Feature amplified voting algorithm for functional analysis of protein superfamily

Case study: Imidase-related proteins in amidohydrolase superfamily

In this study, rat imidase was the target sequence, and DRPs (Group II proteins) were classified into ~A proteins. Bacterial hydantoinases (Group III enzymes) were classified as A proteins. Although Dihydroorotase, allantoinase and other amidohydrolases (Group IV) were also classified as A proteins, they differ from the Group II enzymes in their functional correlation to target sequence (rat imidase). Following the above classification, the clustered sequences were subjected to FAVAT analysis, and two sets of scores were obtained for each residue of the target sequence. In the experiments, Groups II, III and IV were aligned using the MUSCLE tool adopted by the National Center for Biotechnology Information (NCBI) for protein database alignment.

Voting scores of imidase by FAVAT analysis

Figure 1 shows the FAVAT analysis results. The T-score suggests the importance of each residue of the target sequence (rat imidase) after accumulating the total voting scores (V-scores), each V-score is calculated from the target sequence, an A protein sequence and a ~A protein sequence. For each rat imidase residue, two sets of T-scores were obtained from the Group II-Group III votes and the Group II-Group IV votes. Ten amino acids (Ala34, His67, His69, Ala134, Lys159, His248, Met297, Arg302, Asp326 and His459) were further analyzed after merging the two sets of the higher T-scores over 60. Table 1 summarizes the amino acid residues selected by FAVAT and MSA analyses and their corresponding locations in two proteins with known crystal structures [PDB:1GKQ, PDB:1KCX].

Comparison of FAVAT and MSA results

Biologists often use MSA to conjecture important residues of proteins or enzymes of interest among their related sequences. Figure 2 shows the MSA fragments of rat imidase, hamster dihydroorotase domain, yeast allantoinase and Bacillus sp. D-hydantoinase. The MSA analyses revealed that one aspartate and four histidines were highly conserved. The results in Fig. 2 are consistent with data published previously [30–32].

In their studies [30–32], residues Asp57, His59, His61, His183 and His239 were hypothesized to be critical amino acids for the metal coordinators and function of D-hydantoinase of Thermus sp. Notably, the study [33] reported that a crystal structure of hydantoinase (1GKQ), in which a carboxyl-lysine is responsible for metal binding and is important for enzyme activity, was not revealed by MSA analysis. Another residue, Asp57, which was incorrectly identified as a metal coordinator in a previous study that applied MSA method as the corresponding residue in 1GKQ, revealed no involvement in metal coordination. Table 1 shows that, in the current study, FAVAT successfully identified all known important residues in rat imidase. The Lys159, His67, His69, His248 and Asp326, which correspond to Lys150, His59, His61, His239 and Asp315 in hydantoinase (1GKQ), are metal ion coordinators (boxed amino acids in Fig. 3).

The corresponding locations of FAVAT-selected residues in 1GKQ and 1KCX

The possible functions of imidase amino acids selected by FAVAT were further analyzed using 1GKQ and 1KCX, which are known structures of imidase related proteins. The former is the crystal structure of a D-hydantoinase that represents an A protein (Group III) in FAVAT analysis. The latter is the crystal structure of a dihydropyrimidinase-related protein (CRMP1) that represents a ~A protein (Group II) in FAVAT analysis. Figure 3 shows their corresponding sequences and secondary structures. The similar β/α core structures were observed in the wiring diagrams of 1GKQ and 1KCX. The significant difference in these structures is that 1GKQ forms a typical (β/α)₈ domain, but 1KCX does not. The FAVAT-selected amino acids may reflect both the structure feature and metal requirement that are responsible for the different functions of the A and ~A proteins. The corresponding locations of Ala34 and His459 in 1GKQ (Arg30 and Trp448 in the N-terminal and C-terminal β-Sheet, respectively) and in 1KCX (Gln44 and Met465 in the N-terminal and C-terminal β-Sheet, respectively) were domains in which they interact with another monomer to form a quaternary structure in both hydantoinase and CRMP1 [34, 35]. Residues His67, His69, Ala134, Lys159, His248 and Asp326 (His59, His61, Ala126, Kcx150, His239 and Asp315 in 1GKQ; His73, Tyr75, Asp139, Gln165, Lys254 and Gly332 in 1KCX) are located in the β/α core region.

Figure 4a is a topological view of the corresponding FAVAT-selected amino acids in 1GKQ. Thermus sp. D-hydantoinase (1GKQ) contains two divalent metal ions in active site (Fig. 4b). The central binuclear zinc center is bridged by the carboxylated lysine residue (Kcx150) and a hydroxide ion. Residues His59, His61, His183, His239 and Asp315 correspond to the active site zinc ion (Fig. 4b). The corresponding residues were conserved in other members of the amidohydrolase superfamily [36, 37]. However, in CRMP1, four of the five corresponding residues are diverse [38]. These residues were all identified by FAVAT analysis. Other candidates recognized by FAVAT analysis, Ala134, Arg302, Ala34, Met297 and His459, were also found to reside in positions critical to protein function. The corresponding residue of Ala134 in 1GKQ (Ala126), was located near the active site (Fig. 4b). Residues Arg302 and Met297 (Arg292 and Met287 in 1GKQ, respectively) were located at the helix-loop domain outside the (β/α)₈ catalytic domain. This implies that they may be important for maintaining structure or stabilizing the protein conformation (Fig. 4a). These preliminary findings merit further detailed study.

Although the metal coordinators of imide-hydrolyzing enzymes in this case study were dispersed sequentially, almost all the known metal coordinators in 1GKQ were identified by FAVAT except His 183 (His 192 for rat imidase). This residue is conserved in CRMP1 but lacks metal and amidohydrolytic activity. The role of this histidine needs further study. The major difference between bacterial hydantoinase and mammalian imidase is their metal content. The former contains two metal ions while the later contains only one metal ion [39, 40]. Fewer metal coordinators may be needed for mammalian imidase, and residue His 192 may not be required as a coordinator of metal ions in rat imidase. A mammalian imidase was crystallized recently [41]. The difference between mammalian imidase and non-mammalian imidase is expected to be clarified in the near future.

Imidase and sequence clustering in the amidohydrolase superfamily

Hydantoinase activity was first reported in plants and animals [42, 43] to hydrolyze hydantoin derivatives that are not known as physiological metabolites. This enzymatic activity is useful for preparing optically pure amino acids that are precursors for various antibiotics [44]. Due to its industrial application, several hydantoinases have been studied and purified from microorganisms [45, 46]. A dihydropyrimidinase (5, 6-dihydropyrimidine amidohydrolase) partially purified from animal livers was shown to hydrolyze the physiological substrate dihydropyrimidine [47]. A detailed study of a homogenous imide-hydrolyzing enzyme, imidase, which was purified from rat, pig or fish livers [48–51], revealed that it catalyzes a wide spectrum of substrates, including dihydropyrimidines, hydantoins and other imides [52]. Despite the substrate spectra of hydantoinase highly similar to imidase, these imide-hydrolyzing enzymes from bacterial and mammalian sources reportedly have relatively low sequence similarity. Some mammals, flies and C. elegans, reveal proteins with high sequence similarity to dihydropyrimidinase (or imidase). These dihydropyrimidinase-related proteins (DRPs) may be involved in cancer and neuron cells development, but possess no imidase activity [53–55]. Additionally, other enzymes revealed by the studies in evolution of the metabolic pathway are also known to use mechanisms similar to those observed in imidase [56, 57]. These enzymes include dihydroorotase, allantoinase, urease and amidohydrolases, which originate in mammals, plants and fungi [58]. All use distinct substrates that contain similar imide functional groups.

All of the above enzymes can be classified into the amidohydrolase superfamily according to their properties and structures [59]. In this superfamily, some proteins have similar sequences but divergent functions whereas others have similar functions but low sequence similarity. This phenomenon strongly suggests that only a few critical amino acid residues in this superfamily are needed for specific protein functions. Proteins in the amidohydrolase superfamily can be grouped according to their sequence similarity and biochemical properties, and an effective strategy for analyzing these proteins may yield valuable information.

Observation and assumption

Table 2 shows the significant findings of the comparisons of sequence identity and functional properties of the imidase-related proteins in the amidohydrolase superfamily. Sequence-related proteins (Group II) had no imidase activity even though the overall similarity of sequences in this group was higher than 50% [61]. This phenomenon implicated that key amino acid residues for imidase activity have been altered or removed from Group II proteins. In contrast, functionally identical and functionally related enzymes (Group III and Group IV, respectively) had lower sequence identity, but they basically catalyzed the same reaction. For these enzymes, few conserved residues should be needed to provide a similar imidase function. The above observation implies a principle for classifying the sequence or the functional divergent enzymes in a superfamily, which may help to develop a feature amplified voting algorithm for identifying key residues in a target protein. Biologists generally select a sequence of interest as a target and perform BLAST analysis to recover related protein sequences. By definition, a specific function or property (e.g., substrate specificity) of a sequence of interest can be used to cluster the related sequences into different groups. For example, if property A is used to classify these sequences, sequences with property A should be classified into group A. Otherwise, the sequence should be classified into group ~A. Based on the classification of these three groups, a critical assumption can be made. When the target sequence, the A sequence and the ~A sequence belong to a protein superfamily, the conserved residues of the target sequence and the A sequence should correlate to property A. However, the non-conserved residues of the target sequence and ~A sequence should also reflect the negative correlation in property A. The intersection of the residues provides important functional clues. Thus, important residues can be found by comparing these three property groups. Therefore, the three-profile alignment method was developed to optimize the alignment of three groups so that key residues are then revealed by voting algorithm. Figure 5 shows the FAVAT flowchart that was developed to retrieve useful information for these distinct groups. Analysis of the imidase-related sequences in Table 2 indicated that, within this superfamily, the degree of sequence similarity did not necessarily reflect the similarity in biochemical properties. This provides a good opportunity to develop a novel method for extracting functional residues of a target enzyme. In this work, FAVAT was used to examine each residue in mammalian imidase by comparing other sequences in the amidohydrolase superfamily. Given the limited structural information about mammalian imidase, the analytical results of this study should provide important clues for enzymologists to perform further in-depth biochemical analyses of these results.

Algorithm

The FAVAT was performed in two steps. The first step was to align the target enzyme, functionally identical enzymes (A proteins) and sequence-related proteins (~A proteins) using three-profile alignment. The three-profile alignment algorithm, which is based on the dynamic programming three-way alignment approach [62, 63], was designed to align three profiles in a space. As in the FAVAT pre-process, each profile can be generated by multiple sequence alignment tools such as T-COFFEE, HMMER [64] and MUSCLE.

Let P₁, P₂ and P₃ be three profiles, and P_1i, P_2j and P_3k refer to the i th, j th and k th positions in P₁, P₂ and P₃, respectively, starting from 1. The symbol “-” denotes a “gap” in the alignment. Scores for the alignment of two columns are denoted by Sp(α, β). The scoring pair profiles P₁-P₂ are defined as follows:

where Sp₁₂ is the score at the i th andj th columns on P₁ and P₂, respectively. The P₁ has m sequences, and P₂ has n sequences. The W _a and W _b are the sequence weights for sequence a in P₁ and sequence b in P₂, respectively. The residue at i th column for sequence a in P₁ is denoted by r_1(a,i). The M is the value of the substitute matrix for r_1(a,i) and r_2(b,j). Many substitution matrices, such as BLOSUM, have been proposed to improve alignment accuracy [65]. Similarly, the definitions of scoring pair profiles P₁-P₃ and P₂-P₃ are similar to those of pair profile P₁-P₂. Gap penalties are determined by gap opening (GOP) and gap extension (GEP) scores. The best score of the alignments with prefixes P_1i, P_2j and P_3k is denoted by S(i, j, k) if the residues (P_1i, P_2j, P_3k) are aligned; G(i, j, k) is the best score given that (P_1i, P_2j, -) is the last column of the partial alignment, and H(i, j, k) is the best score given that the last column is of the form (P_1i, -, -). E(i, j, k), F(i, j, k), I(i, j, k) and J(i, j, k) are defined analogously. These quantities clearly satisfy the recursions summarized in Fig. 6.

The next step after the alignment is to determine whether amino acid residues critical for imidase activity exist in target and A proteins but are absent in ~A proteins. In the second step, a voting score (V-score) is given based on the previous assumption, and the V-scores are then summed in each comparison. In this step, a substitution matrix (BLOSUM62) is used to give V-score when each sequence of property A and ~A is compared to the target sequence. The V-score is calculated as follows:

V_k(a,b) = M[t _k , A_(a,k)] – M[t _k , ~A_(b,k)],

where V_k(a,b) is the V-score at the k th residues on target sequence, sequence a in A proteins and sequence b in ~A proteins. The A and ~A proteins have m and n sequences, respectively. The t _k is the k th residue of the target sequence. The A(a,k) and ~A(b,k) are the k th residues on sequence a in A proteins and on sequence b in ~A proteins, respectively. The M[t _k , A_(a,k)] is the value of the substitution matrix for t _k and A_(a,k). Figure 7 shows the V-score calculation. The BLOSUM62 substitution matrix is generally used for protein or nucleic acid sequence alignment. In the proposed algorithm, each V-score is given and accumulated to a total score (T-Score) until all sequences of property A and ~A are compared. The T-Score is calculated as follows:

where T _k is the T-Score at the k th residue on target sequence. The T-Score at each residue position is obtained by adding m × n V-scores. The normalization function is used to transform all T-scores into a range from 0 to 100 in FAVAT. The normalization function is as follows:

where Max(T) and Min(T) are the maximum and minimum scores of all T-scores, respectively. The details of the FAVAT algorithm are shown below, and Fig. 8 presents an example of V-Score and T-Score calculations using FAVAT.

Algorithm FAVAT (t, P, Q);

Input: Target sequence t, a set of proteins P without property A, a set of proteins Q with property A. P has p sequences and Q has q sequences.

Ouput: The scores correspond to the residues of t (high T-scores indicate potentially critical residues)

End

The novel feature of the FAVAT algorithm is its use of the sequence and functional properties among target sequence, ~A proteins and A proteins. When voting for reliable critical residue candidates, three relations are considered: the relation between target sequence and A proteins, the relation between target sequence and ~A proteins and the relation between A proteins and ~A proteins. To accurately identify the key residues, some useful alignment tools with physicochemical properties, such as T-COFFEE and HMMER, can be employed in the FAVAT pre-process to align A and ~A proteins separately (profiles). The appropriate alignments of A and ~A proteins can enhance the accuracy of the resulting alignment to the target sequence, A and ~A proteins by three-profile alignment. The most important residues can then be found accurately from the resulting alignment using FAVAT. The FAVAT algorithm was designed to account for the importance of alignment-based voting skill by V-score function. The time complexity of FAVAT is O(m _max ³), and m _max is the length of the resulting alignment by three-profile alignment. To reduce the time complexity of three-profile alignment method, this study developed a parallel version implemented by the MPICH library. The time complexity for the parallel version is O(m _max ³/p), where p is the number of processors. After the voting process, the residue candidates obtain high T-scores. The uncritical candidates can be eliminated by advanced research.