Crystal Structure of Avian Aminoimidazole-4-carboxamide Ribonucleotide Transformylase in Complex with a Novel Non-folate Inhibitor Identified by Virtual Ligand Screening*

Aminoimidazole-4-carboxamide ribonucleotide transformylase (AICAR Tfase), one of the two folate-dependent enzymes in the de novo purine biosynthesis pathway, is a promising target for anti-neoplastic chemotherapy. Although classic antifolates, such as methotrexate, have been developed as anticancer agents, their general toxicity and drug resistance are major issues associated with their clinical use and future development. Identification of inhibitors with novel scaffolds could be an attractive alternative. We present here the crystal structure of avian AICAR Tfase complexed with the first non-folate based inhibitor identified through virtual ligand screening of the National Cancer Institute Diversity Set. The inhibitor 326203-A (2-[5-hydroxy-3-methyl-1-(2-methyl-4-sulfophenyl)-1H-pyrazol-4-ylazo]-4-sulfo-benzoic acid) displayed competitive inhibition against the natural cofactor, 10-formyl-tetrahydrofolate, with a Ki of 7.1 μm. The crystal structure of AICAR Tfase with 326203-A at 1.8 Å resolution revealed a unique binding mode compared with antifolate inhibitors. The inhibitor also accessed an additional binding pocket that is not occupied by antifolates. The sulfonate group of 326203-A appears to form the dominant interaction of the inhibitor with the proposed oxyanion hole through interaction with a helix dipole and Lys267. An aromatic interaction with Phe316 also likely contributes to favorable binding. Based on these structural insights, several inhibitors with improved potency were subsequently identified in the National Cancer Institute Compound Library and the Available Chemical Directory by similarity search and molecular modeling methods. These results provide further support for our combined virtual ligand screening rational design approach for the discovery of novel, non-folate-based inhibitors of AICAR Tfase.

The biological significance of the folate cofactor stems from the key role that folates play as one-carbon carriers in several important metabolic pathways, including de novo purine and thymidylate synthesis. The recognition of the critical role of reduced folates for the synthesis of DNA precursors led to the discovery of folate analogues as anti-metabolite agents. Antifolates typically interfere with the binding of the natural cofactor to key enzymes in these biosynthetic pathways, such as thymidylate synthase, dihydrofolate reductase, glycinamide ribonucleotide transformylase (GAR Tfase), 1 and AICAR Tfase. The first generation of antifolates, aminopterin and methotrexate, heralded the era of anti-metabolite cancer chemotherapy (1). Subsequent design and development of folate analogues has led to new generations of antifolates, some currently under clinical evaluation. For example, pemetrexed (ALIMTA) can bind multiple targets, such as thymidylate synthase, dihydrofolate reductase, GAR Tfase, and AICAR Tfase, and is one of the most promising of these current antifolates. Pemetrexed exhibits antitumor activity in several types of solid tumors (2), and its phase III clinical trial has recently been approved by the Food and Drug Administration.
Nevertheless, general toxicity issues associated with antifolates have posed major challenges for folate analogue development. Folic acid and its cellular derivatives function mainly in their fully reduced form and are involved in many cellular metabolic pathways, such as cell proliferation and amino acid metabolism. Because at least 18 folate-dependent enzymes are involved in 1-carbon transfer reactions, as well as transformation of folate cofactors themselves, in human cells, antifolate agents often inhibit multiple pathways, resulting in adverse cytotoxic side effects. One such example is methotrexate, whose general toxicity stems from the inhibition of several folate-dependent enzymes involved in methionine biosynthesis, thymidylate synthesis, and de novo purine biosynthesis (3). Lometrexol is a selective inhibitor to GAR Tfase (nM) in de novo purine biosynthesis, but can also inhibit AICAR Tfase at higher concentrations (4). Although lometrexol has exhibited promising anti-neoplastic activity, it was suspended from further clinical evaluation because of folate depletion after administration of the drug (4). However, more recent studies have indicated that supplementation with folate can reduce its toxicity (5).
Currently, substantial efforts have been focused on specific tailoring of folate analogues to individual enzymatic targets to achieve greater selectivity, as well as increased potency. Nevertheless, only a limited number of antifolates for cancer chemotherapy have been approved by the Food and Drug Administration for clinical applications, including methotrexate and raltitrexed, a thymidylate synthase inhibitor for treatment of colon cancer in Europe (3). Therefore, inhibitors with novel scaffolds that differ from traditional antifolates are attractive because they are unlikely to produce the same undesired side effects. Moreover, such classes of inhibitors would not necessarily possess a glutamate or multiple glutamates as in traditional antifolates and would eliminate the requirement for active transport, a frequent source of antifolate drug resistance. Nevertheless, the added folate requirement in tumor cells leads to up-regulation of these folate transporters and increased accumulation of folates and antifolates in tumor cells. Thus, bypassing of the active transport system and reliance on passive diffusion raises its own challenges for selective uptake into tumor cells to obtain an efficacious chemotherapeutic agent.
In this study, a virtual ligand screening approach via docking was undertaken to search for inhibitors of AICAR Tfase with novel scaffolds. AICAR Tfase/IMPCH (ATIC), which catalyzes the last two steps in the de novo purine biosynthesis pathway, is a promising target for the development of therapeutic intervention in various types of cancer (6). The de novo purine biosynthesis pathway consists of 10 enzymatic reactions that sequentially convert 5-phosphoribosyl-1-pyrophospate to IMP. The homodimeric ATIC is a bifunctional enzyme composed of two distinct active sites. The high resolution crystal structure of ATIC (7) revealed that the C-terminal domain (residues 200 -593) contains the AICAR Tfase activity, which catalyzes formyl transfer from the cofactor 10-f-THF to AICAR producing 5-formyl-AICAR ( Fig. 1) in the penultimate step of purine biosynthesis. The N-terminal IMPCH domain (residue 1-199) is responsible for the final cyclohydrolase ring closure step that converts 5-formyl-AICAR to IMP (Fig. 1). The IMPCH active site and catalytic mechanism have been elucidated by the structure of ATIC complexed with the IMPCH domain inhibitor, xanthosine 5Ј-monophosphate (8,9). Crystal structures of ATIC with the substrate AICAR (8) and two sulfonylcontaining folate inhibitors, BW1540 and BW2315, have illuminated the cofactor-binding site (10). The crystal structure of a multisubstrate adduct inhibitor ␤-DADF (MAI), mimicking both the substrate AICAR and the cofactor when complexed with ATIC, unambiguously demonstrated that the AICAR Tfase active site was located at the dimer interface and identified putative key catalytic residues (11). Furthermore, the complexed structures revealed that AICAR Tfase possesses an unusually large active site, with unoccupied binding pockets and cavities even when both AICAR and cofactor are bound. Thus, the structural information provides invaluable guide-lines for the rational design of novel inhibitor scaffolds. ATIC has no structural or sequence similarity to any other folate-dependent enzymes in the de novo biosynthesis pathways. In addition, no clinically useful inhibitors for ATIC have been developed to date, which underscores the opportunity to pursue structure-based inhibitor design targeting of ATIC. Molecular docking and de novo design are widely used in the discovery of enzyme inhibitors with the desired selectivity and potency profiles. Virtual screening of the structural database of known and commercially available small molecules can rapidly generate lead compounds with complementary shapes and favorable hydrogen bonding, electrostatic and hydrophobic interactions within the active site of the target (12). Such docking studies also enable discovery of novel ligands with chemical scaffolds that are dissimilar to those already tested. Here, we describe the outcome of the molecular docking strategy for the discovery of novel inhibitors of ATIC, where crystal structures of ATIC provided the necessary entry points for the docking and rational design of inhibitors.
Compounds from the National Cancer Institute Diversity Set were docked into the three-dimensional structure of human ATIC using AutoDock. The screening led to the identification of 44 new compounds with distinct chemical structures, compared with traditional antifolates, that inhibited ATIC (13). In vitro inhibition assays confirmed that 8 of the 16 compounds that had good solubility in water surprisingly inhibited AICAR Tfase activity at the micromolar level. To establish the structural basis for their specificity and affinity, we determined the crystal structure of ATIC with one of the most potent inhibitors identified from the screening effort.
Thus, we report here the discovery of a new class of inhibitors that bind to ATIC with a binding mode that differs from the traditional folate-based inhibitors. The crystal structure of ATIC complexed with one of the novel inhibitors, 326203-A, a commercially available dye named Acid Yellow 54, at 1.8 Å resolution (Fig. 2) has revealed the key enzyme-inhibitor interactions that now provide a valuable structural template for further improvement via rational design.

EXPERIMENTAL PROCEDURES
Materials-LB and agar were obtained from Invitrogen. All common buffers and reagents were purchased from Sigma-Aldrich. The inhibitor candidates were obtained from the National Cancer Institute Open Chemical Repository (Bethesda, MD) and the Sigma-Aldrich Library of Rare Chemicals.
Protein Expression and Purification-The plasmid pET28a, encoding the N-terminal hexahistidine-tagged human ATIC (kindly provided by Dr. G. Peter Beardsley, Yale University), was transformed into Escherichia coli BL21.DE3 cells (Novagen) by heat shock. The E. coli transformants were grown in 2YT medium (Invitrogen) at 37°C to an A 600 value around 0.7 and induced with 0.4 mM isopropyl ␤-D-thiogalactopyranoside (Invitrogen) overnight.
The cells were harvested by centrifugation and were resuspended in 100 ml of solution of 100 mM sodium phosphate, pH 8.0, 300 mM NaCl, 10 mM imidazole, and 5 mM ␤-mercaptoethanol (BME) (Buffer A). The resultant mixture was disrupted by sonication, as described previously (8). The lysate was clarified by centrifugation at 20,000 ϫ g at 4°C for 50 min. The supernatant was incubated with nickel affinity beads (Qiagen) overnight before transferring to a 2.5 ϫ 10 cm Econo column (Bio-Rad) and further washing with 10 column volumes of Buffer A. The protein was eluted with a 10 -250 mM imidazole gradient. Fractions containing protein were analyzed by SDS-PAGE, pooled, and further purified on a Superdex 200 HR column (Amersham Biosciences) equilibrated with 20 mM Tris-HCl, pH 7.5, 150 mM NaCl, 50 mM KCl, 5 mM BME, with or without 5 mM EDTA. The peak fractions were again analyzed by SDS-PAGE and pooled.
The plasmid pET28a, encoding the avian ATIC cDNA with an Nterminal hexahistidine tag (a kind gift from Dr. Stephen J. Benkovic), was transformed into E. coli BL21.DE3 cells (Novagen, Inc.) for overexpression. Avian ATIC was purified in the same manner as human ATIC (10). Virtual Ligand Screening by AutoDock-The human AICAR Tfase holo-template was extracted from its crystal complex with AICAR and BW1540 (Protein Data Bank code 1p4r) (10). All crystallographic water molecules and active site ligands were removed. The protonation states of His residues were assigned as follows: HD1 on His 267 , His 385 , His 469 , His 584 , and His 592 ; HE2 on His 213 , His 290 , His 293 , His 453 , His 470 , and His 591 . Non-polar hydrogens were merged with heavy atoms, and Kollman charges (14) were assigned. The active site in the homodimer that had lower average B values (2 monomers/asymmetric unit) was chosen as the docking site. A 60 ϫ 50 ϫ 66 three-dimensional energy grid with 0.375 Å spacing was calculated for each of the following atom types: carbon, A (aromatic carbon), nitrogen, oxygen, sulfur, hydrogen, fluorine, chlorine, bromine, iodine, iron, phosphorus, and e (electrostatic) using Autogrid3 (15). The docking screenings were carried out with AutoDock3.0.5 (15), and the jobs were distributed to the Scripps Atlas SGI Origin 2000 cluster, the NBCR Meteor, and the UCSD KeckII linux clusters. The National Cancer Institute Diversity Set (dtp.nci.nih.gov/ branches/dscb/diversity_explanation.html) (1990 compounds with a rich structural and pharmacophore diversity) was chosen as the compound library. All-atom Gasteiger charges were added, and non-polar hydrogens were merged with their connecting heavy atoms (16). The simulation parameters were: trials of 100, population size of 150, random starting position and conformation, translation step of 0.5 Å, rotation step of 35°, elitism of 1, mutation rate of 0.02, cross-over rate of 0.8, local search rate of 0.06, and 10 million energy evaluations. Final docked conformations were clustered using a tolerance of 1.5 Å root mean square deviation (rmsd). The top compounds were picked based on the higher binding energies.
Preparation of 10-f-THF-10-f-THF was prepared by a modified procedure of Rowe (17) and Black et al. (18). 10 mg of (6R,S)-5-f-THF (Schircks Laboratories, Jona, Switzerland) was dissolved in 500 l of 3.5 M BME, and 12 M HCl was then added to adjust the pH to 1-2. This solution was flushed with nitrogen and stored at 4°C for 3 days. Yellow precipitate of (6R,S)-5,10-methenyl-THF formed and was harvested by centrifugation at 11,000 rpm for 10 min. The precipitates were resuspended in 5 ml of 0.1 M Tris-HCl, pH 8.0, 0.1 M BME. The conversion of (6R,S)-5, 10-methenyl-THF to 10-f-THF was monitored by the decrease of the UV absorbance at 356 nm (A 356 ) and was usually complete after 3 h. The concentration of 10-f-THF was determined by A 298 using the extinction coefficient of 9.54 ϫ 10 3 cm Ϫ1 M Ϫ1 (18).
AICAR Tfase Inhibition Assay-The human ATIC enzyme was used for the inhibition assay. In an initial screen to determine approximate K i values, 1 mg of each compound was dissolved in 100 l of Me 2 SO and diluted to 1 mM using the assay buffer (32.5 mM Tris-HCl, pH 7.5, 25 mM KCl, 5 mM BME) as described previously (19 -21). The assay buffer was flushed with nitrogen to minimize oxidization of 10-f-THF. The Me 2 SO used in this assay did not have any inhibitory effect on the activity of ATIC. 25 nM of human ATIC, 10 M inhibitors, and different concentrations of cofactor 10-f-THF were mixed in the assay buffer to a volume of 150 l and incubated for 2 min. The reaction was initiated by adding 50 M of substrate AICAR (total volume, 300 l). Using a SpectraMax Plus384 microplate reader, the reaction was monitored at 298 nm by measuring the increase in absorbance corresponding to the formation of THF. The THF was generated in the transfer reaction of the formyl group from cofactor to AICAR to produce 5-formyl-AICAR. IC 50 values were measured at 25 nM ATIC, 50 M AICAR, and 8.5 M 10-f-THF.

FIG. 2. Chemical structures of inhibitor 326203-A and the active site of AICAR Tfase.
A, the structure of 326203-A is shown as depicted in National Cancer Institute Compound library; the Cr 3ϩ chelation site is highlighted in red. B, the crystal structure of the inhibitor did not show any bound Cr 3ϩ but instead an intramolecular hydrogen bond array, which is colored in red. C, stereo view of the AICAR Tfase active site with bound 326203-A. The two different subunits of the dimer that compose the active site are colored in sky blue and pink, respectively. The bound inhibitor is depicted in ball-and-stick representation, with the oxygen atoms in red, carbons in yellow, sulfurs in green, and nitrogens in blue. The 2F o Ϫ F c electron density map (blue) for the inhibitor is contoured at 1.0 . The key interactions of the inhibitor with AICAR Tfase involve three moieties of the inhibitor: benzene ring A, pyrazole ring, and benzene ring B.
The K i was obtained by generating plots of 1/V i versus 1/[S] at several levels of inhibitor concentration, where the slopes of these plots are given by Equation 1.
The apparent K i was obtained from the plot of these slopes versus [I], which generated a straight line with a y intercept of K m /V max and a x intercept of ϪK i (22). Crystallization and Data Collection-Because avian ATIC protein crystallizes more readily than human ATIC, apo avian ATIC was used for crystallization experiments. Apo avian ATIC crystals were grown at 22°C, as described previously (7), via the sitting drop vapor diffusion method by mixing equal volume of avian ATIC (10 mg/ml) with a reservoir solution consisting 18% (w/v) polyethylene glycol 8000, 0.2 M imidazole, pH 7.2, and 5 mM dithiothreitol. Crystals of ATIC/326203-A complex were obtained by soaking a native avian ATIC crystal for 6 h with 10 mM 326203-A in mother liquor (20% polyethylene glycol 8000, 0.2 M imidazole pH 7.2, and 5 mM dithiothreitol). Ethylene glycol at 20% was used as cryo-protectant. The data were collected on an ADSC 3 ϫ 3 CCD detector at Beamline 11-1 at Stanford Synchrotron Radiation Laboratory. The data were processed to 1.8 Å using DENZO-SCALE-PACK (23). The crystal is isomorphous with the apo avian ATIC structure previously reported (7) and was indexed in monoclinic space group Structure Determination and Refinement-The structure of ATIC with 326203-A was determined to 1.8 Å resolution by molecular replacement with AMoRe (CCP4 program suite) (25) using the apo avian ATIC coordinates (Protein Data Bank code 1g8m) (7) as the search model. The initial molecular replacement solution after rigid body refinement yielded a correlation coefficient of 53.4% and an R value of 36.7% for data between 15 and 4 Å. The bound ligand was identified using the initial difference F o Ϫ F c and 2F o Ϫ F c electron density maps. Subsequent refinement was performed with CNS (26) using simulated annealing, conjugated gradient minimization, and individual B value refinement with intervening rounds of manual building using the graphic software O (27). A bulk solvent correction was applied throughout the entire refinement. Two-fold noncrystallographic symmetry restraints were applied for the initial five rounds of refinement and then released completely to accommodate any side-chain structural differences between the two monomers. The final model has R cryst and R free values of 21.3 and 24.0%, respectively (see Table I). Because the electron density for one of the 326203-A inhibitors was weaker in one of the active sites in the dimer, most of the structural analysis presented here was carried out on the active site with the better ordered inhibitor.
The model was analyzed with Procheck (28) and WHATCHECK (29). The program MS (30) was used to calculate buried surface area using a 1.4-Å probe. Hydrogen bonds and van der Waals' interactions were identified with LIGPLOT (31) and CONTACSYM (32). Calculations of rmsd on the complexed and apo avian ATICs were carried out with PROFIT (www.bioinf.org.uk/software/profit). Figs. 2C and 4A were created with Bobscript (34).

RESULTS AND DISCUSSION
Before using AutoDock (15) to screen potential small molecule inhibitors, the substrate AICAR was docked into the active site of AICAR Tfase. Good agreement was found between the modeled conformation in the lowest docking energy binding cluster and the x-ray structure (13). We also docked the natural cofactor 10-f-THF in the AICAR Tfase active site because the complexed structure of 10-f-THF is not available because of the instability of the cofactor. The docked conformation of the cofactor was compared with the complexed crystal structures of the two folate analogues, BW1540 and BW2315 (10). Autodock predicted one binding cluster with significantly lower binding energy compared with the other six clusters; the predicted natural cofactor location and orientation closely resembled the actual antifolate crystal structures.
AutoDock (15) was then used to screen 1990 compound from the National Cancer Institute Diversity Set with diverse pharmacophores. Energy scoring was calculated on the sum of the electrostatic and van der Waals' interaction energy between the ligand and the enzyme. Based on the calculated binding energy, AutoDock retrieved 44 best energy-scoring compounds that have scaffolds that differ substantially from known antifolates. We tested 16 of these compounds that had good solubility in water for their ability to inhibit the formyl transfer reaction catalyzed by ATIC. Remarkably, 8 of the 16 inhibitors bound to ATIC with IC 50 values in the micromolar range. 326203-A displayed an IC 50 of 11.6 M and was a competitive inhibitor of 10-f-THF (Fig. 3), with a K i value of 7.1 M. Further details of the docking, database screening, and enzymatic inhibition assay will be described elsewhere (13). Here, we focus on the binding activities of the compounds identified by virtual screening and on the crystal structure of ATIC with inhibitor 326203-A generated from the initial screen (Fig. 2). This compound now serves as a promising lead for future generations of novel inhibitors for ATIC.
Crystal Structure of Avian ATIC Complexed with Inhibitor 326203-A-Because the active site in the ATIC dimer interface is clearly accessible to solvent in the crystals, and minimal conformational changes have been observed upon inhibitor binding (8,10,11), ligand soaking experiments were performed. Because avian ATIC originally crystallized more readily than human ATIC, soaking experiments were performed on the apo avian ATIC crystals. Avian and human ATIC have 83% sequence identity, and their structures are highly conserved (9). The crystals soaked in the solution of compound 326203-A were indexed in monoclinic space group P2 1 (see "Experimental Procedures"). The structure was determined at 1.8-Å resolution by molecular replacement using the apo avian ATIC structure (Protein Data Bank code 1g8m) (7) as the search model. The final model of the complex included residues 4 -593 for both monomers of the dimer, two 326203-A ligands in the AICAR Tfase active sites, two potassium ions, and 748 water molecules (Table I) Inhibitor Binding-The initial F o Ϫ F c electron density map of the inhibitor was well defined (Fig. 2C). Almost every atom of the inhibitor can be distinguished in the electron density map, except for the anticipated Cr 3ϩ . No Cr 3ϩ chelation is observed, which differs from the chemical structure of the inhibitor provided by the National Cancer Institute (Fig. 2A). In the National Cancer Institute chemical structure, Cr 3ϩ is chelated by the carboxylate oxygen (O22) (labeled in Fig. 4B), the hydroxyl oxygen from the pyrazole ring (O17), and the nitrogen from the diaza moiety (N-19), as well as a hydroxide ligand. The chelation of the free ligand with the Cr 3ϩ creates a rigid and planer inhibitor structure. It is possible that the 5 mM EDTA present in the protein solution reacted with the Cr 3ϩ complex and removed the Cr 3ϩ ion from compound 326203-A. Nevertheless, without Cr 3ϩ , the bound ligand in the crystal complex remains rigid and planar, with the pyrazole ring rotated about 30°from the plane of ring A. It is plausible that a proton replaced the Cr 3ϩ , acting as a chelation site to complex the three lone pairs of electrons on oxygens O-17, O-22, and on nitrogen N-19 to form an intramolecular hydrogen bond network (Fig. 2B).
The bound ligand is located at the ATIC dimeric interface around the AICAR Tfase active site (8,11), as anticipated. However, the ligand surprisingly only occupies part of the natural cofactor (10-f-THF)-binding cleft and does not enter the AICAR-binding subsite (Fig. 5A). This structural result corresponds well with the kinetic data, which show that the ligand is a competitive inhibitor only of the cofactor (Fig. 3), but appears to be a noncompetitive inhibitor of the substrate AICAR (data not shown). The ligand consists of three parts: two benzene sulfonate ring-binding regions (A and B) at each end of the ligand and the middle pyrazole moiety (Fig. 2C). 28% (124 Å 2 ) of the total surface area (437 Å 2 ) of the ligand is exposed to the solvent. The corresponding buried surface areas are 77% for the benzene sulfonate moiety of ring A and 71% for the pyrazole ring component, whereas ring B is slightly more exposed to solvent (64% buried surface area).
The 326203-A ligand represents a new class of inhibitors with a novel scaffold that is dissimilar to folates. Consequently, this inhibitor adopts a binding mode that differs substantially from any ATIC-specific inhibitor, such as a MAI with avian ATIC (11), and two folate analogues (BW2315 and BW1540) in complex with human ATIC (10). Thus, in retrospect, it is not surprising to find a novel mode of binding that engages different active site residues. The substantial number of new hydrogen bonding and hydrophobic interactions (Fig. 4B) must then account for the observed binding potency of the inhibitor. All of the interactions are from residues in the biological dimer; no crystal contacts that might lead to artifactual binding were observed. A relatively "conserved" interaction was noted that arises from binding of the sulfonate moiety of ring A in 326203-A to the N terminus of helix 17 (residues 451-469) and suggests that the helix dipole stabilizes the binding of the negatively-charged sulfonate. Furthermore, this proposed oxyanion hole, which is composed of the main-chain amides from the N terminus of helix 17, has previously been proposed to dominate the location and orientation of two sulfonyl antifolate inhibitors (BW2315 and BW1540) in the active site (10). Similarly, a key hydrophilic interaction between Lys 267 and the benzoyl sulfonate moiety of 326203-A is conserved with the antifolates. Here, O-32 and O-34 of the sulfonate form hydrogen bonds with the side chain of Lys 267 , with distances of 2.9 and 3.3 Å, respectively. The positively charged Lys 267 has been suggested to stabilize the oxyanion transition state of the formyl transfer reaction (8,11). In this complex, the polar interactions involving Lys 267 , as well as with the proposed where F o and F c are the observed and calculated structure factor amplitudes. d R free is the same as R cryst , but for 2.5% of the data randomly omitted from refinement. oxyanion hole, both appear to be the key features that confer selectivity and probably a substantial portion of the binding energy of this compound.
The proposed catalytic His 268 does not form a direct hydrogen bond with this ligand, but instead a water molecule W327 mediates this interaction via a relatively long hydrogen bond (3.6 Å). O-33 of the ligand hydrogen bonds with main chain amide of Arg 452Ј from the opposite subunit. No direct interactions are made between active site residues and the ring A benzoate moiety. Nevertheless, a well-organized water shell forms around the carboxylate oxygens to create an extensive hydrogen bonding network that assists in the interaction be-  Fig. 4B). Otherwise, the remaining interactions are via hydrophobic packing and stacking of the aromatic rings. The benzene ring of Phe 316 forms parallel stacking with ring A with a ring-to-ring distance of 3.5 Å, whereas Pro 544Ј stacks against the opposing face of ring A (Fig. 4A).
The primary interaction between the pyrazole ring and the enzyme involves the side chain of Asn 490 and N-12 of the inhibitor with a hydrogen bonding distance of 2.7 Å. Phe 316 , apart from stacking with ring A, also provides a hydrophobic environment for the pyrazole ring by forming van der Waals' interactions with C-14 and C-15 of the ligand. The sulfonate of ring B is anchored atop the N terminus of ␣-helix 19 (residues  5. Superposition of 326203-A with three ATIC-specific inhibitors, MAI, BW2315, and BW1540, and surface representation of AICAR Tfase active site. A, stereo view of 326203-A and MAI in AICAR Tfase-binding site. The solvent-exposed surface is translucent and colored in light gray and blue for the two different subunits of the dimer. The protein structures were first overlaid, and the ligand shown in the comparison reflects only the protein superposition. 326203-A is colored as in Fig. 2C, and the carbon atoms of MAI are colored pink. The three conserved water molecules (W695, W696, and W697) are colored red. B, stereo view of 326203-A superposed with MAI, BW2315, and BW1540 with ϳ90°rotation relative to the view in A, showing the different binding orientation of their benzene rings (highlighted in black oval). 326203-A and MAI are colored as in A; the carbon atoms of BW1540 are colored red, and BW2315 is aquamarine. 484 -498), and its negative charge also appears to be stabilized by the dipole interaction of the helix. A hydrophilic interaction of the sulfonate of ring B is made between the side chain of Glu 487 and O-10. O-8 and O-9 form electrostatic interactions with the side chains of Lys 484 and Glu 487 through two water molecules W400 and W401. Ala 486 and Val 338 provide a favorable hydrophobic environment above the ring B (Fig. 4A).
AutoDock predicted several different binding clusters for 326203 with similar docking energy ranging between Ϫ12.4 and Ϫ14.4 kcal/mol. However, there is no dominant binding cluster and no preferred conformer over the others (13). Nevertheless, the interaction between Lys 267 and the sulfonate of ring A is conserved both in the crystal structure and in some of the predicted binding configurations. The successful prediction of this important polar interactions is probably due to the dominant role of this sulfonate in enzyme-inhibitor binding, as suggested for the two sulfonyl groups in the two antifolate-ATIC complex structures (10). Indeed, docking identified three other novel sulfonate-containing inhibitors; ongoing crystallization studies revealed superimposable benzene sulfonate positions; these form dominant interactions with Lys 267 . The consensus orientations of these inhibitors were, again, successfully reproduced by docking. 2 What is most likely and consistent with all of the docking, enzymatic and structural data, is the domination of the sulfonate interaction with the oxyanion hole that minimizes the contributions of the other interactions affecting the binding mode. The predicted number of clusters with a similar binding energy is consistent with that observation. However, in almost 50% of the binding modes, the sulfonate occupies the same relative location. Different placement of the rest of 326203 could partly arise from the interactions mediated by ordered water molecules, which were not included in the docking of the ligand, or might be due to electrostatic and interactions that play a more important role in the binding than accounted for in the energy scoring function (13).
Comparison with Folate Analogues-Superimposition of inhibitor 326203-A onto MAI, BW1540 and BW2315 (Fig. 5B) reveals that the only overlapping moiety is the benzene ring A with the benzoyl moieties of the ATIC-specific inhibitors. Although stacking of aromatic rings is predominantly involved in all of these interactions, minor differences are apparent. Ring A rotates about 50°with respect to the benzoyl ring of MAI that enables Phe 316 to form a perpendicular stacking with the benzoyl ring of MAI, but parallel stacking with ring A of 326203-A (Fig. 4A). The benzoyl rings of BW2315 and BW1540 also align in a perpendicular orientation compared with ring A where a small rotation of Phe 315 of the human enzyme, the equivalent residue to avian Phe 316 , enables a similar perpendicular stacking. Another form of stacking involves the imidazole ring of substrate AICAR with the benzoyl rings of the antifolates BW2315 and BW1540 (10), which assists in the orientation of these inhibitors in the vicinity of one of the key catalytic residues, Lys 266 , in the human enzyme. Although AICAR is not present in the complex structure of 326203-A and ATIC, Phe 316 provides a similar function.
As discussed above, the sulfonate moiety of ring A in 326203-A is anchored in a position similar to that of the sulfonyl groups of BW1540 and BW2315, which are adjacent to the N terminus of helix 17. Therefore, the helix dipole and the oxyanion hole in the N terminus of the helix seem to play the same roles of stabilizing the negatively-charged sulfonate and directing the binding of 326203-A, as proposed for sulfonyl antifolates (10). Another common feature of binding for these four ATIC-specific inhibitors is the conserved residue, Lys 267 .
Lysine is essential for the catalytic formyl transfer reaction, as shown by mutagenesis studies (39) and various crystal structures (8,11) and appears to be one of the key elements for high affinity binding of BW1540 and BW2315 (10). Slight differences of the attached benzene ring systems appeared such that their polar functional groups adjacent to the benzene rings (the amide carbonyl oxygen of MAI, sulfonyl for BW1540 and BW2315, sulfonate for 326203-A) occupy similar positions (Fig.  5B) and are within electrostatic interactions range of the Lys 267 (the equivalent of Lys 266 in human enzyme).
Otherwise, 326203-A is accommodated in shallow hydrophobic pockets (Fig. 5A) that are not occupied by MAI, BW2315 and BW1540. Most of these residues, Phe 545Ј , Phe 316 , Met 313 , Val 338 , and Ala 486 (Fig. 4A), are not utilized in folate binding, except for Phe 316 . For the ATIC-specific inhibitors, most of their surface area is buried (77% for MAI, 78% for BW1540, and 86% for BW2315), which consequently leads to more extensive interactions with the enzyme than 326203-A, which explains their higher binding affinities. Nevertheless, this novel binding mode of 326203-A utilizes new binding pockets and additional active site residues that can be further exploited to improve future generations of folate and nonfolate-based inhibitors.
Structural Differences upon Binding-One challenge for automated docking arises from the difficulty in accounting for the degrees of freedom of the enzyme in the inhibitor-enzyme interactions. For example, in E. coli GAR Tfase, isomerism of the folate-binding loop occurs at different pH values and upon ligand induced binding (40 -45). Similarly, in the human enzyme, the substrate-binding loop is pH-dependent and affected by ligand binding (45). However, ATIC provides a relatively rigid protein scaffold that is optimal for docking studies. Examination of known AICAR Tfase complex structures reveals that the protein undergoes minimal conformational changes in the protein main chain, whereas some side chain flipping occurs upon inhibitor binding. In the ATIC complex structures with MAI, BW1540, and BW2315, upon AICAR binding, Arg 208 rotates about 50°into the AICAR-binding site to interact with one of the phosphate oxygens of AICAR. Because no AICAR is present in the 326203-A crystal structure, this arginine residue points away from the active site and hydrogen bonds with Ser 235 at a distance of 3.5 Å. The side chain of Asn 490 adopts a different conformation compared with the apo enzyme by rotating about 80°toward the ligand, acting as a hydrogen bond donor to the nitrogen atom (N-12) in the pyrazole ring (Fig. 4A).
In human ATIC complexes with BW2315 and BW1540, the side chain of Asn 431 rotates more than 90°(relative to the MAI complex structure) toward the AICAR-binding pocket to hydrogen bond with one of the sulfonyl oxygens, as well as the 5-amino moiety of AICAR. This conformational change appears to be induced by antifolate binding (10). Although this same movement occurs for Asn 432 in one of the active sites of 326203-A avian complexes, it is not involved in a direct interaction with the ligand because of its considerable distance from the ligand sulfonate (about 5 Å). Arg 452Ј in the same active site moves about 4 Å toward the AICAR-binding site to hydrogen bond with Asn 432 .
Similar Structure Search for Improved Binding Affinity-Although the affinity of 326203-A to ATIC is modest (M), the compound is intriguing because it possesses an unexplored scaffold, different from conventional antifolates. Furthermore, the ATIC-326203 complex structure assists in identifying interactions that have not been observed in previous ATIC complex structures. To optimize 326203-A as a lead compound for future inhibitor design via chemical synthesis, a further similarity search in Available Chemical Directory or National Cancer Institute Compound Library was performed. Compounds with similar shape and higher complementarity to the active site of ATIC should exhibit higher binding affinity to the enzyme, which could then serve as a better lead. Visual examination of ligand binding reveals that ring A and the central pyrazole ring are mostly buried in the active site and presumably provide most of the binding specificity (Fig. 5). In addition, the molecular mass of the inhibitor already exceeds 500; therefore, ring B was eliminated from the subsequent substructure design. Furthermore, the planar and rigid structure of 326203-A suggested formation of an intramolecular hydrogen bond network (Fig. 2B); it is conceivable that heterocyclic structures such as quinoline or indole, which partially mimic this ring system, could preserve a similar set of interactions with ATIC.
Substructure searches of the Available Chemical Directory with a quinoline or an indole ring as templates were performed and followed by visual inspection that lead to seven candidate compounds. In addition, a similarity search of the National Cancer Institute Compound Library yielded 10 compounds.
AutoDock was then used to score the binding of these compounds to the active site of ATIC, which was followed by in vitro inhibition assays on the top-scoring candidates with good solubility. The most potent of these compounds, SALOR2, has an IC 50 of 1.4 M, which is about eight times more potent than 326203-A (Table II). Although our initial substructure and similarity searches were relatively crude, we believe that a more systematic substructure search, combined with rational design, will lead to more potent inhibitors containing these novel scaffolds.
Implications in Future Lead Discovery-326203-A represents a completely new chemical scaffold that differs from known folate analogues and possesses a novel binding mode. It has already served as a good structural template for the rational design and optimization of 326203-A as a lead compound. Initial exploitations of small scale similar chemical structure searches in the National Cancer Institute Compound Library and Available Chemical Directory have generated compounds with increased binding affinity. Inhibitor 326203-A and its derivatized compounds contain one or more sulfonate groups, which notably contribute most to the binding affinity and specificity. However, maintaining such a negatively-charged functional group in the inhibitor will pose severe difficulties for cellular uptake. A surrogate of the sulfonate, such as a sulfonamide or carbonate, common drug components, could be suitable alternatives. Another structural feature of the ATIC 326203-A complex structure that could be exploited further is the aromatic stacking interaction of Phe 316 with the sulfabenzoyl ring that assists in anchoring the sulfonate group of 326203-A in the vicinity of Lys 267 . Frequently, optimization of specific hydrophobic interactions has been attributed to enhanced inhibitor binding affinity. For example, aromatic stacking formed between the pyrimidine ring of dUMP and benzoquinazoline ring of antifolate, BW1843U89, has been suggested to be the main inhibition mechanism in thymidylate synthase (46) by blocking the key interaction between dUMP and the catalytically essential cysteine residue. The increased affinity of the BW1843U89 over its predecessor is attributed to the enhanced aromatic properties of the inhibitor. Similarly, comparison of the more potent compound SALOR2 and inhibitor 326203-A revealed SALOR2 possesses a naphtho-thiophene ring and an indole ring. Hence, the enhanced aromatic properties and rigidity of compound SALOR2 over inhibitor 326203-A may also be responsible for the improved binding potency. Therefore, one future direction is to generate different fused heterocyclic ring systems to harness the aromatic interactions to improve the affinity. The original sets of hydrogen bonding interactions present in the ATIC-326203 complex should also be maintained, especially the interaction with Lys 267 . A charged hydrogen bond can contribute up to 4.7 kcal/mol of binding energy, which is equivalent to a 3000-fold increase in affinity (47).
The active site of AICAR Tfase has been well characterized (8,11), which provides an excellent platform for iterative cycles of structure-based inhibitor design. Traditional folate analogues are developed based on the natural folate cofactor scaffold. The structural similarities of these folate analogues pose a great challenge for improving their binding specificity to the target folate-dependent enzymes. It is certainly likely that binding specificity and potency profile could be improved by exploring the surprisingly large binding pockets and cavities of AICAR Tfase that are not occupied by substrate or cofactor. Binding of inhibitor 326203-A leaves the AICAR-binding site essentially empty. In addition, the benzene ring B of the inhibitor extends toward the other subunit, where it occupies a binding site not used by folates that has not been explored previously in folate analogue design.
Although direct hydrophobic and electrostatic interactions involving the active site residues are important for attaining the binding potency of an inhibitor, hydrogen bonds with active site water molecules also need to be taken into account. A well organized water shell (water molecules W695, W696, and W697) is seen in the AICAR-binding pocket (Fig. 5A) mediated through main chain interactions, including Gly 210 , Ile 239 , and Asn 240 . These water molecules have been observed in most of the apo ATIC and complex structures, except for ATIC-BW2315 complex, where the water molecules W695 and W696 are absent. Although these water molecules do not seem to be involved in the catalytic process, they assist in anchoring the substrate AICAR by mediating hydrogen bonding between AICAR and enzyme (8). They appear to act as the integral part of the protein structure in the absence of substrate or ligand. Displacement of the active site water molecules has been recognized as one important design strategy to gain additional binding efficiency (33). The appropriate selection of functional groups that provide favorable steric and polar interactions could potentially replace the water molecules and enable analogous interactions directly with the protein.
In conclusion, this proof of concept study not only revealed the structural basis for the binding specificity and affinity of this novel inhibitor of ATIC but also offers an attractive platform for future lead discovery and subsequent structure-based lead optimization efforts for ATIC.