Figure 1: Sequence comparison of human CDK2 and starfish CDK1. The CDK2 sequence was taken from the crystal structure (PDB entry 1FIN). The CDK1 sequence was taken from Asterina pectinifera mRNA (GenBank entry D79982). The sequence regions corresponding to the PSTAIRE helix and the T-loop are indicated with magenta lines. The regions that contact the respective cyclin molecule in the structures are outlined in blue.




Figure 2: Sequence comparison of human cyclinA and starfish cyclinB. The cyclinA sequence was taken from the crystal structure (PDB entry 1FIN). The cyclinB sequence was taken from Asterina pectinifera (GenBank entry A37350). The sequence regions corresponding to those portions of the structure that make contact with the PSTAIRE helix in the respective CDK molecule are indicated with magenta lines. Other regions that contact the respective CDK molecule in the structures are outlined in blue.

CDK2/cyclin A Energy Refinement

Hydrogens were added to the CDK2/cyclin A structure with ATP bound (PDB entry 1FIN) and CFF91 potentials were assigned using molecular editor software (4). Initially, all heavy atoms were fixed in the model and molecular mechanics minimization (5) was performed on the hydrogens alone. This consisted of a full Fletcher Powell minimization (6) until the norm of the gradient reached 0.01. ATP was replaced with 9-cyano paullone using a previously described complex as a template (7) and the structure was refined using molecular mechanics and hydropathic terms. The resulting energy refined CDK2/cyclin A complex was analyzed for interaction between CDK2 and cyclin A using HINT (8, 9).

Preliminary CDK1/cyclin B Model

A preliminary model of CDK1/cyclin B was formed using a template consisting of the energy refined CDK2/cyclin A crystal structure with 9-cyano paullone occupying the adenine portion of the ATP binding site as previously described (7). Initial three-dimensional coordinates for identical residues were taken directly from the CDK2/cyclin A template. Placement of the non-identical residues was performed using the LOOK 3.0 software (10), which utilizes a segment matching method for geometry assignment (11). Hydrogens were added to this CDK1/cyclin B model and this was followed by assignment of potentials in CFF91 (12, 13). This preliminary model of CDK1/cyclin B contained numerous steric conflicts consisting of atomic overlaps > 0.5 Å.

CDK1/cyclin B Model Refinement

All heavy atoms in the model were tethered to the corresponding template atoms. In the case of non-identical residues, only backbone and beta carbons were tethered. As before, molecular mechanics minimization was first performed on the hydrogens only. This was followed by a relaxation protocol of constraints and local geometry searches in which the hydrogens were free to move. This protocol consisted of the gradual removal of heavy atom constraints (from 5000 to 0 kcal/mol·Å) from their initial positions over the course of an extended conjugate gradient minimization procedure (6). Typically, tethering forces were reduced by a factor of 0.8 on each cycle of the minimization. Minimization continued on each cycle until the norm of the gradient reached 2.0. These parameters usually resulted in ∼50 cycles of the relaxation procedure, at which point the tethering force approached zero. Finally, an unconstrained minimization was performed until a norm of the gradient of 0.01 was achieved.

Hydropathic Evaluation and Further Refinement

In addition to the evaluation afforded by the molecular mechanics potential energy terms, we employed hydropathic analysis using the HINT program (9). Specifically, we applied the evaluation of all atom pair interactions to the interface between the cyclin B and the CDK1 models. This provided a description of both favorable and unfavorable atom contacts. The favorable contacts consist of hydrogen bonding, hydrophobic, and acid/base interactions. Unfavorable contacts consist of base/base and hydrophobic/polar interactions.

Following identification of unfavorable contacts, corrections to geometry to reduce or eliminate unfavorable hydropathic interactions were made by manually adjusting multiple side chain dihedral angles so that interatomic contacts were within steric limits of 0.25 Å. The resulting geometries were then optimized by repeating the refinement protocol described in the previous section, assuring that any manual adjustments were still subjected to the force field minimization procedure. During optimizations, the spatial disposition of the amino acids in the ATP binding pocket were maintained to fit the structure activity relationship for the congeneric series of paullones (7). This process was repeated until both the norm of the gradient was 0.01 and the hydropathic interaction constants on individual residues was comparable to that of the CDK2/cyclin A crystal structure.

Introduction

Cyclin-dependent kinases (CDKs) are a family of structurally homologous serine threonine kinases composed of a catalytic subunit bound to an activating cyclin molecule (14). They are attractive targets for the development of cancer therapeutics (15) due to the observation that CDK regulators are frequently altered in malignancies (16, 17, 18). CDKs and cyclins perform essential functions governing the eukaryotic cell cycle (19). CDK1 (previously known as p34^cdc2)/cyclin B governs transitions from G2 to M; CDKs 4 and 6 complexed to cyclin D govern progression from G1; and CDK2/cyclin A or E govern transition to S phase and progression through S phase.

Full activation of CDKs typically results after the catalytic subunit forms a complex with the cyclin subunit. This is followed by phosphorylation of a key residue on the T loop by a CDK-activating kinase (14) and a loss of inhibitory phosphorylations on Thr14 and Tyr15 near the ATP binding pocket. The PSTAIRE helical domain and T loop of CDKs are reoriented after these events. This results in changes in the shape of the ATP-binding catalytic cleft (20) which, in one of our previous investigations, provided useful insight into rationalizing the optimization of the paullone inhibitors (7). At the conclusion of that study, we determined that structure-based design efforts may benefit greatly from the availability of the full 3-D coordinates for cyclin-CDK complexes.

At present, the majority of deposited CDK structures in the Protein Data Bank consist of CDK2 complexed with ATP (21), CDK2 structures complexed with inhibitors (22), CDK2 structures complexed with cyclin A (1), and CDK6 complexed with INK (23). Two model structures of CDK1 have been reported (24, 25), however neither of these are in association with a cyclin molecule. There are no published crystallographic structures of CDK1. In the present paper, we describe the homology generation of a CDK1/cyclin B molecular model developed from the deposited CDK2/cyclin A crystallographic structure.

Results & Discussion

Characterization of Refined CDK1/cyclinB Model

Our CDK1/cyclin B model was found to contain the expected major secondary structural features that are present in the CDK2/cyclin A crystal structure (Figure 3). For example, with respect to the cyclin B portion of the model, the relative positions of the cyclin helices α1-α5 to cyclin A were retained. As shown in Figure 3, the RMS deviation between both complexes of the CDK Cα atoms is 1.76 Å and in the PSTAIRE region the Cα RMS is 1.64 Å.



Figure 3: A 15 Å backbone ribbon rendering of the energy refined structures of the CDK2/cyclin A X-ray crystal and the CDK1/cyclin B homology model. Regions of interest include: a, CDK PSTAIRE helix; b, CDK ATP binding site; and c, cyclin α5 helix. The RMS deviation between both complexes of the CDK backbone atoms is 1.76 Å; for the Cα atoms is 1.76 Å, for the PSTAIRE Cα atoms is 1.64 Å, and for the PSTAIRE heavy atoms is 3.31 Å.

The CDK1/cyclin B complex modeled in this study contains similar intermolecular contacts between the cyclin and kinase as those observed in the CDK2/cyclin A structure (see Table I and Figure 4). For example, regarding secondary structure, the interface regions of the two protein complexes are similar. These regions consist of the C-terminal portion of the α3 helix and the C-terminal portion of the α5 helix in both cyclins. Additionally, the hydropathic analysis of surface features (Figure 4) of the interface domains reveals that the chemical nature of the interface surface is relatively conserved in both CDK/cyclin complexes. That is to say, prominent regions of hydrophobicity and polarity are consistent between the structures.

Figure 4: Comparison of the CDK/cyclin interfaces from the crystal structure for CDK2/cyclinA and the CDK1/cyclinB model. Both molecules are viewed from the perspective of the cyclin interface with the cyclin molecules omitted. The peptide backbone of the CDK molecules is visible behind the interface surface. A Connolly surface generated from the hydropathic grid (computed using Hint software) is shown for the interface region of the CKDs and cyclins. A continuum of polar to hydrophobic surface contacts spans from yellow to blue. Features of interest based on the sequence comparison are also displayed.

PSTAIRE Region Contacts

Previous investigations indicate that the conserved PSTAIRE helix in the CDK family of proteins is the primary site of intermolecular recognition in cyclin binding (1). It is also known that upon cyclin binding at PSTAIRE a conformational change ensues which is propagated to the ATP binding site. The details of the PSTAIRE/cyclin B interaction for the model generated from this study reveal a set of intermolecular contacts similar to those of the PSTAIRE/cyclin A crystallographic structure, with the caveat that differing amino acid contacts are contributed from cyclin B. In our model, the overall pattern of stabilizing interactions contained in the region of cyclin B bound to PSTAIRE resembles that of the CDK2/cyclin A crystal structure.

Starting at a position in the turn (residue Glu42) located to the N-terminal side of PSTAIRE, both the CDK2/cyclinA crystal and the CDK1/cyclinB models contain extensive hydrogen bonding networks. Both CDK?s contain the sequence Glu42-Gly43-Val44. In the CDK1 crystal, Glu42 and Val44 backbone carbonyls both accept hydrogen bonds from donors located on the cyclin molecule. Our modeling suggests that the Gly43 residue in CDK1 and CDK2 may be critical at this position since no other residue would apparently permit the dihedral angles necessary for the observed orientation of the flanking amino acids. The CDK2 Val44 and Glu42 backbone carbonyls accept hydrogen bonds from cyclin A Lys266 ε-amino group and the CDK2 Val44 backbone N-H amide donates a hydrogen bond to the Glu295 carboxylate of cyclin A. Although CDK1 also contains the same Glu-Gly-Val triplet, cyclin B lacks the lysine in the position homologous to cyclin A Lys266. In fact, the relevant sequences for these two segments consist of LIAA-YEEM for cyclin B, while cyclin A has LLASKFEEI (see Figure 2). Interestingly, it appears that the gap in cyclin B (where the missing Lys266 would be) fulfills a steric requirement enabling this sequence of cyclin B to maintain a similar spatial disposition to the analogous sequence of cyclin A. Specifically, this is accomplished through the incorporation of a hydrogen-bonding network for CDK1 that consists of the CDK1 Ser39 and Glu40 backbone carbonyls that accept hydrogen bonds from the cyclin B Arg289 side chain. The Arg289 guanidinium also forms a hydrogen bond with CDK1 Glu41 carboxylate. The CDK1 Glu41 carboxylate, in turn, receives a hydrogen bond from cyclin B Lys285 ε-amino.

Within the PSTAIRE domain itself, the next prominent set of interactions occurs with hydrogen bonds between CDK1 Ser46 Oγ and the backbone carbonyl of cyclin B Met268 (analogous to CDK2 Ser46 Oγ/cyclin A Lys266 backbone carbonyl). This is flanked on one side by the hydrophobic cluster of CDK2 Val44 Cγ with cyclin A Leu299 Cδ and on the other side by the interaction of CDK2 Ile49 Cδ with cyclin A Leu306 and Leu263 Cδ?s. In our model of the CDK1/cyclin B complex, the corresponding two hydrophobic contact regions to CDK2/cyclin A are: CDK1 Val44 Cγ with cyclin B Leu296 Cδ on the N-terminal side, and CDK1 Ile49 Cδ with cyclin B Leu262 Cγ and Leu303 Cδ.

In the middle of the PSTAIRE helix in the CDK2/cyclinA crystal structure there is a hydrogen bond between CDK2 Arg50 Nε and cyclin A Lys266 backbone carbonyl. Also, near the end of the PSTAIRE domain, there is a CDK2 Lys56 ε-amino in a position that permits hydrogen bonding with either cyclin A Thr303 backbone carbonyl or Asp305 Oδ. A similar arrangement is observed in our CDK1/cyclinB model since the CDK1 Lys56 ε-amino is situated to donate hydrogen bonds to both cyclin B Asp300 backbone carbonyl and Asn302 Oδ.

Overall, it is apparent that despite significant differences in the amino acid composition between the cyclin molecules, the nature of the tertiary interaction with the respective PSTAIRE region is the same between our CDK1/cyclinB model and the reported CDK2/cyclinA structure.

Features Connecting the Cyclin Interface with the ATP Binding Site

A hydrophobic patch that anchors the surface of helix α5 (see Figure 3) in the cyclins to a turn in the CDK molecule between the anti-parallel strands 1 and 2 from β-sheet A (26). Strand 1 consists of residues 66-73 and strand 2 contains residues 74-81. This β-sheet A connects the cyclin binding surface with the ATP binding site in both CDK1 and CDK2. Specifically, CDK2 His71 Cε interacts favorably with cyclin A His296 Cε. In a similar manner, our CDK1 model contains the analogous hydrophobic interaction consisting of CDK1 Met71 Cε with cyclin B Leu296 Cδ. This region interacts with a hydrophobic patch on the PSTAIRE helix in both CDK1 and CDK2 consisting of a CDK Ala48 and Ile49. Thus, these regions may serve to stabilize interactions with β-sheet A that spans the distance between the PSTAIRE helix and the ATP binding domain.

Upon cyclin binding, observed shifts (20) in the position of β-sheet A directly effect the position of the first neighbors in the nucleotide binding site in the CDK molecule. The amino acids of the adenine pocket, which are part of β-sheet A, are conserved between CDK1 and CDK2. Specifically, they are: Val64 (two residues before the N-terminal end of strand 1 in β-sheet A), Phe80 (located on strand 2 of β-sheet A), Phe82 (one residue beyond the C-terminal end of strand 2 in β-sheet A), Ile10 (one residue beyond the C-terminal end of strand 5 in β-sheet A), Ala31 (located on strand 3 of β-sheet A), and Val18 (located on strand 4 of β-sheet A). In addition, there are backbone carbonyls from residues Thr84 and Glu81 that are associated with β-sheet A. These residues are of interest because they serve as a connection between the surface of the ATP binding domain and the cyclin binding domain. These details could provide insight for future studies of this region regarding CDK activation due to cyclin binding and the understanding of mutants with phenotypes associated with altered cyclin activation.

Comparison of the CDK2 and CDK1 ATP Binding Sites

In the simplest description, the ATP binding site is composed of an adenine pocket that is characterized by a hydrogen bond recognition motif and a hydrophobic portion occupied by the nucleotide base. There is also a hydrolytic portion of the binding site that positions the Mg⁺⁺-bound triphosphate of ATP.

Regarding the nucleoside hydrogen bond recognition site, hydrogen bonds form between the adenine base N and the backbone N-H of Leu83, and between the substituted amine and the Glu81 backbone carbonyl. This is found in both CDK1 and CDK2. The RMS deviation of the backbone of these residues in the CDK2 crystal and our model of CDK1 is only 0.23 Å. This slight deviation is due to the fact that the paullones used during the formation of our CDK1 model did not utilize the Glu81 backbone carbonyl for hydrogen bonding. This enables Glu81 to have a significant amount of freedom during minimizations. These observed differences in the hydrogen bond recognition site are also within the level of resolution for the reported CDK2/cyclin A crystal structure. Therefore, the hydrogen bond nucleoside recognition site remains relatively unchanged in our CDK1 model compared with CDK2 crystal structure.

The triphosphate portion of ATP binds among residues 142-148 and 30-36 in CDK2 and according to our model, in CDK1 as well. For the most part, there are only small differences observed in this region, evidenced by the difference between the backbone atoms of our CDK1 model and those of the CDK2 crystal structure (RMS deviation = 1.29 Å). Critical to the hydrolytic functionality of this region are residues Asp145, Lys33, and Glu51, the latter of which is contributed from the PSTAIRE helix.

Originally, we constructed a preliminary model of CDK1 without cyclin B based on a backbone superimposition of CDK1 with the CDK2/cyclin A crystal structure. In that model, the PSTAIRE helix lacked a stabilizing interaction with the cyclin B-CDK1 interface. As a result, upon refinement of that model, the CDK1 surface normally associated with cyclin B binding resided in a local minimum at too great a distance from the ATP site. By contrast, the optimization and formation of the CDK1/cyclin B complex resulted in a stabilized PSTAIRE interaction and surface compatible with CDK2/cyclin A structure in the vicinity of the ATP binding site. Furthermore, this stabilization results in a shift in the position of the electrostatic triad between Glu51, Lys33, and Asp145. Critical to the conformation of the triphosphate-binding site, was the displacement of Glu51 on PSTAIRE towards the catalytic cleft. This shift in turn enables Asp145 to be available to coordinate a Mg atom and form the pre-catalytic complex with the terminal two phosphate groups of a bound ATP. Lys33 of our structure is also positioned to coordinate α and β phosphate groups of an ATP as a result of the Glu51 conformation. Thus, the documented shift in the interaction of Lys33 away from Asp145 and towards the Glu51 side chain, a critical step in the activation of the CDK upon cyclin binding (20), is also present in our model of CDK1/cyclin B.

Our previous work demonstrated that the position of the aliphatic chain of Asp145 in CDK models and X-ray structures without a cyclin bound was in steric clash with 10-bromo-paullone, an analogue with some inhibitory activity (7). In fact, a complete 3D-QSAR was only possible when the cyclin B bound form of our CDK1 model was used in docking experiments. This observation supports the hypothesis that the cyclin bound form of the kinase is the appropriate form to use in a structure-based ligand design effort.

Of particular interest to our investigation is the inclusion of a single solvent molecule that is believed to be hydrogen bonded to the aromatic side chain of Phe80 in the CDK1/cyclin B model. This feature was a direct extension of the water reported in the CDK2/cyclin A structure. Indeed, both 9-cyano paullone and 9-nitro paullone (alsterpaullone) utilize the water molecule for an additional hydrogen bond which is purported to be a salient aspect of enhanced affinity of these two ligands (7). The fact that this feature is apparently conserved in both kinases suggests that the Phe80 side chain may function as a hydrophobic gate enabling a single column of water molecules to enter the catalytic portion of the ATP site in a highly regulated manner. Perhaps this controlled solvent movement is critical to the positioning of the ambident nucleophilic water for the final stage of hydrolytic cleavage. This postulate is also based on the notion that a single water would be positioned behind the Mg⁺⁺ triphosphate, closer to the Asp145 and Lys33 catalytic residues, and capable of being presented to this region in a regulated manner.

In addition to the hydrogen bond recognition site and the hydrolytic region, there are amino acid sequence differences between CDK1 and CDK2 in the ATP binding site. For example, CDK2 Val29 and CDK1 Ile29 both interface with Phe82 and Lys20 aliphatic carbons of both proteins. The additional steric bulk provided by the aliphatic carbons of Ile29 in CDK1 results in a slight displacement of the side chain of Lys20. According to our modeling, this has the effect of increasing the hydrophobic surface area of the CDK1 ATP pocket in the vicinity of Lys20 and Phe82 relative to CDK2. Additionally, CDK2 and CDK1 differ at amino acid position 84. The position of this residue is on the solvent accessible surface of both proteins. In CDK2, the His84 imidazole side chain may form a boundary for solvent entrance into the adenosine-binding region. This is possibly due to the fact that the polar Ns of the His84 imidazole side chain can form hydrogen bonds with up to two solvent molecules, while at the same time the aliphatic and aromatic portions of His 84 form a hydrophobic deterrent to solvent entrance into the pocket. In a similar manner, CDK1 utilizes a Thr at position 84. The hydroxyl portion of CDK1?s Thr84 is oriented in the same direction as the polar half of the CDK2 His84 ring. Thus, a comparable mechanism of solvent ?patrolling? is afforded by rotation of Thr84?s side chain OH that serves to hydrogen bond solvent, while the aliphatic portion of side chain barricades the pocket entrance. The proposed function of these residues may be critical, due to the proximity of the solvent-attracting backbone carbonyl of Thr84 (about 4.6 Å). Furthermore, the carbonyl of Thr84 is also near the N-H and carbonyl of Leu83 (6.0 and 3.3 Å, respectively), the implication being that in the absence of a ligand, these three backbone polar groups can form a favorable electrostatic attraction for the formation of a hydrogen-bonded network of solvent.

CDK1 also differs from CDK2 at positions 85 and 90, which are located just beyond (1 and 5 Å, respectively) the hydrophobic area utilized by the adenine base of ATP. CDK2 employs a Gln85 and Phe90 in these positions, whereas CDK1 utilizes Met85 and Tyr90. In the CDK2 crystal structure, the Gln85 amide side chain serves to stabilize the C-terminal domain (residues 292 to the end) of CDK2 by hydrogen bonding with the backbone of Asp86, and the side chain of Arg297. The Arg297 guanidinium side chain is also hydrogen bound to the His295 backbone carbonyl. Our modeling suggests the possibility that the relatively nonpolar side chain of Met85 in CDK1 may be isosteric with the aliphatic side chain carbons of CDK2 Gln85. However, since Met85 does not possess the hydrogen bonding capability observed for CDK2 Gln85, it is possible that the phenolic OH of CDK1 Tyr90 (in place of CDK2 Phe90) may compensate by hydrogen bonding with the C-terminal residues 292-295. The implication being that CDK1 stabilizes this region with the Met85-Tyr90 pair which may be a functional substitution for the Gln85-Phe90 pair of CDK2. It is interesting to note that the CDK1 Met85-Tyr90 pair is conserved in CDK1?s from human (GenBank Accession NP_203698), bovine (GenBank Accession P48734), mouse (GenBank Accession A36074), chicken (GenBank Accession P13863 ), and frog (GenBank Accession 2005165A) sources.

A series of paullones (27) as well as Roscovitine (28) have been shown to utilize additional amino acid residues outside of the traditional ATP pocket for binding. CDK2 residues Glu8, Ile10, Lys20, and Leu298, have been shown to be involved in the binding with the benzoamine moiety of Roscovitine (2927). The side chains of these surface residues are within the envelope of secondary interactions with the docked 10-bromo-paullone in the CDK1/cyclin B model. The RMS deviation of our CDK1 model Glu8, Ile10, Lys20, and Leu298, peptide backbone atoms from those in the CDK2/cyclin A crystal structure is 1.35 Å. The heavy atom deviation for the same residues is 3.36 Å, which represents significant movement away from the template residues in CDK2. Again, this supports the contention that there are critical differences exhibited in this CDK1/cyclin B model and that these may prove to be helpful in design of specific inhibitors for CDK1 or CDK2.

Conclusion

The CDK1/cyclinB model described in this study was used to successfully model the binding of the paullone series of inhibitors for a number of years. Key features of the interface between CDK?s and cyclins have been identified in this study. In the preliminary CDK1 model (constructed and refined in the absence of an associated cyclin-data not shown) the adenine portion of the ATP pocket retains the same general character exhibited by the ATP pocket in CDK2 unbound to cyclin A.

This molecular modeling study supports the similar finding reported for CDK2/cyclin A. Specifically, that upon binding to CDK1, cyclin B induces conformational changes resulting in a distinctly different shape of the ATP binding pocket?s molecular volume. Furthermore, these results relate to the spatial disposition of amino acid side chains of over 80 paullone inhibitors docked in models where the cyclin molecule has been included. Additionally, there are differences in the CDK1 amino acid sequence and protein structure near and in the ATP pocket. These differences suggest that it is possible to design inhibitors with enhanced specificity for CDK1, CDK2, or CDK5.

Supplemental Electronic Material

The coordinates of the model described in this work are available from the PDB with the identifier 1T30.

Acknowledgments

We wish to thank the Advanced Biomedical Computing Center at NCI-Frederick for providing computational resources. This work was supported in part by the National Cancer Institute, NIH, under contract number N01-CO-12400 with SAIC-Frederick.

References and Footnotes

1. P. D. Jeffrey et al. Nature 376, 313-320 (1995).
2. S. B. Needleman and C. D. Wunsch. J. Mol. Biol. 48, 443-453 (1970).
3. J. L. Risler, M. O. Delorme, H. Delacroix, and A. Henaut. J. Mol. Biol. 204, 1019-1029 (1988).
4. Insight 2000 Molecular Modeling Workbench, 2000. Accelrys (2000).
5. Discover Molecular Simulation Program, 2.97. Accelrys (1997).
6. R. Fletcher. Practical Methods of Optimization, Vol. 1. John Wiley & Sons, New York (1980).
7. R. Gussio et al. Anti-Cancer Drug Design 15, 53-66 (2000).
8. G. E. Kellogg, S. F. Semus, and D. J. Abraham. J. Comput. Aided Mol. Des. 5, 545-552 (1991).
9. HINT, 2.01. eduSoft (1993).
10. GeneMine/Look, 3.5.2. Molecular Applications Group (1999).
11. M. Levitt. J. Mol. Biol. 226, 507-533 (1992).
12. J. R. Maple, T. S. Thacher, U. Dinur, and A. T. Hagler. Chemistry, Design, and Automation News 5, 5 (1990).
13. Class II CFF91 Forcefield, 2.0. Biosym Technologies Incorporated (1992).
14. N. P. Pavletich. J. Mol. Biol. 287, 821-828. (1999).
15. M. Knockaert, P. Greengard, and L. Meijer. Trends Pharmacol Sci. 23, 417-425 (2002).
16. L. Meijer. Chemical Inhibitors of Cyclin-dependent Kinases. In Progress in Cell Cycle Research, Vol. 1. Eds., L. Meijer, S. Guidet, and H. Tung. Plenum Press, New York (1995).
17. L. Meijer. Trends in Cellular Biology 6, 393 (1996).
18. L. Meijer and S. H. Kim. Methods Enzymol 283, 113-128 (1997).
19. D. O. Morgan. Annu. Rev. Cell Dev. Biol. 13, 261-291 (1997).
20. A. A. Russo, P. D. Jeffrey, and N. P. Pavletich. Nat. Struct. Biol. 3, 696-700 (1996).
21. U. Schulze-Gahmen, H. L. De Bondt, and S. H. Kim. Journal of Medicinal Chemistry 39, 4540-4056 (1996).
22. N. S. Gray et al. Science 281, 533-538 (1998).
23. D. H. Brotherton et al. Nature 395, 244-250 (1998).
24. A. Cavalli, C. Dezi, G. Folkers, L. Scapozza, and M. Recanatini. Proteins 45, 478-485 (2001).
25. K. Narayana, R. P. K. Surendra, and J. A. R. P. Sarma. Protein Data Bank (2002).
26. P. D. Jeffrey, L. Tong, and N. P. Pavletich. Genes Dev. 14, 3115-3125. (2000).
27. C. Kunick et al. Bioorganic & Medicinal Chemistry Letters 10, 567-569 (2000).
28. L. Meijer et al. European Journal of Biochemistry 243, 527-536 (1997).
29. W. F. De Azevedo et al. European Journal of Biochemistry 243, 518-526 (1997).