With large volumes of sequence and structural data now available, functional characterization of proteins has become the rate-limiting step in putting biological information to practical use. Large-scale functional annotation efforts have focused on automated strategies, as more traditional methods, such as experimental characterization of gene function and manually curated analysis of gene sequence and structure, can only be used efficiently on small subsets of the available data.

While this scale-up of the analysis process is required to handle the sheer volume of new information, automated analysis strategies possess inherent and serious limitations. For example, simple pairwise comparisons have been shown to be inadequate for functional classification of proteins with less than 30% to 40% identity 123. Utilizing information from multiple related sequences, especially via probabilistic methods such as sequence profiles or hidden Markov models 456, the number of true evolutionary relationships found between proteins with less than 30% identity can be tripled 13. Unfortunately, even when true homologous relationships are detected, direct transfer of functional annotation is not often possible at low levels of sequence identity 2789.

Even when direct transfer of the full functional annotation is not possible, evolutionarily related proteins usually share some functional relationship. To determine what this relationship is, we must start by examining the type of evolutionary linkage between the proteins. Here we have concentrated on enzymes because they have a well-defined biochemical function - the catalysis of a particular reaction.

Horowitz suggested that ligand binding is the dominant constraint guiding enzyme evolution 1011. According to his theory, biochemical pathways evolved backwards. When the substrate for the final enzyme in the pathway was depleted, a new enzyme evolved from this enzyme, via gene duplication and divergence, to produce the needed substrate from an available precursor. While the reaction mechanism of the new enzyme was allowed to drift away from that of the original enzyme, the ability to bind the common substrate/product was retained. Although this theory appears to apply to some groups of enzymes, for example HisA/HisF in the histidine biosynthesis pathway and TrpF/TrpC in the tryptophan biosynthesis pathway 12, it does not appear to be the dominant mechanism governing enzyme evolution 13. Furthermore, the model typically applies only to pairs of divergent enzymes.

Chemistry-driven evolution 141516, an alternative theory that appears to represent a substantial proportion of enzymes 13, identifies a chemical step or capability as the dominant constraint guiding enzyme evolution. According to this model, a newly evolved enzyme retains a fundamental chemical capability of its progenitor. The newly evolved enzyme may catalyze a reaction similar to its progenitor with only an altered substrate specificity, or it may catalyze a quite different overall reaction while still retaining some chemical capability common to its progenitor 12.

A group of related enzymes that share a common chemical capability mediated by conserved catalytic elements but catalyze different overall reactions has been termed a mechanistically diverse superfamily 12. A mechanistically diverse superfamily can be subdivided into families, where a family is defined as a group of related enzymes whose members catalyze the same overall reaction via conserved catalytic elements. Each of these mechanistically diverse superfamilies may contain hundreds or even thousands of proteins, representing many different overall functions and utilizing a wide range of substrates.

Mechanistically diverse superfamilies pose an especially difficult problem for automated functional classification methods due to the complexity of their underlying biology. For example, a newly sequenced superfamily member may not catalyze the same overall reaction as its closest relative in the superfamily, but may instead be related to other superfamily members by a more subtle conserved chemical capability. If the superfamily itself has not been characterized, the conserved chemical capability may not be immediately obvious. It is thus useful to subdivide a superfamily into families containing enzymes that catalyze the same overall reaction.

Sequence and structural similarity alone cannot be used to cluster sequences into families because different families evolve at different rates 17 (M.E. Glasner, R.A. Chiang, N. Fayazmanesh, M.P. Jacobsen, J.A.G, P.C.B., unpublished data; J.L.S., L.P. Wackett, P.C.B. unpublished data). Consequently, the boundaries between different families within a superfamily are uneven in sequence and structure space; in some cases, even very highly similar sequences may perform different reactions. In the mechanistically diverse amidohydrolase superfamily, for example, melamine deaminase and atrazine chlorohydrolase share 98% sequence identity, but catalyze different reactions 18.

Likewise, functional information alone cannot be used to cluster proteins into superfamilies and families, due to convergent evolution, in which nature has evolved more than one structural strategy to perform a given chemical reaction 192021. For example, George et al. 21 found that 69% of the functions described by three digit EC numbers are found in multiple Structural Classification of Proteins database (SCOP) 22 superfamilies, suggesting, at least for some of these, independent evolutionary origins. Further, some functions are found in multiple SCOP fold classes, providing further evidence that they have evolved via convergent evolution 2021. Thus, although enzymes in these groups catalyze the same overall reaction, they likely utilize different mechanisms.

Even within a single superfamily, the same function may have evolved more than once 23. For example, the ability to hydrolyze an organophosphate appears to have evolved on at least two separate occasions within the common lineage of the amidohydrolase superfamily (J.L.S., L.P. Wackett, P.C.B., unpublished data). The distinct evolutionary origins of the aryldialkylphosphatase family and the phosphotriesterase family are reflected in an extremely low overall sequence identity between the two families and by subtle differences in the constellation of active site residues used to catalyze the common reaction.

To address these issues and provide a useful test set for benchmarking and development of tools for functional inference, we have constructed a new gold standard set of mechanistically diverse enzyme superfamilies. Most importantly, these proteins are clustered according to rigorous and systematic definitions of family and superfamily. Because these definitions map specific elements of protein sequence and structure to specific elements of function, gold standard families and superfamilies are especially useful for developing tools for elucidation of function of uncharacterized members. Moreover, because they represent related proteins whose functions have diverged, sometimes substantially, they may serve as a challenging test set for automated superfamily clustering methods based on either sequence or structure. To further enhance the utility of the gold standard set as a test set for evaluation of automated superfamily clustering methodologies, evidence codes, based on those developed by the Gene Ontology consortium 24, are provided for all functional assignments.

As of August 2005, our five gold standard superfamilies include four distinct fold classes and contain a total of 91 families, 4,887 sequences and 282 structures (Table 1). For the purposes of this paper, we have defined two different types of families. Gold standard families contain only sequences with either experimentally determined functions or sequences that are highly similar to them, that is, show highly significant BLAST e-values (≤ 1 × 10^-175) to experimentally characterized sequences. In addition, each of the sequences in a gold standard family is required to conserve all family-specific catalytic residues identified from the literature. Silver standard families contain all the sequences from the corresponding gold standard family, but may also contain additional sequences that have not been experimentally characterized, show an e-value between 1 × 10^-20 and 1 × 10^-175 to a characterized family member, and meet other relevant criteria (see Materials and methods).

Table 2 gives a detailed view of the gold and silver standard families that make up each superfamily. As shown in this table, these families catalyze a wide variety of reactions, spanning five of the six EC classes. The superfamily sequence sets represent different diversity levels, as described further in the Discussion. All of the gold standard superfamilies have been rigorously studied, and their structure-function relationships extensively interpreted, providing detailed information, including reaction mechanisms, superfamily-specific catalytic residues, and family-specific catalytic residues (see J.L.S., L.P. Wackett, P.C.B., unpublished data, and 252627282930313233343536 and references therein, for reviews and general descriptions of these superfamilies.) We have compiled this information (as well as information on additional superfamilies) into a publicly available database that explicitly links enzyme sequence, structure and function in the manner described above 373839. (Structure-Function Linkage Database (SFLD) superfamilies correspond to gold standard superfamilies in this paper. SFLD families correspond to the silver standard families in this paper.)