Conformation Mining: An Algorithm for Finding Biologically Relevant


Conformation Mining: An Algorithm for Finding Biologically Relevant...

0 downloads 79 Views 265KB Size

J. Med. Chem. 2005, 48, 3313-3318

3313

Conformation Mining: An Algorithm for Finding Biologically Relevant Conformations Santosh Putta,* Gregory A. Landrum, and Julie E. Penzotti Rational Discovery LLC, 555 Bryant Street #467, Palo Alto, California 94301 Received November 18, 2004

Discovering essential features shared by active compounds, an important step in drug-design, is complicated by conformational flexibility. We present a new algorithm to efficiently mine the conformational space of multiple actives and find small subsets of conformations likely to be biologically relevant. The approach identifies chemical and steric similarities between actives, providing insight into features important for binding when structural data are absent. Validation studies (thrombin and CDK2 data) produce alignments similar to protein-based alignments. 1. Introduction In the absence of structural data to use in developing structure-activity relationships (SARs), biological screening data is often used to derive insight into the binding of ligands to proteins. Among the many strategies that have been developed to do this, pharmacophores1-4 and shape-feature methods5-8 have played important roles in drug discovery. These approaches, which attempt to identify molecular shapes (containing chemical-feature locations) or pharmacophores that are associated with activity, typically use an ensemble of low-energy conformations to account for the flexibility of drug-like compounds. Published studies, comparing ligand conformations generated with conformational analysis tools to conformations from X-ray structures, have demonstrated that the implicit assumption that one (or more) of these conformations is similar to the binding conformation is often valid.9,10 Some model building methods search through all conformations directly in order to identify pharmacophores and/or shape-feature configurations that are associated with active compounds (and, in some cases, that are not associated with inactives1). Other methods represent each compound as a binary fingerprint where bits are set if particular pharmacophores or shape-feature configurations are matched.2,3,5,11 These fingerprints typically contain the union of the information present in all of a compound’s conformations. Ensembles of bits that appear frequently in active compounds have been used to screen databases and virtual libraries in order to find new leads.2,12 This approach can be useful when screening large numbers of compounds. However, conformational flexibility and biases in the numbers and/or chemical diversity of the actives and inactives composing a data set make it quite common to find pharmacophores and shapes that appear frequently in active compounds but that do not represent features essential for binding. An alternative machine-learning approach, known as multiple-instance learning, has been applied to address this problem.13,14 In this framework, a compound is considered to be a collection of multiple instances, one * Corresponding Author: S. Putta, [email protected], tel: 510-742-0486.

for each conformation. An active compound contains one unknown active instance (the remaining conformations are considered to be inactive) while all instances of an inactive compound are considered to be inactive. Multiple-instance learning has shown some success in building models for activity;13,14 however, it is not clear how the active instances discovered by the method compare to binding conformations. Furthermore, the information obtained from these studies is specific to individual compounds; no alignment of the multiple active compounds is provided. Here we present a new approach that can be used to mine the conformational space of multiple active compounds for a target without using any X-ray crystal structure information in order to find the following: three to five of the conformations from each active compound that are most likely to be similar to its binding conformation, and alignments of these conformations that highlight the steric and chemical-feature overlap across multiple actives. This conformation-mining algorithm is not a structurebased approach; however, we demonstrate that the resulting conformations and alignments are very similar in terms of both molecular shape and chemical-feature locations to the bound conformations from X-ray crystal structures. Conformation mining provides a visual means to inspect both the steric and chemical features that are important for binding. The algorithm can also be used to generate multimolecule alignments and to provide information-rich inputs to model-building algorithms; the mined conformations and their alignments can reduce the “noise” associated with using the entire conformational model. 2. Method The conformation-mining algorithm discovers and exploits the common steric and chemical features of a set of active compounds, each of which has multiple conformations. Figure 1 illustrates the idea that motivated our use of shapes and features to mine biologically interesting conformations. Each compound in Figure 1 is shown with multiple shapes (one per conformation) and feature points. Assuming that all of the actives bind to the same site, we would expect their binding conformations to have similar shapes. Comparing A to B, one notices that they share two common shapes (the arc

10.1021/jm049066l CCC: $30.25 © 2005 American Chemical Society Published on Web 04/09/2005

3314 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9

Putta et al.

Figure 1. Illustration of the steric and chemical-feature similarity between multiple active compounds. A, B, C represent compounds, and the shapes in each of the boxes are conformations. The circles on the shapes indicate locations of various chemical-feature types. and the rectangle); one (or both) of these shapes is likely important for binding. Now looking at compound C, it is clear that one of these common shapes (the arc) is shared by C, further reinforcing the importance of this shape. All three compounds share two common chemical features (highlighted with black outlines in Figure 1). Our conformation-mining approach sifts through the collection of shapes for each compound looking for such steric and feature commonalities. Clearly these assumptions about similarity need not apply to active compounds that bind in different ways to the target; for the purposes of this work we will limit ourselves to compounds that bind similarly. Future research will focus on extending the conformation-mining method to detect cases where there may be multiple binding modes. Earlier efforts,15,16 which constrained conformational space using a pharmacophore (collection of “functional groups”) derived from previously examined active compounds and/or structural information, require prior knowledge of a biologically relevant pharmacophore as well as its mapping onto each active molecule. In our algorithm, no prior knowledge of the important chemical features is needed. It is also not necessary for all of the active compounds in the data set to contain a common pharmacophore. The conformation-mining algorithm, described in detail below, requires methods to align two conformations and to score the resulting alignments. Two previously published approaches, subshape alignment and feature maps, are used in this study. 2.1. Subshape Alignment. To align conformations from various active molecules, we use subshape alignment; details of this method have been previously published.17 The subshape alignment algorithm was developed to enable the alignment of molecules of different sizes. The methodology also provides a means of computing multimolecule alignments: a probe molecule can be aligned to the union shape from a collection of molecules (a combined shape). Subshape alignment has no difficulty finding the steric similarities between small probe molecules and these large combined shapes. The unique capabilities of subshape matching, such as combining multiple conformations into a single shape or aligning differently sized shapes to each other, are used in the conformation-mining algorithm presented below. However, with some modifications, other molecular alignment techniques currently available in the literature6,18-20 could also be used to provide the alignments for our conformation-mining algorithm. 2.2. Feature Maps. The subshape method provides alignments between conformations; feature maps, collections of chemical feature points in three-dimensional space, are used to score these alignments.17 Each point in the feature map is represented by a Gaussian function whose position is derived from the conformation used to generate the target shape. The score for an alignment of a probe conformation is computed by summing the overlaps of Gaussian functions centered at each of the probe’s feature locations with the closest Gaussians of a similar feature type in the map. Thus a large score results from close overlays of the feature points in the conformation with those in the map. A standard set of six feature types was used: positive and negative ionizable groups, hydrogen-bond donors and acceptors, aromatic groups, and hydrophobic centers.21 As in the case of a combined shape, a combined feature map can be constructed from the feature maps of a set of aligned

Figure 2. Illustration of the conformation-mining approach. See the text for discussion. conformations. To combine two feature maps, the weights of the Gaussians at each feature point are increased based on the distance to the closest feature point of the same type in the second map. Feature points of the same type that lie within a cutoff distance are coalesced to form a single new point with an averaged weight and location. Further details on feature maps and the method for combining them are presented in the Supporting Information. 2.3. Conformation Mining. The input to the conformationmining algorithm is a list of active compounds, each of which has multiple (tens to hundreds) low-energy conformations. Given the large number of conformations for each compound, a brute-force search for similar conformations would result in a computationally intractable, combinatorial problem. In earlier work on multimolecule alignment, Mestres et al. examined each pair of conformations from all the compounds, thereby limiting both the number of compounds and number of conformations per compound that could be handled. 22 We instead adopt a “learn as we go”, iterative approach, where at each stage the conformations from a single new compound are added, and the steric and chemical similarities are learned and stored. This steric and chemical-feature information is then used to align and rank the conformations of the next compound. The overall algorithm is illustrated in Figure 2. All pairs of conformations between the initial compounds A and B are aligned (using subshape alignment) and ranked (using feature map scores). The shape and feature information from the top ranking alignments, the white boxes in Figure 2, is then combined (by forming combined shapes and combined feature maps) and used to align and rank the conformations of a new compound C. The process is repeated for each of the remaining training compounds in turn. The conformation-mining algorithm is order dependent; the best results are obtained when the initial compounds are large (relative to the other actives) and as dissimilar as possible. Although the process illustrated in Figure 2 can theoretically be repeated for as many actives as one wishes, combined shapes, which are formed using the union of individual shapes, tend to become globular as the number of compounds becomes larger. These globular shapes are not particularly selective and give poor-quality alignments when used in subshape matching. As a result, in this study we limit ourselves to three to four diverse actives for each dataset; we will demonstrate that this is sufficient to discover biologically relevant conformations. It is important that the actives selected for conformation mining be relatively diverse. The algorithm mines the conformational space to find conformations that are sterically and chemically similar to each other; if two compounds are very similar, every conformation of one will have a close match on the other. This causes the search for similar conformers to become too promiscuoussmost pairs of conformations will satisfy the similarity criteriasmaking it difficult to extract any useful information. The work described here was performed using a modified version of the CombiCode software licensed from DeltaGen Inc. Conformational models were generated using CONAN.10 2.4. Validation Data. Although the conformation-mining algorithm is designed to be used in the absence of crystalstructure data, for the purposes of validation we have chosen two well-studied targets where bound-ligand crystal structures

Conformation Mining

Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3315

Table 1. Comparison of the Conformational Models of the Thrombin Ligands and Their Crystal Conformersa number of closest farthest Ki molecule CONAN conformations (Å) (Å) (nM) ref 1ETS 1ETR 1A61 1UVT

100 100 200 46

1.67 1.29 1.26 1.36

4.28 3.76 4.10 3.13

7 1 2 23

22 22 23 24

aThe root-mean-square deviation (RMSD) between the CONAN conformer and the bound conformer in the crystal structure is listed for the closest and farthest CONAN conformer.

Figure 5. The top-ranking alignment from conformation mining thrombin ligands (center) versus the protein-based alignment of the X-ray crystal structure conformations (left). Also shown is a high-ranking alignment from a control experiment (right), where the X-ray crystal structure conformation for each of the ligands was added to its CONAN conformations. The 1A61 thiazole ring (circled in blue) does not overlap with the other ligands in the protein-based alignment. However, the shape-based component of the conformation-mining approach results in a closer alignment in this region as seen in the top-ranked alignment and control experiment results. Figure 3. Structures of the thrombin inhibitors used in this study.

Figure 4. Structures of the CDK2 inhibitors used in this study. Table 2. Comparison of the Conformational Models of the CDK2 Ligands and Their Crystal Conformers number of closest farthest IC50 molecule CONAN conformations (Å) (Å) (µM) ref 1FVV 1OIR 1H01 1KE8

17 94 90 29

0.796 0.66 1.08 0.78

2.59 3.94 4.27 2.39

0.01 0.01 1.0 1.0

25 26 27 28

are available in the PDB:23 thrombin and cyclin-dependent kinase 2 (CDK2). 2.4.1. Thrombin. Compounds were chosen from a set of 25 X-ray crystal structures. The four compounds selected were chosen to be as diverse as possible. The PDB codes for the thrombin structures bound to these four compounds are 1ETS, 1ETR, 1A61, and 1UVT (Table 1 and Figure 3). 2.4.2. CDK2. Four compounds were selected from a collection of eleven aligned crystal structures. The four diverse compounds chosen are shown in Table 2 and Figure 4.

3. Results & Discussion 3.1. Thrombin. To perform conformation mining for thrombin, 1A61 and 1ETS were chosen as the initial pair of ligands. The conformations of 1ETR and 1UVT were then added, in that order. The top ranking alignment is shown in Figure 5. Many of the high-ranking alignments for the four compounds are quite similar to the crystal structure

alignment obtained by aligning the protein structures. Since subshape matching generates alignments using steric similarities, the conformation-mined alignments tend be more compact than those from the crystal structures. For example, the location of the thiazole ring in 1A61 (circled in Figure 5), which does not overlap with portions of the other ligands from the remaining crystal structures, is not well reproduced in the results. Some of the other differences between the conformationmining results and the crystal structures can be attributed to weaknesses in the conformational model: the relatively large RMSD values in the Closest column of Table 1, compared to those in Table 2, indicate that the CONAN conformations are not particularly close to the crystal conformers. To further investigate this issue, a control experiment was conducted by adding the X-ray crystal-structure conformation of each ligand to its list of CONAN conformations. The conformation-mining algorithm was then applied to these expanded conformational models. While the top-ranked alignments from this experiment consist of CONAN conformations that allow tighter overlap of the ligands than observed in the protein-based alignments, crystal-structure conformations of the ligands also appear in very highly ranked alignments. Figure 5 shows a high-ranking alignment (ranked 20th out of 1164 total alignments) from this experiment that consists of the crystal conformers for 1ETS, 1UVT, and the closest CONAN conformation for 1ETR. The conformation of 1A61 that appears in this alignment is not very close to the crystal conformation (RMSD ) 2.2 Å), an unsurprising result considering the poor steric overlap between 1A61’s crystal conformer and the other ligands. These results indicate that, within the limits of the steric-alignment approach used, conformation mining is capable of discovering crystal conformers when they are present in the data. The similarities between the crystal structures and high-ranking conformation-mining alignments are not limited to shape; chemical feature locations are also well reproduced. Figure 6 shows the final feature maps constructed from the crystal structures and the topranking conformation-mining alignment (see Supporting Information for the feature locations and weights in the

3316

Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9

Putta et al.

Figure 6. Comparison of the feature maps generated using the crystal structures and the top-ranking conformationmining alignments, shown in Figure 5. Important regions of the crystal-structure feature map are highlighted and labeled following Banner et al.24

top-ranking map). The feature points are drawn as spheres with radii proportional to their weights (indicating the degree of feature overlap across the four compounds). Several of the features that are important for thrombin binding are recognized as high-overlap features by the conformation-mining method. For example, the positive charge feature in the P1 pocket shows good overlap across all four ligands in both the crystal structures and the conformation-mining results. The strong acceptor interaction at P3 and hydrophobic interactions in the D pocket are also reproduced well. The absence of a well-defined P pocket in the conformation-mining results can be explained by the weaknesses in our conformational model and the steric biases introduced by the alignment method discussed above. The feature map from our control experiment (where crystal conformations were included in the conformational model), which does show a collection of feature points in the P pocket area, supports this conclusion. To ensure that our scoring methodology is valid and that the promising results shown in Figure 6 are not due to chance, we performed an additional experiment where alignments were ranked randomly (instead of using feature-map scores). As expected, the feature maps obtained from this process did not resemble those from the crystal structure: key feature groups, such as the positive ionizable group in the P1 pocket, were not identified by this randomized experiment (see Supporting Information for additional details and figure). The conformation-mining algorithm is formally order dependent. To test the impact of this dependence on the results, the above process was repeated starting with 1ETR instead of 1A61. Results of the same quality were obtained (see Supporting Information). In addition, the conformations that appear in the top ranking alignments were repeated across the two runs, indicating that the order dependence of the method does not have a strong impact on the results for thrombin. 3.2. CDK2. The conformation-mining algorithm was applied to the following three CDK2 ligands in order: 1FVV, 1OIR, and 1H01. The top-ranking alignments are shown in Figure 7 for these compounds. Once again, there is substantial agreement between the conformation-mined and crystal-structure alignments and the resulting feature maps are quite similar (see Supporting Information for the feature locations and weights in the top-ranking map). All three regions

Figure 7. Comparison of the crystal structure alignments with the top-ranked conformation-mining results. Regions of the crystal-structure feature map that have been identified as important for binding are labeled following the literature.25,26

of the crystal-structure feature map that have been identified as important for binding25,26 are well reproduced in the conformation-mining results. The acceptor and donor features of the ligands in the vicinity of Leu83 are clearly identified in the conformation mining results; hydrogen bonds between ligands and the carbonyl O and/or amide N of Leu83 are commonly observed interactions in the kinase/inhibitor complexes. The hydrophobic/aromatic regions are also in good agreement with the crystal-structure feature map. The major differences occur in the Asp86 region (at the bottom of Figure 7), an area near the opening of the binding cleft where variation in the type of ligand-protein hydrogenbonding interactions is observed. Sulfonamide oxygen atoms of 1FVV and 1KE8 form hydrogen bonds to the carboxylate group of Asp86,27,28 while for 1H01, a bridging water mediates a hydrogen bond from Asp86 to the phenoxy oxygen.29 In the crystal structure of 1OIR, no hydrogen bonds are observed between the ligand and Asp86, and the ligand is poorly ordered beyond the hydroxyl group.25 The conformation-mining algorithm, with its emphasis on steric-similarity, finds conformations that are more tightly clustered in the Asp86 region. In the kinase-ligand complexes 1FVV and 1KE8, the pyridyl and thiazole substituents on the sulfonamide groups (respectively) are exposed to solvent. Without structural information, it would be difficult for any computational method to accurately predict the hydrogen bonding interactions and conformations in this solvent-accessible region. Similar results were obtained from an additional experiment where 1H01 was replaced with 1KE8 (see Supporting Information). 3.3. Quantitative Evaluation of Conformation Mining Alignments. A final experiment was carried out to evaluate the utility of the information contained in the top-ranking alignments from conformation min-

Conformation Mining

Figure 8. Enrichment plot for picking thrombin actives using crystal-structure alignments (dashed red line) and conformation-mining results (solid blue line).

ing for developing predictive models. The combined shape and feature map from the top-ranking alignments of the thrombin actives 1A61, 1ETS, 1ETR, and 1UVT were used to screen a set of 74 compounds: 37 thrombin actives obtained from the literature30 and 37 compounds chosen randomly from a screening library of drug-like molecules31 that were assumed to be inactive. The inactives were constrained to contain a positively charged group and to be approximately the same size as the actives. Compounds were scored by aligning each of their CONAN conformations to the combined shape and scoring the alignments using the combined feature map; the highest scoring alignment was selected. The compounds were ranked by score, and an enrichment plot was generated. To establish a baseline for accuracy, the experiment was repeated using the combined shape and feature map from the crystal structures (in the proteinaligned orientation) for 1A61, 1ETS, 1ETR, and 1UVT. The enrichment plots for the two runs, Figure 8, are quite similar: both show about the same (significant) improvement over randomly selecting compounds. This simple experiment has several shortcomings. Because we do not have access to a large thrombinscreening data set, we had to assume that compounds randomly selected from a general-purpose screening library are inactive. We would expect a general screening library to contain a very small number of thrombin actives (and most likely weak actives); however, we do not know their identity. Also, in this scheme the scores were not normalized to account for factors such as the size of the compounds or the number of features they have. Normalization of the scores may improve the enrichment as well as the diversity of actives identified by this method. Despite these shortcomings, the results of this experiment illustrate that the conformationmining approach provides information that is useful for identifying active compounds and suggests several new approaches to descriptor generation and model development that may lead to a virtual screening tool. 4. Conclusions In the validation studies presented here, the conformation-mining algorithm was used to sift through the conformational spaces of sets of active molecules and identify a small collection of conformations that are likely to be biologically active. In both studies, the resulting multimolecule alignments, generated without the benefit of X-ray structural data, produced feature

Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3317

maps that provide information about the requirements for ligand-protein binding. The resulting feature maps are very similar to maps derived from the crystal structures. An additional experiment, using a simple predictive model to select thrombin actives, demonstrated that the conformation-mining feature map contains information that can be utilized in virtual screening. The aim of our future work is to improve several aspects of this approach. Our results suggest that the quality of the alignments can be further improved by using directionality constraints; we will increase the descriptive power of feature maps by adding directionality to some chemical features (acceptor, donor, and aromatic groups). Alternate approaches to generating combined shapes will also be explored so that larger numbers of actives can be used. Since the top-ranking alignments seem to contain useful information for identifying new actives, some of our future effort will be dedicated to developing novel descriptors and predictive models for activity, based on the feature maps. Last, shape and feature alignments of actives that belong to the same binding mode exhibit notable differences when compared to alignments of actives that adopt different binding modes. We will pursue several approaches to the identification of these distinctions and automatic perception of multiple binding modes. Acknowledgment. We would like to thank David Spellmeyer (IBM) and Michelle Lamb (Astra Zeneca) for their helpful comments and suggestions. In addition, the authors would like to thank Erin Bradley (Sunesis Inc.) for providing the aligned CDK2 crystal structures. Supporting Information Available: Details on feature maps and the method to combine multiple feature maps. In addition, conformation-mining results for an alternate ordering of thrombin ligands and an alternate selection of molecules for CDK2 are presented. Finally, coordinate and feature weight information for the top-ranking maps shown in Figures 6 and 7 are also provided. This material is available free of charge via the Internet at http://pubs.acs.org.

References (1) Barnum, D.; Greene, J.; Smellie, A.; Sprague, P. Identification of Common Functional Configurations Among Molecules. J. Chem. Inf. Comput. Sci. 1996, 36, 563-571. (2) Mason, J. S.; Morize, I.; Menard, P. R.; Cheney, D. L.; Hulme, C.; Labaudiniere, R. F. New 4-Point Pharmacophore Method for Molecular Similarity and Diversity Applications. J. Med. Chem. 1999, 42, 3251-3264. (3) Bradley, E. K.; Beroza, P.; Penzotti, J. E.; Grootenhuis, P. D. J.; Spellmeyer, D. C.; Miller, J. L. A Rapid Computational Method for Lead Evolution: Description and Application to R1-Adrenergic Antagonists. J. Med. Chem. 2000, 43, 2770-2774. (4) Putta, S.; Lanctot, K.; Srinivasan, J. S.; Lamb, M. L.; Lemmen, C.; Eksterowicz, J. Mitigating the Sensitivity of Pharmacophore Fingerprints. In Machine Learning In Computational Chemistry, 225th American Chemical Society National Meeting, New Orleans, LA, 2003. (5) Putta, S.; Lemmen, C.; Beroza, P.; Greene, J. A Novel ShapeFeature Based Approach to Activity Model Building. J. Chem. Inf. Comput. Sci. 2002, 42, 1230-1240. (6) Goldman, B. B.; Wipke, W. T. Quadratic Shape Descriptors. 1. J. Chem. Inf. Comput. Sci. 2000, 40, 644-658. (7) Hahn, M. Three-Dimensional Shape-Based Searching of Conformationally Flexible Compounds. J. Chem. Inf. Comput. Sci. 1997, 37, 80-86. (8) Jain, A. N.; Dietterich, T. G.; Laterop, R. H.; Chapman, D.; Critchlow, R. E.; Bauer, B. E.; Webster, T. A.; Lonzano-Perez, T. Compass: A shape-based machine learning tool for drug design. J. Comput.-Aided Mol. Des. 1994, 8, 427-440. (9) Bostrom, J.; Greenwood, J. R.; Gottfries, J. Assessing the performance of OMEGA with respect to retrieving bioactive conformations. J Mol. Graph. Model. 2003, 21, 449-462.

3318

Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9

(10) Smellie, A.; Stanton, R.; Henne, R.; Teig, S. Conformational analysis by intersection: CONAN. J. Comput. Chem. 2003, 24, 10-20. (11) McGregor, M. J.; Muskal, S. M. Pharmacophore fingerprinting. 1. Application to QSAR and focused library design. J. Chem. Inf. Comput. Sci. 1999, 39, 569-574. (12) Lanctot, J. K.; Putta, S.; Lemmen, C.; Greene, J. Using ensembles to classify compounds for drug discovery. J. Chem. Inf. Comput. Sci. 2003, 43, 2163-2169. (13) Dietterich, T. G.; Lathrop, R. H.; Lozano-Per, T. Solving the Multiple Instance Problem with Axis-Parallel Rectangles. Artif. Intell. 1997, 89, 31-71. (14) Dooly, D. R.; Zhang, Q.; Goldman, S. A.; Amar, R. A. MultipleInstance Learning of Real-Valued Data. J. Machine Learning Res. 2002, 3, 651-678. (15) Dammkoehler, R. A.; Karasek, S. F.; Shands, E. F.; Marshall, G. R. Constrained search of conformational hyperspace. J. Comput.-Aided Mol. Des. 1989, 3, 3-21. (16) Mayer, D.; Naylor, C. B.; Motoc, I.; Marshall, G. R. A unique geometry of the active site of angiotensin-converting enzyme consistent with structure-activity studies. J. Comput.-Aided Mol. Des. 1987, 1, 3-16. (17) Putta, S.; Eksterowicz, J.; Lemmen, C.; Stanton, R. V. A Novel Subshape Molecular Descriptor. J. Chem. Inf. Comput. Sci. 2003, 43, 1623-1635. (18) Lemmen, C.; Lengauer, T.; Klebe, G. FlexS: A Method for Fast Flexible Ligand Superposition. J. Med. Chem. 1998, 41, 45024520. (19) Mestres, J.; Rohrer, D. C.; Maggiora, G. M. MIMIC: A molecularfield matching program. Exploiting applicability of molecular similarity approaches. J. Comput. Chem. 1997, 18, 934-954. (20) Kramer, A.; Horn, H. W.; Rice, J. E. Fast 3D molecular superposition and similarity search in databases of flexible molecules. J. Comput.-Aided Mol. Des. 2003, 17, 13-38. (21) Greene, J.; Kahn, S.; Savoj, H.; Sprague, P.; Teig, S. Chemical Functional Queries for 3D Database Search. In J. Chem. Inf. Comput. Sci., 1994; pp 1297-1308. (22) Mestres, J.; Rohrer, D. C.; Maggiora, G. M. A molecular-fieldbased similarity study of nonnucleoside HIV-1 reverse transcriptase inhibitors. 2. J. Comput.-Aided Mol. Des. 2000, 14, 3951. (23) Berman, H. M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T. N.; Weissig, H.; Shindyalov, I. N.; Bourne, P. E. The Protein Data Bank. Nucleic Acids Res. 2000, 28, 235-242. (24) Banner, D. W., Hadvary, P. Crystallographic analysis at 3.0A resolution of the binding to human thrombin of four active sitedirected inhibitors. J. Biol. Chem. 1991, 266, 200085-200093.

Putta et al. (25) Anderson, M.; Beattie, J. F.; Breault, G. A.; Breed, J.; Byth, K. F.; Culshaw, J. D.; Ellston, R. P.; Green, S.; Minshull, C. A.; Norman, R. A.; Pauptit, R. A.; Stanway, J.; Thomas, A. P.; Jewsbury, P. J. Imidazo[1,2-a]pyridines: a potent and selective class of cyclin-dependent kinase inhibitors identified through structure-based hybridisation. Bioorg. Med. Chem. Lett. 2003, 13, 3021-3026. (26) Beattie, J. F.; Breault, G. A.; Ellston, R. P.; Green, S.; Jewsbury, P. J.; Midgley, C. J.; Naven, R. T.; Minshull, C. A.; Pauptit, R. A.; Tucker, J. A.; Pease, J. E. Cyclin-dependent kinase 4 inhibitors as a treatment for cancer. Part 1: identification and optimisation of substituted 4,6-bis anilino pyrimidines. Bioorg. Med. Chem. Lett. 2003, 13, 2955-2960. (27) Davis, S. T.; Benson, B. G.; Bramson, H. N.; Chapman, D. E.; Dickerson, S. H.; Dold, K. M.; Eberwein, D. J.; Edelstein, M.; Frye, S. V.; Gampe Jr, R. T.; Griffin, R. J.; Harris, P. A.; Hassell, A. M.; Holmes, W. D.; Hunter, R. N.; Knick, V. B.; Lackey, K.; Lovejoy, B.; Luzzio, M. J.; Murray, D.; Parker, P.; Rocque, W. J.; Shewchuk, L.; Veal, J. M.; Walker, D. H.; Kuyper, L. F. Prevention of chemotherapy-induced alopecia in rats by CDK inhibitors. Science 2001, 291, 134-137. (28) Bramson, H. N.; Corona, J.; Davis, S. T.; Dickerson, S. H.; Edelstein, M.; Frye, S. V.; Gampe, R. T., Jr.; Harris, P. A.; Hassell, A.; Holmes, W. D.; Hunter, R. N.; Lackey, K. E.; Lovejoy, B.; Luzzio, M. J.; Montana, V.; Rocque, W. J.; Rusnak, D.; Shewchuk, L.; Veal, J. M.; Walker, D. H.; Kuyper, L. F. Oxindolebased inhibitors of cyclin-dependent kinase 2 (CDK2). J. Med. Chem. 2001, 44, 4339-4358. (29) Breault, G. A.; Ellston, R. P.; Green, S.; James, S. R.; Jewsbury, P. J.; Midgley, C. J.; Pauptit, R. A.; Minshull, C. A.; Tucker, J. A.; Pease, J. E. Cyclin-dependent kinase 4 inhibitors as a treatment for cancer. Part 2: identification and optimisation of substituted 2,4-bis anilino pyrimidines. Bioorg. Med. Chem. Lett. 2003, 13, 2961-2966. (30) Srinivasan, J.; Castellino, A.; Bradley, E. K.; Eksterowicz, J. E.; Grootenhuis, P. D.; Putta, S.; Stanton, R. V. Evaluation of a novel shape-based computational filter for lead evolution: application to thrombin inhibitors. J. Med. Chem. 2002, 45, 2494-2500. (31) Bradley, E. K.; Miller, J. L.; Saiah, E.; Grootenhuis, P. D. Informative library design as an efficient strategy to identify and optimize leads: application to cyclin-dependent kinase 2 antagonists. J. Med. Chem. 2003, 46, 4360-4364.

JM049066L