key: cord-0985388-e09fyykq
authors: Pang, Y P
title: In Silico Drug Discovery: Solving the “Target‐rich and Lead‐poor” Imbalance Using the Genome‐to‐drug‐lead Paradigm
date: 2006-12-14
journal: Clin Pharmacol Ther
DOI: 10.1038/sj.clpt.6100030
sha: d339a29590408a1bc58f331953725bdaabcb34c2
doc_id: 985388
cord_uid: e09fyykq

Advances in genomics, proteomics, and structural genomics have identified a large number of protein targets. Virtual screening has gained popularity in identifying drug leads by computationally screening large numbers of chemicals against experimentally determined protein targets. In that context, there continues to be a “target‐rich and lead‐poor” imbalance, reflecting an insufficiency of chemists pursuing drug discovery in academia, the challenge of engaging more chemists in this area of research, and a paucity of available protein target structures. This imbalance in manpower and structural information can be ameliorated, in part, by adapting a “genome‐to‐drug‐lead” approach, in which chemicals can be virtually screened against computer‐predicted protein targets, within the context of the US National Science Foundation's petascale computing initiative. This approach offers a solution to reduce manpower requirements for more chemists to experimentally search for drug leads, which represent one of the greatest limitations to drug discovery and better exploits the extensive availability of drug targets at the gene level, ultimately improving the success of moving discoveries from the laboratory to the patient. Clinical Pharmacology &d Therapeutics (2007) 81, 30–34. doi:10.1038/sj.clpt.6100030

The completion of the Human Genome Project in 2003 and recent advances in proteomics and the Structural Genomics Initiative have identified a large number of human proteins as drug targets whose activities can be specifically affected by traditional small organic molecules (chemicals). [1] [2] [3] The human genome has advanced our understanding of the scientific basis of individual variations, and those variations caused by single-nucleotide polymorphism have further increased the number of potential drug targets to an estimated 5,000 (http://www.bio-itworld.com/archive/100902/ firstbase.html). At the same time, the number of chemicals generated by traditional and contemporary approaches has increased dramatically. In theory, there could be as many as 10 47 quadrillion chemicals that can be made to interact with human protein targets. 4 To test this myriad of chemicals, computational screening (virtual screening) can be pursued by iteratively docking each chemical into the active site of a protein target to identify drug leads. [5] [6] [7] [8] Identification is based on the evaluation of the fitness between the two molecules in terms of their shapes and charges. Virtual screening has gained popularity in identifying drug leads with potencies of less than 100 mM by screening chemicals against a protein structure determined by single-molecule X-ray crystallography or nuclear magnetic resonance spectroscopy. In theory, virtual screening is scalable computationally and could yield the desired balance between available therapeutic targets and the identification of drug lead compounds. However, there remains a ''target-rich and lead-poor'' imbalance. Why?

One obvious reason is that, relative to biologists, there is a paucity of organic/medicinal chemists in academia who are supported to experimentally identify drug leads. This situation could be ameliorated by policies directing support to this endeavor. However, a change in policy to support more chemists pursuing drug research will not immediately correct the imbalance. Skilled organic/medicinal chemists are expensive and require time for adequate training, reflecting their need for tacit, rather than explicit, knowledge to create chemicals as potential drugs. It typically takes 4-6 years of training to acquire the tacit knowledge required for drug design and organic synthesis. Therefore, in the context of the shortage of organic/medicinal chemists to search experimentally for drug candidates, computational approaches offer a solution to better balance the availability of therapeutic targets and the identification of drug leads. Technically, virtual screening appears to offer a viable solution to optimize therapeutic drug candidate discovery. A one-teraflops computer is able to perform 31.536 Â 10 18 float point operations per year. In a highly simplified scenario, a dedicated one-teraflops computer in a year could screen 200 million chemicals for each of the 5,000 protein targets, in a year. This scenario assumes that it takes 31.536 Â 10 6 float point operations to screen one chemical against a protein target at a resolution of 1.0-Å translational increment in a 3 Â 3 Â 3-Å 3 docking box and 101 of arc rotational increment in the x, y, and z directions. 9 According to the US National Science Foundation's petascale science and engineering initiative (http://www.nsf.gov/pubs/2005/nsf05625/nsf05625. htm), in 2010 a one-petaflop computer will perform 31.536 Â 10 21 float point operations, with a theoretical capacity to screen 200 billion chemicals for each of the 5,000 protein targets, in a year. Although 200 billion chemicals are a small fraction of 10 47 quadrillion chemicals, this is more than the number of chemicals tested for the development of any clinical therapeutic developed to date.

Although this computational approach clearly offers a solution to better balance the availability of drug targets and the identification of therapeutic lead candidates in the current drug discovery/development paradigm, there are limitations to this virtual screening model. Many human protein targets, especially those with variations caused by single-nucleotide polymorphism, currently do not have three-dimensional (3D) structures defined. This lack of 3D protein structure for targets prohibits the application of virtual screening to identify lead drug candidates. In that context, a new approach is required.

Whereas several approaches can be employed to define 3D protein structures, the primary method is to experimentally determine structures of globular proteins bearing unique folds through the Structural Genomics Initiative. 3 A complementary method is to predict 3D protein structures from their sequences. By combining improved low-and highresolution conformational sampling methods, 1.5-Å -resolution structure prediction has been achieved for small protein domains with less than 85 amino acids. 10 This advance and our own protein modeling experience described below suggest that virtual screening can be expanded to screen chemicals against protein targets whose active site-containing domain or subdomain is predicted computationally, an ambitious approach we term ''genometo-drug-lead''. 11 To illustrate feasibility, we built a dedicated 1.1 teraflops computer ( Figure 1 ) to run multiple molecular dynamics simulations (MMDSs) in parallel. The stochastic sampling of protein conformations achieved by MMDSs is more efficient than sampling by a single long molecular dynamics simulation. [12] [13] [14] [15] [16] [17] The efficiency of the stochastic sampling is demonstrated by MMDSs of the ubiquitin E2 variant domain of human tumor susceptibility gene 101 protein in complex with a peptide ligand in explicit water. 18 Here, the MMDSs comprise 200 different 10-ns molecular dynamics simulations (2 Â 10 6 snapshots) of the complex for which nuclear magnetic resonance data are available. 18 The trajectories obtained during the first 5 ns period of the MMDSs reproduce B92% of the protein-protein nuclear Overhauser effects and B85% of the protein-peptide nuclear Overhauser effects (YP Pang and P Dasgupta, unpublished data). Given this sampling efficiency, MMDSs could refine a low-resolution protein domain, which is readily obtained from homology modeling or threading, to a high-resolution protein domain. For example, MMDSs refined a homology model, provided by the Protein Structure Prediction Centre (http://predictioncenter.org/caspR/), to a computer model that was nearly identical to the corresponding crystal structure (Protein Data Bank ID: 1XE1). Relative to the 1XE1 crystal structure, the alpha carbon root mean square deviation of the computer model was 1.7 Å , whereas the alpha carbon root mean square deviation of the homology model was 4.6 Å (Figure 2 , unpublished work of Pang).

In the context of this advanced performance, we applied homology modeling and MMDSs to predict a 3D model of a chymotrypsin-like cysteine proteinase (CCP) from a severe acute respiratory syndrome-associated coronavirus. 11,17 CCP is an ideal drug target for treating severe acute respiratory syndrome viral infection because it is required for viral replication and transcription. Here, 200 different molecular dynamics simulations of monomeric CCP in explicit water (4.0 ns for each simulation with a 1.0 fs time step and different initial velocities) were executed to refine the homology model. 11, 17 Then, we screened 361,413 chemicals against the CCP model refined by MMDSs and identified 12 chemicals for antiviral testing. Of the 12 chemicals tested in cell-based inhibition assays, one inhibited the human severe acute respiratory syndrome-coronavirus Toronto-2 strain with a concentration of ligand that produces half of the maximum response of 23 mM and four others exhibited 13-17% inhibition at a drug concentration of 32 mM. 11 The most potent inhibitor lead overlays well with a reported substrate fragment (ATVRLQ p1 A p1' ) bound in the active site of CCP (Figure 3) 17 . These results demonstrate that, given target information at the gene level only, virtual screening can identify chemicals that penetrate and rescue cells from viral infection. It is noteworthy that this genome-to-drug-lead approach leapfrogs the requirements for experimental determination of protein target structure and cell-free assays to confirm molecular interactions.

Interestingly, CCP exists in a homodimer in which only one of the two monomers is active. 19 Thus, simulation of monomeric CCP may lead to a structure that is not representative of the active CCP. In fact, many protein targets are functional only in a multimeric form. With target information at the gene level only, it is difficult to deduce the precise multimeric form required for the function of the protein target, let alone the challenge of simulating proteins in their multimeric forms. This problem appears to mitigate against the use of the genome-to-drug-lead approach. However, information concerning ternary structure is not required if virtual screening is searching for inhibitors (not activators) of protein targets. Indeed, an inhibitor lead identified from the inactive, monomeric CCP binds to the monomeric CCP, and possibly to the dimeric CCP as well. While binding to dimeric CCP can certainly inhibit CCP, binding to monomeric CCP also can inhibit CCP because the dimer is in equilibrium with the monomer and binding to the monomer can convert the active dimeric CCP to the inactive monomeric CCP. This explains why a 23 mM inhibitor lead was successfully identified using the inactive monomeric CCP in virtual screening.

While virtual screening identified compounds that inhibited CCP activity in cell-based assays, this approach did not empirically validate model predictions by examining direct molecular interactions in cell-free systems. However, a screen using the same CCP model but contracted by 12% ( Figure 4) failed to identify the 23 mM inhibitor ( Figure 5 ). 11 Moreover, it identified two weak inhibitors that are structurally very similar to the 23 mM inhibitor. 11 These observations demonstrate that the identification of a drug lead is sensitive to a change of the structure used in virtual screening, implying the interaction of the lead with CCP, and thereby confirms the validity of virtual screening using the inactive monomeric CCP. Further, it demonstrates that leapfrogging the cell-free assay has the advantage of avoiding identification of both toxic inhibitors and inhibitors that have high affinities for CCP but cannot penetrate cells.

The best CCP inhibitor lead has a concentration of ligand that produces half of the maximum response of only 23 mM, which raises the question of its utility. Using the traditional trial-and-error approach, it is certainly difficult to improve the potency of a 100 mM lead by several orders of magnitude. For this reason, the definition of a drug lead is commonly defined as a chemical possessing an inhibitory potency less than 50 mM. However, using MMDSs to guide structural modification, we improved an inhibitor lead of a zinc endopeptidase in botulinum neurotoxin serotype A from 15% inhibition at a drug concentration of 100 mM to 19% inhibition at 2.5 mM. 20 This demonstrates that the 23 mM inhibitor is useful as a drug lead, especially because its potency was determined by a cell-based assay. It also suggests that a drug lead can be re-defined as a chemical possessing an inhibitory potency less than 100 mM.

To further appreciate the value of drug leads obtained from virtual screening, it is worth discussing the goal of virtual screening because this goal may differ among research groups and change over time. In 2000, our goal of virtual screening was to identify a subset of chemicals enriched in active inhibitors 5 based on the gigaflops (10 9 floating point operations per second) computing technology available then. In 2006, we have the same goal, even though 3.8 teraflops (3.8 Â 10 12 floating point operations per second) computing technology has become available. The goal has not changed because terascale (10 12 floating point operations per second) computers remain insufficiently fast to identify drug candidates. Drug discovery relies on organic/medicinal chemists who have the tacit knowledge to create chemicals as drug candidates with the aid of fast computers. We do not anticipate that virtual screening can identify drug candidates that are ready for preclinical studies. Rather, we expect that virtual screening can offer drug leads as building blocks for drug candidates. Virtual screening cannot create new chemicals, but organic chemists can use leads as building blocks to create new chemicals. That is the value of drug leads obtained from virtual screening.

An unusually large computing resource was used to search for CCP inhibitors. Would the genome-to-drug-lead approach be practical for a typical academic research laboratory? Indeed, it is practical for several reasons. First, the cost A relatively small database of 361,413 chemicals was used to identify CCP inhibitors. However, this success of identifying CCP inhibitor leads from a relatively small library of chemicals does not address the feasibility of docking targets with 200 billion chemicals. In that context, data storage is a common problem in bioinformatics. Do we have enough disk space to store 200 billion chemicals to be screened as discussed in the previous section? The answer is yes. The average disk space to hold one chemical with a molecular weight in a range of 380-420 is 973 bytes, according to the database designed by this author. A welldesigned database of 200 billion chemicals is estimated to In summary, given advances in protein structure prediction and the change in computing speed, from gigascale in the past, to terascale in the present, and to petascale in the near future, it is clear that the genome-to-drug-lead approach is feasible and has broad application in drug discovery.

The potential impact of the genome-to-drug-lead approach on clinical pharmacology and therapeutics is the increase in the number of drug candidates that can be identified and moved from the laboratory into clinical trials. However, the impact goes further. The genome-to-drug-lead approach permits the docking of one drug candidate against an array of human proteins to predict drug interactions and toxicity, and effectively address individual variations of a drug target caused by single-nucleotide polymorphism for personalized medicine. A slight modification of the genome-to-drug lead approach can dock the computer-identified drug candidate against human serum albumin to predict protein binding and, by extension the distribution, of drug candidates.

A large number of drug targets and a paucity of organic/ medicinal chemists in academia pursuing drug discovery research have created a ''target-rich and lead-poor'' imbalance. Skilled organic/medicinal chemists are expensive and take time to train and it is difficult to engage more academic organic/medicinal chemists in drug research to remediate this imbalance in the short term. Virtual screening can successfully identify drug leads against experimentally determined drug target structures. Given current terascale computers, and petascale computers in the near future, virtual screening can be expanded to screen chemicals against target structures predicted from genes by computers, a paradigm termed ''genome-to-drug-lead''. This approach can reduce the formidable manpower requirements for more chemists to search experimentally for lead compounds, help resolve the imbalance between disease targets and therapeutic agents, and, ultimately, enrich the drug development pipeline to move discoveries from the laboratory into patients.

The sequence of the human genome

Proteomics: new perspectives, new biomedical opportunities

Structural genomics: beyond the Human Genome Project

Drug discovery -contemporary small molecule drug discovery -Tutorial: stacking the deck in favor of drug-like leads

Successful virtual screening of a chemical database for farnesyltransferase inhibitor leads

Chemical database techniques in drug discovery

Protein-ligand docking: current status and future challenges

Receptor-ligand binding sites and virtual screening

EUDOC: A computer program for identification of drug interaction sites in macromolecules and drug leads from chemical databases

Toward high-resolution de novo structure prediction for small proteins

From genome to drug lead: identification of a small-molecule inhibitor of the SARS virus

Locally accessible conformations of proteins -multiple molecular dynamics simulations of crambin

Assessing equilibration and convergence in biomolecular simulations

Absolute comparison of simulated and experimental protein-folding dynamics

Simulation of folding of a small alpha-helical protein in atomistic detail using worldwidedistributed computing

Modeling domino effects in enzymes: molecular basis of the substrate specificity of the bacterial metallo-beta-lactamases IMP-1 and IMP-6

Three-dimensional model of a substrate-bound SARS chymotrypsin-like cysteine proteinase predicted by multiple molecular dynamics simulations: catalytic efficiency regulated by substrate binding

Structure of the Tsg101 UEV domain in complex with the PTAP motif of the HIV-1 p6 protein

Only one protomer is active in the dimer of SARS 3C-like proteinase

Serotype-selective, small-molecule inhibitors of the zinc endopeptidase of botulinum neurotoxin serotype A

The author's work described here was supported by the Defense Advanced Research Projects Agency (DAAD19-01-1-0322), the US Army Research Office (DAAD19-03-1-0318), the US Army Medical Research Acquisition Activity (W81XWH-04-2-0001), the National Institutes of Health (5R01AI054574-03 and 5R01GM061300-06), the IBM Blue Gene Life Sciences Center of Excellence, the Mayo Clinic-IBM Center of Excellence, the High Performance Computing Modernization Program of the US Department of Defense, the San Diego Supercomputing Center, the University of Minnesota Supercomputing Institute, the Compaq Medical Sciences Group, the Jay and Rose Phillips Family Foundation, and the Mayo Foundation. The opinions or assertions contained herein belong to the author and are not necessarily the official views of the US Army, the US Department of Defense, or the National Institutes of Health.

The author declared no conflict of interest.& 2007 American Society for Clinical Pharmacology and Therapeutics