof relativedifferencesin thebind-ing free energies of HIV1 protease in-hibitors: a thermodynamic cycle pertur-bation approach. J. Med. Chem. 37:11451152

80. Rami Reddy M, Bacquet RJ, Zichi D,Matthews DA, Welsh KM, et al. 1992.Calculation of solvation and binding freeenergy differences for folate-based in-hibitors of the enzyme thymidylate syn-thase. J. Am. Chem. Soc. 114:1011722

81. Wlodek ST, Antosiewicz J, McCam-mon JA, Straatsma TP, Gilson MK,et al. 1996. Binding of tacrine and6-chlorotacrine by acetylcholinesterase. Biopolymers 38:10917

82. Marrone TJ, Straatsma TP, Briggs JM,Wilson DK, Quiocho FA, McCammonJA. 1996. Theoretical study of inhibi-tion of adenosine deaminase by 8R-coformycin and 8R-deoxycoformycin. J. Med. Chem. 39:27784

83. Damewood JR Jr. 1996. Peptide mimeticdesign with the aid of computationalchemistry. In Reviews in ComputationalChemistry , ed. KB Lipkowitz, DB Boyd,9:179. New York: VCH Publishers

84. Kollman PA. 1993. Free energy cal-culationsapplications to chemical andbiochemical phenomena. Chem. Rev.93:2395417

85. Cannon WR, Madura JD, Thummel RP,McCammon JA. 1993. Molecular recog-nition: effect of rotational isomers onhost-guest binding. J. Am. Chem. Soc.115:87984

86. McCammon JA. 1994. Free energy andbinding selectivity. In Computational Ap- proaches in Supramolecular Chemistry ,ed. G. Wipff, 15:51517. Dordrecht:

Kluwer Academic


20/20


87. McCammon JA, Harvey S. 1987. Dynam-ics of Proteins and Nucleic Acids. Cam-bridge: Cambridge Univ. Press

88. Karplus M, Petsko GA. 1990. Moleculardynamics simulations in biology. Nature347:63139

89. Frauenfelder H, Wolynes PG. 1994.Biomolecules: where the physics of com-plexity and simplicity meet. Physics To-day 47:5864

90. CumminsPL,Gready JE.1995. Theinu-ence of starting coordinates in free energysimulations of ligand binding to dihy-drofolate reductase. Mol. Simul. 15:15575

91. Jorgensen WL. 1991. Computational in-sights on intermolecular interactions andbinding in solution. Chemtracts Org.Chem. 4:91119

92. Chipot C, Maigret B, Pearlman DA, Koll-man PA. 1996. Molecular dynamics po-tential of mean force calculations: a studyof the toluene-ammonium -cation in-teractions. J. Am. Chem. Soc. 118:29983005

93. Luty BA, van Gunsteren WF. 1996. Elec-trostatic interactions using the particle-particle particle-mesh method with non-periodic long-range interactions. J. Phys.Chem. 100:2581

94. Halgren TA. 1995. Potential energy func-tions. Curr. Opin. Struct. Biol. 5:20510

95. Clark TW, Hanxleden RV, McCammonJA, Scott LR. 1994. Parallelization usingspatial decomposition for molecular dy-namics. Scalable High Perform. Comput.

Conf. , pp. 95102. IEEE Comput. Soc.96. Zacharias M, Straatsma TP, McCammon

JA. 1994. Separation-shifted scaling, anewscalingmethod for Lennard-Jones in-teractions in thermodynamic integration. J. Chem. Phys. 100:902531

97. Liu H, Mark AE, van Gunsteren WF.1996. Estimating the relative free energyof different molecular states with respectto a single reference state. J. Phys. Chem.100:948594

98. Gibas CJ, Subramaniam S. 1996. Explicitsolvent models in protein pKa calcula-tions. Biophys. J. 71:13847

99. Antosiewicz J, McCammon JA, GilsonMK. 1996. The determinants of pK a s inproteins. Biochemistry 35:781933

100. Resat H, McCammon JA. 1996. Free en-ergy simulations: correcting for electro-static cutoffs by use of the Poisson equa-tion. J. Chem. Phys. 104:7645-51

101. Resat H, Mezei M. 1994. Grand canon-ical Monte Carlo simulations of watersin crystal hydrates. J. Am. Chem. Soc.116:745152

102. Resat H, Mezei M. 1996. Grand canoni-cal ensemble Monte Carlo simulation of the dCpG/proavine crystal hydrate. Bio- phys. J. 71:117990

103. Cagin T, Pettitt BM. 1991. Moleculardynamics with a variable number of molecules. Mol. Phys. 72:16975

104. Wade RC, Bohr H, Wolynes PG. 1992.Prediction of water binding sites on pro-teins by neural networks. J. Am. Chem.Soc. 114:828485

SB Drug Design

Documents

Transcript of SB Drug Design