`
`Current Opinion in Drug Discovery and Development 1998 Vol 1 No 1
`
`Combinatorial and computational approaches in structure-based drug
`design
`Hugo Kubinyi
`
`Address
`Combinatorial Chemistry and Molecular Modelling
`ZHF/G - A30
`BASF AG
`D-67056 Ludwigshafen
`Germany
`Email: HUGO.KUBINYI@MSM.BASF-AG.DE
`
`Current Opinion in Drug Discovery and Development 1998 1(1):16-27
`© Current Drugs Ltd ISSN 1367-6733
`
`The increasing number of protein 3D structures and the success of
`structure-based approaches has led to the development of several
`experimental and theoretical techniques for the rational design of
`protein ligands. Combinatorial chemistry significantly speeds up
`the synthesis of potential new drug candidates. Diversity
`considerations, as well as the use of 3D structural information of
`the biological targets, reduce the size of huge libraries to a
`reasonable number of rationally-designed ligands. New NMR
`techniques (SAR by NMR) allow the construction of high-affinity
`ligands
`from small molecules with much
`lower affinities.
`Computer-aided drug design uses building, linking, and/or rigid
`docking procedures to search for ligands for a certain binding site.
`Scoring functions provide a rank order of the designed ligands
`according
`to
`their
`estimated binding affinities. Further
`developments in computer-aided drug design are automated
`approaches for the flexible alignment of molecules, the flexible
`docking of ligands to their binding sites, and the stepwise
`assembly of
`synthetically
`easily accessible
`ligands
`from
`combinatorial libraries of fragments.
`
`Introduction
`Structure-based ligand design has adopted a growing
`importance in pharmaceutical research, especially in the
`search for new drugs [1,2•,3•,4••,5••]. The application of
`these techniques is supported by an exponential increase in
`the number of experimental protein 3D structures [6••]. The
`design of new ligands is performed in several cycles, most
`often only by visual
`inspection
`and qualitative
`interpretation of
`the
`ligand-binding site
`interactions.
`Correspondingly, there is an urgent need for more rational
`techniques. Several experimental and theoretical approaches
`that have been developed to aid the design process will be
`reviewed
`in this article. Approaches of the greatest
`importance are
`the rational design of combinatorial
`libraries, the SAR by NMR method for the construction of
`high-affinity ligands, flexible ligand docking, and de novo
`drug design methods.
`
`Combinatorial techniques for structure-based
`ligand design
`Classical drug research depends on a combination of
`working hypotheses, synthesis, and testing of potential drug
`candidates, as well as good luck. Combinatorial chemistry
`and high-throughput screening have added a new
`dimension to the direction of random searching as opposed
`to rational design [7,8]. Such a view, however, is valid only
`at first sight. Combinatorial chemistry [9•,10••] began with
`
`the
`libraries of mixtures and
`the concept of huge
`deconvolution of biologically active mixtures to detect new
`leads. Nowadays, the automated parallel synthesis of
`specially designed and focused small libraries, made up
`from single compounds, is at the forefront of research.
`
`Rational design and validation of
`combinatorial libraries
`In addition to synthetic accessibility, diversity is the most
`important property of combinatorial
`libraries. Many
`different descriptor sets have been used to characterize the
`diversity of combinatorial libraries. There is an ongoing
`discussion of whether 2D or 3D descriptor sets are superior
`[11•,12•]. A logical explanation for the observed weakness
`of 3D descriptors might be that 2D descriptors have
`undergone much more extensive development. An
`additional issue is whether diversity considerations should
`be restricted to the scaffolds and the building blocks or
`should be applied to the resulting compounds of a library.
`Diversity profiling was applied to select diverse subsets
`from structural databases [11•,12•,13•]. HARPick (Rhône-
`Poulenc Rorer) is a program that selects reagents to build a
`library on product-based diversity calculations
`[14].
`Combinatorial libraries have also been designed using a
`genetic algorithm
`to optimize
`the distribution of
`physicochemical or any other properties of a library [15•].
`
`There is, however, no objective definition of diversity. If
`diversity is understood to be the lack of similarity, one has
`to be aware that compounds that are closely related
`chemically might show significantly diverse biological
`activities [16]. Books [17•,18•] and reviews [19,20a,20b] have
`been published on molecular diversity considerations in
`combinatorial chemistry, and can be referred to for further
`background information.
`
`An interesting approach to the determination of the 'drug-
`likeness' of series of organic molecules [21] has been
`pursued by two industrial groups [Ajay, Vertex Pharmaceu-
`ticals, personal communication; Sadowski J, BASF AG,
`personal communication]. Simple structural parameters and
`scoring values of 0 and 1 were used to train a neural net
`with sets of chemicals (eg, from the Available Chemicals
`Directory) and drugs (eg, from the Derwent World Drug
`Index). The discrimination of the relatively small training
`sets as well as the predictions for the rest of the huge
`databases are in the range of 75 to 80%. Surprisingly good
`results are even obtained if whole series of biologically
`active compounds (eg, all cardiovascular drugs or all
`hormones) are eliminated from the training sets. Whilst the
`'drug-likeness' assignment of a single compound may be
`incorrect, the method allows a reasonable ranking within
`large in-house, external, combinatorial, and virtual libraries.
`In this manner, financial resources are focused on sets of
`compounds of general biological interest.
`
`Argentum Pharm. v. Research Corp. Techs., IPR2016-00204
`RCT EX. 2051 - 1/12
`
`
`
`Combinatorial and computational approaches Kubinyi
`
`17
`
`• Combinatorial libraries for the SH3 domain of Src
`tyrosine kinase were designed at the Howard Hughes
`Medical Institute, Harvard University, USA, in cycles, by
`multidimensional NMR spectroscopy investigation of
`the few highest affinity ligands [35,36].
`• A potent, non-peptide GPIIb/IIIa receptor antagonist
`(collagen-induced platelet aggregation, IC50 = 92 nM)
`was developed at the Life Science Research Center,
`Nippon Steel Corporation, Japan, from combinatorial
`libraries based on the Arg-Gly-Asp sequence (the RGD
`motif) of the natural ligand, fibrinogen [37].
`• A selective a
`3 integrin receptor antagonist (IC50 = 1.1
`nM) was designed at DuPont Merck, USA, as a focused
`RGD peptidomimetic library, based on an amine or
`guanidine group to mimic the arginine side chain, a
`variable linking group, and b -alanine to mimic the
`aspartate of the RGD motif [38].
`
`vb
`
`Self-assembly of ligands
`In principle, one could imagine that an enzyme could be
`inhibited by two (or even more) small ligands, binding at
`different pockets of the protein. The laws of thermo-dynamics
`are, however, against
`this concept. Translational and
`rotational degrees of
`freedom are
`lost on binding.
`Correspondingly, the affinity of a ligand which connects two
`fragments in an optimal geometry, and which itself does not
`interfere with the binding, is much higher than the affinity of
`the two fragments as separate ligands.
`
`Episelection (Arris Pharmaceutical Corporation, USA) is a
`new strategy in structure-based ligand design. The reaction
`of various alcohols with a boronic acid trypsin inhibitor
`produces a series of esters. These are selected either by
`preferential binding to the protein (epitaxial selection) or
`assembled at the enzyme surface (epitaxial reaction) [39•].
`
`Huc and Lehn (Université Louis Pasteur, Strasbourg, France)
`formulated a general concept for the dynamic generation of
`virtual combinatorial libraries, in which molecular diversity is
`produced by self-assembly of protein ligands, eg, enzyme
`[40••]. This
`inhibitors,
`from appropriate components
`approach has been applied to the selective induction of
`carbonic anhydrase inhibitors by reversible combination of
`amines and aldehydes; the presence of the enzyme favors the
`formation of those analogs, which are expected to have high
`affinities to the protein.
`
`Another example of spontaneous self-assembly was recently
`observed. Physiological concentrations of zinc ions convert
`low-affinity, metal-chelating ligands into selective, high-
`affinity serine proteinase inhibitors [41••]. In the absence of
`zinc ions, bis(5-amidino-2-benzimidazolyl)methane (BABIM)
`inhibits human and bovine trypsin with a K i = 19 •M. The
`addition of 100 nM of Zn2+ increases the affinity for human
`trypsin to K i= 90 nM, and for bovine trypsin, by more than
`four orders of magnitude, to Ki = 5 nM. An even greater effect
`is observed for keto-BABIM, where the affinity to bovine
`trypsin increases by a factor of 19,000 to Ki < 1 nM. Further
`structural variation led to analogs with improved selectivities
`versus trypsin, tryptase, and thrombin (Figure 1) [41••].
`
`Structure-based design of combinatorial
`libraries
`The integration of structure-based design into combina-
`torial chemistry for new pharmaceutical discovery has been
`reviewed [4••,22] and critically commented upon [23••].
`There are many examples of the discovery of enzyme
`inhibitors and other protein ligands, without considering
`protein 3D structures, through combinatorial chemistry
`[24•,25]. Some recent examples of combinatorial libraries
`that were designed by using information from protein or
`ligand 3D structures are discussed below.
`
`• Structural variation of
`the P3 position of a
`peptidomimetic thrombin inhibitor was performed, at
`Merck Research Laboratories, USA, by rapid, multiple
`analog synthesis. Out of > 2,200 commercially and in-
`house available acid components, 200 were selected and
`coupled to resin-bound prolyl trans-4-aminocyclohexyl-
`methyl amide, resulting in the orally available, potent
`and selective thrombin inhibitor, L-372460 (Merck & Co;
`Ki thrombin = 1.5 nM, Ki trypsin = 860 nM) [26•]. Novel
`potent thrombin inhibitors were also discovered by
`solid-phase synthesis using different, nonbasic P1
`building blocks [27].
`• Bis-phenylamidine factor Xa inhibitors were designed, at
`DuPont Merck, USA, by docking and minimizing small
`fragments in the P1 and P4 binding sites; subsequently,
`these fragments were connected with a tether, resulting
`in a potent factor Xa inhibitor (Ki = 34 nM) [28•].
`• A library of potential inhibitors of the aspartyl proteinase,
`cathepsin D, was designed at the University of California,
`Berkeley, USA, using 3D
`structural
`information.
`Approximately 6 to 7% of the analogs were active at 1 m M
`concentrations, the most potent analog having a Ki of 73
`nM. A second-generation library resulted in the rapid
`identification of further potent nonpeptide inhibitors (Ki =
`9-15 nM) [29].
`• The design of matrix metalloproteinase inhibitors, at
`DuPont Merck, USA, led to combinatorial libraries from
`which a specific, low molecular weight, MMP-8 inhibitor
`(MMP-3, Ki = 148 nM; MMP-8, Ki = 1.9 nM) resulted; an
`unexpected alternative binding mode was observed.
`Minor structural modification led to a high-affinity
`MMP-3 inhibitor (Ki = 9 nM) [30].
`• A structure-based library design of kinase inhibitors, at
`Howard Hughes Medical
`Institute, University of
`California, Berkeley, USA, produced a 10-fold increase in
`the
`inhibitory potency of
`the natural product,
`olomoucine [31].
`• A library of 4-amino-4H-pyran-6-carbonamides, struct-
`urally related to the anti-influenza drug, zanamivir
`(Monash University, Biota/Glaxo Wellcome), was
`prepared at Glaxo Wellcome, UK, from a 4-amino-
`Neu5Ac-2en-derived carboxylic acid and 80 primary and
`secondary amines; several aliphatic N-dialkylamides and
`N-phenethyl-N-alkylamides proved to be nanomolar
`inhibitors of influenza A virus neuraminidase [32,33].
`• A targeted library of phosphatase inhibitors was derived
`at the University of Pittsburgh, USA, from a rational
`backbone design and random side chain variation [34].
`
`Argentum Pharm. v. Research Corp. Techs., IPR2016-00204
`RCT EX. 2051 - 2/12
`
`
`
`18
`
`Current Opinion in Drug Discovery and Development 1998 Vol 1 No 1
`
`Figure 1.
`
`N
`
`N
`
`NH
`
`1c
`
`N
`
`NH2
`
`NH
`
`N
`
`NH
`
`1b
`
`NH2
`
`NH
`
`O
`
`N
`
`NH
`
`N
`
`NH
`
`NH2
`
`NH
`
`1a
`
`Ki (trypsin) m M
`Ki (tryptase) m M
`Ki (thrombin) m M
`
`1a
`no Zn2+
`87.5
`5.7
`> 1,000
`
`plus Zn2+
`0.005
`0.05
`0.1
`
`NH2
`
`NH
`
`1b
`no Zn2+
`> 1,000
`358
`> 1,000
`
`plus Zn2+
`136
`0.3
`10.5
`
`1c
`no Zn2+
`31.2
`8.8
`31
`
`plus Zn2+
`22.5
`54.5
`0.04
`
`O
`
`NH
`
`NH
`
`N
`
`N
`
`2+
`
`Zn
`
`His-57
`
`N
`
`NH
`
`O
`
`H
`
`Ser-195
`
`NH2
`
`NH
`
`Asp-189
`
`2
`
`In the presence of zinc ions, the BABIM analog, 1a, becomes a fairly selective trypsin inhibitor, 1b, a selective tryptase inhibitor, and the
`analog, 1c, a highly selective thrombin inhibitor (all Ki values refer to human enzymes). The lower part of the diagram shows the
`experimental binding mode of the Zn2+-keto-BABIM complex to bovine trypsin, as determined by protein crystallography. The zinc ion
`coordinates to the benzimidazole nitrogen atoms of keto-BABIM, 2, the His-57 nitrogen atom and the Ser-195 oxygen atom [41••].
`
`These results correspond
`the activation of GDP
`to
`complexes of various G-proteins
`in the presence of
`aluminum and fluoride ions, which otherwise only takes
`place in the presence of GTP. Protein crystallography
`confirmed the hypothesis on the mode of action of this
`- ion mimics the
`serendipitous discovery, in which the AlF4
`outer phosphate group of GTP [42-44].
`
`Although the principle of self-assembly of inhibitors in the
`binding site looks attractive, it is probably too early to
`decide whether general principles for drug design may
`result from such single observations.
`
`Experimental methods for combinatorial drug
`design: SAR by NMR
`Ligand design based on the combination of fragments which
`bind to proximal subsites of a certain protein has already
`been realized. Stephen Fesik (Abbott Laboratories) has
`developed an elegant approach for this purpose, ie, the SAR
`by NMR (structure-activity relationships by nuclear magnetic
`resonance spectroscopy) method [45••,46•]. In this important
`new experimental technique for structure-based drug design,
`
`libraries of typically a thousand small molecules are screened
`against a certain protein. The binding of ligands to a subsite is
`observed by shifts of the corresponding amide proton signals
`of the 15N-labeled protein. In the next step, the protein is
`saturated with the highest affinity ligand for this site and a
`different library is screened for ligands which bind to another,
`proximal subsite. If this second step is also successful, both
`ligands are combined with an appropriate tether. In this
`manner, high-affinity ligands can be constructed within a
`short time. The first successful application of the SAR by NMR
`method was the construction of a high-affinity FK-506 binding
`protein (FKBP) ligand (Kd = 19 nM), by combining two small
`molecules (Kd = 2 and 100 m M, respectively) with a linker
`[45••]. Other applications included the discovery of potent
`nonpeptide
`inhibitors of
`the matrix metalloproteinase,
`stromelysin (Figure 2) [47•,48•], and of inhibitors which block
`the DNA binding of a certain Papillomavirus protein [49].
`
`Despite the elegance of this approach, SAR by NMR has
`several limitations:
`
`• The molecular weight of the protein must be < 35 to 40 kD.
`
`Argentum Pharm. v. Research Corp. Techs., IPR2016-00204
`RCT EX. 2051 - 3/12
`
`
`
`Combinatorial and computational approaches Kubinyi
`
`19
`
`protein and the solvent molecule complexes are determined
`by protein crystallography in order to detect specific
`binding sites. Although such measurements take only a few
`days, there is no clear evidence available to suggest that this
`approach could be as widely applicable as the SAR by NMR
`method. In addition, no inhibitor has yet resulted from the
`application of
`this
`technique without
`independent
`information from other sources.
`
`Computer-aided ligand design
`Whereas structure-based design can be regarded as the
`predominant strategy of the last decade [1,2•,3•,4••,5••],
`several computer-assisted methods have been developed
`more recently. If several thousands of candidates, from large
`structural databases, are to be tested for their suitability as
`ligands of a certain binding site, molecular modeling [56••]
`can no longer be performed manually. The design process
`needs to be automated. The methods of choice for this
`purpose are
`computer programs
`that
`superimpose
`molecules by a flexible alignment to derive pharmacophoric
`patterns and/or quantitative structure-activity relationships,
`dock molecules to the surface of a protein 3D structure or to
`a hypothetical pseudoreceptor, or construct new ligands
`within a predefined binding site [57•,58••].
`
`Automated flexible superposition of
`molecules
`Methods for the alignment of rigid molecules are well
`established. A simple strategy to perform the alignment of
`flexible molecules involves the generation of multiple
`conformations of each compound by a knowledge-based
`approach (using torsion angle libraries from small-molecule
`crystal structures), to rank them by an energy function, and
`to superimpose all of the different pairs of low-energy
`conformations [59]. Different molecular property fields,
`such as electrostatic, steric, hydrophobic, hydrogen bond
`acceptor and donor fields, as well as their weighted
`combinations, have been used to achieve a fully automated
`alignment of the molecules. MIMIC (Pharmacia & Upjohn,
`USA) is a program that matches steric and electrostatic
`fields to guide the superposition; in a preprocessing step,
`similar conformations of a molecule are clustered [60].
`MIMIC has also been extended to multimolecule alignments
`[61]. Another approach for the consideration of ligand
`flexibility starts from conformationally rigid ligands using
`different template conformations for the superposition of
`the molecules [62].
`
`The much more demanding flexible superposition of one
`molecule onto another has been achieved only recently. The
`GASP program (University of Sheffield, UK) uses a genetic
`algorithm [63•] to consider conformational flexibility in the
`optimization of the alignment of a set of molecules [64]. A
`recent development for time-efficient flexible superposition
`of pairs of molecules is the computer program, FlexS
`(German National Research Centre
`for
`Information
`Technology (GMD), Germany), which resulted from a
`modification of the docking program, FlexX (see the section
`on docking in this review). A test ligand is superimposed
`onto a rigid template molecule (which is considered to be in
`its receptor-bound conformation, eg, as determined by
`
`Figure 2.
`
` 3
`Kd = 17,000 m M
`
` 4
`Kd = 20 m M
`
`N
`
`C
`
`N
`
`OH
`
`O
`
`NH
`
`OH
`
`2+
`
`Zn
`
`O
`
`O
`
`NH
`
`OH
`
`2+
`
`Zn
`
` 5
`Kd = 0.015 m M
`
`SAR by NMR identifies ligands that bind to proximal subsites of a
`protein. Acetohydroxamic acid, 3, and 3-(cyanomethyl)-4'-
`hydroxybiphenyl, 4, are
`low-affinity
`ligands of
`the matrix
`metalloproteinase, stromelysin. Combining them with an appropriate
`linker produces the high-affinity inhibitor, 5 [47•].
`
`• Large amounts (> 200 mg) of pure 15N-labeled protein
`are required.
`• Sufficient aqueous solubility (~
` 2 mM) and stability of
`the protein, also in the absence of an inhibitor (which is
`sometimes a problem, especially for proteinases), are
`preconditions for the NMR measurements.
`• The ligands must have sufficient aqueous solubility and
`stability.
`• The assignment of the -NH- signals can take weeks or
`even months (the 3D structures of the proteins need not
`be known).
`• Ligands for different subsites must be discovered.
`• The second subsite should be closely adjacent to the first
`subsite in order to avoid linkers which are too large.
`• A linker which connects the two low-affinity ligands in a
`relaxed conformation must be designed.
`• The linker itself must not have any negative influence on
`binding affinity.
`
`Alternatives to the SAR by NMR method for large proteins
`are 1D NMR methods that exploit the changes in relaxation
`or diffusion rates of small molecules upon binding to
`unlabeled proteins [46•,50-52]. Different organic solvents
`have been used to identify specific ligand binding sites on
`protein surfaces by observing the transfer NOEs to the
`protein [53].
`
`Another alternative to the SAR by NMR method is the
`multiple solvent crystal structures
`(MSCS; Brandeis
`University, MA, USA) approach [54,55•]. Protein crystals are
`soaked with different solvents, eg, acetonitrile, ethanol,
`hexenediol, isopropanol, dimethylformamide and acetone.
`Differences in electron densities between the unliganded
`
`Argentum Pharm. v. Research Corp. Techs., IPR2016-00204
`RCT EX. 2051 - 4/12
`
`
`
`20
`
`Current Opinion in Drug Discovery and Development 1998 Vol 1 No 1
`
`protein crystallography) by dissecting the test molecule into
`rigid fragments, selecting a base fragment to start the
`alignment, and re-assembling the molecule in a low-energy
`conformation which fits the template molecule [65••]. The
`alignment
`is
`speeded up by
`first
`searching
`for
`correspondences of intermolecular interaction centers. A
`further acceleration comes from the transformation of the
`Gaussian property functions into Fourier space [66]. FlexS
`gives reasonable alignments of highly flexible molecules
`within a few minutes [65••], ie, at least one order of
`magnitude faster than most other automated programs for
`flexible alignment.
`
`Following the concepts of a genetic algorithm alignment
`program [63•] and some strategies of FlexX, the program
`GOLD
`(Genetic Optimization
`for Ligand Docking;
`Cambridge Crystallographic Data Centre, UK) was
`developed. For 100 ligand-protein complexes, extracted
`from the Brookhaven Protein DataBank, GOLD achieved a
`71% success rate in identifying the experimental binding
`mode [78•]. DOCK was also extended to a program that
`explores ligand flexibility by selecting an anchor fragment
`of a ligand, positioning it in the binding site, and adding the
`other parts of the molecule to generate the ligand in a low-
`energy conformation that fits the binding site [79•].
`
`Docking
`Several computer programs for molecular docking have
`been described within the last years [63•,67•,68••]. The first
`computer-assisted approach to the discovery of ligands for a
`given binding site was the program DOCK (UCSF, CA,
`USA) [69•]. In its original version, DOCK searched in 3D
`databases for ligands that would fit into a binding cavity
`based merely on the geometric properties of a certain rigid
`conformation. Later,
`the
`complementarity of other
`properties was considered. DOCK frequently permits the
`discovery of micromolar ligands that can serve as lead
`structures for further development. The latest refinement to
`DOCK was a significant speed-up of the program [69•].
`Molecular docking to ensembles of 3D structures of the
`same protein allows an indirect consideration of target
`flexibility [70]. In a recent application, selective micromolar
`inhibitors of Pneumocystis carinii dihydrofolate reductase
`were derived from a DOCK database search, including >
`50,000 molecules from the Fine Chemicals Directory (now
`Available Chemicals Directory, MDL, CA, USA) [71].
`
`The computer program GRID (University of Oxford, UK)
`calculates
`interaction energies between proteins and
`different probes that are positioned around the surface of
`the protein [72•]. Porphyrins were superimposed by using a
`new option of the program that takes into account the
`flexibility of the propionic acid side chains. Each of the
`investigated analogs could be correctly placed into the heme
`binding site of myoglobin [73].
`
`FlexX (GMD, Germany) is an efficient and fast docking
`program [74••,75••] that starts by dissecting the ligands into
`rigid fragments. One or several base fragments are selected,
`either manually [76] or automatically [77], and positioned in
`favorable orientations within the binding site. Other
`fragments are added in the next steps, using a tree-search
`technique for placing the ligand incrementally into the
`binding site (Figure 3). Only low energy conformations are
`created, and the different results are ranked according to
`favorable interaction energies using the scoring function of
`the de novo design program LUDI (see the section on de novo
`ligand design). The program FlexX has been validated by
`the successful reproduction of the experimental binding
`modes of 19 ligand-protein complexes [74••,77]. Further
`extensions will include the combinatorial design of ligands
`from series of building blocks [Lengauer T, GMD, personal
`communication].
`
`Different search algorithms for molecular docking have
`been compared; the results indicate that several different
`approaches are effective and give satisfactory performance
`[80a,80b]. An interesting endeavor was discussed during the
`docking session of the Second Meeting on the Critical
`Assessment of Techniques for Protein Structure Prediction in
`Asilomar, California, in December 1996. A total of 77
`predictions were made by nine groups for the docking of
`seven small molecules into their binding sites. Overall
`results were good, with at least one prediction for each
`target within 3 Å root-mean-square deviation (RMSD), and
`within 2Å RMSD for over half the targets [81••]. Four
`groups were invited to describe their experiences in the
`competition, in separate publications [82-85].
`
`De novo ligand design
`De novo design methods have been extensively reviewed
`[86••,87••,88•,89••,90•]. The first de novo design program
`GROW (Upjohn Laboratories, USA) [89••] started from a
`simple seed fragment, eg, an amide group that is capable of
`interacting with the binding site, and continued by adding
`different amino acids in different conformations, to this
`fragment. Only the best candidates were selected and the
`same procedure was repeated several times until a peptide
`of a certain size had been generated in the binding site.
`
`The de novo design program LUDI (BASF, Germany and
`MSI, CA, USA) [87••] constituted a significant improvement
`and nowadays, it is the most widely distributed software for
`computer-aided ligand design. After the definition of a
`binding site region by the user, the program automatically
`identifies all of the hydrogen bond donor and acceptor sites,
`as well as the aliphatic and aromatic hydrophobic areas, of
`this region of the protein surface. From the program-
`implemented information on the geometry of interaction of
`such groups with a ligand, the program creates vectors and
`regions in space where complementary groups of a ligand
`should be located. In the next step, LUDI searches databases
`of 3D structures of small and medium-sized molecules for
`potential ligands. Each candidate is tested in all possible
`different orientations and interaction modes. After a rough
`evaluation by counting the number of interactions and by
`checking for unfavorable van-der-Waals overlap between
`the ligand and the protein, the remaining candidates are
`prioritized by a simple but efficient scoring function which
`estimates interaction energies on the basis of charged and
`neutral hydrogen bonding energies, hydrophobic contact areas,
`
`Argentum Pharm. v. Research Corp. Techs., IPR2016-00204
`RCT EX. 2051 - 5/12
`
`
`
`Figure 3.
`
`Combinatorial and computational approaches Kubinyi
`
`21
`
`Base fragment
`
`NH
`
`NH2
`
`O
`
`N
`
`NH
`
`O
`
`NH
`
`S
`
`O O
`
`Tree-search technique:
`
`Selection of base fragment(s)
`
`Alternative positions of base fragment
`
`Next fragments
`
`Next fragments
`
`The program FlexX dissects a ligand into rigid fragments. One or several base fragment(s) are selected manually or automatically and
`favorable binding geometries are generated for these fragments. After ranking by a scoring function, the best ones are kept and the next
`fragments are added, using a tree-search technique for incrementally placing the ligand to the binding site; open circles indicate unfavorable
`solutions that are not considered for further ligand building [74••,75••].
`
`and the number of rotatable bonds of the ligand. In the final
`step, the program is capable of attaching groups, fragments
`and/or rings to a hit or to an existing lead structure [87••].
`
`LUDI has been further developed for the automatic
`combinatorial design of synthetically accessible protein
`ligands, such as amides, peptides, and peptidomimetics
`[91,92]. An
`interesting realization of
`the concept of
`combinatorial docking
`is
`the MCSS
`(multiple copy
`simultaneous search; Harvard University, MA, USA)
`method [93•,94]. This approach searches for energy minima
`of ligand-protein interactions, ie, for preferred locations of
`specific functional groups or small ligands in the binding
`site. The corresponding positions are analyzed and selected
`ligand orientations are connected with alkane linkers to
`build molecules whose structures are optimized within the
`binding site. Recently, the MCSS method has been applied
`to the design of ligands binding to a new class of
`Picornavirus coat proteins [95]. A related computer program
`searches for binding sites by coating the protein surface with
`molecular fragments that could potentially interact with the
`protein; high affinity clusters are used as computational
`binding pockets for docking [96]. The method was validated
`by successfully docking a number of ligands to their protein
`binding sites.
`
`Dedicated computer programs for the structure-based
`design of ligands, by combinatorial docking from series of
`building blocks, are being developed within several
`companies and research institutes, such as Hoffmann-La
`Roche [Böhm HJ, personal communication], Agouron
`Pharmaceuticals [Virtual SAR by NMR; Rose PW, Cuty BA,
`Marrone TJ, personal communication] and GMD [Lengauer
`T, personal communication].
`
`In the meantime, more than 20 different programs for the
`computer-assisted construction of
`ligands have been
`developed and are used
`in de novo drug design
`[86••,87••,88•,89••]. Most of them follow, more or less, the
`concepts of DOCK, GROW and LUDI. Although it is
`difficult to draw firm conclusions on the specific merits of
`different de novo design programs, for practical applications
`in medicinal chemistry, a computer-assisted approach
`should include the following functionalities:
`
`• Searches in large 3D databases for potential ligands.
`• The consideration of conformational flexibility, at least
`of the ligand.
`• The option to create new ligands or to modify existing
`leads by fusion of groups, fragments and rings.
`• A scoring function which is appropriate to evaluate and
`sort the hits.
`
`Argentum Pharm. v. Research Corp. Techs., IPR2016-00204
`RCT EX. 2051 - 6/12
`
`
`
`22
`
`Current Opinion in Drug Discovery and Development 1998 Vol 1 No 1
`
`Due to its combinatorial complexity, the flexible treatment of
`whole ligand-protein complexes still remains unsolved. A
`more serious implication for successful de novo design is our
`lack of knowledge on the energetics of ligand-protein
`[97••]. Thermodynamic data of complex
`interaction
`formation are urgently needed for a better understanding of
`ligand binding. Microcalorimetric measurements seem to
`offer the best chance in this respect. Whilst hydrophobic
`interactions always contribute to binding affinity, the
`influence of hydrogen bonds on the ligand affinity depends
`on
`the balance of solvation-desolvation energies.
`In
`addition, hydrogen bonds have significantly different
`strengths, as can be seen
`from an
`inspection of
`intermolecular crystal contacts [98•]. Unfortunately, such
`statistics of nonbonded contacts are not representative of an
`aqueous environment. Water molecules do not only have a
`significant
`influence on
`the affinity contribution of
`hydrogen bonds, they also have to be considered as possible
`ligands between the functional groups of the binding site
`and the active molecule [4••,99,100]. In addition to the
`empirical scoring function implemented in LUDI, which is
`also used in some other programs, several alternative
`procedures for the estimation of ligand affinities have been
`developed [101••,102•,103-111].
`
`Conclusions
`Can ligands be rationally designed [112•]? Yes, they can.
`Structure-based drug design is supported by numerous
`experimental and theoretical approaches. Several methods
`have been developed, such as SAR by NMR, LUDI and the
`MCSS approach, that use combinatorial principles to
`construct new
`ligands. Further developments
`in this
`direction are to be expected. Of greatest value are
`computational approaches which consider, in addition to
`affinity, the synthetic accessibility of a new
`ligand.
`Compared to various experimental techniques, including
`combinatorial chemistry, the correct ranking of the results
`obtained seems to be the largest unsolved problem of
`computer-aided design
`techniques. Experimental and
`theoretical approaches complement each other, especially in
`the e