Genome Informatics 17(1): 13{22 (2006) 13

Computational Methods for Functional Site

Identi�cation Suggest a Substrate Access Channel in

Transaldolase

Michael Silberstein1 Melissa R. Landon1 Yaoyu E. Wang1

mikesilb@bu.edu mlandon@bu.edu yew@bu.edu

Andras Perl2 Sandor Vajda3

perla@upstate.edu vajda@bu.edu

1 Graduate Program in Bioinformatics, Boston University, Boston, Massachusetts02215

2 Department of Medicine and Microbiology and Immunology, College of Medicine,State University of New York, Syracuse, NY 13210

3 Department of Biomedical Engineering, Boston University, Boston, Massachusetts02215

Abstract

The homozygous deletion of Serine 171 results in the catalytic inactivation of the human transal-dolase. Since Ser171 is in an outside loop, whereas the catalytic site is inside of the �/�-barrel ofthe protein at least 15 �A away, the loss of activity is di�cult to explain. Two distinct computationalmethods are used to elucidate the potential origin of inactivation. Computational solvent mapping,which moves small organic molecules as probes around a protein surface and �nds favorable bind-ing positions, identi�es the region around Ser171 as an important binding site. Three-dimensionalcluster analysis, based both on a reference structure and multiple sequence alignment, shows thata patch of functionally important residues extends from Ser171 toward the catalytic site. Based onthe �ndings of these two methods, we propose a novel ligand access path connecting these speci�csites to the enzyme's active site. We also suggest that this mechanism may be aided by a signi�-cant conformational change involving the separation of two helices, �D and �G, in order to createan easy-access channel between the Ser171-related site and the active site. Further experimentalprocedures will be necessary to examine the biological feasibility of this proposed ligand shuttlingpath.

Keywords: molecular recognition, substrate binding, computational solvent mapping, 3D clusteranalysis

1 Introduction

The enzyme transaldolase is involved in the well-studied pentose phosphate metabolic pathway [10].The importance of this pathway is two-fold. First, ribose-5-phosphate, a key precursor in the synthesisof DNA and RNA, is a metabolic end product of the pathway. Second, reducing equivalents, in the formof NADPH, are formed in order to allow multiple anabolic pathways to progress toward the biosynthesisof key macromolecular components such as amino acids, lipids, and nucleic acids. Transaldolase(E.C. 2.2.1.2) catalyzes the reaction that converts sedoheptulose-7-phosphate and glyceraldehydes-3-phosphate to erythrose-4-phophate and fructose-6-phosphate [6, 13]. The enzyme accomplishesthis task by removing a three-carbon moiety o� of sedoheptulose-7-phosphate via a covalent Schi�base mechanism and subsequently inserting this three carbon unit onto glyceraldehyde-3-phosphate,in order to generate a key six-carbon glycolytic intermediate molecule [13]. The formation of the

14 Silberstein et al.

covalent Schi� base intermediate is facilitated by Lys142, located near the center of the enzyme's �/�barrel structure [6, 15].

The homozygous deletion of three nucleotides coding for Ser171 of the human transaldolase genewas observed in a 9-year-old girl su�ering from cirrhosis of the liver [5, 18] It was subsequentlydetermined that the mutant protein with no Ser171 was e�ectively transcribed and translated in

vitro, but had no detectable activity [5]. Despite the e�ects that this deletion exhibits, the structuralmechanism that directly causes the inactivation and eventual degradation of transaldolase is notunderstood. Ser171 is located in a loop connecting the secondary structure elements, �5 and �5.Mutations in surface loop regions typically do not result in large conformational changes, and onewould expect the protein's structure to adapt to a slightly shorter loop without compromising theintegrity of the overall structure. The assumption that the native and mutant proteins have similarstructures implies that Ser171 must play some role in substrate recognition, although it is locatedmore than 15 �A away from the known active site [13].

We have used computational methodologies to investigate the potential roles of residue Ser171in substrate binding to transaldolase. Since the deletion of Ser171 does not unfold the protein andyet inactivates the enzyme, this residue is likely to play a role in the binding of the substrate and,consequently, may be part of a functional site. Two di�erent methods that have been developed forthe identi�cation and characterization of protein functional sites were used to test this hypothesis.The �rst method, computational solvent mapping (CS-Map), is based on the experimental work byRinge and coworkers [1, 9] who determined protein structures in aqueous solutions of organic solvents;in each of their studies, only a limited number of solvent molecules were bound to the resulting crystalstructure. Furthermore, when �ve or six structures of a protein determined in di�erent solvents weresuperimposed, the organic molecules tended to cluster in the active site, forming \consensus" sitesthat delineated important subsites of the binding pocket [9]. All other bound solvent molecules wereeither in crystal contact, occurred only at high ligand concentration, or were in small, buried pocketswhere only a few types of solvent molecules clustered as compared to the active site. The CS-Mapalgorithm [4, 7, 14] moves and minimizes small organic molecules (\probes") on the protein surfacein order to �nd the most favorable binding positions. Probe conformations are clustered and rankedbased on the average free energy of the cluster. We note that all bound ligands and water moleculesare removed prior to mapping, and thus the results are based only on the structure of the protein.CS-Map has reproduced correctly the available experimental solvent mapping results [4].

The second method utilized in the identi�cation of potential functional sites of transaldolase is thethree-dimensional cluster analysis developed by Eisenberg and co-workers [8]. While solvent mappingis based entirely on the structural and chemical properties of the target protein and ligand probes, thethree-dimensional cluster analysis combines the information coming from multiple sequence analysiswith the structural proximity of various residues within a given site in order to predict the regions offunctional and sequential conservation for a given protein [8].

Both solvent mapping and three-dimensional cluster analysis suggest that Ser171 is part of awell-de�ned functional site in transaldolase. Mapping results for transaldolase yielded the largestconsensus site close to the Ser171 position, and a second site approximately 11 �A away from Ser171,on the opposite end of helix �5. The three-dimensional cluster analysis shows signi�cant conservationfor Ser171 and its neighboring residues, in addition to residues located within the protein's bindingsite. Based on the above observations, we propose a novel access channel for the entry of the substrateentry into the binding site involving Ser171 that may involve a signi�cant conformational change.Further experimental veri�cation will be necessary to con�rm or refute this proposed access channel.

Computational Methods for Functional Site Identi�cation 15

2 Materials and Methods

2.1 Protein Structures

The crystal structure �le 1F05 of human transaldolase from the Protein Data Bank (PDB) [2] wasutilized in our studies. While this structure is a homodimer, the enzyme is active as a monomerin solution [12]; since the monomer-monomer interface does not involve any residue of interest, allanalyses were performed on a single subunit, chain A. The 1F05 structure is the only one availablefor human transaldolase, and it does not include any bound ligand. However, nine more transaldolasestructures from bacteria are available, derived primarily from E. coli. One of the E. coli structures,1UCW, has been co-crystallized with the reduced Schi�-base intermediate [6]. The location of theSchi� base provides information on the catalytic residues, and one can also reconstruct the positionof the sedoheptulose 7-phosphate in the binding site. The bacterial structures were used only forcomparison.

2.2 Computational Solvent Mapping

Seven small molecules (acetone, acetonitrile, tert-butanol, dimethylsulfoxide, phenol, isopropanol, andurea) were used as probes [14]. The �ve steps of computational solvent mapping have been describedpreviously [4, 7, 14]; a summary of the four steps used in this study is given here, since the sub-clustering step was omitted from this analysis.

(1). Rigid body search. For each probe, 2,000 docked conformations were generated using the rigid-body docking algorithm GRAMM [16, 17]. We used a 1.5 �A grid step for translations and 15�

increments for rotations.

(2). Minimization and re-scoring. The free energy of each of the 2,000 complexes, generated inStep 1, is minimized using the free energy potential �G = �Eelec + �Evdw + �Gdes, where�Eelec, �Evdw, and �Gdes denote the electrostatic, van der Waals, and desolvation contributionsto the protein-probe binding free energy [7]. The sum �Eelec + �Gdes is obtained by theAnalytic Continuum Electrostatic (ACE) model [11] as implemented in version 27 of Charmm [3]using the parameter set from version 19 of the program. The model includes a surface areadependent term to account for the solute-solvent van der Waals interactions. The minimization isperformed using an adopted basis Newton-Raphson method as implemented in Charmm. Duringthe minimization the protein atoms are held �xed, while the atoms of the probe molecules arefree to move. At most 1,000 minimization steps are allowed, although most complexes requirefar fewer steps to achieve convergence.

(3). Clustering and ranking. The minimized probe conformations from Step 2 are grouped intoclusters based on Cartesian coordinate information. The method creates an appropriate numberof clusters such that the maximum distance between a cluster's hub and any of its members (thecluster radius) is smaller than half of the average distance between all the existing hubs. Clusterswith less than 10 members are excluded from consideration. For each retained cluster, wecalculate the probability pi = Qi=Q, where the partition function Q is the sum of the Boltzmannfactors over all conformations, Q = Sj�exp(��Gj=RT ), and Qi is obtained by summing theBoltzmann factors over the conformations in the ith cluster only. The clusters are ranked on thebasis of their average free energies < �G >i= Sjpij�Gj , where pij = exp(��Gj=RT )=Qi, andthe sum is taken over the members of the ith cluster.

(4). Determination of consensus sites. Mapping is primarily used to �nd \consensus" sites, e.g.locations on the protein at which many di�erent probe molecules cluster. In order to �nd theconsensus sites, we select the minimum free energy conformation in each of the �ve lowest average

16 Silberstein et al.

free energy clusters for each solvent. The structures are superimposed, and the position at whichmost probes of di�erent types overlap is de�ned as the main consensus site.

2.3 Three-Dimensional Cluster Analysis

Three-dimensional cluster analysis, developed by Landgraf et al. [8] o�ers a method for the predictionof functional residue clusters in a family of proteins. The method requires a reference structure and amultiple sequence alignment as input data, and it is based on evaluating the residue conservation ateach position of the reference structure. We used the PDB structure 1F05 as reference. Homologoussequences were identi�ed in the NCBI nonredundant protein database by a FASTA search. Multiplesequence alignment was based on the full-length sequence of the reference protein, and it was performedusing ClustalW on the sequences pertaining to the following transaldolase structures: 1ONR, 1I2R,1UCW, 1I2Q, 1I2O, 1I2P, 1I2N, 1VPX, and 1WX0. The conservation was evaluated for the 3D regionsurrounding residue x as de�ned by the spatial neighbors of x in the reference structure. The regionalconservation, CR(x), was calculated and assigned to residue x. This score represents the di�erence inconservation between the structural neighbors of residue x versus the protein as a whole. The higherthe CR(x) score of a residue, the better its neighborhood is conserved as compared to the protein asa whole, and this measure of conservation was used as the primary predictor of functional sites.

3 Results

3.1 Computational Solvent Mapping

In order to investigate the potential role of Ser171 in transaldolase, we mapped the crystal structureof the wild type protein. As shown summarized in Table 1, mapping yields two large consensus sites,each with probe clusters of �ve di�erent types (Sites 1 and 2 in Table 1). Site 1 is most relevantto our analysis, as it is located less than 3 �A from Ser171. This site includes the lowest free energycluster for acetonitrile, the third lowest free energy cluster for isopropanol, and the fourth lowest freeenergy clusters for acetone, DMSO, and phenol. The residues that surround this consensus site, inthe order of decreasing number of contacts with the probes, are Asp301, Phe172, Leu297, Ser171,Ser216, Pro212, Ala173, Glu300, and His197. In total, the probes that cluster at Site 1 make 179non-bonded contacts with Ser171, indicating the signi�cance of this residue in the predicted bindingsite. Hydrogen bonds are formed between the probes and some of the above named residues, but notwith Ser171.

In the orientation shown in Figure 1, the long helix on the left side of the protein is helix �G. The

Table 1: Free energy ranking of probe clusters within the consensus sites for 1f05 using GRAMMa.

Consensus ProbeSite acetone acetonitrile isopropanol DMSO phenol urea t-butanol

1 4 (3.17) 1 (2.93) 3 (2.74) 4 (2.93) 4 (1.96)2 1 (11.46) 1 (11.63) 2 (11.35) 2 (10.58) 4 (11.05)3 2 (31.20) 3 (31.01) 2 (31.21) 1 (30.52)4 3 (21.97) 3 (21.18) 2 (20.29)5 5 (23.10) 4 (22.44) 1 (21.63)6 3 (32.55) 5 (30.20)7 2 (17.03) 1 (16.32)8 5 (30.48) 3 (30.20)

aThe number in the parenthesis shows the distance of each cluster center from the nearestatom of S171.

Computational Methods for Functional Site Identi�cation 17

Figure 1: Computational solvent mapping results on human transaldolase. Lys142 and Ser171 areshown for reference and are colored blue and magenta, respectively. Shown here are the top �ve lowestfree energy cluster center positions for each of the following probes: acetone (green), acetonitrile (red),isopropanol (grey), phenol (light blue), DMSO (hot pink), urea (orange), and t-butanol (yellow).

active site of the enzyme is indicated by the side chain of Lys142, shown in blue. Consensus site 1 islocated in a pocket that is de�ned by the turn between helices �F and �G from the right, helix �6from the right, and the loop between �5 and from the back, where the latter helix is perpendicularto the plain of the �gure and hence is shown as a circle behind the consensus site. From the top, thesite is surrounded by helix �D. As we will discuss, this helix is connected to �6 by a long loop. Thisfactor will be important, as we will needs to assume that upon substrate binding helix �6 can slightlymove to the right and the C-terminal helix �G moves to the left, creating an access channel from Site1 at Ser171 toward the active site.

The second consensus site, containing 5 di�erent probe clusters (Site 2 in Table 1), is located atthe distal end of helix �5, approximately 11 �A from Ser171. In Figure 1, Site 2 is in the background,slightly left from Site 1. The residues surrounding this site are Glu180, Trp147, His288, Leu289,Ala181, Gln151, Ala177, and Lys154. Hydrogen bonds are formed with residues Gln151, Lys154,Ala177, Glu180, His288, and Leu289. Sites 3 though 6 are surface pockets, unrelated to any knownsite. The only clustering of probes inside the �-barrel is at Site 7, consisting of the lowest free energycluster of urea and the second lowest free energy cluster of acetonitrile.

It is surprising that no large consensus site is found near the active site of transaldolase. In fact,in most previous applications of the mapping to enzymes, the highest or second highest number ofdi�erent probe clusters occurs in a subsite of the active site [14]. We have seen exceptions for two

18 Silberstein et al.

systems so far. The �rst such enzyme is haloalkane dehalogenase, which binds very small substratesin a narrow channel. Thus, the larger probes are unable to get to the active site, and they cluster atthe two ends of the channel. Since the transaldolase structure is wide open, this problem is unlikely tooccur here. The second case in which no clustering of the probes occurs at the active site is the familyof mammalian cytochrome P450 (CYP) proteins, but only when mapping the unbound state. Theseproteins have a very large cavity for substrate binding above the heme groups that could bind severalsubstrates simultaneously; yet binding is normally accompanied by conformational changes that createsmall crevices, in turn binding both the substrate and the probes. The large binding pockets prior tosubstrate binding seem to lack the structural features required for the small molecules used as probesin the docking calculations; however, given the nature of binding in the P450 proteins, it is reasonableto conjecture that transaldolase may undergo a signi�cant conformational change upon ligand binding.

3.2 Three-Dimensional Cluster Analysis

The results of the three-dimensional clustering analysis are shown in Figure 2. The colored residuesall have CR(x) values exceeding the stringent background threshold of CR(x) = 1:8, which impliesless than 1% expected false positives. The most important result is a well de�ned strip of functionallyimportant residues from the pocket at the Ser171 location toward the active site, represented inFigure 2 by the side chain of Lys142, which is itself a well-conserved residue (red). It is interestingthat the CR(x) values are relatively high for a number of residues on the 168{171 loop. Overall, theseresidues include many residues that line up the beta barrel and typically are within �3 residues fromknown active site residues, such as Asp27, Asn45, Glu106, Lys142, Asn164 Thr166, and Ser186. Asa result, this method was able to identify regions of known substrate binding and catalytic activity.Additionally, regions containing residues of high CR(x) values extend beyond the regions of the betabarrel and towards the loop containing Ser171 as well as the neighboring helix �D (residues 189{199),and the loop connecting �7 and �7. These regions are all within close proximity of Ser171 and providefurther con�rmation of the functional importance of this residue.

4 Discussion

Previous studies have shown that deletion of the loop residue, Ser171 of human transaldolase can causeinactivation of the enzyme, as determined in a 9-year-old girl su�ering from liver cirrhosis. However,the biomolecular details of this occurrence have not been clear to date, in particular how a loop residuethat is far from the enzyme's active site is directly involved in the mechanism of inactivation.

In order to shed some light on mechanism of transaldose inactivation upon deletion of Ser171,we employed two computational methodologies for functional site elucidation. Computational solventmapping addresses the issue of binding site determination primarily on a structural and biochemicalbasis by assessing where the most diverse set of small molecular probes bind along the surface of theprotein. Three-dimensional clustering provides supporting evidence of the functional site predictionsby the comparison of multiple sequence alignments consisting of the residues in a speci�c structuralregion to the multiple sequence alignment of the entire protein. As a result, these two methodologiescan be used to complement each other towards determining novel functional sites. The solvent mappingfound that the highest concentration of di�erent probes cluster in a pocket adjacent to the residueSer171, indicating that the region may serve as binding site for one of the substrate, most likely atransient one since it is located more than 15 �A from the active site. Furthermore, three-dimensionalclustering reveals a string of functionally important residues from the just discovered pocket containingSer171 toward the active site.

The above results suggest a potential mechanism of substrate entry involving the Ser171 regionas a primary binding site, followed by transition of the molecule along the path based on the 3Dclustering results. However, the existence of such an entrance route assumes a substantial, albeit

Computational Methods for Functional Site Identi�cation 19

Figure 2: Three-dimensional clustering analysis results on human transaldolase. Residues whichcontain CR(x) values greater than or equal to 1.8 have been mapped onto the structure. The colorscheme for these residues is the following: Residues with CR(x) values from 1.80{2.00 are colored blue,2.00{2.25 are slate, 2.25{2.50 are cyan, 2.50{2.75 are green, 2.75{3.00 are yellow, 3.00{3.25 are orange,3.25{3.50 are salmon, and 3.50{3.75 are red. For reference, the sidechains for Lys142 and Ser171 havebeen included in the color indicative of their CR(x) values.

localized, conformational change. The surface representation of transaldolase in Figure 3 shows thatif the site containing Ser171 has potential to be involved in ligand binding, it is primarily separatedfrom the active site cleft by a `ridge' that is largely caused by helix �D protruding outward. However,a simple helix motion which would move helices �D (colored green) and �F/G (colored red) slightlyapart would remove this `ridge' and create an easy-access channel from the active site to Ser171, andextending potentially father towards the second highly occupied consensus site, located 11 �A away fromSer171. The feasibility of this helix separation increases signi�cantly when we consider two importantfactors. First, the loop that separates helix �F/G from the rest of the structure is unusually long,consisting of 24 residues (266{290), with only a single helical turn interspersed within this long loop.This exceptionally long loop may provide the capability for facilitating such a conformational change.Second, adjacent to helix �D is an eleven residue loop (residues 200{211) that has signi�cantly highB-factors (>50). Apart from the two termini of the protein sequence, B-factors this high are not seenanywhere else throughout the structure. These two structural features may supply enough exibilityto provide for the conformational change necessary to assist in the shuttling of substrates and/orproducts between the active site and these putative binding locations along the protein surface.

20 Silberstein et al.

The results that we describe here can provide a novel hypothesis as to a detailed path of ligandmovement towards and/or away from the active site of human transaldolase. In order to providefurther veri�cation of this hypothesis, additional experimental studies will be necessary to probe theissues regarding the protein's exibility and conformational rearrangements. Information providedfrom further site-directed mutagenesis studies of the various consensus sites as well as the interspersedloop regions may shed further light as to the relative a�nity of speci�c consensus sites toward ligandinteractions, in addition to as the potential importance of the conformational changes that have beenproposed here.

Figure 3: Surface depiction of the region of transaldolase where the two primary consensus sites havebeen observed. The coloring scheme for the probe molecules is identical to Figure 1. Additionally,the active site lysine, Lys142 (colored blue) and Ser171 (colored magenta) have been included forreference. As described in the text, a ligand potentially bound at the Ser171-based consensus sitehas to pass over a `ridge' created primarily by the helix, �D (green) protruding outward, in orderto arrive at the active site. However, by moving helices �D and �G (red) apart from each other viaa conformational rearrangement, an access channel can form allowing easier access to and from theactive site.

5 Conclusions

We have applied two unique computational methodologies to the structure of human transaldolase inorder to provide further biomolecular explanations as to the reason for inactivation of catalytic activityas a result of Ser171 deletion. Signi�cant signals are observed from both methodologies concerning

Computational Methods for Functional Site Identi�cation 21

the potential functional importance of this mutated residue. We have also noticed speci�c connectionsbetween the Ser171 related site with other regions of the protein, including the active site. As a result,we have proposed a novel path for ligand entry and/or exit that appears to be potentially linked toa conformational change, allowing easy access between the above site and the enzyme's active site.The feasibility of the proposed path needs to be further tested experimentally to provide furthercon�rmation of this ligand-binding mechanism.

Acknowledgments

We thank Dr. Lawrence Brown for constructing and investigating the homology model of the mutantprotein with Ser171 deleted and Deepa Rajamani for performing further molecular dynamics simula-tions. This work is funded by grants DBI 0213832 from the National Science Foundation and RO1GM64700 from the National Institute of Health to S.V., and by grant RO1 DK 49221 from from theNational Institute of Health to A. P.

References

[1] Allen, K. N., Bellamacina, C. R., Ding, X., Je�ery, C. J., Mattos, C., Petsko, G. A., and Ringe,D., An experimental approach to mapping the binding surfaces of crystalline proteins, J. Phys.Chem., 100:2605{2611, 1996.

[2] Berman, H. M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T. N., Weissig, H., Shindyalov,I. N., and Bourne, P. E., The Protein Data Bank, Nucleic Acids Res., 28:235{242, 2000.

[3] Brooks, B. R., Bruccoleri, R. E., Olafson, B. D., States, D. J., Swaminathan, S., and Karplus,M., CHARMM: A program for macromolecular energy, minimization, and dynamics calculations,J. Comp. Chem., 4:187{217, 1983.

[4] Dennis, S., Kortvelyesi, T., and Vajda, S., Computational mapping identi�es the binding sites oforganic solvents on proteins, Proc. Natl. Acad. Sci. USA, 99:4290{4295, 2002.

[5] Grossman, C. E., Niland, B., Stancato, C., Verhoeven, N. M., van der Knaap, M. S., Jakobs,C., Brown, L. M., Vajda, S., Banki, K., and Perl, A., Deletion of serine 171 causes inactivation,proteasome-mediated degradation, and complete de�ciency of human transaldolase, Biochem. J.,382:725{731, 2004.

[6] Jia, J., Schorken, U., Lindqvist, Y., Sprenger, G. A., and Schneider, G., Crystal structure of thereduced Schi�-base intermediate complex of transaldolase B from Escherichia coli: mechanisticimplications for class I aldolases, Protein Sci., 6:119{124, 1997.

[7] Kortvelyesi, T., Dennis, S., Silberstein, M., Brown, L. III, and Vajda, S., Algorithms for compu-tational solvent mapping of proteins, Proteins, 51:340{351, 2003.

[8] Landgraf, R., Xenarios, I., and Eisenberg, D., Three-dimensional cluster analysis identi�es inter-faces and functional residue clusters in proteins, J. Mol. Biol., 307:1487{1502, 2001.

[9] Mattos, C. and Ringe, D., Locating and characterizing binding sites on proteins, Nat. Biotechnol.,14:595{599, 1996.

[10] Mayes, P. A., The pentose phosphate pathway & other pathways of hexose metabolism, Harper'sBiochemistry, 23 ed., 201{211, Appleton & Lange, Murray, R. K., Granner, D. K., Mayes, P. A.,and Rodwell, V. W. eds., 1993.

22 Silberstein et al.

[11] Schaefer, M. and Karplus, M. A., A comprehensive analytical treatment of continuum electro-statics, J. Phys. Chem., 100:1578{1599, 1996.

[12] Schorken, U., Jia, J., Sahm, H., Sprenger, G. A., and Schneider, G., Disruption of Escherichia colitransaldolase into catalytically active monomers: evidence against half-of-the-sites mechanism,FEBS Lett., 441:247{250, 1998.

[13] Schorken, U., Thorell, S., Schurmann, M., Jia, J., Sprenger, G. A., and Schneider, G., Identi�ca-tion of catalytically important residues in the active site of Escherichia coli transaldolase, Eur.J. Biochem., 268:2408{2415, 2001.

[14] Silberstein, M., Dennis, S., Brown III, L., Kortvelyesi, T., Clodfelter, K., and Vajda, S., Identi-�cation of substrate binding sites in enzymes by computational solvent mapping, J. Mol. Biol.,332:1095{1113, 2003.

[15] Thorell, S., Gergely, P. Jr., Banki, K., Perl, A., and Schneider, G., The three-dimensional struc-ture of human transaldolase, FEBS Lett., 475:205{208, 2000.

[16] Vakser, I. A. and A alo, C., Hydrophobic docking: A proposed enhancement to molecular recog-nition techniques, Proteins, 20:320{329, 1994.

[17] Vakser, I. A., Matar, O. G., and Lam, C. F., A systematic study of low-resolution recognition inprotein{protein complexes, Proc. Natl. Acad. Sci. USA, 96:8477{8482, 1999.

[18] Verhoeven, N. M., Huck, J. H., Roos, B., Struys, E., Salomons, G. S., Douwes, A. C., van derKnaap, M. S., and Jakobs, C., Transaldolase de�ciency: liver cirrhosis associated with a newinborn error in the pentose phosphate pathway, Am. J. Hum. Genet., 68:1086{1092, 2001.