Genome scale prediction of substrate specificity for acyl adenylate superfamily of enzymes based on active site residue profiles

Phylogenetic and HMM profile analysis

Figure 2 shows a dendrogram obtained from a set of sixty representative sequences from various subfamilies. As can be seen, the sequences utilizing similar substrates have clustered together in the phylogenetic tree. This indicates that despite the large sequence divergence within each subfamily, the sequences of a given subfamily have distinct features which distinguish them from other subfamilies, such that they cluster in a subfamily specific manner. Therefore profile based methods like HMMs[49] can be used to devise a prediction protocol for assigning the substrate specificity to an uncharacterized protein belonging to acyl adenylate superfamily. For each subfamily, the sequences were randomly divided into training set and test set. The training set was used to derive subfamily specific HMM profiles and the test set was used to benchmark the prediction accuracy. The sequences of the training set were aligned by CLUSTALW and the resulting multiple sequence alignment was given as input to the HMMER package[49] for building HMM profiles for each subfamily. All the six HMM profiles were compiled in the form of a HMM library. Each sequence in the test set was matched with all six profiles in this library of HMM's using the program 'hmmpfam'. The test sequence was scored against each subfamily HMM and substrate preference was assigned to the test sequence based on best scoring HMM i.e. if the query sequence scores best with HMM for 4-Coumarate:CoA ligases, then the specificity assigned to the query would be Coumarate. By comparing predicted substrate specificity to the known substrate specificity, sensitivity (Sn) and specificity (Sp) were calculated to determine the overall prediction accuracy of the method. Table 2 summarizes the results for HMM based substrate specificity predictions for the acyl-adenylate superfamily. The sensitivity ranged from 0.91 for MCS to 1.0 for luciferases. The specificity was 0.96 for LCS, 0.99 for MCS and 1.0 for other subfamilies. Thus based on whole sequence HMM profiles it is possible to correctly group acyl adenylate superfamily of enzymes into various subfamilies as per their substrate specificity. However, this whole sequence HMM failed to predict correctly the substrate specificities for the enzymes, CG6178, At4 g05160, At5 g63380 and At1 g62940 which show sequence homology to luciferase and 4CLs, but utilize long chain or medium chain fatty acyl substrates (Table 3). Rather than predicting them as long chain or medium chain CoA synthetases, the HMM profiles predicted them as luciferases and 4CLs respectively.

Prediction of substrate specificity based on SDRs

It is possible that, the enzymes CG6178, At4 g05160, At5 g63380 and At1 g62940 which show anomalous substrate preference have acquired specificity for long chain substrates as a result of specific mutations in the active site pocket residues, while rest of the protein shows similarity to luciferase or 4CLs. To test the hypothesis, we attempted to develop a prediction protocol based on a limited set of binding pocket residues responsible for substrate selection. In an earlier study on adenylation domains of NRPS, Stachelhaus et al have demonstrated that predictions based on binding-pocket residues were significantly better than whole sequence based approaches [32]. Based on the crystal structure of adenylation domain of gramicidin synthetase (PDB code:1AMU) in complex with the substrate phenylalanine, it had been proposed that the substrate selectivity of NRPS adenylation domain is governed by the residues forming the substrate binding pocket [32]. Since the phenylalanine binding site on 1AMU is in close proximity of the conserved AMP binding site and the catalysis by the superfamily of enzymes involve formation of covalent bond between AMP and acyl group, it was reasonable to assume that other subfamilies will have the acyl group binding site at a position structurally analogous to the Phe binding pocket of 1AMU. Therefore in our work, all amino acids which fall within 6 Å distance from the C_β of the substrate phenylalanine were identified as putative substrate specificity determining residues (SDRs) (Figure 3). By this method, 12 amino acids positions were identified. These include 7 out of the 10 residues which are identified in 1AMU as the substrate binding pocket residues [18]. The remaining three, (position number 239, 278 and 299) were also included in our analysis, even if they were at a distance slightly larger than 6 Å. Hence, a total of 15 amino acid positions were defined as the SDRs. The sequential order of these 15 SDRs was defined as the active site profile (ASP) for each enzyme in the training set (Figure 3). Multiple sequence alignments were obtained for each subfamily by CLUSTALW and separate position specific scoring matrices PSSMs were built for each subfamily as described in methods. The PSSM for the ASPs for each subfamily were used to score the ASP of a query sequence and the substrate preference was assigned based on the highest scoring PSSM. The prediction accuracy of this protocol was also evaluated using the above mentioned test and training set approach. Table 2 also shows the results for prediction of substrate specificity based on 15 active site pocket residues. It was encouraging to note that for the acyl adenylate superfamily, the substrate specificity can also be predicted with high accuracy using a limited number of residues which define the putative substrate binding pocket. The 15 SDRs were able to correctly predict the substrate preference for CG6178 to be long chain:CoA ligase, rather than a luciferase. Similarly for At4 g05160 and At5 g63380 the PSSM-15 based method correctly identified the substrate preference as long chain fatty acid (Table 3). Our SDR profiles indicates that, At1 g62940 is not a 4CL in agreement with experiments. However, by our PSSM-15 based method this sequence was predicted as long chain CoA ligase, rather than medium chain CoA ligase (Table 3). The origin of this discrepancy might be due to the ambiguity in literature regarding definition of acyl chain length which distinguishes medium chain CoA ligase from long chain CoA ligases. Thus, our analysis suggests that, for this superfamily of enzymes a limited number of residues lining the substrate binding pocket are involved in substrate selection. In some cases, the enzymes having whole sequence homology to one subfamily, have evolved to acquire specificity for completely different substrates by mutating a limited number of active site residues. Prediction of substrate specificity based on limited number of specificity determining residues (SDR) has the additional advantage that, based on the identified residues site directed mutagenesis experiments can be designed to alter substrate specificity by protein engineering.

Even though PSSM-15 method could successfully predict for cases in which enzymes showed so called anomalous specificity, the overall prediction accuracy for the PSSM-15 method was slightly lower compared to the prediction by whole sequence HMMs. The prediction accuracy fell from 0.98 to 0.93 for LCS, from 0.91 to 0.83 for MCS and from 0.98 to 0.92 for 4CLs. This could be due to the fact that these 15 SDRs were identified based on a smaller substrate phenylalanine. For larger substrates like luciferases, long chain and medium chain fatty acids the binding pocket could be larger and they can comprise of more number of SDRs. To confirm this hypothesis, SDRs were increased from 15 to 44 by increasing the distance cut-off for defining putative SDRs from 6 Å to 10 Å. The PSSM for each subfamily were rebuilt for 44 positions and the prediction accuracy recalculated. As can be seen, from Table 2 the prediction accuracy with larger binding site was comparable to that obtained by whole sequence based method (HMMER).

Analysis of conservation pattern of SDRs

Since these 15 residues could successfully predict the substrate preference, it was intriguing to see the amino acid conservation pattern at these positions across the different subfamilies. Figure 4 shows the conservation pattern of the 15 SDR positions (consensus active site profile). The conservation pattern was identified using an approach is similar to that used in the evolutionary trace method, which has been used in earlier studies to identify functional sites on proteins [50]. As can be seen in Figure 4, eight out of these 15 positions clearly change in a subfamily specific manner. The binding pocket of all subfamilies is majorly composed of hydrophobic amino acids. The consensus active site profile of LCS, MCS and luciferases are very similar. This is also reflected in the substrates of the enzymes. LCS and MCS only differ in the chain length of the fatty acid that they use as substrate. There are reports where luciferases have been known to catalyze formation of long chain fatty acyl:CoA [42, 51]. Many biochemical and mutagenesis experiments have been carried out to determine the residues governing substrate specificity in this superfamily. These experiments also independently show that, many of these 15 SDRs are indeed important for substrate selection. The position 324 (numbering according to 1AMU) has a conserved tryptophan in AcCS (substrate C₂), whereas in all other subfamilies which utilize larger substrate (>C₂), this position has a conserved glycine. Analysis of the substrate binding pocket in the available crystal structures, it clearly shows that bulkier tryptophan makes the substrate binding cavity smaller whereas a glycine will allow accommodating larger substrates (Additional file 1, Figure S2). Various structural and mutagenic studies have confirmed the role of tryptophan in controlling the substrate preference of the enzyme [25, 38, 40, 52]. Similarly, position 234 has a conserved tryptophan in AcCS, a conserved histidine in 4CL, LCS, MCS and luciferases and a conserved phenylalanine in NRPSs. We hypothesize that this histidine may play a role in attracting the carboxylic acid substrate into the deep seated binding pocket. In NRPSs this role is played by negatively charged aspartic acid (adjacent position i.e. 235) which interacts with the NH³⁺ group of amino acids [18]. The conserved tryptophan in AcCS might provide a suitable chemical environment for the substrate acetate which allows it to easily move in and out of the binding pocket. Hence it is hypothesized that this histidine may act as an attracting hook which pulls the larger substrates (like coumaric acid, fatty acid chains, luciferin) into the active site; however for small substrates like acetic acid only chance entrapment can catalyze the reaction. Mutagenic experiments in different subfamilies have shown the importance of this histidine [39, 53–56].

Structural similarity between members of ACS superfamily

It is known that, various members of ACS superfamily adopt a highly conserved three dimensional fold despite low sequence similarity. We analyzed the extent of structural similarity between available crystal structures of ACS superfamily of enzymes. Since, the orientation of the C-terminal domain relative to the N-terminal domain is known to change upon ligand binding [26], the structures were aligned only over the N-terminal domain and RMSD was calculated. Simultaneously, the percentage similarity between the sequences is also calculated by Needleman-Wunsch program. The Table S1 in Additional file 2summarizes the results. It clearly shows that luciferase (PDB code:1LCI) and acetyl:CoA ligase (PDB code:1PG4) which show only 17% similarity at sequence level, are found to have almost identical folds with RMSD of 1.6 Å. Also luciferases from Photinus pyralis (PDB code: 1LCI) and Luciola cruciata (PDB code: 2D1R), which show higher similarity (64%) show RMSD in the comparable range (1.1 Å). During the protein structure comparison studies it was noticed that, the large N-terminal domain of the ACS fold, can be further subdivided into two subdomains, namely subdomain A and B. The subdomain A stretches from amino acid 67 to 203 and B subdomain from amino acid 204 to 428 (numbering according to 1AMU). These two domains are relatively independent and there are very few contacts between the two (Figure 5). The B subdomain contains all the substrate binding residues. As seen from the multiple sequence alignment (Figure 5) most of the insertions and deletions are confined to the subdomain A and the subdomain B is relatively more conserved. This may be due to the functional constraint of the superfamily of catalyzing the same reaction but with different substrates. During superimposition the N-terminal domains of various structures of the superfamily, it was found that there are relative movements between the subdomain A and B in different structures. Thus our structural analysis helped in explaining how enzymes of this superfamily can adopt a highly conserved tertiary fold despite large divergence in primary sequence. It is also interesting to note that, the core of the conserved structural fold consists of the subdomain B of the N-terminus domain. Since substrate binding pocket is primarily confined to subdomain B, binding pockets of the various members of this superfamily can be modelled using the limited number of available crystal structures as templates.

Effect of the choice of structural templates on SDRs

At the time when the work was carried out, crystal structures were available for adenylation domain of NRPS (PDB code:1AMU), luciferases (PDB code:1LCI) and acetyl:CoA ligases (PDB code:1PG4). For long chain:CoA ligases, medium chain CoA ligases and luciferases, ligand free luciferase structure 1LCI was used as the template. However, the SDRs for all subfamilies were identified based on superposition of the modeled structure on 1AMU. With the availability of substrate bound crystal structures for all the subfamilies, it is important to understand whether SDR profiles for various subfamilies would change with the choice of structural template. Currently substrate bound structures are available for five out of the six sub families of acyl adenylate superfamily discussed in this work. They are 1PG3 (acetyl CoA ligase), 3EQ6 (medium chain CoA ligase), 1V26 (long chain CoA ligase), 2D1R (luciferase) and 1AMU (adenylation domain of NRPS). For the enzymes 1PG3, 3EQ6, 1V26 and 2D1R, active site residues were identified by three different methods and compared.

In method 1, the sequences were submitted as query to our web server http://www.nii.res.in/pred_acs_substr.html which models long chain:CoA ligases, medium chain CoA ligases and luciferases using the ligand free luciferase structure 1LCI as the template, but identifies binding pocket residues based on superposition of the modeled structure on 1AMU. In method 2, no structural modelling is carried out, instead the crystal structures 1PG3, 3EQ6 and 1V26 are used, but binding pocket residues are identified based on superposition of the crystal structures on 1AMU. In method 3, binding pocket residues are identified from the ligand bound crystal structures based on amino acids which are within a distance of 6 Å from the ligand. This will correspond to the true binding pocket as seen in experimentally determined structure. The SDRs identified by these three different methods for the four enzymes 3EQ6, 1PG3, 2D1R and 1V26 are listed in Additional file 2, Tables S2-S5.

It is interesting to note that, for 3EQ6, the medium chain CoA ligase 12 out of the 15 binding pocket residues were common in the three different methods for binding pocket identification (Additional file 2, Table S2). The binding pocket in the medium chain fatty acid bound crystal structure had only four additional residues which were not predicted by our web server. Similarly, in case of 1PG3, the acetyl CoA ligase, which has a smaller ligand compared to our reference structure phenylalanine bound 1AMU, 8 out of the 15 binding pocket residues were found to be common in the three different methods and the true binding pocket in the acetyl bound structure had only 2 extra residues (Additional file 2, Table S3). In case of 2D1R, the luciferase, all the 15 amino acids identified by the web server were common to the three different methods, but because of the larger size of luciferin 10 additional residues were identified in the actual luciferin bound crystal structure (Additional file 2, Table S4). Similarly, in case of 1V26, the long chain CoA ligase, 13 out of the 15 binding pocket residues identified by our web server were present in the true binding pocket, but additional residues lined the actual binding pocket in view of the larger chain length of the bound ligand (Additional file 2, Table S5). Thus the binding pocket residues identified based on 1AMU agree very well with the true binding pocket as seen in crystal structure and only in case of enzymes having larger substrates additional residues constitute the binding pocket. In view of the high degree of structural conservation in the superfamily, modeled structures (method 1) and actual crystal structures (method 2) give almost identical binding pocket residues.

Our analysis demonstrates that, in view of the high degree of structural conservation in the acyl adenylate superfamily, active site based substrate specificity can be done using information from single structural template. The identification of SDRs based on single structural template has certain advantages for prediction of substrates for acyl adenylate superfamily without knowing a priori the subfamily information. In absence of prior information about subfamily, it will be difficult to choose subfamily specific structural templates. This becomes a major problem, specially in those cases which show anomalous substrate specificities i.e. sequences showing high homology to 4CL but utilizing long chain acyl substrates. Secondly, use of different structural templates from different subfamilies would lead to different number of binding pocket residues and this will make comparison of scores across subfamilies for matches to active site profiles a difficult task. Therefore, in cases where subfamily information is not known, a better approach for prediction of substrate specificity would be would be to first identify the subfamily using single structural template and then align the sequences with the respective templates for more accurate identification of binding pocket residues.

Ligand docking

We next proceeded to carry out docking of substrates to their respective proteins and understand how exactly the identified SDRs stabilize the cognate substrates in their respective binding sites. In view of the large sequence divergence within each subfamily and availability of only a limited number of crystal structures, structural features of various members of this superfamily can only be obtained by homology modeling approach. We decided to carry out all the docking studies using homology models based on PheA (PDB code:1AMU) and compare the results to actual crystal structures wherever available (Additional file 1, Figure S3). A representative sequence from each subfamily was chosen. Acetyl:CoA ligase from Salmonella enterica (GenBank:31616027), long chain:CoA ligase from Thermus thermophilus (GenBank:55980573), luciferase from Photinus pyralis (GenBank:126501), Coumarate:CoA ligase from Arabidopsis thaliana (At4CL2) (GenBank:5702188) and medium chain:CoA ligase from Pseudomonas oleovorans (GenBank:416605) were chosen for model building. The structure of adenylation domain of gramicidin synthetase (PDB code:1AMU) was used as a template and its alignments with the query sequence were obtained from GenTHREADER [57]. Based on these alignments MODELLER[58] package was used to build the homology models. During docking simulations the docking grid was defined based on cavities found adjacent to the AMP binding pockets. The cavities were identified in the models, using the program ACSITE [59]. For each model, a docking grid was built to encompass the whole cavity (Figure 6). During docking AMP was kept in the AMP binding site as a fixed ligand.

For AcCS, MCS and LCS, C₂ (acetic acid), C₁₀ (decanoic acid) and C₁₄ (myristic acid) fatty acids were used as substrates respectively. At4CL2 has been shown to utilize coumarate and caffeate efficiently [60] and hence caffeate was chosen for docking on the model of At4CL2 (Additional file 1, Figure S3). The conformations of the ligand generated after docking runs were clustered at different RMSD cut off value. For AcCS, this was set at 1 Å and for larger substrates; it was set at 1.7 Å. The final orientation of the ligand which was selected as the docked conformation was based on three criteria; the energy rank, cluster rank and the distance of the carbonyl carbon and the donor phosphate of AMP. The energy rank is based on the docking energy and the cluster rank is based on the number of conformations in the cluster. For the catalytic activity, the carbonyl group of the substrate should be close to the donor oxygen of AMP. Hence, the conformation having lower energy rank and cluster rank and minimum distance between the carbonyl carbon and phosphate of AMP is selected as the docked complex.

Table 4summarizes the results of docking for each subfamily, while Figure S4 (A-E) in Additional file 1shows the final docked orientations for the cognate substrates for each of the different subfamilies. As can be seen from Table 4, except for 4CL, in all other cases the final docked conformation has lower energy as well as cluster rank. Since crystal structure [24] of myristic acid in complex with long chain CoA synthetase was available, we compared it with the docked conformation for long chain fatty acid (C₁₄) obtained by our docking simulation. Interestingly, the ligands had a RMSD of 2.2 Å (Figure 6) and 80% of the binding pocket residues identified by docking studies matched with those identified from actual crystal structure. Apart from long chain CoA synthetase, similar comparisons of docked complexes with the available crystal structures of enzyme-substrate complexes were also carried out for acetyl CoA synthetases, luciferase and medium chain CoA ligase. Figure 7 shows the superposition of the docked complexes on to the crystal structures of the corresponding protein-ligand complexes. For acetyl CoA ligase the comparison of the docked complex with the acetyl group bound crystal structure 1PG3 showed that, the ligand had an RMSD of 1.7 Å and 85% of the binding pocket residues in the crystal structure matched with the docked complex (Figure 7A). Figure 7B shows the superposition of the docked complex for medium chain CoA ligase on the medium chain fatty acid bound crystal structure 3EQ6. In our docking studies medium chain CoA ligase from Pseudomonas oleovorans (GenBank:416605) was docked with decanoic acid. However, the human medium chain CoA ligase crystal structure (3EQ6) had the substrate butyric acid bound in the active site. The ligands were compared over the first four carbon atoms and the RMSD was 3.3 Å and 78% of the binding pocket residues matched. For luciferase the docked complex was compared with the luciferase from Luciola cruciata (PDB code: 2D1R) which had the ligand luciferin bound in the active site. The ligand RMSD was 2.3 Å and 75% of the binding pocket residues matched (Figure 7C).

Hence for this enzyme superfamily, inspite of low sequence similarity, one is able to build approximate homology models, dock the ligand and identify the correct binding pocket. Therefore, the best docking solution from each subfamily was analyzed to identify additional functionally important residues apart from the 15 SDRs. The residue at position number 210 was found to be important for long and medium chain:CoA ligases. The long chain:CoA ligases have a conserved asparagines, whereas MCS have a conserved histidine (Figure 4). This position could potentially be the one controlling substrate chain length. In medium chain:CoA ligases, the presence of a positively charged histidine in the fatty acid binding tunnel will disfavour binding of longer fatty acyl substrates. However, in long chain:CoA ligases, the presence of partially hydrophobic asparagine might favour binding of fatty acids having longer chain length. Interestingly, in luciferases this position has a conserved arginine which has been suggested in earlier studies to form the base of the cavity for luciferin binding [39, 40, 61].

Thus the results from our docking studies as well as conservation profile analysis of binding pocket residues indicate that, in various subfamilies of acyl adenylate superfamily, the amino acids in the binding pocket have changed collectively to accommodate their respective substrates. The amino acids forming the binding pocket of acetyl:CoA ligases are larger in size than the ones forming the binding pocket of other subfamilies, e.g. position numbers 234, 278, 301, 322. (Figure 4). These observations from our in silico analysis are also in agreement with experimental studies. In Methanothermobacter thermautotrophicus ACS, the mutation of valine (corresponding to position number 301) to smaller amino acids (alanine) have shown to lengthen the binding pocket to accommodate longer substrates like propionate [52]. The glycine corresponding to position 301 in firefly luciferase (G315) has been found to be very important for proper luciferin binding [40].

Identifying new members of acyl adenylate superfamily from NR database

Since the benchmarking studies on the test set of ACS sequences clearly demonstrated the power of SDR based prediction of substrate specificity, we developed a user friendly web based tool for genome scale prediction of substrate preference for acyl CoA synthetase superfamily of enzymes. Figure 8 shows the results of a typical analysis using this web server available at http://www.nii.res.in/pred_acs_substr.html. This server permits classification of acyl AMP superfamily of enzymes into various subfamilies using both whole sequence HMM profile as well as PSSM based on 15 SDRs. We used this novel tool to identify new members of the AMP superfamily from NR database. NR database was downloaded. HMMER profiles built for each subfamily were used to pick up homologous sequences. Duplicated sequences and sequences which were part of the training set were removed. This way a total of 5480 sequences were fetched out. Sequences with length ranging from 300-750 amino acids were considered. Table 5 shows the number of proteins which could be assigned to various substrate specific subfamilies. It is important to note that, using our computational protocol it is possible to assign specificity to a large number of proteins which were originally annotated as just AMP ligases or hypothetical/unknown proteins. Interestingly apart from classifying various AMP ligases as per their chain length specificity, our method has identified several coumarate:CoA ligases and NRPS adenylation domains. These sequences deserve special attention for experimental studies as they do not show obvious sequence homology to coumarate:CoA ligases or NRPS adenylation domains, but might utilize coumaroyl moiety or amino acids as substrates. Thus our analysis demonstrates that, using our computational method it is not only possible to identify ACS superfamily of enzymes in various genomes, but also get valuable clues about the biosynthetic pathways they could possibly be involved based on in silico assignment of their substrate specificity.

Compilation of the sequences

Since, sequences with experimentally confirmed specificities were crucial for our analysis the dataset had to be curated carefully. All the sequences of adenylation domain of NRPSs were extracted from NRPSDB http://www.nii.res.in/nrps-pks.html[34] which contains manually curated substrate specificities for adenylation domains based on experimentally characterized NRPS gene clusters. Since adenylation domains often occur in multiple copies on a single polypeptide chain of a multifunctional enzyme, it was easy to obtain confirmed specificities of a large number of NRPS adenylation domains based on experimentally identified chemical structure of the nonribosomal peptide product. However, for other families like 4CL, luciferase, acetyl CoA ligase, medium chain CoA ligase and long chain CoA ligase such large numbers of experimentally characterized sequences were not available. Based on extensive literature survey we collected at least 20 biochemically characterized sequences for AcCS, LCS, 4CL and at least 10 characterized sequences for MCS and luciferases. Since, such small numbers of sequences are not adequate for HMM and PSSM analysis using a training and test set approach, we had to include additional sequences in each group using the experimentally characterized sequences as seed sequence. The additional members were selected based on the criteria of low e-value, alignment over the whole length, and same annotation as the seed sequence. Finally, a data set comprising of 608 protein sequences with known substrate specificity (120 AcCS, 130 LCS, 119 4CL, 25 luciferases and 29 MCS) were collected by keyword search and repeated BLAST searches[47] against the NR database of NCBI [48].

Sequence analysis

Pairwise sequence comparisons were carried out by Needleman and Wunsch alignment program of the EMBOSS package [62]. BLAST was used to perform local alignment [63]. BLOSUM62 scoring matrix and default values for gap penalties were used for sequence alignments. Multiple alignments and phylogenetic dendrograms were also constructed for each of the domain types using CLUSTALW program [64].

Building PSSM

For building the PSSM, for each subfamily columns corresponding to 15 active site positions were extracted from the alignment. The score of amino acid 'a' at column/position 'p' is defined as:

Freq(a, p) refers to the frequency of residue 'a' at position 'p' in the alignment. 0.05 represents the probability of random occurrence of any amino acid.

If any amino acid occurs with frequency of '0', then a large fixed negative score of -3.00 is assigned to the amino acid at that position. Hence for each subfamily a 15 × 21 position specific scoring matrix (PSSM) (corresponding to 15 positions and 20 amino acids + gap) was calculated, which quantifies the occurrence of amino acids at each position.

Calculation of prediction accuracy

For the data which are predicted as positive, the actual positive ones are called true positives (TP), while the others are called false positives (FP). For the data which are predicted as negative, the actual positive ones are called false negatives (FN), while the others are called true negatives (TN). The sensitivity and specificity are defined as:-

Structure analysis

For model building, the alignments with the template structure 1AMU were generated by program GenTHREADER [57]. The models of proteins were built by MODELLER 6V2[58]. The model structures were optimized by conjugate gradient energy minimization and molecular dynamics with simulated annealing provided in the MODELLER program. The ligands for docking were built using the builder module of InsightII. The ligands were minimized before docking using the discover module of InsightII.

Protein ligand docking

Dockings were carried out by molecule docking evaluation program, AUTODOCK 3.0 [65]. Dockings were done by Lamarckian genetic algorithm. Each docking experiment consisted of a series of 100 simulations, each producing a docking solution. The docking experiment was performed 250 times to give 250 possible solutions. The solutions having an RMSD over all atoms of < 1 Å (for small substrate) or <1.7 Å (for large substrate) were grouped together in a single cluster. Parameters for the docking used were: a population size of 50, a random starting position and conformation, a maximal mutation of 2 Å in translation and 50° in rotations, an elitism of 1, a mutation rate of 0.02, a crossover rate of 0.8 and a local search rate of 0.06. Simulations were performed with a maximum of 1.5 million energy evaluations and 27000 generations. The center of the ligand was chosen as the root. The 60 × 60 × 60 grid with grid points separated by 0.375 Å was build centered on the ligand.