INTRODUCTION

TOP ABSTRACT INTRODUCTION RESULTS DISCUSSION METHODS REFERENCES

Metabolic enzymes represent one of the most important and extensively studied class of proteins. Consequently, enzymes have been used extensively to address various issues of protein sequence/function relationships. Known metabolic enzyme families exhibit complex patterns of divergent and convergent evolution  many enzyme families usually catalyze a range of biochemical reactions (Jensen and Gu 1996), whereas some of these reactions may also be catalyzed by members of apparently unrelated protein families. A deeper understanding of such subtleties of the sequence-to-function relationship may shed light into the processes of molecular evolution of proteins (Petsko et al. 1993). These studies enable us to address fundamental questions, such as the origins and evolution of biochemical networks, and practical issues, such as function assignment by homology.

For instance, a case study of the distribution of the ()₈ barrel fold into different pathways appears to suggest a "patchy" mode of evolution for metabolic pathways (Copley and Bork 2000). Another study covering the Enzyme Commission (EC) hierarchy and the distribution of EC numbers across protein fold types addresses the problem of annotation transfer (Hegyi and Gerstein 1999; Wilson et al. 2000). These studies have used protein structure similarities, as three-dimensional structure is generally more conserved than the primary sequence. Although structural similarities permit the detection of very distant homologies, these analyses are confined to homologs of proteins of known structure.

In the absence of protein structure information, sequence comparisons provide a less sensitive but much more comprehensive way of detecting protein function at a genome-wide scale. Function prediction by sequence similarity provides useful hints for the potential cellular roles of proteins in entire genomes. Currently, an average of 60% of the encoded proteins for any genome can be functionally characterized by homology to proteins of known function (Iliopoulos et al. 2000). This assignment procedure, however, often overlooks issues of evolutionary divergence, whereby homologous sequences may have different functions. Such effects may lead to error propagation in sequence databases (Karp 1998). One of the most suitable sets of proteins to address some of these issues is the set of known metabolic enzymes, because of available classification schemes for their functional properties, in terms of the EC reaction hierarchy and pathway participation.

We have performed an extensive correlation of enzyme sequence and function, using the full known metabolic complement of Escherichia coli (Karp et al. 2000). Sequence relationships are represented by the membership of enzymes into protein families on the basis of sequence similarity. Function properties are represented by reaction capabilities and pathway involvement of the corresponding enzymes. We examined (1) functional versatility within enzyme families, that is, the association of enzymes with distinct reactions and pathways and (2) molecular diversity of protein function, that is, the distribution of reactions and pathways across enzyme families. We call these aspects of our analysis the sequence-to-function and the function-to-sequence problems, respectively. The detected patterns of functional versatility and molecular diversity across enzyme families, reactions, and pathways allow the first genome-wide overview of these complex, many-to-many relationships.

	RESULTS

TOP ABSTRACT INTRODUCTION RESULTS DISCUSSION METHODS REFERENCES

We have examined the sequence-to-function problem through the mapping of functional versatility of enzyme families to reaction types (EC numbers) and pathway involvement. The degree to which enzymes span different reactions (section 1) and pathways (section 2) corresponds directly to the structural plasticity and divergence within enzyme families. Furthermore, we have examined the function-to-sequence problem, namely the molecular diversity of protein function in terms of reactions (section 3) and pathways (section 4), using criteria for family membership. The extent to which biochemical reactions and pathways are associated with distinct protein families possibly indicates functional properties that have been invented more than once during evolution.

The 548 small-molecule metabolic enzymes were clustered into 405 enzyme families, with 316 single-member families and the remaining 232 enzymes classified into 89 families. Only 30 families containing 47 enzymes did not match any EC number, therefore mapping these families to reaction types was not possible. Overall, the 548 enzymes were found in 132 pathways and encoded for 422 unique EC numbers, representing 90 oxidoreductases, 133 transferases, 49 hydrolases, 82 lyases, 35 isomerases, and 33 ligases. These six enzyme classes correspond to the first level of the EC hierarchy, respectively.

Functional Versatility of Enzyme Families in Reaction Space

Interestingly, 75% of the total number of enzyme families appear to contain monofunctional enzymes, that is, enzymes known to catalyze a single enzymatic reaction (Fig. 1). This is a very important observation with direct applications to function prediction by sequence similarity, because it suggests that within these families, direct transfer of annotation by similarity can be reliable (desJardins et al. 1997). These monofunctional families include well-known, homologous isozymes such as the Fe/Mn superoxide dismutases (McCord 1976), gluconokinases, and L-asparaginases (Table 1). There is also a number of heteropolymeric enzyme subunits sharing sequence similarity, such as glutamate decarboxylase, formate dehydrogenase, nitrate reductase, and others (Table 1).

View larger version (35K): [in this window] [in a new window]   Figure 1   Functional versatility of enzyme families as assessed by reaction types. Frequency distribution of enzyme families (Y-axis) in relation to the number of unique EC numbers (X-axis) these families span.

An additional 16% of enzyme families contain two unique EC numbers, as judged by the number of reactions that the family members are known to catalyze. The remaining 9% of enzyme families contain three or more unique EC numbers. There is one enzyme family that contains eight enzymes with 11 unique EC numbers, representing different oxidoreductase reaction types (Fig. 2; Table 2). The common domain of these enzymes corresponds to the pyrroline-5-carboxylate (P5C) dehydrogenase family (Ling et al. 1994).

View larger version (112K): [in this window] [in a new window]   Figure 2   Multiple sequence alignment of the common domain of the enzyme family members listed in Table 2. SWISS-PROT accession numbers are given at left. Residue numbers corresponding to the entire protein sequence length are also shown. Identical residues (>50%) are boxed; similar residues (PAM250, >25%) are shaded.

This extreme concentration of unique EC numbers within enzyme families may reflect some strong property of function conservation in the set of known metabolic enzymes of E. coli (Labedan and Riley 1995). There are two potential artifacts, however, that may contribute toward this pattern. First, most of the monofunctional families (93%) correspond to single-member families. In addition, there may be other enzyme homologs in the genome that have not been characterized or included in the EcoCyc database. Second, the EC classification scheme does not always account for the reaction specificity of enzymes. Examples are reaction 1.3.99.1 (Table 1), representing two enzyme complexes catalyzing the reversible interconversion of succinate to fumarate, and reaction 4.1.1.15 (Table 1), representing two subunits of a single enzyme. In fact, when the resolution is reduced (for example by ignoring the fourth, third, or second level of the EC hierarchy), the one-to-one correspondence between families and EC numbers is even more pronounced (Fig. 3).

View larger version (48K): [in this window] [in a new window]   Figure 3   Functional versatility of enzyme families at different levels of the EC hierarchy. Frequency distribution of enzyme families (Z-axis) in relation with the number of unique EC numbers (X-axis) these families span and the four levels of the EC hierarchy (Y-axis). The effect observed in Figure 1 is enhanced as the resolution of the EC scheme becomes lower.

Functional Versatility of Enzyme Families in Pathway Space

Surprisingly, when the resolution is further reduced by characterizing enzyme function using pathway involvement instead of reaction type, the distribution of enzymes is widened toward multifunctional families (Fig. 4). Still, the majority of enzyme families corresponding to 59% of the total, appear to participate in a single metabolic pathway (Fig. 4). This indicates a sharing of structural types across pathways and may correspond to an evolutionary signature of pathway origins (Jensen 1976). We define enzyme families as "confined" if the number of pathways their members span is less than the number of family members and "promiscuous" if the number of pathways is more than the number of family members.

View larger version (36K): [in this window] [in a new window]   Figure 4   Functional versatility of enzyme families as assessed by pathway involvement. Frequency distribution of enzyme families (Y-axis) in relation to the number of unique pathways (X-axis) in which these families appear.

Examples of confined enzyme families include gene products AroG/AroH/AroF (EC 4.1.2.15), involved in the first committed step of the aromatic amino acid biosynthesis pathway, and gene products MurC (EC 6.3.2.8), MurD (EC 6.3.2.9), MurE (EC 6.3.2.13 ), and MurF (EC 6.3.2.15), involved in successive steps of the peptidoglycan biosynthesis pathway. Examples of promiscuous enzyme families include malate dehydrogenase (EC 1.1.1.37), which is known to be involved in six pathways (Ouzounis and Karp 2000), and the cluster of eight members (Table 2) mentioned previously, involved in a total of 10 pathways, including proline biosynthesis, proline utilization, fucose catabolism, methylglyoxal metabolism, rhamnose catabolism, 4-aminobutyrate degradation, and fermentation. A full table of these cases is available on the above-mentioned web site.

A potential artifact in this step may involve the somewhat arbitrary definitions of biochemical pathways (Karp 2000) as well as the threshold values in sequence clustering. We have, however, observed many cases of distinct enzyme families that are known to be involved only in a limited number of pathways, such as the aro gene group mentioned above, suggesting that this pattern is a genuine property of the metabolic map.

Molecular Diversity of Reaction Types in Sequence Space

When the reverse relationship of reaction types to enzyme families was examined, similar patterns are observed. The majority of reaction types, 86% of total, are known to be catalyzed by a single enzyme family (Fig. 5), signifying a low dispersion of catalytic activity across homology groups. Overall, this pattern suggests that sequence largely determines the known functional attributes in term of reaction capability of an enzyme type.

View larger version (37K): [in this window] [in a new window]   Figure 5   Molecular diversity of reaction types as assessed by enzyme families. Frequency distribution of unique EC numbers (Y-axis) in relation to the number of enzyme families (X-axis) they span.

The remaining 14% of the EC numbers span more than one enzyme family, which may represent sufficiently divergent sequence clusters or biochemically convergent enzyme types. A well-known case of distinct enzyme families are the class I (FumA and FumB, sharing 80% sequence identity) and class II (FumC) fumarases (Mohrig et al. 1995), all sharing the same EC number (EC 4.2.1.2). Another striking example is a group of four oxidoreductase reactions (EC numbers 1.1.1., 1.3.99.1, 1.10.2., 1.18.99.1) with members belonging to six different families. A potential explanation is the bias in the EC classification scheme, which provides a high-resolution breakdown of oxidoreductases (EC class 1, with 79 three-level classifications), compared, for instance, with lyases (EC class 4, with 12 three-level classifications) (Bairoch 2000). Another deficiency of the EC classification scheme produces the two extreme examples of a single EC reaction belonging to multiple families (Fig. 5). These reactions correspond to assignments of the same EC number across different (nonhomologous) subunits of enzyme complexes, not necessarily involved in the same catalytic action. The first example is nitrate reductase (EC 1.7.99.4), present in seven families and the second example is NADH dehydrogenase (EC 1.6.5.3), present in 13 different families (Fig. 5). Allowing for these exceptions, the pattern of one reaction-one enzyme family becomes even more pronounced.

Molecular Diversity of Biochemical Pathways in Sequence Space

Finally, and in analogy to the sequence-to-function problem, we examined the extent to which pathways employ members of different enzyme families. The detected enzyme types provide an estimate for the average amount of essential building blocks for each biochemical pathway, in terms of homology groups. Overall, we observe that 63% of the metabolic pathways (83 out of 132) in E. coli employ up to four unique enzyme family types (Fig. 6).

View larger version (65K): [in this window] [in a new window]   Figure 6   Molecular diversity of biochemical pathways as assessed by enzyme families. Frequency distribution of small-molecule metabolic pathways (Y-axis) in relation to the number of enzyme family types (X-axis) they use.

In contrast to reaction types, only 12% of pathways (16 out of 132) span a single enzyme family. Some of these pathways contain more than one step, although there is a number of single-step catalytic cascades that are defined as a single pathway (e.g., transaminase reactions) (Ouzounis and Karp 2000).

Due to the low number of counts, it is not possible to provide a normalization scheme for this observation. For instance, weighting by the number of steps within pathways does not yield a meaningful pattern (data not shown). With more pathways from different organisms and a deeper understanding of the topological properties of metabolic maps, a re-evaluation of these data may become possible in the future. Currently, we can only describe the number of steps per pathway in relation to the enzyme families these pathways span (Fig. 7).

View larger version (84K): [in this window] [in a new window]   Figure 7   The relationship of enzyme families (X-axis) to the number of individual steps per pathway (Y-axis). If there were a one-to-one correspondence of enzyme family types and biochemical reactions, the dominant pattern would be on the diagonal of this matrix. () The nine pathways spanning 12 or more enzyme families (mentioned in the Results section); () pathway counts (Z-axis, data not shown) and most contain more than one pathway count.

In that respect, it is interesting that 7% of pathways (nine out of 132) span 12 or more enzyme families. These pathways correspond to some of the most complex and extensively studied biochemical cascades (the number of families these pathways span is shown in parentheses): gluconeogenesis (12), purine biosynthesis (13), nucleotide metabolism (14), tricarboxylic acid (TCA) cycle (18), variants of anaerobic and aerobic respiration (21, 26, 27), and fermentation (28). It is worth noting that these pathways appear to span a significant number of enzyme families because of the presence of multiple heteropolymeric enzyme complexes involved in their various catalytic steps (Fig. 7).

	DISCUSSION

TOP ABSTRACT INTRODUCTION RESULTS DISCUSSION METHODS REFERENCES

We have analyzed the sequence-to-function and function-to-sequence problems, using enzyme families. It is instructive to compare these results with a previous study conducted similarly for single enzymes (Ouzounis and Karp 2000). It is striking that the numbers of monofunctional enzymes are comparable, 83% for single enzymes (Ouzounis and Karp 2000) and 75% for enzyme families (this study, as noted above). The reverse relationship also yields very similar percentages for the number of reactions catalyzed by a single enzyme, 91% for single enzymes (Ouzounis and Karp 2000) and 86% for enzyme families (this study, as noted above). These patterns imply that despite currently held views, the catalytic activities of various enzyme types are highly concentrated within enzyme families and there is a dominant one-to-one relationship between sequence and biochemical function.

To characterize biochemical function of metabolic enzymes, we have employed the EC hierarchy assignment and the biochemical pathway involvement. Other, more coarse-grained, functional classification schemes such as the EcoCyc functional classes (Riley 1993) and the Drosophila-derived Gene Ontology scheme (Ashburner et al. 2000) exist. These schemes may be very useful for future analyses of this kind that are not restricted to the metabolic pathway complement but encompass the full spectrum of cellular roles.

Another pattern emerging from the present study is a direct comparison of function divergence within enzyme families (25% of which are multifunctional) to function convergence across families (14% of reactions are catalyzed by more than one family). As pointed out previously, the process of function prediction by homology relies on these principles. Potential errors in this procedure may rise from the class of multifunctional enzyme families. Conversely, the failure to identify specific enzyme types that catalyze a specific reaction may be explained by our inability to detect functionally convergent enzymes.

Using pathways as function descriptors, the one-to-one correspondence between sequence and function is less prominent (see above). A direct comparison of single enzymes (Ouzounis and Karp 2000) and enzyme families with pathway involvement is not possible, because the individual enzymes are highly specific to the corresponding pathways. Our correlation of enzyme families with protein pathways (see above) also sheds some light on the possible mechanisms of pathway evolution.

Early theories for the evolution of biochemical catalysts have suggested that pathways have evolved backward: upstream reactions became possible by the diversification of enzymes that catalyzed the reactions downstream, as compounds became depleted from the environment (Horowitz 1945). This hypothesis allows the development of complex pathways with small modifications of existing ones. It also predicts the accumulation of homologous enzymes within individual pathways, as it would be more probable that similar enzymes are used to handle similar metabolites. Fifty-nine percent of enzyme families are confined within a single metabolic pathway (Fig. 4). Given the significant number of single-member enzyme families, however, the support for the Horowitz hypothesis is rather weak. It appears that only very few pathways exhibit this evolutionary mode.

A competing hypothesis states that biochemical pathways evolve by the recruitment of enzymes in a more opportunistic manner (Jensen 1976). This hypothesis implies that novel biochemical activities have evolved by the use of pre-existing structures (Gerlt and Babbitt 2000). Organisms gained selective advantage first by employing enzymes available through gene duplication to catalyze other reactions and later by further enzyme evolution to fine-tune the enzyme activity. This hypothesis predicts the existence of a high number of homologous enzymes across biochemical pathways. The remaining 41% of enzyme families span multiple pathways, indicating a considerable degree of enzyme recruitment during protein and pathway evolution. The reverse function-to-sequence relationship reveals a high number of pathways (88%) spanning more than one enzyme families, implying that biochemical pathways require a small number of different enzyme types to accomplish the chain of chemical transformations essential for life. This study reflects the knowledge currently available on the metabolic pathways of a single species. It will be interesting to perform similar analyses for other species, when such databases become available, to assess the generality of these observations.

	METHODS

TOP ABSTRACT INTRODUCTION RESULTS DISCUSSION METHODS REFERENCES

First, the entire set of small-molecule metabolism enzymes for the known metabolic complement of E. coli was extracted from the EcoCyc database (Tsoka and Ouzounis 2000). EcoCyc describes the genome and pathways of E. coli solely on the basis of experimental information (Karp et al. 2000). The enzyme set was obtained by formulating a complex query that extracts all proteins catalyzing a reaction whose reactants and products (1) are small molecules and (2) differ. The first condition excludes protein modification and other types of reactions with large molecules, whereas the second condition excludes intracellular transport reactions with no chemical transformation. EcoCyc is currently the only database that allows the extraction of this type of information with such high fidelity. This procedure identified 548 enzymes involved in small-molecule metabolism.

Second, functional descriptions for each enzyme in the form of type of reaction catalyzed (i.e., EC numbers) as well as pathway participation were obtained using EcoCyc (Karp et al. 2000). All sequences of the corresponding entries were extracted using the appropriate pointers in EcoCyc. In addition, protein names and accession numbers were also obtained from SWISS-PROT (Bairoch and Apweiler 2000). Structural descriptions for each enzyme were obtained by first matching each individual sequence to the corresponding PDB entry using BLAST (Altschul et al. 1997) (with an E-value threshold of 10^-6) and then extracting protein fold information using SCOP (Lo Conte et al. 2000). Only 39% of these sequences (214 out of 548) have a homolog of known structure. Because of this low coverage, we have based our analyses on sequence similarity alone.

Third, all enzyme sequences were automatically clustered on the basis of sequence similarity, using the GeneRAGE algorithm (Enright and Ouzounis 2000). The algorithm employs a fast sequence similarity search algorithm such as BLAST (Altschul et al. 1997) and represents similarity information between proteins as a binary matrix. Compositionally biased regions are masked with CAST (Promponas et al. 2000). This matrix is subsequently processed through successive rounds of the Smith-Waterman dynamic programming algorithm (Smith and Waterman 1981) to detect inconsistencies, which may represent false-positive or false-negative similarity assignments (Enright and Ouzounis 2000). All parameters for these programs were set to their default values. The resulting clusters comprise protein families with information reflecting the domain structure of proteins. In this analysis, only six two-domain proteins were identified, increasing the effective number of single-domain enzyme entries to 554.

Fourth, the effect of BLAST E-value thresholds for clustering was investigated to identify an optimal threshold value for clustering. Permissive thresholds result in fewer and larger clusters, compared with more stringent values. We have performed full clustering at E-value thresholds ranging from 10⁴ to 10¹⁰⁰ and found that the number of enzyme families did not vary significantly (data not shown), indicating that the clustering procedure is quite robust. All results reported here were obtained with the BLAST E-value thresholds of 10⁶.

Finally, the analysis involved the detection of functional versatility of the E. coli enzyme families in terms of reaction and pathway properties of the individual family members. This study represents a continuation of previous work, where various metrics for the characterization of an entire metabolic complement were proposed (Ouzounis and Karp 2000). Herein, the additional enzyme family information introduces an evolutionary perspective on the structure and function of biochemical pathways. All of our results are available at http://www.ebi.ac.uk/research/cgg/pathways/families/ and as supplementary data at www.genome.org.