INTRODUCTION

TOP ABSTRACT INTRODUCTION RESULTS DISCUSSION REFERENCES

The ultimate goal of the genome projects is to determine the structure and function of all the newly identified gene products. Fundamentally, this will be carried out via annotation transfer, transferring the structural and functional annotation from an experimentally characterized protein (as in a model organism such as Escherichia coli) to a predicted protein in a newly sequenced genome that shares similarity in sequence. The degree of annotation transferred will depend on the degree of sequence similarity. This process is shown schematically in Figure 1. In this paper, we aim to address this major question in bioinformatics, specifically focusing on multi-domain proteins, as they make up the bulk of the proteome in eukaryotic organisms (Gerstein 1998).

View larger version (29K): [in this window] [in a new window]   Figure 1   Schematic illustrating annotation transfer. This figure illustrates the process of annotation transfer for a group of hypothetical TIM barrel proteins. The leftmost panel represents sequence comparisons between idealized barrel domains from a number of organisms. The next panel shows analogous results for structural comparison, and the panel after that, functional comparison. The rightmost panel represents sequence comparisons between idealized multi-domain proteins that match over a single domain, the subject of much of this paper.

Our work is a direct outgrowth of two previous analyses of ours that concentrated on single-domain proteins. In an earlier paper, we found that the different structural classes of the scop classification system have different propensities to carry out certain types of function (Hegyi and Gerstein 1999). In particular, while the alpha/beta folds were disproportionately associated with enzymes and all-alpha and small folds with non-enzymes, the alpha + beta structures had an equal tendency for both enzymatic and non-enzymatic functions. Wilson et al. (2000) compared a large number of protein domains to one another in a pair-wise fashion with respect to similarities in sequence, structure, and function. Using a hybrid functional classification scheme merging the ENZYME and FlyBase systems (Gelbart et al. 1997; Bairoch 2000), they found that precise function is not conserved below 30-40% identity, although the broad functional class is usually preserved for sequence identities as low as 20-25%, given that the sequences have the same fold. Their survey also reinforced the previously established general exponential relationship between structural and sequence similarity (Chothia and Lesk 1986).

Other Work on Establishing Relationships between Sequence, Structure, and Function

Several other groups have studied the relationship between sequence, structure, and function in detail, attempting to determine the extent to which functional transference between matching proteins is feasible (Shah and Hunger 1997; Martin et al. 1998; Thornton et al. 1999, 2000; Zhang et al. 1999; Shapiro and Harris 2000; Todd et al. 2001). Orengo et al. (1999) analyzed protein families in the CATH database and concluded that > 96% of the folds in the PDB are associated with a single homologous family. By investigating enzymatic folds they also found that more than 95% of homologous families show either single or closely related functions. Pawlowski et al. (2000) studied the relationship between sequence and functional similarity in the twilight zone of 10%-15% sequence similarity and found a clear correlation between the two, with functional similarity based on the E.C. classification of enzymes.

Russell et al. (1997) analyzed binding sites in proteins with similar 3D structures and estimated that 90% of new remote homolog have common binding sites and similar functions. Eisenstein et al. (2000) evaluated the first results from the structural genomics projects and found that in many instances the protein structure itself offers an important clue to its biological function. Stawiski et al. (2000) found that function could be predicted rather successfully for just the proteases. Devos and Valencia (2000) presented a critical view of function transference between similar sequences, highlighting the limitations of this process due to errors in databases and the inherent complexity of the relationship between protein sequence-structure and function that does not allow "simplistic interpretations." They also found that binding sites are the least conserved features between related proteins while the catalytic activity of enzymes is the most conserved one.

Multi-Domain Proteins with Divergent Functions: How Common?

Most of these previous investigations focused on single-domain proteins or did not distinguish between single- and multi-domain ones. It is not clear how the multi-domain proteins with various functions behave with respect to functional conservation; namely, whether they are more or less conserved than their single-domain counterparts. In particular, as shown in Figure 1, if one multi-domain protein shares a single domain fold with another one, it is not clear the degree to which the functional conservation of these proteins is constrained by the shared part, and to what degree it is influenced by other domains that are not shared.

Specific groups of proteins that have the same combination of structural domains but dramatically different functions illustrate this situation. One example is the combination of the SH3-domain (scop superfamily identifier 2.24.2) and the P-loop containing NTP hydrolase (3.29.1). While in higher organisms this combination is associated with presynaptic and tumor suppressor functions (SWISS-PROT names SP02_HUMAN and DLGI_DROME, respectively), in the lower Dictyostelium it was found in myosin (MYSP_DICDI). Another example is the combination of the FAD/NAD(P)-binding superfamily and FAD-linked reductases C-terminal superfamily (3.4.1 and 4.12.1 superfamilies, respectively). In one group of proteins they appear in enzymes of the oxidoreductase group (e.g. OXDA_CAEEL or PHHY_PSEAE), while in another they are found in a dissociation inhibitor (e.g. GDIA_HUMAN). It should be noted that the proteins are not covered completely by the structural matches, so it is quite possible that the rest of them contain totally different domains that are responsible for the dramatically different functions. However, do these two examples show a rather rare or a more frequent phenomenon? How often do multi-domain proteins, sharing the same structural domain composition, differ in their functions?

In this paper, we attempt to provide a comprehensive answer to this question. This is particularly timely given that most of the unknown proteins in eukaryotic genomes are multi-domain. We use the same approach as in our previous analyses, comparing the sequences of the structural domains in scop to those of SWISS-PROT using BLASTP. We focus on the functional divergence of single and multi-domain proteins, extending previous investigations of single-domain proteins. Also, in comparison to previous work, we focus more on non-enzymatic functions and scop structural superfamilies, instead of folds.

	RESULTS

TOP ABSTRACT INTRODUCTION RESULTS DISCUSSION REFERENCES

Our Approach to Functional and Structural Assignment

We used the BLASTP program (version 2.0) (Altschul et al. 1997) to identify the scop 1.39 (Murzin et al. 1995) structural domains in SWISS-PROT (version 37) (Bairoch and Apweiler 2000) with e = 10⁴. We removed the hypothetical and fragment proteins. This resulted in two sets of proteins.

Single-Domain

Multi-Domain

Overall Distribution of the Matches

Figure 2 shows the most commonly observed multi-domain combinations in a set of recently sequenced genomes. The occurrences of further combinations are available from the Web site. Clearly, the distribution is very skewed, with certain combinations, such as 3.29-2.32, and 2.29-4.61 tending to predominate.

View larger version (94K): [in this window] [in a new window]   Figure 2   Distribution of multi-domain combinations amongst the genomes. The figure shows the occurrence of multi-domain fold combinations in a number of genomes, indicating its great variability. Each row indicates a particular combination of scop fold pairs (using scop 1.39), where a fold pair is defined as two distinct folds occurring in tandem in a protein. Each column represents a different genome, using the four-letter codes in the PartsList system (Qian et al. 2001): Aaeo, Aquifex aeolicus; Aful, Archaeoglobus fulgidus; Bbur, Borrelia burgdorferi; Bsub, Bacillus subtilis; Cele, Caenorhabditis elegans; Cpne, Chlamydia pneumoniae; Ctra, Chlamydia trachomatis; Ecol, Echerischia coli; Hinf, Haemophilus influenzae Rd; Hpyl, Helicobacter pylori; Mthe, Methanobacterium thermoautotrophicum; Mjan, Methanococcus jannaschii; Mtub, Mycobacterium tuberculosis; Mgen, Mycoplasma genitalium; Mpne, Mycoplasma pneumoniae; Phor, Pyrococcus horikoshii; Rpro, Rickettsia prowazekii; Scer, Saccharomyces cerevisiae; Syne, Synechocystis sp.; Tpal, Treponema pallidum. The numbers in each intersection cell indicate the number of times the fold pairs occur in a genome. Only the 20 most common fold pair combinations are shown here; the remainder are shown on the Web site (http://partslist.org/func). If a cell is greater than 6, it is shaded black; between 3 and 6, gray; and below 3, white. The blank spaces show instances in which one of the pairs does not occur in the organism at all (indicated by a value of -1 in the data table on the Web site). The fold assignments are done in a fashion consistent with those in PartsList and associated systems (Gerstein 1997; Lin et al. 2000; Drawid et al. 2001; Harrison et al. 2001; Qian et al. 2001).

Figure 3 shows the overall distribution of the single-domain and multi-domain matches in the different structural classes. The distribution of matches between enzymes and non-enzymes in multi-domain proteins largely agrees with that in the single-domain proteins. The multi-domain matches follow the overall tendency of the alpha/beta folds to be associated with enzymes to a larger extent and the all-alpha and small folds with non-enzymes. However, the values for the multi-domain matches are generally less extreme than for single-domains; for example, the 10-fold difference between single-domain alpha/beta enzymes and non-enzymes decreases to about twofold in multi-domain proteins. Another significant difference is the reduction in the number of multi-domain non-enzymes in the all-beta and alpha + beta structural classes compared to the single-domain matches. Altogether, there are more enzymes than non-enzymes among the multi-domain proteins (2805 enzymes vs. 1958 non-enzymes) whereas for single-domain proteins, the opposite is true (850 enzymes vs. 968 non-enzymes).

View larger version (38K): [in this window] [in a new window]   Figure 3   Distribution of proteins amongst broad structural and functional classes; the distribution of the matches among the seven structural and two functional classes in single- and multi-domain proteins. The single-domain and multi-domain matches each total 100%, independently of each other. The horizontal axis indicates the seven scop classes, which are (from 1 to 7): all-alpha, all-beta, alpha/beta, alpha + beta, multi-domain, membrane, and small protein.

Table 1 summarizes the distribution of superfamilies and superfamily combinations among the major functional classes, i.e. whether they have only enzymatic, only non-enzymatic or both enzymatic and non-enzymatic functionality. Altogether, 215 superfamilies were found in single-domain proteins and 310 in multi-domain ones. As 70 superfamilies were found in both, altogether 455 distinct structural superfamilies matched a SWISS-PROT protein with our required coverage criteria (described above). Similarly, we apportioned the 281 superfamily combinations observed in multi-domain proteins amongst different broad functional categories.

In single-domain proteins there are about as many superfamilies with exclusively enzymatic functionality as there are those with exclusively non-enzymatic functions (82 vs. 78). In contrast, in multi-domain proteins this ratio increases to almost threefold (135 vs. 56). This agrees with the notion that most enzymes are multi-domain. Another difference between single and multi-domain proteins appears in the ratio of superfamilies with a single function compared to multifunctional ones. As it is apparent from Table 1, about a quarter of the superfamilies matched single-domain proteins with different functions (55 of 215), whereas in the multi-domain proteins, this ratio increased to more than a third (119 of 310).

Single-Domain Proteins

Table 2 lists the two functionally most diverse structural superfamilies in single-domain proteins with some representative functions. The most diverse superfamily, the 3.38.1 Thioredoxin-like, has 11 different functions associated with it, most of them with an oxidoreductase mechanism. For instance, THIO_BPT4 is a small disulphide-containing thioredoxin that serves as a general disulphide oxidoreductase, while TDX2_BRUMA is almost twice as long (199 aa) and serves as a thiol-specific antioxidant that acts against sulfur-containing radicals. Another interesting example of functional diversity is provided by the Scorpion toxin-like superfamily (7.3.6). While BRAZ_PENBA is a small protein that is known to be 2000 times sweeter than sucrose, the other members of the superfamily are associated with different host-defense mechanisms. In insects the superfamily possesses antifungal activity (DMYC_DROME) or acts as a toxin (SCX5_BUTEU). Interestingly, in plants it can also act as an antifungal (AF2B_SINAL) or as an inhibitor of insect alpha-amylases (SIA1_SORBI). It appears that many single-domain proteins are toxins or allergens, or are related in other ways to a host-defense response.

Based on the data we can also determine the probability of two single-domain proteins that match domains in the same superfamily category also carrying out the same function. Using Bayes' theorem:

(1)

where S is the probability that two proteins share the same superfamily, F is the probability that two proteins have the same function, and ^~F is the probability that two proteins do not have the same function. Rearranging and simplifying the equation we get:

(2)

where N is the number of times that the two events in the parentheses occur together in our database of 1818 single-domain proteins. This results in
That is, the probability that two single-domain proteins that have the same superfamily structure have the same function (whether enzymatic or not) is about 2/3.

Multi-Domain Proteins

Table 3 lists the combinations of superfamilies that have been associated with the greatest number of different functions in multi-domain proteins, with representative entries in SWISS-PROT. The combination with the greatest number of different functions is that of 1.95.1 and 7.33.1. Although it has twice as many different functions as the most diverse superfamily in the single-domain proteins (22 vs. 11, respectively), careful examination reveals that all the proteins in this category are DNA-binding and most of them act as hormone receptors.

The second entry listed in the table is the combination of the 3.4.1 and 4.48.1 superfamilies associated with the FAD/NAD(P)-linked reductases. It is an all-enzymatic combination and always carries out an oxido-reductase function. All the proteins in this category are completely covered by matches with these two superfamilies. The 1.78.1-2.1.1 hemocyanin-immunoglobulin combination seems also to be fairly conserved; although the proteins in this category are called by eight different names, most of them turn out to be extracellular larval storage proteins, except for the copper-containing oxygen carrier hemocyanin itself (HCY_PALVU).

Following the same logic, we can also determine the probability that two proteins that have the same superfamily combination share the same function, viz:
This means that we have significantly greater certainty in determining the function of a multi-domain protein with a particular superfamily combination than that of a single-domain protein containing a particular superfamily. We also determined a similar probability for those proteins that have an almost complete coverage with exactly the same type and number of superfamilies, following each other in the same order. The probability that the functions are the same in this case was 91%, a considerably higher value than above. However, if two multi-domain proteins share only a single superfamily, the probability that they share the same function drops to only 35%! This greater functional certainty from sharing a combination of superfamilies rather than just one is also reflected in Table 1. While one-fourth of the single-domain proteins and one-third of singularly matching superfamilies in multi-domain proteins have multiple functions, only about one-fifth of the multi-domain combinations possess multiple functions (60 of 281). It is also clear from the data that domains in larger proteins often lose their original function and no longer have an autonomous function.

Seventy Common Superfamilies and Their Functions Compared in Single-Domain and Multi-Domain Proteins

As mentioned above, of the 455 superfamilies in our analysis, only 70 occur in both single- and multi-domain proteins. Even more surprising is the small number of structural superfamilies (14) that have the same function in both single- and multi-domain proteins. These are listed in Table 4; 12 of them have enzymatic function, supporting the notion that enzymes are more conserved during evolution than non-enzymes. The two non-enzymatic superfamilies are the 4.29.1 ribosomal superfamily and the 5.4.1 superfamily in penicillin-binding proteins.

Table 5 presents several examples of the converse situation, shared superfamilies that have different functions in single and multi-domain proteins. Comparing parts A and B of the table highlights the fact that although both superfamilies in a multi-domain protein are often present in single-domain form as well, the functions in the different settings are only vaguely related. One example is the combination of the lipocalin superfamily (2.45.1) with that of the BPTI-like or Kunitz inhibitor (7.7.1), which in higher organisms forms a complex protein called alpha-1-microglobulin (AMBP_RAT). Another interesting example is the combination of the 2.5.1 Cupredoxin (occurring in the single-domain blue-copper protein, SOXE_SULAC) and the 6.5.1 Membrane all-alpha (single-domain representative: BACT_HALVA, a sensory rhodopsin) superfamilies into a component of the respiratory chain, cytochrome C oxidase II (COOX_ZOOAN). All these examples demonstrate the evolutionary advantage of a domain fusion event, which creates a function that is more complex than either of the components.

Multifunctionality vs. Sequence Similarity

Previously, we presented a variety of graphs that show how the probability that two domains would share the same function varied with respect to sequence similarity (Hegyi and Gerstein 1999; Wilson et al. 2000). Figure 4 shows a similar graph with the calculations extended to multi-domain proteins. The figure shows that the functional divergence of a single domain in multi-domain proteins dramatically increases, more than twofold, compared to the single-domain ones. This reinforces our findings above, based only on superfamily content, that the certainty with which we can predict the function of a protein based on its sequence similarity with a domain in another multi-domain protein, is considerably less than for a comparable single-domain situation.

	DISCUSSION

TOP ABSTRACT INTRODUCTION RESULTS DISCUSSION REFERENCES

Here we built on our previous studies on the relationship between protein structure and function to develop new results related to multi-domain proteins. Throughout the paper, we focused on superfamilies instead of folds, as the members of a superfamily are presumably of common evolutionary origin (Murzin et al. 1995).

We found that the 4763 multi-domain and 1818 single-domain proteins that met our selection criteria have about the same distribution of structural classes, with more enzymatic functions associated with the alpha/beta structural classes and more non-enzymatic ones with the all-alpha and small classes. We identified more than three times as many multi-domain proteins that were enzymes than single-domain ones (2805 and 850, respectively) and, conversely, about twice as many multi-domain proteins as single-domain ones that were non-enzymes (1958 vs. 968).

We focused on the functional divergence of the two groups and found that about a quarter of the superfamilies in single-domain proteins are associated with multiple functions, whereas only about a fifth of the multi-domain superfamily combinations are. Therefore, we can conclude that a combination of specific superfamilies results in a more specific functional assignment for a particular protein. However, about one-third of the superfamilies in the multi-domain proteins were associated with multiple functions, underlining the lesser autonomy of a domain function in multi-domain protein.

This latter finding was also supported by the difference in functional divergences between the two groups of proteins based on particular sequence similarities between the domains and SWISS-PROT proteins. As is shown in Figure 4, the average functional divergence of a single domain is much larger (more than twofold) in multi-domain proteins than in single-domain ones.

View larger version (27K): [in this window] [in a new window]   Figure 4   Divergence in function with respect to sequence similarity. Relative number of matching domains with multiple functions, as the function of e-value threshold. Diamonds represent single-domain proteins, squares multi-domain ones (matching just for a single domain), respectively. The first value on the X-axis starts at 4 (corresponding to an e-value=104).

We also found that only 70 of a total of 455 superfamilies are shared between the multi-domain and single-domain proteins and only a small fraction (14) share their functions. This was rather surprising to us, and should be taken into consideration in functional characterization and annotation of new gene products. When the functions were related in single- and multi-domain proteins, we could observe an increasing functional complexity with the appearance of large multi-domain proteins.

Altogether, with the recent sequencing of the human genome and the genomes of other model organisms, we hope that this work can contribute to the successful annotation of the individual gene products, and will help to avoid some pitfalls associated with the functional characterization of large, complex proteins.

The publication costs of this article were defrayed in part by payment of page charges. This article must therefore be hereby marked "advertisement" in accordance with 18 USC section 1734 solely to indicate this fact.