The Many Roles of Computation in Drug Discovery

William L. Jorgensen
Science 303, 1813 (2004);
DOI: 10.1126/science.1096361

This copy is for your personal, non-commercial use only.

Permission to republish or repurpose articles or portions of articles can be obtained by
following the guidelines here.
The following resources related to this article are available online at (this information is current as of June 12, 2012 ):
Updated information and services, including high-resolution figures, can be found in the online
version of this article at:
A list of selected additional articles on the Science Web sites related to this article can be
found at:
This article has been cited by 242 article(s) on the ISI Web of Science
This article has been cited by 27 articles hosted by HighWire Press; see:
This article appears in the following subject collections:

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by the
American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. Copyright
2004 by the American Association for the Advancement of Science; all rights reserved. The title Science is a
registered trademark of AAAS.

Downloaded from on June 12, 2012

If you wish to distribute this article to others, you can order high-quality copies for your
colleagues, clients, or customers by clicking here.

Dev. R. D. Molecular libraries are screened. C. the drug discovery process is driven by multidisciplinary teams made up of the very best experts in each discipline. for example.. A. Watt. A. The reality is that the use of computers and computational methods permeates all aspects of drug discovery today. and although chemists often end up as group leaders of discovery efforts. Today 6. Eng. Despite large research budgets. 9. A. U. 435 (1991). and the resulting leads are optimized in a cycle that features design. 4607 (1998). However. M. It is often a very weak binder and is likely to have a nonoptimized pharmacokinetic profile. New Haven. This example highlights the difficulty of and resources needed to optimize the ancillary properties of potent inhibitors/antagonists so that they can become safe. Lipinski. Dev. O. Chem. Snider et al. development of the underlying molecular mechanics (MM) force fields. 17. the rule of five (2) predicts that poor absorption or permeation of drugs is more likely when a drug molecule possesses either (i) more than 5 hydrogen bond donors. In large pharmaceutical companies. and industrial research in the development of new drug entities. Most drugs now arise through discovery programs that begin with identification of a biomolecular target of potential therapeutic value through biological studies including. evaluation of druglikeness. (ii) 10 hydrogen bond acceptors. 3 (1997). 51 (2003). synthesis and assaying of numerous analogs. i. In the drug discovery process. The success of SBDD is well documented (1. C. J. Minimally. Toxicol. Hale et al. There is usually much “tweaking” toward the end of the preclinical period of drug discovery when a series of compounds www. R. government. 16. a “hit” is defined as a nonoptimized structure obtained from some screening process on a target protein. 75 (2003).. viable medicines. it has contributed to the introduction of ⬃50 compounds into clinical trials and to numerous drug approvals. 2). Crystal structure determination for complexes of some analogs with the biomolecular target is often possible. Ebneth. SCIENCE VOL 303 19 MARCH 2004 1813 Downloaded from www. Today 8. Feeney. (iii) a molecular weight greater than 500. Pongs. funded with public monies. Drug Discov. Dev. Drug Discov. Dev. which typically selectively bind to the molecular target and interfere either with its activity as a recep- tor or enzyme. A. W. we must also be cognizant of the interrelatedness of academic. Adv. 41. 5. This process requires the very best organic chemistry skills. These teams have ready access to experts in other areas of biomedical science. Mass Spectrom. D. Chem. McCoy. that usually occurs after much experience has been gained in the drug discovery process. 15. 4. In this discussion. and we must continue to provide funding in the university system for training in these core skill sets to chemists in their graduate and postdoctoral studies if we are to continue to provide the very best in medicines for what is becoming an aging population. de novo design. It is also inconceivable that a human with or without computational tools could propose a single chemical structure that ends up as a drug. but in order to use these tools effectively. 59 (2003). but will likely not have been fully optimized for pharmacokinetic properties or undesirable off-target activities. 5. Rapid Commun. Often a good lead will have shown some proof-of-concept activity in an in vivo pharmacological model. although still not fully optimized. druglike compounds that are ready for human clinical trials. Med. “lead” is defined as a structure that has been derived from an early “hit” and. News 81. P. 30 (2003). the role of computation here is in the structure refinement using simulated annealing (3). or (iv) a calculated logP greater than 5. Those who are most proficient with the computational tools have the advantage for delivering new drug candidates more quickly and at lower cost than their competitors. Particular emphasis is placed on virtual screening. the inquisitor could go back to the lab with the reassurance that his or her choice to avoid learning about computational chemistry remains wise. The recent advances discussed above have put more tools in the chemist’s toolkit. Chem. and building and MM evaluation of analogs. 14. J. Drug Discov. a large percentage of which likely exhibited subnanomolar potency at the NK-1 receptor. In the discovery setting. 50 (2002). it invariably comes down to the ability to make the absolutely “correct” molecule in a timely and cost-effective manner. Nicoll-Griffith. “Is there really a case where a drug that’s on the market was designed by a computer?” When asked this. See (17) for an excellent review of the cardiovascular effects manifested by QT interval prolongation and the evaluation of drug candidates for this parameter. Uetrecht. T. Res. Chem. J. 22 (2003). has been shown to have some appropriate characteristics to be a precursor of a drug entity. and animal studies. All of the above discussion speaks to one of the most . Science 251. A multidisciplinary project team is then assembled with the goal of finding clinical candidates. Dev. News 81. USA. Koppal. jorgensen@yale. and advanced methods for determining protein-ligand binding. 15. the phrasing of the question Department of Chemistry. analysis of mice with gene knockouts. Koppal. E-mail: william. Evans. Baillie. However. 13.”).DRUG DISCOVERY References and Notes 1. 11. Eng. Dominy.. Netzer.e. 17. M. 12. A. All top pharmaceutical companies have substantial structural biology and computational chemistry groups that are intertwined and participate on the project teams. while sensing that the desired answer is “no”. Bischoff. but only one has made it to market. 6. Drug Discov. REVIEW The Many Roles of Computation in Drug Discovery William L. A 2. First. Drug Discov. J. 15 (2003). CT 06520-8107. DePalma. T. A. 2012 important issues facing discovery medicinal chemistry today: the continuing need for excellent synthetic chemists. I invoke the professorial mantra (“All questions are good questions. Drug Delivery suggests misunderstanding and oversimplification of the drug discovery process. Jorgensen An overview is given on the diverse uses of computational chemistry in drug discovery. Then. 335 (2001). and chemistry is one key element in this. A. Lombardo. A. 10. which enables “structure-based drug design” (SBDD) and the efficient optimization of leads. Korfmacher et al. 3. Drug Discov. there are far too many hurdles and subtleties along the way.-F. 832 (2003). F. 78 (2001). W. J. structure display. 6. the biomedical research carried out by pharmaceutical companies still represents only a small percentage of the overall worldwide research effort on diseases and approaches to their treatment.M. 3 (2004). A. it is the rare case today when an unmodified natural product like taxol becomes a drug. Academic and government laboratories. 6. B. 23. 8. J. SPECIAL SECTION the past decade as substance P antagonists. Tremblay. often provide much basic research and fundamental insight into diseases that can direct researchers toward novel ways of attacking on June 12. P. See the cover story in Drug Discov. Rouhi. Yale University. 6.sciencemag. 7. T. they are rarely organized (nor is it their mission) to embrace the drug discovery process in the multidisciplinary fashion outlined above that is the modern paradigm by which new hits or leads are first identified and then get transformed into new viable medicines.. DePalma.

too much metabolic activity reduces bioavailability. threedimensional (2D to 3D) structure conversion.000 compounds in the 2001 Mayin software comparisons owing to lingering bridge catalog with the addition of 20 nonfavoritism. As an example. usually from x-ray crystallography or predicted by homology modeling. Although the production of primary metabolites can be well predicted. it is easy to see Scheme 2. Drug discovery is complex: Successful teams and companies need to be congratulated. A related activity is to build computationally all members of a proposed combinatorial library and then select the final reagents for purchase or synthesis based on the diversity of the resultant library members (14). and. so. as computed by QikProp (27). Instead.g. When the structure of the target is unknown. respectively. flexibility of the host. duration of action. nephrotoxicity. and the ventricular irregularity known as long Q-T syndrome (6). There are many approaches to docking and related software packages (15).the lead to candidate progression should be highlighted. database entry-management-query. including the action of efflux pumps. Even simple considerations for a set of active compounds can greatly reduce the search on June 12.. Nuthe solvent-accessible surface area (SASA) merous successes of this approach for obtainfor each compound. with adequate potency has been identified and the remaining concerns focus on differences in pharmacological issues relating to bioavailabilty. a reactive hydrogen may be replaced by a halogen to block a metabolic process (5). finally there are the differences between humans. The binding site is relatively small and hydrophobic. and scored for potential activity. potential drug-drug interactions (e. comparison of new versions of scriptase (NNRTIs) in red. binding to plasma proteins. true objectivity is difficult to realize for the ⬎ Downloaded from www. active transport. extensive lead optimization is typically needed to lower this value to the 1 to 10 nM range. 13). conformational searching. experimental high-throughput screening (HTS) requires a library of compounds and an assay for activity (8). that inhibitors for a given target cluspounds can then be purchased or synthesized ter in a limited region. general. Such computational screening for similarity is common in selecting compounds for purchase from commercial libraries. Library Screening and De Novo Design Scheme 1. associated biased choices for the nucleoside inhibitors of HIV-1 reverse trantest data sets. there is systematic use of wideranging computational tools to facilitate and enhance the drug discovery process. Sometimes there are no hits (9). including efavirenz (Sustiva) and UC-781 (Scheme 2). drug distribution in blood and tissue. optimally positioned in the binding site. Tweaking is common as well to improve solubility or cell permeability. hepatotoxicity. the methyl group in celecoxib (Celebrex) makes the compound a weaker COX-2 inhibitor than the unsubstituted parent or chloro analog (Scheme 1). and the combination of high costs and low hit rates has put the large-scale approaches of the early 1990s out of favor (10). Although conformational sampling of the ligand is now performed by all of the programs. In a plot such as Fig. In the more common case. quantum chemistry. Computational virtual screening (11–13) can be performed on libraries of known or constructed compounds and requires either measured activities for some known compounds or a structure of the biomolecular target. some widely used ones are DOCK (18). what’s the best docking program? In (Å2) versus log P. The return can be expected to be low for investing in compounds outside these ranges in search of new NNRTIs. remaining issues for these programs to overcome include the generation of optimal structures of the complexes. Libraries of potential ligands are built in the computer. Glide (20). and biomolecular structure refinement. don’t take antacids and fluoroquinolones including ciprofloxacin because metal ion chelation reduces their absorption). however. The top-scoring com- At the lead-generation stage. which lead to the future of pharmacogenomics (7). Some particularly notable examples tion coefficient (log P). molecular dynamics.sciencemag. so the solvent-accessible surface area and the log P for an inhibitor should fall in the 450 to 650 Å2 and 2 to 6 ranges. FlexX (19). and details of the scoring functions including treatment of solvation. which are also central to bioavailability. When the structure of the target is known. Aside from the universal and now takenfor-granted use of software developed by chemists for structure drawing. the details of the kinetics cannot. microbial resistance. the activity data can be used to construct a pharmacophore model for the positioning of key features like hydrogen-bonding and hydrophobic groups. 15). whereas the search for one responsible individual or computer program is counterproductive. and GOLD (21). Docking. 2012 SPECIAL SECTION DRUG DISCOVERY . The complex differences between animals and humans are addressed mainly in human clinical trials. recent advances in several specific areas in 1814 nonnucleoside inhibitors of HIV-1 reverse transcriptase (NNRTIs). The ordinate shows and submitted for experimental testing. 19 MARCH 2004 VOL 303 SCIENCE www. Such a model can be used as a template to select the most promising candidates from the library. The list of other issues and concerns is long.sciencemag. The biomolecular binding sites for inhibitors have associated characteristics of spatial extent and overall polarity. for example. which is a measure of are the discovery of DNA gyrase inhibitors hydrophobicity. 1. There is not going to be a “voila`” moment at the computer terminal. and toxicities of many forms such as mutagenicity. molecular visualization. A successful hit would have a concentration for 50% inhibition (IC50) of ⬃10 ␮M. Fig. The specific example is for after HTS failed (9) and the findings of much higher hit rates for docking than HTS in two comparative studies for tyrosine phosphatase 1B (16) and dihydrodipicolinate reductase (17). and the abscissa shows ing lead compounds are well documented the predicted log of the octanol/water parti(12. 1. the most common virtual screening approach is molecular docking (12. Plot of solvent accessible surface area So. it also introduces a site for metabolic oxidation that then leads to acceptable clearance of the drug (4). and toxicity. molecular mechanics.

SPECIAL SECTION one program with older versions of another. Tyr188. www. Results of similarity searching for NNRTIs seeded into the Maybridge catalog. and co-workers at Pfizer ture generator. Therefore.000 non-druglike molecules from the 2001 Maybridge catalog and then added the 20 NNRTIs of Fig. which emphasize similarity of surface areas. and a combinatorial library of analogs is created by appending up to four substituents from a list of more than 500 typical drug fragments. Similarity searching. A similar notion has been formalized in the SYNOPSIS program by only considering molecules that can be built from known precursors through successive application of appropriate reactions from a set of 70 possibilities (31). hydrogen bonding. 2. hydrogen bond counts. current individual scoring functions are not optimal. 3). and uneven skill levels with different programs. The analogous procedure was performed for 20 nonsteroidal anti-inflammatory drugs. Lys101. the analog. Thus. This file was then processed with QikSim (28) to compute the similarity of each compound to the average of the active compounds with simple Euclidian and Tanimoto measures of similarity (26). respectively. I removed ⬃20. which were ranked among those most similar to the average NNRTI.e. Fig. distribution. application of multiple scoring functions to the same structures (22–25). As a recent example.sciencemag. e. serious potential users need to invest some effort in testing alternatives in their own environment. Leu100.sciencemag. 1.Phe227. 65% (Euclidian) and 50% (Tanimoto) of the active molecules were ranked as being among the top 5% in similarity to the average NNRTI. i. Our approach is to build libraries with a parallel synthesis plan in mind that centers on standard reactions such as amide bond formation.. Examples of two compounds from the Maybridge catalog. BOMB is fully integrated with QikProp. 2. and excretion) considerations early in preearly examples of the alternatives are LUDI (29) and SPROUT (30). 2012 Scheme 3.Fig. though only sition and orientation of selected protein atoms are illustrated here. Trp229.e. However. Docking proPrediction of Properties and grams can also be used for this purDrug-likeness pose by coupling them with a strucLipinski. which may be as simple as methane. The final negatively and positively charged areas. ments or by growing substituents on a SCIENCE VOL 303 19 MARCH 2004 1815 Downloaded from on June 12. which is reflected in limited transferability of scoring functions between different classes of protein targets. so each analog that is built is filtered for drug likeness. and log P. Molecules can be built in five topologies. and Gold and found that the hit rates of ⬃10% from a single scoring function could rise to 25 to 40% and 65 to 70% with double and triple scoring.. The weakest point is the scoring. The corresponding random result would be that only 5% of the active molecules (1 molecule) would emerge in the top 5% of processed molecules (0. are RJC 02097 and RJC 03066. It can then be used with BOMB in construction mode to design new inhibitors (Fig. An NNRTI built by BOMB in the HIV-1 RT binding site starting from ammonia as the core. 3. One scoring function has been trained on several hundred complexes for HIV-RT and COX2. one general observation is that improved results are obtained using consensus scoring. The remaining 50684 compounds were processed by QikProp (27) to compute 36 descriptors and properties for each compound that are stored in a spreadsheet. my contribution in this area is the growing program BOMB (biochemical and organic model builder). and the common over-the-counter alternatives.g. Interaction between all team members including the computational and synthetic chemists is clearly important. If a developer or vendor is doing the comparisons. They can be compared to efavirenz and UC-781. and cycloaddition chemistry. The typical test is to seed a library of inactive compounds with active ones and then determine the number of actives that are predicted to be among the ⬃5% of top-scoring compounds. The user positions a core. metabolism. The protein surface is hedral angles for protein colored according to its electrostatic potential. For example. structures are again scored with consideration of the interaction energy. Using the default weights. De novo design. respectively (Scheme 3). design of inhibitors from scratch given the target binding site. and some di. usually by positionlike (11. Full atomic detail is provided. All reasonable conformers of each analog are constructed and optimized with the OPLS-AA force field with variation of the dihedral angles. and solvation energy changes.. Murcko. the strains on objectivity are obvious and they have the last shot at fine-tuning algorithms to remove problem cases. Another general issue with de novo design is the synthetic accessibility of the constructs. i. in the binding site. po. Further comparisons also need to be made with simpler alternatives. overall shape (globularity). From the broader performance comparisons of docking programs.DRUG DISCOVERY against actives (26). including Tyr181. Bissantz et al. and Val106.05 ⫻ 50684 ⫽ 2534 molecules). FlexX. red and blue for side chains. 32–35). The results for the Euclidian analysis are illustrated in Fig. For thymidine kinase and estrogen receptor. A related activity is de novo design (15). coupling reactions. (23) applied Dock. a variety and Vertex deserve much credit for raising of more specialized programs have awareness about properties and structural feabeen developed for building ligands tures that render molecules more or less drugin binding sites. by sprouting from a central core or by linking groups sequentially to yield more elongated constructs. The motivation is to apply ing and connecting molecular fragADME (absorption. such as similarity matching . 50% (Euclidian) and 75% (Tanimoto) of the active compounds emerged in the topranked 5%. including celecoxib. rofecoxib (Vioxx).

40 (0. ADME which through evolution and use of activecule to be smoothly mutated to the other by software can help thorough predictions for transport mechanisms are effective antibactechanging the force field parameters and gea proposed analog series and early identirial agents. brain/blood partitioning. which promotes long Q-T syndrome. plexes and representation of the aqueous protein binding.74 0.28 –5.80 –1. although it is potent in vitro. The well-known rule-of-five (33) points out that most orally administered drugs have a molecular weight (MW) of 500 or less. The amount of data and poor cell permeability for this relatively reliability of the corresponding predictions polar molecule. five or fewer hydrogenbond donor sites.sciencemag. interesting from the standpoint of potential Such software is often implemented in corCNS penetration. The log BB predictions are also QikProp is 3 min on a 3 GHz laptop PC.09 tion (TI) calculations then provide formally the hydrophilic comlog Khsa 0. whereas the oplog S –4.13 The input is a 3D molecular structure and can also be informative.02 –3. For biomolecule-ligand affinities. conversion to a 3D structure compute the effect of a protein mutation on the mostly come from regression equations or with Chem3D.74 0. which yield offer modules for computation of ADMEability of only 1% (2). Classical force ical patterns. Predicted and experimental (in parentheses) ADME-related freedom for the biomolecule-ligand comP. The difference in free energies of fication of potential problems. e. as a topical microbicide rather than as an oral the processing time for the 71.05 0.04 –5. cell permeability.sciencemag.41 0. The conversions involve gle of solubility versus a coupling parameter that causes one molepermeability. Some results of a QikProp analysis for Advanced Treatments of the antipsychotic drug chlorpromazine Protein-Ligand Binding (Thorazine) and the antiadrenergic atenolol To progress beyond scoring functions.70 able functionality. yses for classes of compounds MKC-442 3. Caco-2 and MDCK cell permeabilities (42). I released one of the earliest genfilter for library screening. Table 2.10) –0.52 0. and predicted to be the most hydrophobic with log IC50 for HERG K⫹-channel blockage poor solubility and high serum-protein bind(44).11 age. QikProp (27).Scheme 4.31 In 2000. analTMC125 2. The two Monte Carlo (MC) statistical mechanics or drugs approach the hydrophobic and hydromolecular dynamics (MD) simulations are philic extremes. primary metabolites. but Chlorpromazine Atenolol surroundings with hundreds to thousands of good cell permeabildiscrete water molecules.79 0. its pursuit and can be executed for large libraries. including lack of reactive functionality except in on June 12. The structural features include surface area components and hydrogen-bonding potentials.16) log BB 0.28 –5. against a database of known drugs.80 –2. rules invite exceptions. and high serumproperties. experimental data.08 –0. and there is always someone who is quick to note cases such as macrolide antibiotics like erythromycin (MW 734). are listed in Table 1 (Scheme 4). and the methodology alhave relatively poor lows extensive sampling of all degrees of solubility. for CNS side effects. DPC083 4. the peptide-like thrombin in(Scheme 6). Cost-efficient modern drug disometry. and processing by QikProp. high log Table 1.78 (1.5 (–5.43 the output is a profile of structural feaance of the extremes in a class is Efavirenz 3. some predictions from QikProp and comparisons of the same quantities are shown in Table 2 for NNRTIs 1816 19 MARCH 2004 VOL 303 SCIENCE www. log BB (40). The predictions ChemDraw. and. pounds and the obvious use as a Delavirdine 2. hibitor (Scheme 5) had to be abandoned formed to convert X to Y unbound in water and Most chemical software companies now late in development due to its oral bioavailcomplexed to the biomolecule.25 eral programs of this type.13 –5. log Khsa in order of increasing log P.500 comdrug is undoubtedly related to these properpounds in the 2001 Maybridge catalog with ties (45). binding of one inhibitor by making P the inhibneural networks (39) that are trained on reveals that the problem appears to be itor and X and Y the two proteins.57 0. and they illustrate the typtypically applied (46–48). and the properties include octanol/water and water/gas log Ps.05 0. Free-energy perity. ADME-relevant properties.83 0.61 –3. ADME software is generally very fast ing. As reviewed elsewhere (37).92 0. which can be beneficial for porate Web-based systems for ready processattack on latent HIV reservoirs or a concern ing of individual compounds or libraries. an acceptable level of solubility is also critical to permit dissolution and absorption. Two series of mutations are perresources to take on the higher risks.19) 0.59 –5.10 –0. a log P no higher than 5. for central nervous system (CNS) compounds.61 (–1. Hydrophobic compounds fields are used. virtually all drugs have aqueous solubilities above 10⫺6 M (log S ⬎ – 6). omy is responsible PMDCK (nm/s) 1425 33 (18) perturbations are made to convert one ligand for the classic leadCNS Activity ⫹⫹ –– to another using the thermodynamic cycle optimization strugshown in Scheme 6..52 0.clinical development to avoid costly latestage preclinical and clinical failures (36). Avoid8-chloro-TIBO 3.06) –1. Selected predicted properties for NNRTIs.79 (–0. log S (37). The selection of such properties and features and their allowable ranges can be adjusted to provide filters for specific applications such as selection of screening compounds or purchase of reagents for a combinatorial library.74 (1. This dichotPCaco (nm/s) 2003 66 (33) changes.80 (5.29 tures.01) –0. undesira safe strategy.30) turbation (FEP) and thermodynamic integraposite is true for log P 4. 2012 SPECIAL SECTION DRUG DISCOVERY . There are other important issues for drug likeness. A quick sketch with ⌬GF and ⌬GC.52 –0. The same cycle can be used to related properties (38).61 –3. For Downloaded from www. all CNS activity (41). Of course.g. UC-781 is for human serum albumin binding (43).52 UC-781 5. and 10 or fewer hydrogenbond acceptor sites (N and O atoms). with predicted varies from excellent for log P and solubilPCaco and PMDCK values near ity to more limited reliability for properties 10 nm/s.48) rigorous means to compute free-energy pounds. over. such as log BB (brain/blood partitioning) Beyond the use of ADME ⫹ NNRTI log P log BB log S log Khsa and the IC50 for HERG K -channel blocksoftware for individual comNevirapine 2. As an examcovery has a substantial statistical component binding for the ligands X and Y then comes ple of a promising compound that went to it: Envelope pushers need to have the from ⌬⌬Gb ⫽ ⌬GX – ⌬GY ⫽ ⌬GF – ⌬GC over the line.

1347 (1997). Med. Bradshaw. for cavity formation (59). A. and results tein mutations on the activity of have since been obtained for much anti-HIV drugs (52–54). Other approximate methods have networks around nonpolar groups (55. work included predictions for the showing that better accuracy can be structures of the complexes of achieved by considering alternate efavirenz and TMC125 with descriptors in the LR equations. or 180° reorientation of the ligand. and the water molecules and a few protein ticularly notable because it demonmolecules in these studies certainly residues are presented as stick images. Med. Maryanoff. An approximate It can be expected that these more advanced FEP and TI approaches are the rigor. design in the future. and in many of the com. McGovern. Med. Lahana. 39. exten. ed. duced to a few hours per compound. P. Some most viable for a series of relatively of our own recent contributions rigid analogs. Chem. Pharmacokinetics – Processes. Scheme 6. differences between ligands as in changing also been introduced (64–66) that typically Optimization of the water structure for the core structure (chemotype). Drug Discov. Water molecules complete. J. 56). 2. Eds. E. the sampling and and simulation technology. 4. 6. G. proportional to the solute’s solventanalysis of large substituent efaccessible surface area. Welling. Licinio. M. Discov. K. simulations at the endpoints of a mutation are 9.DRUG DISCOVERY www.-J.sciencemag. Sci. Drug. DC. it also elucidated with COX-2 (60). D 54. 4.. Brunger et al. 447 (1999). P. P. tions. Today 4. so the computing demands are re10. J.. J. Annu. For binding a li(2003). Chem. 439 (2002). 769 (2004). and 148 in796. R. 4). gand to a protein. Comput. ⬃250 NNRTIs design principles for minimizing Fig. which might Opin. L. and it MD or MC simulations on the protein-ligand hibitors and proteins as well as clathrate-like remains difficult to handle large structural complexes. In spite of the approximations includ14. Murcko. 3. B. The latter work is parThe explicit inclusion of water is shown as a CPK model. A. 2664 (2000). V. Biol. LR methods can be viewed as a merger of QSAR inhibitors and protein hosts is well illustrated per compound versus a minute with scoring (quantitative structure-activity relationships) in x-ray structures. the differences in the inter2. A major Today 3. J. and the detailed insights that introduced by Åqvist et al. 43. The inhibitor kinase (63). Rep. when bound water mole. The last larger protein-inhibitor data sets. Chem. L. 169 (1999). J. Faster. Med. as an index fects on Factor Xa inhibition (51). bound to p38 the effects of protein mutations mitogen-activated protein kinase. and p38 on drug activity. Other strengths of the sampling are being pursued. 47. W. A. for two kinases yielded excellent The relevance of discrete water molecules finities. Med. Chem. The 5. Malikayil. D. Chem. V. B.. as reresults (58). Shoicket. Wong.el. P. ing the neglect of intramolecular energetics. Stahl. treat structurally diverse ligands. W. F. D. and they have the cules are resolved. . B. Pharmacogenomics: The in water. Curr. Herzberg. Walters. Chem. binding can be approximated and that only 445 (2000). S. It is expected to be viewed elsewhere (46–48). T. have addressed substituent optiAn early extension of the LR mization for selective COX-2 method to calculate free energies of inhibitors (49). Penning et al. Lck. 11. 1997).convergence of the results are not always advantage of using powerful descriptors inputational studies (Fig. an estimate of the free energy of binding. especially for significant backcluding energy components and estimates of are individual entities that form specific. water and for the protein-inhibitor complexes 7. more approxivent through Poisson-Boltzmann or generalbecome a more common part of inhibitor mate alternatives that still include substantial ized Born/surface area (GB/SA) procedures. the free energy of interaction of a solute with its environment is given by one-half the electrostatic (Coulombic) energy plus the van References and Notes der Waals (Lennard-Jones) energy scaled by 1.-L. R. Drug Discov. J. Opin. The drawbacks are that they are predictions for the third kinase (the predictive that form hydrogen-bonded bridges between comparatively time-consuming (⬃2 days correlation coefficient q2 ⫽ 0. an empirical parameter. 2012 SPECIAL SECTION There have been numerous the approach has yielded promising successful applications. Key advantages over FEP and TI Search for Individualized Therapies (Wiley-VCH. 1099 (2001). M. Pope. The techand elucidation of effects of pronology has evolved. feature simplified representation of the solprotein-ligand complexes can be expected to Hybrid methods. Blake.function approaches). Wei. Mathematics. J. In this modcome more widely used in drug discovery. plus for the LR approach is that it is easy to 12.. 15. Biol. T. methods are that absolute free energies of 8. an anti-inflammatory clinical candidate (67). R. Computed structure from a Monte Carlo simulation for BIRB with HIV-RT (61. Curr. E. from MC or MD simulations for inhibitors in Washington. Acta Crystallogr. required. J.71). 160 (1998). 2002). A. J. can be obtained on variations in binding af. M. the origin of hydration incorporated a third term COX-2/COX-1 selectivity (50). were built using the activity data molecular systems more realistic. J. ␣. H. on June 12. S. Irwin. SCIENCE VOL 303 19 MARCH 2004 1817 Downloaded from www.sciencemag. Chem. Green. Bo¨hm et al. bone and side-chain conformational changsolvation energies that are extracted from the hydrogen-bonding interactions with the in. the remaining protein residues strated that the LR models that makes the representation of the are rendered in the surface.. Fifty-six amino acid residues and hibitors with CDK2. J. 305 be impractical to tackle with FEP calcula(2003). T.approach based on linear response theory was techniques will continue to evolve and besive sampling. 905 actions between the ligand in the unbound (1998). Hester et al. Inf. 13. 62). (57). J.54 to 0. J. 74 water molecules nearest the inhibitor are illustrated. 40. which were subseSome recent large studies have quently confirmed by x-ray crysconsidered 45 celecoxib analogs tallography. state and bound in the complex then provide 4. Curr. Willett. and Applications (American Chemical Society. 38. Laird. Chem. required energy components are obtained 6. HIV-1 RT. Gillet. Scheme 5.

46. 123 (2003). J. Comput. However. Chem. 27. 37. C. lipid emulsions. Allen1* and Pieter R. 41.. 30. 26. L. 44. Am. J. Chem. 28. M. Opin. L. A. 2994 (2002).ca 1818 ticles with diameters of ⬃200 nm or less) for parenteral (primarily intravenous) applications. 66. Med. 6. 151 (2002). J. 33.sciencemag.. Luco. 57. Chem. Comput. Charifson. A. J. 4759 (2000). L. 531 (2002). late-phase clinical trials (table S1). Canada. J. 46. Ostrovsky. Chem. or to function as drug reservoirs (i. polymer microspheres. Dominy. J. 43. 2773 (1990). 16. 49. Jorgensen. L. Chem. A. 10388 (2001). Because each liposome can entrap up to tens of thousands of drug molecules (6). L. H. 396 (1999). M. J. M. R. 7. A. 727 (1997). B. Rev. Trends Biochem. composed primarily of lipids and/ or polymers. 1163 (2003). Med. Kollman. Rizzo. R. J. G. 4370 (2001). 2003). P. Jewsbury. 25. Sci. DDS are designed to alter the pharmacokinetics (PK) and biodistribution (BD) of their associated drugs. even the relatively high carrying capacity of liposomes becomes problematic for very large therapeutic molecules such as proteins. 68. 983 (1998). Chem. P. 2970 (2002). Baldwin. A. For some types of DDS that can carry only a few molecules of a drug (such as immunotoxins and immunoconjugates) or a few tens of molecules (such as polymer conjugates). Sci. Wang. and charge are also important. H. S. Kramer. F. Doman et al. 54. British Columbia V5J 5J8.. 12. Chem. on June 12. 2Department of Biochemistry and Molecular Biology.e. Jorgensen. par- 19 MARCH 2004 VOL 303 SCIENCE www. 2012 SPECIAL SECTION DRUG DISCOVERY . Chem. Email: terry. 63. or biodegradability. Phys. and lipid-drug complexes. D. Chem. 18. P. J. T. Canada. De Ponti. Soc.. Med. Angew. 6016 (2003). 66. Bioorg. 2395 (1993). D. Chem. Drug Deliv. L.. Kollman. C. 50. 3 (1997). 19. 449 (1997). Rev. Additional properties such as stability. 56. Duffy. Sci. J. Tirado-Rives. M.. Wang. 22. J. Many of the pharmacological properties of conventional (“free”) drugs can be improved through the use of drug delivery systems (DDS). W. Murcko. 38. E. Gillet et al. Res. Chem. M. 112. J. 35. Drug Deliv. Udier-Blagovic´. C. J. Rev. Chem. Jorgensen. 45. Chem. 267. 42. Jorgensen. Kuntz. J. 44. C. Chem. VIEWPOINT Drug Delivery Systems: Entering the Mainstream Theresa M. in press. Med. 43. Murcko. W. L. Soc. I. Adv. P. J. Adv. 45. 64. Chem. Taskinen. Simonson. 302 (2003). 32. solubility. Many more DDS are in early. 430 (2002).-P. J. Drug Deliv. Phys. W. Muegge. Med. Jr. Chem. 8133 (1999). D.. Jorgensen. W. W. I. Zhou et al. Lavandera. 62. J. 9455 (2000).allen@ualberta. 61 (1992). Bemis.. 20. Inf. Many of the early problems that hindered the clinical applications of particulate DDS have been overcome. New York. Lett. K. Aided Mol. Bioorg. 1. 60. Med. M. 52. Rognan. R. drug potency is less of an issue for this type of carrier. University of Alberta. 3844 (2002).org Downloaded from www. Chem. Chem. R. Colmenarejo. Biophys. 17. G. J. Klebe. G. A. Folkers. M. W. Gohlke. QikProp. W. Soc. 3337 (2003). then the more potent the drug must be. Biol. Med. 255 (2002). Med. D. Med. Medina. Watkins. L. Mattos. Egan. Feeney. pegylated liposomal doxorubicin (Doxil/Caelyx). Jones et al. L. 23. C. Several DDS have reached the market (Table 2). Methods 44. C. J. Bo¨hm. 61. Chem. 2287 (2003). J. Rarey. J. Essex. L. Adv. These include liposomes and other lipid-based carriers such as micelles. Med. 44. Comput. 34. Edmonton T6G 2H7. J. 2003). M. Chem. v. Vancouver V6T 1Z3. P. Irvine et al. Vinkers et al. Ortiz. J. M. J. Wesolowski. H. F. Chem. 2. 40.or polymer-based nanoparticles can be designed to improve the pharmacological and therapeutic properties of drugs administered parenterally. 385 (1994). T. J. 207 (1994). W. R. E. Ajay. Walters. approved DDS typically undergo additional clinical trials in which they substitute for the free drug in combination chemotherapy. 28 (1999). Karplus. We will not address the use of larger scale systems such as drug-releasing implants or systems used as vaccines or immunostimulants. Med. Udier-Blagovic´. 42. L. 470 (1996). Soc. V. J. University of British Columbia. 65. Am. Tirado-Rives. A. 12898 (2000). 38. 2003). M. Des. J. 122. Lipinski. Lett. Y. 16. Acc. Lesney. Biochim. B. 51. Lombardo. The use of unreasonably high quantities of the carrier can lead to problems of carrier toxicity.-L. 5100 (1999). Several ligand-targeted therapeutics have also received approval (Table 2). Willett. 43. Med. Med. R. Taylor. How can we decide whether a particular therapeutic is suited to delivery in a DDS? Is one type of DDS more suited than another for particular classes of drugs? One of the more important drug properties to consider is potency. L. Res. Drug Deliv. Chem. 21. there is considerable interest in exploiting the advantages of DDS for in vivo delivery of new drugs derived from proteomics or genomics research and for their use in ligand-targeted therapeutics. 3768 (2001). J. 34. 36. New York. N. W. J. A. Protein Eng. G. Rev. T. J. Jorgensen. Merz. W. Pharmacol. J. W. Chem. Jorgensen. Jorgensen. Jorgensen. Ed.. L. Regan et al. 29. 355 (2002). Chem. 1175 (1997). J.e. Curr. Brandsdal et al. P. 2765 (2003). J. J. Rizzo.. Bioorg. 93.. J. Rev. W. Med. and various ligandtargeted products such as immunoconjugates (1–5). 55. J.. J. De Clercq. Rev. Jorgensen. Lamb. Lett. Mol. 11. D. 46. Med. Corkery. Tominaga. W. Mol. with several DDS formulations of anticancer and antifungal drugs now approved for clinical use. 23. 47.3 Drug delivery systems (DDS) such as lipid. Canada. Jorgensen. Toxicol. P. 18.-J. Biol. Chem. the lower the drug: carrier ratio). Bissantz. Lett. 125. Res. Chem. W. J. G. D. G. Murcko. O. Am. Med. Am. Chem. A. 10667 (1995). 27 (2004). Jorgensen. 121. J. Tirado-Rives. W. 53. 27. 4942 (1999). Med. S. Wang. 45. Glide (Schro¨dinger Inc. Chem. Åqvist. Cavalli. 122. J. Furthermore. 39. Plount-Price. 99. K. J. Chem. 2. J. the fewer molecules that a DDS can carry (i. Med. J. G. Des. QikSim. Adv. W. Comput. Table 1 gives examples of problems exhibited by free drugs that can be ameliorated by the use of DDS. J. Chem. Med. J. Prot. 31. 3Inex Pharmaceuticals Corporation. Int. Soc. 1 *To whom correspondence should be addressed. Walters. Med. J. The author is grateful to his co-workers for their dedication and to the NSF and the NIH for research support. J. 59. J. Tirado-Rives. Acta 1545. Chem. D. M. J. S. E.15.1 (Schro¨dinger Inc. Archontis. Udier-Blagovic´. Burnaby. B. R. Chem. 41. 267 (2002). 67 (2001). Comput. M. M. Aided Mol. One such product is a PEG-stabilized liposome. Plount-Price. v. Lengauer. 2644 (2002). P. Adv. W. The majority of the DDS currently approved for parenteral administration fall into the category of liposomal or lipidbased formulations or therapeutic molecules linked to polyethylene glycol (PEG). Cullis2. W. 48..1 (Schro¨dinger Inc.. New York. M. As a general rule. Rizzo et al. Chem. Massova. 55. I. 58. Yliruusi. A. 45. 11. Lipinski. M. C. Inf. which include particulate carriers. Tirado-Rives. J. Here we analyze the opportunities and problems associated with the use of smallscale DDS (nanoparticles and microparDepartment of Pharmacology. J. J. 54. B 105. and their associated therapeutics.-E. Inf. 261. J. Poluzzi. Carlson. 235 (2000). M.-P. or both. L.. as sustained release systems). A. 2791 (2001). W. A. J. Sammuelsson. Although most of the approvals are for DDS used as a monotherapy. S. Today’s Chemist at Work. Recanatini. also included are polymer-drug conjugates. Jorgensen. 45. Chem. 54. 67. Rev. Pharm. T. 24. metabolism and elimination. L. 88. H.. Sci. size (molecular weight). 42. 35. 23. 13. M. 39. Biol. 1541 (2001). Engl. L. Klebe. 2213 (2002). Comput. A. drugs with higher potencies are needed in order to deliver therapeutically relevant amounts of drug (5). 22. Sci. M. 203 (2002). Perez. Bioorg. C.. 46. C. E. Am. 3867 (2000).sciencemag. 5691 (2003). Paiva et al.