Abstract Ligand-protein inverse docking has recently been
introduced as a computer method for identification of potential protein targets of a drug. A protein structure database is searched to find proteins to which a drug can bind or weakly bind. Examples of potential applications of this method in facilitating drug discovery include: (1) identification of unknown and secondary therapeutic targets of a drug, (2) prediction of potential toxicity and side effect of an investigative drug, and (3) probing molecular mechanism of bioactive herbal compounds such as those extracted from plants used in traditional medicines. This method and recent results on its applications in solving various drug discovery problems are reviewed.
Index Terms—Drug binding, computer-aided drug design,
ligand-protein interactions, molecular modeling, adverse drug reactions, therapeutic effects, medicinal plant drug, natural product.
I. INTRODUCTION
nowledge of protein targets of drugs is useful in solving various problems in drug discovery process 1,2. For instance, therapeutic effects of a drug generally result from its interaction with one or more proteins or nucleic acids critical in a disease process 1,2. These proteins and nucleic acids are called therapeutic targets. The therapeutic targets of a number of clinical drugs are still unknown. Some drugs, such as aspirin, are known to have multiple therapeutic effects. For many of these drugs, not all of the targets associated with the respective multiple effects are determined. Lack of a complete knowledge about these targets hinders the effort to fully explore the novel mechanism of these therapeutic agents in the design of new drugs.
The adverse reaction of a drug with human body is often induced by its interaction with some of the proteins critically important in the normal processes 3. Identification of the interaction with these proteins, known as toxicity and side effect targets, has potential application in facilitating the prediction of side effect and toxicity of an investigative drug in early stage of drug development. A substantial portion of the $350 million and 12 years spent on average for commercial drugs have been squandered on many drug candidates that failed to ever reach the market 4, 5. Given a low success rate of drug candidates, detection of potential toxicity and side effect in early stages of drug development can potentially save money and time by focusing resources on drug leads and candidates likely to be safe to patients.
Manuscript received March 31,2002 (e-mail: yzchen@cz3.nus.edu.sg).
Drug target identification also finds potential application in facilitating the study of molecular mechanism and pharmacology of medicinal plants. Herbal medicines have gained increasing popularity 6 and they have been extensively used in traditional medicines 7. While herbal medicines are commonly used and the chemical constituents of a large number of medicinal herbs have been extracted and analyzed 8, the molecular mechanism as well as the basic and clinical pharmacology is known for only a relatively small number of medicinal herbs 9. Lack of knowledge about mechanism and pharmacology is one of the main obstacles for more extensive clinical application of medicinal herbs particularly those from traditional medicines. It also hinders efforts for new drug discovery based on novel therapeutic approaches of herbal medicines. Given the accumulated knowledge about their chemical constituents, the mechanism and pharmacology of medicinal herbs may be derived from investigation of the protein targets of these constituents and their collective actions.
Information about protein targets of drugs have traditionally been generated experimentally. Thus far, only one computer method, ligand-protein inverse docking, has been introduced 10, 11. This inverse docking strategy requires a sufficient number of proteins of known 3D structure. At present there are 13,976 protein entries in Protein Data Bank (PDB) and the number increases at a rate of well over 100 per month 12. About 17% of these have unique sequence 13. Introduction of high-throughput methods is expected to enable structural determination of 10,000 proteins with unique sequence within 5 years 14. Thus the number of proteins is approaching a meaningful level to cover a diverse set of potential targets. Ligand-protein inverse docking involves a computer-automated search of a protein database to identify potential proteins into which a small molecule can be docked 10. The software, INVDOCK, has been developed for such a purpose 10. INVDOCK uses a flexible docking algorithm to attempt to dock a molecule to each cavity of each of proteins in a database. A method for automated identification of binding cavities and the development of cavity models and related database have also been developed along with INVDOCK 10. Several studies have been conducted for testing the performance of the INVDOCK in identification of therapeutic and toxicity targets of clinical drugs and natural products from medicinal plants 10, 11, 15. These developments are reviewed in this paper.
Computational Method for Drug Target Search
and Application in Drug Discovery
Yuzong Chen , Zerong Li , and C. Y. Ung
II. COMPUTATIONALMETHOD A. Ligand-protein inverse docking procedure
All small molecule drugs appear to bind to a cavity of a protein or nucleic acid target. Thus, to facilitate computer-automated inverse docking, a protein cavity database has been developed along with INVDOCK. This database is composed of models of individual cavities in each protein entry of PDB. The model of a cavity is a cluster of overlapping spheres that fill-up that cavity 10.
INVDOCK is designed for automated search of every entry of this protein cavity database to find potential protein targets of a small molecule 10. Because of large number of cavity entries, cross-the-board specification of possible binding sites within each cavity is difficult. Hence all parts of each cavity are subject to docking. Docking to sites inside each cavity starts from known ligand binding sites, followed by more interior sections and then remaining part. To save CPU time, the program proceeds to the next protein entry when first successful dock is obtained without exhaustive search for optimum binding mode within each cavity. Although the optimum binding-mode is not specifically sort, the prioritized search strategy seems to dock a ligand to a site reasonably close to the experimentally observed positions in most of the ligand-protein structures studied. All cavity entries are subject to search unless the related protein has been identified as a putative target.
A molecule is flexibly docked into a cavity by an approach similar to the multi-step strategy approach proposed by Wang et al 16. Ligand conformation is sampled at a resolution similar to that of Wang et al. Docking of a particular conformer to a cavity is by the following steps: First the ligand is aligned within the selected site by matching the position of each ligand atom with the center of spheres. Because of the relatively low-resolution nature of ligand conformation sampling, certain degree of structural clash is allowed at this stage. A molecular-mechanics conformation optimization is then conducted by a limited torsion space sampling of rotatable bonds in the ligand and those in the side-chain of the receptor amino acid residues at the binding site. Each rotatable bond is sampled in the range of ±15o. This is followed by 50 iterations of Cartesian coordinate energy minimization on all ligand and protein atoms at the binding site so as to further optimize the ligand-protein complex. Energy minimization is by a steepest decent method.
In both torsion optimization and energy minimization, AMBER force fields 17 are used for covalent bond, bond angle, torsion, and non-bonded Van der Waals and electrostatic interactions. A large number of crystal structures in PDB contain only non-hydrogen atoms. To save computing time and to avoid the difficulty in modeling hydrogens, Morse potential 18, which is a function of donor-acceptor distance, is used to represent hydrogen bond terms. This potential has been shown to give reasonable description of hydrogen bond energy and dynamics in biomolecules 19, 20. The energy function is:
V = 1/2 Sbonds Kr (R - Req)2 + 1/2 Sangles Kq (q - qeq)2 +
1/2 Storsions Vn [ 1 - cos(n(f-feq)) ] + SH bonds [ V0 (1-e-a(r-r0) )2 - V0 ] + Snon bonded [ Aij/rij12 - Bij/rij6 + qiqj /er rij]
In this function, R, q, and f denotes bond length, angle and torsion angle respectively; Req , qeq and feq are taken as equilibrium bond length, angle and torsion angle respectively and their values are from the original PDB structure and the structure of the drug respectively; Kr and Kq are covalent and bond angle bending force constant respectively; Vn and n are torsion parameters; r is hydrogen bond donor-acceptor distance, and V0, a and r0 are hydrogen bond potential parameters.
B. Scoring
Scoring of docked molecules is based on a ligand-protein interaction energy function DELP composed of the same hydrogen bond and non-bonded terms as those used for structure optimization 10. Analysis of a large number of PDB ligand-protein complexes shows that the computed DELP is generally below DEThreshold = -aN kcal/mol, where N is the number of ligand atoms and a is a constant ~1.0 10. The exact value of a can be determined by fitting the linear equation DEThreshold = -aN to the computed DELP for a large set of PDB structures. This statistically derived energy value can be used empirically as a threshold for screening likely binders. A polynomial form of DEThreshold involving more parameters may also be introduced to derive an energy threshold. DELP can be required to be lower than DEThreshold when selecting successfully docked structures.
Ligand binding is competitive in nature. A drug is less likely to be effective if it binds to its receptor non-competitively against natural ligands and, to some extent, other drugs that bind to the same receptor site. This binding competitiveness may be partially taken into consideration for cavities known as ligand-bound in at least one PDB entry. Ligands in PDB structures are known binders. Therefore PDB ligands bound to the same receptor site as that of a docked molecule may thus be considered as "competitors" of that molecule. In INVDOCK selection of a putative protein target, the computed DELP is not only evaluated against DEThreshold but also compared to the ligand-protein interaction energy of the corresponding PDB ligands that bind to in the same cavity in this or other relevant PDB entries. The Ligand-protein interaction energy for the relevant PDB structures is computed by the same energy functions as that for the docked molecule. In addition to the condition that it be lower than DEthreshold, DELP of a docked molecule is required to be lower than a "competitor" energy threshold DECompetitor when selecting a putative target. DECompetitor can be taken as highest ligand-protein interaction energy of the corresponding PDB ligands multiplied by a factor b. In order to be able to find weak binders as well as strong binders, a factor b£1 is introduced to scale the ligand-protein interaction energy of PDB ligands. This is because a weak binder may have slightly higher interaction energy than that of a PDB binder. No experimental data has been found to determine the value of b. Hence b has been tentatively determined by an analysis of the computed
energy for a number of compounds. A value of 0.8 has been suggested for b leads to reasonable results statistically 10.
III. TESTINGOF INVDOCKDOCKINGALGORITHM The docking algorithm in INVDOCK has been tested on nine ligand-protein complexes from the PDB 10. Most of these structures have been used in testing different docking programs 16, 21, 22. The rest are complexes of well-known therapeutic agents. Therefore selection of these structures is useful in evaluating INVDOCK. Table 1 gives the computed root mean square deviation (RMSD) of each of the docked ligand with respect to the corresponding ligand in the original PDB structure. Six of the ligands are docked into the binding site with a RMSD in the range of 0.80 to 2.05Å, which is comparable to those obtained in other docking studies 16, 21, 22.
Although docked to the same position as that in the corresponding crystal structure, the RMSD of the three of the tested ligands is found to be substantially larger than obtained by other studies. Each of these docked molecules is found to largely overlap with the crystal ligand. However they are either flipped or with one end oriented in a different direction with respect to that in the crystal structure. The reason for such a deviation in the orientation of docked molecule is because INVDOCK does not attempt an exhaustive search for optimum binding mode in a cavity. As a result, when docked into a location with a similar steric contact as that of a PDB ligand, the ligand-protein interaction energy of a molecule may well pass the scoring threshold. This problem may be resolved by the introduction of the search of optimum binding mode and refinement of scoring function in the INVDOCK algorithm. Non-the-less, our results seem to suggest that INVDOCK algorithm is capable of docking a molecule to a site reasonably close to the original binding location.
Docking quality has been shown to improve by an extended INVDOCK algorithm that includes the search of optimum binding mode. As shown in Table I, the RMSD of the three ligands, with a larger RMSD in the original INVDOCK run, is significantly improved from the range of 3.56 - 6.55 Å in the original INVDOCK computation to that of 0.97 - 2.41 Å in the new computation. The same computation on the other six ligands shows no significant change between RMSD of optimum binding mode and that from the original INVDOCK computation. However, from Table I, the ligand-protein interaction energy of all the ligands in the optimum binding mode is substantially decreased with respect to that derived from original INVDOCK computation. In most cases, this energy is closer to that in the original PDB structure. This seems to indicate that INVDOCK optimization and scoring scheme is capable of guiding the system to a configuration fairly close to the native state.
IV. INVDOCK PREDICTIONOF DRUGTARGETS The effectiveness of INVDOCK prediction of drug targets can be demonstrated from a recent study on a few clinical agents 10. In this paper the results of one drug, tomoxifen, is presented. Tamoxifen is an anticancer drug widely used for treatment of breast cancer 23 and it has been approved as the first cancer preventive drug. Tamoxifen metabolite 4H-tamoxifen is believed to be the major contributor to the anti-oestrogenic effects of tamoxifen inside human body 23. Potential human and mammalian protein targets of 4H-tamoxifen identified by INVDOCK are given in Table 2 along with the respective clinical implications from experiments. A number of known protein targets of tamoxifen are found in the Table. These include estrogen receptor23, protein kinase C 24, collagenase 25, 17b hydroxysteroid dehydrogenase 26, Alcohol dehydrogenase 27, and prostaglandin synthetase 28. It has been observed that two other INVDOCK identified proteins glutathione transferase and 3a hydroxysteroid dehydrogenase exhibited altered activity by tamoxifen 29, 30, which may be indicative of direct binding of tamoxifen to these proteins. Experiments showed that the level of another two identified proteins dihydrofolate reductase and immunoglobulin is changed by tamoxifen 31, 32. Ligand binding is known to self-regulate protein level in certain cases 33. It remains to be seen whether these two proteins are also the target tamoxifen as implicated by INVDOCK search. It is noted that known target of tamoxifen such as calmodulin 24 is not identified by INVDOCK. One possible reason might be that the available PDB structures of calmodulin are not sufficiently close to tamoxifen-bound conformation. None of these PDB structures is bound by a ligand similar in structure to tamoxifen. The conformation of calmodulin is known to change significantly by binding of ligands 34. Because of the intrinsic flexibility of this protein, it is highly likely that ligand binding to this protein involves induced-fit. Our analysis of two PDB structures of calmodulin bound by a different ligand (PDB id: 1a29 and 2bbm) shows that the conformation of this protein is dependent on its binding ligand. In a recent molecular docking study, a ligand was docked into calmodulin by consideration of conformation changes that mimic an induced fit 35. We also finds that 4H-tamoxifen can be placed into calmodulin with slightly less favorable steric interaction than allowed by the INVDOCK scoring function. An appropriate conformation change in calmodulin might allow for the removal of this un-favorable interaction.
Limited number of protein entries available in our cavity database is also expected to result in missed hit. For instance, known tamoxifen metabolizing protein cytochrome P450 36 is not identified in this work because no corresponding human or mammalian entry is available in the cavity database. A search of bacterial proteins in the database identified this protein (PDB Id: 5cp4 and 1cpt) as a putative target.
As shown in Table 2, apart from the 10 putative protein targets that have been implicated or confirmed experimentally, INVDOCK identified 10 other proteins as putative targets of 4H-tamoxifen. These include aldose reductase, Arginase,
carbonic anhydrase, macrophage migration inhibitory factor, purine nucleoside phosphorylase, DNA polymerase b, hypoxanthine-guanine phosphoribosyltransferase, retinoic acid oxidase, sepiapterin reductase, and tyrosine 3-monooxygenase. No literature has been found to link tamoxifen to each of these proteins. There is also no report that indicates each of these proteins is not a target of tamoxifen or its analogs. Further investigation is therefore needed to determine whether or not 4H-tamoxifen can bind to these proteins.
V. INVDOCKPREDCITIONOFTOXICITYANDSIDE EFFECTTARGETS
INVDOCK has been used to predict the toxicity and side effect targets of eight drugs including aspirin, gentamicin, ibuprofen, indinavir, neomycin, penicillin G, 4H-tamoxifen, and vitamin C. These drugs were selected because they have been subjects of extensive investigation including the probing of toxicity and side effect protein targets of these drugs. However, these drugs were selected before the literature of the toxicity or side effects was searched.
As shown in Table 3, there are a total of 68 protein targets for the eight therapeutic agents that have been confirmed or implicated by available experimental studies. The detailed target list for each drug can be found in an earlier publication 11. INVDOCK predicted 38 of these. In addition, there are 29 INVDOCK predicted targets that lack experimental validation. Because of a limited scope of experimental study of toxicity targets of these drugs, it is not expected that all toxicity targets have been determined experimentally. Thus it is difficult to give a complete assessment about which of the predicted targets are false and how many targets are missed without the knowlgedge of all the targets of these drugs. The evaluation given here should therefore be considered as a preliminary assessment based on currently available experimental data. It is noted that some of the known toxicity targets of these drugs are not predicted by INVDOCK. Table 11 shows that there are 30 such INVDOCK missed targets. Of these missed targets, 22 are either without available ligand-bound structure of human or mammalian origin, or whose structure is incomplete (e.g. contains only an irrelevant section/domain of the protein). As this work only searches ligand-bound structures of human or mammalian origin, these 22 targets are beyond the capability of our inverse docking algorithm. In addition, there are 3 other INVDOCK missed targets to which the binding drug is linked by a covalent bond. As our docking algorithm and force field are designed for ligand-protein binding involving intermolecular hydrogen bond and non-bonded interactions only, INVDOCK may not be suitable for the prediction of targets involving a covalent bond. Therefore the miss of these three targets is not unexpected. Non-the-less, there are five INVDOCK missed targets whose structure is available.
The rate of correct predictions from this study may be estimated from the ratio between the number of predicted targets and that of experimentally known targets. Given that the number of INVDOCK predicted and experimentally
determined targets is 38 and 68 respectively, the rate of "correct prediction" is 56%. It is noted, however, the relevant 3D structure of 22 of the experimental targets is unavailable. Therefore these targets are outside the scope of INVDOCK and should be excluded in a more appropriate estimate of rate of correct prediction. This way the rate becomes 83%. If one further excludes the 3 other targets that involve drug-target covalent bonding, the rate is changed to 89%.
Likewise, the rate of missed targets can be estimated from the ratio between the number of INVDOCK missed targets and that of experimentally determined targets. This rate is 5%. The rate of "false-positive" may also be estimated as the ratio between the number of INVDOCK targets without experimental validation and that of the experimental targets, which is 63% when those experimental targets without relevant structure are excluded. Because the number of experimental targets is incomplete, the computed rate of correct prediction and that of the "false-positive" here does not add up to 1. This suggests that, without complete knowledge of all toxicity targets, these computed rates might not fully describe the quality of INVDOCK computations.
VI. INVDOCKPREDICTIONOFTHERAPEUTICAND TOXICITYTARGETSOF ACTIVECOMPOUNDSFROM
MEDICINALPLANTS
INVDOCK has also been tested for its applicability in the prediction of therapeutic and toxicity targets of active compounds extracted from medicinal plants 15. One example is catechin, also known as cyanidol, an active compound from green tea. It has been shown to inhibit the growth of human breast cancer cells 37 and prostate cancer cells 38 partly because of its inhibition of cyclin-dependent kinases 39. The antitumor activity of this compound may also arise from its inhibition of tyrosine phosphorylation of PDGF beta- 40, induction of apoptosis 41 and inhibition of matrix metalloproteinases 42. Catechin exhibits anti-inflammatory as well as cancer chemopreventive effects 43. Some of these effects of catechin are in part from its inhibition of TNF-alpha and NF-kappaB. This compound also have antiplaque and hepatoprotective effects via reduction of membrane fluidity 44. It has been reported that this compound has antioxidative action mediated by the activation of glutathione peroxidase 45.
INVDOCK search produced seventeen potential therapeutic targets, which are given in Table 4. Seven of these targets have been confirmed by experiments, which showed that catechin inhibits each of them. These include cyclin-dependent kinase and FGF receptor 39, neutrophil collagenase 42, protein kinase C 46, CAMP-dependent protein kinase 47, TNF-alpha and NF-kappaB p65 43. Inhibition of each of the first five proteins has potential anticancer implication, binding to the sixth protein may produce anti-inflammatory effect, and the interaction with the seventh and eighth proteins may lead to anti-inflammatory as well as cancer chemopreventive effects.
Available experimental data also seems to implicate another five of INVDOCK identified therapeutic targets. Catechin has been found to be an activator of cathepsin B and insulin 43. It is possible that activation of each protein is by direct binding of catechin. Activation of cathepsin may procude anticancer effect, while activation of insulin may help reducing glucose level and thus have implication in diabetes treatment. Catechin is known to inhibit both the ras-transformed cells and the activation of p38 mitogen-activated protein kinase 48, which may have implication in anticancer property. One possible reason for these inhibitory effects are due to the binding of catechin to ras p21 protein and MAP kinase p38 respectively as predicted by INVDOCK. Catechin is also known to have immuno-enhancing effect on T and B cell functions 49, which may also result from binding of catechin to immnoglobulin lambda light chain as predicted by INVDOCK. Further investigation is needed to determine whether these proteins are targets of catechin.
Moreover, INVDOCK identified four additional potential therapeutic targets. These are aldose reductase, X11, ganylyl cyclase, and C-AMP-dependent protein kinase. No experimental information has been found to either implicate or invalidate them. Hence, further study is needed to determine whether these proteins are targets of catechin. The potential therapeutic effect of the binding of catechin to each of these proteins is diabetes treatment for the aldose reductase, anti-clotting for X11, and anticancer for ganylyl cyclase and C-AMP-dependent protein kinase respectively.
Some known therapeutic targets of catechin are not found by our INVDOCK search. These include matrix metalloproteinase-2, matrix metalloproteinase-9, matrix metalloproteinase-12 and glutathione peroxidase. This occurs because of a lack of relevant structures in the database. Our database does not yet have 3D structure of matrix metalloproteinase-2, matrix metalloproteinase-9 and matrix metalloproteinase-12. Although the 3D structures of glutathione peroxidase are available in the database, these are ligand-free structures that may not be a suitable system for accurate analysis of the binding of a compound that affects the function of that protein. Thus these structures are not used in INVDOCK study.
INVDOCK search also identifies seven potential toxicity and side effect targets of catechin. One target, prostagladin H2 synthase 1, has been confirmed by experiments 50, 51. Inhibition of this protein may induce gastrointestinal toxicity. As to the other six potential toxicity targets, no experimental data have been found to either implicate or invalidate them. Therefore further study is needed to determine whether or not they are targets of catechin. Two such potential targets are carbonic anhydrase II and V. Binding to each protein may potentially induce acidity disorder in red blood cells. Another potential target is alcohol dehydrogenase. Liver toxicity may arise from catechin binding to this protein. The fourth potential target is epididymal retinoic acid-binding protein. The potential effect of catechin on this protein is the possible interference with retinoic acid transport, which may lead to cancer and cardiovascular disorder. The fifth and sixth potential targets are 1, 6-bisphosphatase and fructose-2, 6-bisphosphatase. Red blood cell disorder and thus anemia
may potentially be induced by catechin binding to each protein.
VII. FACTORSAFFECTINGTHEQUALITYOF INVDOCK
Overall, more than 50% of INVDOCK identified potential therapeutic protein targets and more than 27% of identified toxicity targets have been implicated or confirmed by experiments 10, 11, 15. Several reasons might contribute to the discrepancy between INVDOCK screening results and experimental data. It is not expected that exhaustive experiments have been done to determine all protein targets of a drug. Thus, discrepancy might arise for those identified proteins that lack relevant experimental information. Some of the PDB structures may be of little relevance to binding study for a particular molecule. These include entries containing an incomplete section or a chain, protein mutants that are structurally different from the corresponding proteins investigated in experiments, ligand-bound proteins whose conformation is relevant only to a specific set of ligands, and macromolecular complexes unrelated to a particular biological process studied experimentally. Docking of a molecule to such an irrelevant structure may thus generate a false hit. Anticipated rapid progress in structural genomics 14 is expected to provide a more diverse set of relevant structures. Knowledge from study of protein functions 52 also facilitates the selection of relevant structures in determination of putative protein targets related to a particular cellular or physiological condition.
Another possible cause of discrepancy between INVDOCK screening results and experimental data is a lack of consideration of protein profiles such as gene expression pattern and protein levels. Some experimental studies of ligand-protein interactions are based on the investigation of cell lines or other assays. Observation of molecular events related to a particular ligand-protein interaction requires that the protein under study be at a sufficient level in the system being investigated. If such a level is not reached at a particular setting, then the corresponding experiment is not useful in probing the binding of a molecule to that protein. Proteins not expressed or at too low levels in a particular biological process are unlikely to be therapeutically meaningful targets. Advance in proteomics is providing rapidly growing information about the profiles of proteins inside cells 53. Incorporation of this information into an inverse docking procedure can enable the prediction of more relevant protein targets.
Pharmacokinetic and metabolic profile of a molecule may also need to be considered in an inverse docking procedure. The action of a therapeutic molecule requires it to achieve an adequate concentration in the fluid bathing the target tissue. The concentration of a molecule is determined by its pharmacokinetic and metabolic profile. Therefore information about this profile is important in prediction of putative protein targets that a small molecule can reach at sufficient concentration. Development in pharmacokinetics and drug
metabolism 54 is providing more and more information in this regard.
The capability of an inverse docking approach in identifying putative protein targets of a small molecule is constrained by the relatively limited number of available protein 3D structures. This is particularly true for membrane-bound receptor proteins that are key therapeutic and toxicity targets for a variety of diseases or symptoms. This problem can be partially alleviated by structural information generated from rapid progress in high-throughput X-ray crystallography of protein structures and in the development of new structural determination methods 14.
The quality of an inverse docking procedure also depends on the algorithm and force fields used in docking and scoring. Ideally, in an inverse docking procedure, optimum binding mode should be searched in a cavity. Although more CPU-time demanding, such a search may become practical the over-all inverse docking search-speed can be improved. For instance, the search-speed of INVDOCK can be significantly improved by distributing database search onto multiple computers. Cavity models may also be further refined to reduce the search space. CPU time saved from these and other improvements may be used for searching optimum binding modes. There has been progress in refining docking/scoring algorithm and force fields particularly in the areas of protein flexibility 55 as well as ligand flexibility 16, 21, 22. Knowledge and new methods/force fields generated from these studies can also be incorporated into an inverse docking procedure.
In a ligand-protein docking study, putative ligands are typically selected from the lowest energy docked structures. On the other hand, in an inverse docking procedure, putative protein targets are selected based on an energy threshold. Docked structures with ligand-protein interaction energy lower than that threshold are considered as putative targets. However, this energy threshold is more difficult to define. A stricter condition on the energy threshold can easily result in missed hits. Likewise, a too relaxed energy threshold may lead to the over production of false hits. The empirical formula for the energy threshold used in this study seems to give reasonable results for the two agents studied. However, additional study is needed to more extensively evaluate and further refine the energy threshold.
VIII. CONCLUSION
Ligand-protein inverse docking appears to be a useful approach for computer-aided identification of potential protein targets of a small molecule. Testing results have shown the potential of this approach in facilitating drug target identification, toxicity prediction, and the probing of molecular mechanism of compounds extracted from medicinal plants. Performance and applicability of ligand-protein inverse docking program need to be further enhanced by refinement of docking and scoring algorithms, development of more efficient cavity models, and by incorporation of new information from advances in structural genomics,
proteomics, protein function, pharmacokinetics and drug metabolism. Further development of this inverse docking approach may bring into reach applications such as identification of unknown and secondary therapeutic targets of drugs and investigative drugs, toxicity and side effect prediction, and the study of molecular mechanism and pharmacology of medicinal plants.
REFERENCES
[1] Drews, J. (2000) Drug discovery: A historical Perspective. Scinece 287, 1960-1964.
[2] Ohlstein, EH., Ruffolo Jr., R.R., and Ellroff, J.D. (2000) Drug discovery in the next millennium. Annu. Rev. Pharmacol. Toxicol. 40, 177-191.
[3] Royer RJ. Mechanism of action of adverse drug reactions: An overview. 1997; Pharmcoepidemiology and drug safety 6 (Suppl.):S43-S50.
[4] DiMasi JA, Bryant, NR, Lasagna L. (1991). New drug development in the United States from 1963 to 1990. Clin Pharmacol Ther. 50, 471-486.
[5] Drews J. (1997). Strategic choices facing the pharmaceutical industry: A case for innovation. Drug. Discov. Today 2, 72-78.
[6] Ang-Lee M.K., Moss J., Yuan C.-S. (2001). Herbal medicines and perioperative care. JAMA 286,
[7] Cheng JT. Review: Drug therapy in Chinese traditional medicine. (200). J Clin Pharmacol. 40:445-450.
[8] Zhu DY, Bai DL, Tang XC. (1996). Recent studies on traditional Chinese medicinal plants. Drug Dev. Res. 39:147-157.
[9] Gong X, and Sucher NJ. (1999). Stroke therapy in traditional Chinese medicine (TCM) prospects for drug discovery and development. Tends Pharmacol Sci. 20:191-196.
[10] Chen Y.Z., and Zhi D. G., (2001). Ligand-Protein Inverse Docking and Its Potential Use in Computer Search of Putative Protein Targets of a Small Molecule. Proteins, 43, 217-226
[11] Chen Y.Z., and Ung C. Y. (2001). Prediction of Potential Toxicity and Side Effect Protein Targets of a Small Molecule by a Ligand-Protein Inverse Docking Approach. J. Mol. Graph. Mod., 20, 199-218.
[12] Berman, HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN, Bourne PE. The Protein Data Bank. Nucleic Acids Research, 2000; 28:235-242.
[13] Rost B, Sander C. Bridging the protein sequence-structure gap by sequence-structure predictions. Ann. Rev. Biophys. Biomol. Struct. 1996; 25:113-136.
[14] Sali, A. 100,000 protein structures for the biologist. Nature Struct. Biol. 1998; 5:1029-1032
[15] Chen Y.Z., and Ung C. Y. (2002). Computer Automated Prediction of Putative Therapeutic and Toxicity Protein Targets of Bioactive Compounds from Chinese Medicinal Plants, Am. J. Chin. Med., 30, 139-154. (2002).
[16] Wang, J, Kollman, PA, Kuntz, ID. Flexible ligand docking: A multistep strategy approach. Proteins. 1999; 36:1-19.
[17] Cornell, WD, Cieplak P, Bayly CI, Gould IR, Merz KM Jr, Ferguson DM, Spellmeyer DC, Fox T, Caldwell JW, Kollman PA. A second generation force field for the simulation of proteins and nucleic acids. J. Am. Chem. Soc. 1995; 117:5179-5197.
[18] Baird, NC. Simulation of hydrogen bonding in biological systems: Ab initio calculations for NH3 and NH3-NH4+. Int. J. Quantum Chem. Symp. 1974; 1: 49-53.
[19] Chen, YZ, and Prohofsky, EW. The role of a minor groove spine of hydration in stabilizing poly(dA).poly(dT) against fluctuational interbase H-bond disruption in the premelting temperature regime. Nucleic. Acids. Res. 1992; 20: 415-419.
[20] Chen YZ, Prohofsky EW. Premelting base pair opening probability and drug binding constant of a daunomycin--Poly d(GCAT)-Poly d(ATGC) complex. Biophys. J. 1994; 66:820-826.
[21] Lorber DM, Shoichet BK. Flexible ligand docking using conformational ensembles. Protein Sci. 7:938-950. 1998. [22] Baxter CA, Murray CW, Clark DE, Westhead DR, Eldridge MD. Flexible docking using Tabu search and an empirical estimate of binding affinity. Proteins 33:367-382. 1998.
[23] Favoni RE, and Cupis AD. Sterioidal and nonsteroidal oestrogen antagonists in breast cancer: basic and clinical appraisal. Trends Pharmacol Sci. 1998; 19:406-415.
[24] Rowlands MG, Budworth J, Jarman M, Hardcastle IR, McCague R, Gescher A. Comparison between inhibition of protein kinase C and antagonism of calmodulin by tamoxifen analogues. Biochem. Pharmacol. 1995;50:723-726.
[25] Abbas Abidi SM, Howard EW, Dmytryk JJ, Pento JT. Differential influence of antiestrogens on the in vitro release of gelatinases (type IV collagenases) by invasive and non-invasive breast cancer cells. Clin Exp Metastasis 1997;15:432-439
[26] Santner SJ, Santen RJ. Inhibition of estrone sulfatase and 17 beta-hydroxysteroid dehydrogenase by antiestrogens. J Steroid Biochem Mol Biol. 1993;45:383-90.
[27] Messiha FS. Leu-enkephalin, tamoxifen and ethanol interactions: effects on motility and hepatic ethanol metabolizing enzymes. Gen Pharmacol 1990; 21: 45-48. [28] Ritchie GA. The direct inhibition of prostaglandin synthetase of human breast cancer tumor tissue by tamoxifen. Recent Results Cancer Res. 1980;71:96-101.
[29] Nuwaysir EF, Daggett DA, Jordan VC, Pitot HC. Phase II enzyme expression in rat liver in response to the antiestrogen tamoxifen. Cancer Res 1996;56:3704-3710. [30] Lax ER, Rumstadt F, Plasczyk H, Peetz A, Schriefers H Antagonistic action of estrogens, flutamide, and human growth hormone on androgen-induced changes in the activities of some enzymes of hepatic steroid metabolism in the rat. Endocrinology 1983;113:1043-1055
[31] Levine RM, Rubalcaba E, Lippman ME, Cowan KH. Effects of estrogen and tamoxifen on the regulation of dihydrofolate reductase gene expression in a human breast cancer cell line. Cancer Res 1985;45:1644-1650.
[32] Paavonen T; Aronen H; Pyrhonen S; Hajba A; Andersson LC. The effect of toremifene therapy on serum
immunoglobulin levels in breast cancer. APMIS 1991;99:849-853.
[33] Schmidt TJ, Meyer AS. Autoregulation of corticosteroid receptors. How, when, where, and why? Receptor. 1994 Winter;4:229-257.
[34] WE Meador, AR Means and FA Quiocho. Modulation of calmodulin plasticity in molecular recognition on the basis of x-ray structures. Science. 1993; 262:1718-1721.
[35] Sandak B, Wolfson HJ, Nussinov R. Flexible docking allowing induced fit in proteins: insights from an open to closed conformational isomers. Proteins 1998;32:159-174 [36] Crommentuyn KM, Schellens JH, van den Berg JD, Beijnen JH. In-vitro metabolism of anti-cancer drugs, methods and applications: paclitaxel, docetaxel, tamoxifen and ifosfamide. Cancer Treat Rev. 1998;24:345-366.
[37] Damianaki A, Bakogeorgou E, Kampa M, Notas G, Hatzoglou A, Panagiotou S, Gemetzi C, Kouroumalis E, Martin PM, Castanas E. Potent inhibitory action of red wine polyphenols on human breast cancer cells. J Cell Biochem. 78:429-441. 2000.
[38] Sakamoto K. Synergistic effects of thearubigin and genistein on human prostate tumor cell (PC-3) growth via cell cycle arrest. Cancer Lett. 151:103-109. 2000.
[39] Liang YC, Lin-Shiau SY, Chen CF, Lin JK. Inhibition of cyclin-dependent kinases 2 and 4 activities as well as induction of Cdk inhibitors p21 and p27 during growth arrest of human breast carcinoma cells by (-)-epigallocatechin-3-gallate. J Cell Biochem. ;75:1-12. 1999.
[40] Sachinidis A, Seul C, Seewald S, Ahn H, Ko Y, Vetter H. Green tea compounds inhibit tyrosine phosphorylation of PDGF beta-receptor and transformation of A172 human glioblastoma. FEBS Lett. 471:51-55. 2000.
[41] Gupta S, Ahmad N, Nieminen AL, Mukhtar H. Growth inhibition, cell-cycle dysregulation, and induction of apoptosis by green tea constituent (-)-epigallocatechin-3-gallate in androgen-sensitive and androgen-insensitive human prostate carcinoma cells. Toxicol Appl Pharmacol. 164:82-90. 2000. [42] Demeule M, Brossard M, Page M, Gingras D, Beliveau R. Matrix metalloproteinase inhibition by green tea catechins. Biochim Biophys Acta. 1478:51-60. 2000.
[43] Ahmad N, Gupta S, Mukhtar H. Green tea polyphenol epigallocatechin-3-gallate differentially modulates nuclear factor kappaB in cancer cells versus normal cells. Arch Biochem Biophys. 376:338-346. 2000.
[44] Tsuchiya H. Effects of green tea catechins on membrane fluidity. Pharmacology. 59:34-44. 1999.
[45] Nagata H, Takekoshi S, Takagi T, Honma T, Watanabe K. Antioxidative action of flavonoids, quercetin and catechin, mediated by the activation of glutathione peroxidase. Tokai J Exp Clin Med. 24:1-11. 1999.
[46] Komori A, Yatsunami J, Okabe S, Abe S, Hara K, Suganuma M, Kim SJ, Fujiki H. Anticarcinogenic activity of green tea polyphenols. Jpn J Clin Oncol. 23:186-190. 1993. [47] Polya GM, Foo LY. Inhibition of eukaryote signal-regulated protein kinases by plant-derived catechin-related compounds. Phytochemistry. 35:1399-1405. 1994.
[48] Chung JY, Huang C, Meng X, Dong Z, Yang CS. Inhibition of activator protein 1 activity and cell growth by purified green tea and black tea polyphenols in
H-ras-transformed cells: structure-activity relationship and mechanisms involved. Cancer Res. 59: 4610-4617. 1999. [49] Brattig NW, Diao GJ, Berg PA. Immunoenhancing effect of flavonoid compounds on lymphocyte proliferation and immunoglobulin synthesis. Int J Immunopharmacol. 6:205-215. 1984.
[50] Noreen Y, Serrano G, Perera P, Bohlin L. Flavan-3-ols isolated from some medicinal plants inhibiting COX-1 and COX-2 catalysed prostaglandin biosynthesis. Planta Med. 64:520-524. 1998.
[51] Min KR, Hwang BY, Lim HS, Kang BS, Oh GJ, Lee J, Kang SH, Lee KS, Ro JS, Kim Y. (-)-Epiafzelechin: cyclooxygenase-1 inhibitor and anti-inflammatory agent from aerial parts of Celastrus orbiculatus. Planta Med. 65:460-462. 1999.
[52] Bork P, Dandekar T, Diaz-Lazcoz Y, Eisenhaber F, Huynen M, Yuan Y.J. Predicting function: from genes to genomes and back. J. Mol. Biol. 1998; 283:707-725.
[53] Persidis A. Proteomics. An ambitious drug development platform attempts to link gene sequence to expressed phenotype under various physiological states. Nature Biotech. 1998; 16:393-394.
[54] Lin, JH., Lu AY. Role of pharmacokinetics and metabolism in drug discovery and development. Pharmacol. Rev. 1997; 49:404-449.
[55] Carlson HA, McCammon JA. Accommodating protein flexibility in computational drug design. Mol. Pharmacol. 2000; 57:213-218.
