The procedure used to classify polytopic membrane domains is based mainly on family classification schemes (Pfam-A and Pfam-B) and is shown in Figure 1a. We identified families of polytopic membrane domains in Pfam [15] by allocating TM-helices annotated in SWISS-PROT [16] to proteins in Pfam. After conservatively picking 183 Pfam-A and 152 Pfam-B families, we conducted an analysis of loops that connect TM-helices. It was shown that the loops tend to be short, with most of them (> 95%) having fewer than 80 amino acids. We therefore took 80 residues as the maximal intra-domain loop between TM-helices to define polytopic membrane domains. Though the 80-residue cutoff may not apply to a small portion (around 5%) of integral membrane proteins, it diminished the chance of including soluble domains within membrane domains, given that the average soluble domain has about 170 residues [17].

Using TMHMM, a membrane protein prediction program based on a hidden Markov model [6], TM-helices of membrane proteins in 26 genomes were predicted. Polytopic membrane domains were identified using the loop size between TM-helices as a guide. These domains were then classified into 231 Pfam-A and 318 Pfam-B families either by direct SWISS-PROT ID matching or by sequence similarity matching using FASTA [18]. Of the aligned domains, most of their TM-helices also aligned well, especially in Pfam-A families, which have alignments based on manually crafted hidden Markov models. Unclassified domains were clustered into 121 families by their sequence similarities. For each family, a profile was constructed, as shown in Figure 1b. This included: an averaged hydrophobicity plot of all members in the family based on the Goldman-Engelman-Steitz (GES) scale [8]; a consensus sequence of the family, represented by a sequence logo plot [19]; and consensus sequences of the TM-helices. By analyzing the hydrophobicity plots, we can locate TM-helices in the aligned sequences in protein families, and assign a number of TM-helices to each family. Some families, including 3 in Pfam-A and 20 in Pfam-B, were eliminated at this step, owing to the ambiguity of TM-helices observed in the plot. From this process, we identified 228 Pfam-A, 298 Pfam-B and 121 clustered families for our analyses, with approximately 95% domains classified in Pfam families.

After assigning a number of TM-helices to each family, we conducted a survey of the assigned numbers of TM-helices in 228 Pfam-A families of polytopic membrane domains (Figure 2). Pfam-A families are manually classified families that have well-aligned protein domains, and most of them have a well-defined number for TM-helices. We also picked families in solute transport systems that are annotated as transporters, symporters and channels, and analyzed the number of TM-helices for these families (Figure 2).

In general, most Pfam-A families tend to have a small number of TM-helices. For those with seven or fewer TM-helices, the number of families does not vary significantly with helix number, although there are more families with two or four TM-helices than with three, five, six, or seven. For families with more than seven TM-helices, the number of families decreases sharply as the number of TM-helices increases. Families with 12 TM-helices are the exception, however; they have a small peak in numbers against the overall downward slope of the plot. We also carried out the same kind of analysis on Pfam-A families that are annotated as transporters, symporters, and channels, and found that 12-TM-helix families are preferred by transporter-like families. In addition, most (11 out of 12) Pfam-A families with 12 TM-helices are transporter-like families. There seems to be a tendency for the transporter-like families to have an even number of TM-helices, because families with 2, 4, 6, 8, and 12 TM-helices have a relatively higher occurrence than those with a neighboring odd number of TM-helices.

We selected 168 families from Pfam-A that had more than 20 members. For each of these families, we then generated consensus sequences with conservation value (R_sequence) using the Alpro program [19]. Relatively conserved amino acids in the consensus sequences (R_sequence value > 3.0, representing the top 15% R_sequence value of all amino acids) and in TM-helical regions were analyzed for their composition as well as for pair motifs.

We compared the amino-acid composition of the TM-helices in general with the composition of only the conserved positions in TM-helices in the 168 families (Figure 3). We noticed that some amino acids are considerably more prevalent in the conserved positions, such as glycine (8% average composition in TM-helices versus 19% composition in conserved positions of TM-helices), proline (4% versus 9%) and tyrosine (3% versus 5%). In contrast, isoleucine (10% versus 4%), valine (8% versus 4%), methionine (4% versus 1%) and threonine (7% versus 4%) are less prevalent in conserved positions.

As might be expected, the changes in prevalence of certain amino acids reflect their conservation in the consensus sequence. Therefore, glycine, proline and tyrosine are relatively conserved residues in TM-helical regions, and isoleucine, valine, methionine and threonine have relatively high mutability. This result correlates very well with the mutation data matrix (MDM) for multi-spanning transmembrane regions in membrane proteins [10]. In the MDM of multi-spanning transmembrane α helices, isoleucine, methionine and valine are found to have relatively high mutability as hydrophobic residues, and serine and threonine also rank high in mutability as polar residues. In the matrix, proline appears to be highly conserved. Our results confirm these findings; in addition, we find that glycine and tyrosine are also highly conserved residues in polytopic TM-helices.

We also analyzed the consensus sequences of 168 Pfam-A families for significant amino-acid pair motifs and compared our findings with previous studies. Table 1 shows three pair lists: one includes the top 50 pairs of Senes et al. with their significance [12]; the second includes the top 50 pairs with their occurrences from randomly generated pairs; and the third includes the top 50 pairs with their occurrences using Senes et al.'s top 200 most significant pairs. Of the three lists, the GxxxG pair always ranks first, highlighting its significance in TM-helices [12,13,14]. In the last list, which contains top-ranked pairs in the first two lists, we observed some interesting pair-motif patterns that are associated with glycine. Amino-acid pairs such as ZxxxZ and ZxxxxxxZ (Z represents glycine, alanine, or serine - residues with a small side chain) are highly ranked in the last list. It is known that amino acids are positioned with an average of 3.6 residues per turn in TM-helices [20]. Two residues that are separated by three or six residues are thus oriented in the same direction. Therefore, it was suggested that these motifs are favored for TM-helix packing [12,14]. Our results are in good agreement with the pair motifs that are formed with small residues, but do not favor pairs with β-branched aliphatic residues (isoleucine and valine). This is probably because isoleucine and valine are highly mutable residues in TM-helices.

Of all the 168 Pfam-A families of polytopic membrane domains we studied, 45 are classified as transporters, channels, and symporters, representing 27% of the total families. We studied GxxxG and GxxxxxxG pairs, and found that they tend to be associated within transporter/channel-like membrane proteins (Table 2). When one or both glycines is mutated to a small residue such as serine or alanine, this association is weakened. Therefore, GxxxG and GxxxxxxG pairs are relatively conserved in transporter/channel-like membrane proteins. By comparing the amino-acid composition of conserved residues in the TM-helices of the transporter-like families with that of the rest of the Pfam-A families (Table 3), we found that glycine is two times more conserved in the transporter-like families, reflecting the favored GxxxG and GxxxxxxG pairs in these families. Proline and asparagine are also among the conserved residues favored in transporter-like families, whereas cysteine, histidine, isoleucine, leucine, methionine, and valine are unfavored.

Classified polytopic membrane protein domains represent from 40% to 81% of the total polytopic membrane domains in the genomes studied, with an average coverage of 61% (Figure 4a). We kept the family classification relatively conservative instead of aiming for a high overall coverage with a less careful classification. To avoid including falsely predicted families, we based our analysis on families with no fewer than four members. However, a higher proportion of polytopic membrane domains could be classified if smaller families were considered (Figure 4a).

We classified polytopic membrane domains into Pfam-A, Pfam-B and self-clustered families. Figure 4b shows the distribution of these three kinds of families in all the genomes. Most of the classified polytopic membrane domains belong to Pfam-A and Pfam-B, which cover 95% of classified domains.

Classified polytopic membrane domains and their families were studied in relation to the number of ORFs in each genome. Figure 5a shows the number of classified polytopic membrane domains versus the number of ORFs in all the genomes, and Figure 5b shows the same relation in genomes of single-celled organisms. A rough linear relation seems to exist between the number of classified polytopic membrane domains and the number of ORFs in each genome. However, it is interesting that C. elegans is an obvious outlier in the trend. To try to explain this, we took a closer look at the biggest families of polytopic membrane domains in C. elegans (Figure 5c). The three biggest families in C. elegans are PF01604, PF01461, and PB000009, which are described as 7-TM chemoreceptor families. (The annotation of PB000009 is from PD000148 in Prodom [21].) These families are almost unique to C. elegans, as most of their members in Pfam are from C. elegans. These families contain well-amplified membrane domains, with total numbers of 289, 250, and 216, respectively. Those numbers are more than double the biggest family in Drosophila melanogaster, which is PF00083 (Sugar (and other) transporter) with 108 members. By removing the number of proteins in these three families (a total of 754), we can see a better fit of C. elegans to the trend line. So the unusually large number of polytopic membrane domains is likely to be caused by protein amplification in a few families.

This hypothesis was supported by analysis of Figure 5d, which shows the number of families of polytopic membrane domains in relation to the number of ORFs in studied genomes. The number of families seems to have a logarithmic relation in all studied genomes, including C. elegans. Given that C. elegans has an unusually large number of polytopic membrane domains but a normal number of families, the amplification of polytopic membrane domains is limited to a few families.

In this study, the following databases were used: SWISS-PROT (release 39 and updated to 19 December, 2000) [16], which contains 91,132 protein entries; Pfam (release 6.1) [15], which contains 2,727 protein families in Pfam-A and 40,230 families in Pfam-B; Proteome Analysis Database [28], where complete non-redundant proteomes were downloaded. We selected eight genomes from archaea: Archaeoglobus fulgidus (AF), Aeropyrum pernix K1 (AP), Halobacterium sp. (HS), Methanococcus jannaschii (MJ), Methanobacterium thermoautotrophicum (MT), Pyrococcus abyssi (PA), Pyrococcus horikoshii (PH), and Thermoplasma acidophilum (TA); 14 genomes from bacteria: Aquifex aeolicus (AA), Borrelia burgdorferi (BB), Bacillus subtilis (BS), Chlamydia pneumoniae strain AR39 (CP), Chlamydia trachomatis (CT), E. coli strain K12 (EC), Haemophilus influenzae (HI), Helicobacter pylori strain 26695 (HP), Mycobacterium tuberculosis (MyTu), Mycoplasma genitalium (MG), Mycoplasma pneumoniae (MP), Rickettsia prowazekii (RP), Synechocystis sp. (SS), and Treponema pallidum (TP); four genomes from eukaryotes: Saccharomyces cerevisiae (SC), D. melanogaster (DM), C. elegans (CE), and Arabidopsis thaliana (AT).

Figure 1a shows our complete classification procedure. We extracted 8,301 protein entries in the SWISS-PROT database containing no less than two TRANSMEM annotations in the FT field. In these proteins, a total of 52,636 transmembrane (TM) regions were allocated to proteins in the Pfam database. By analyzing the location of TM regions in protein domains of each Pfam family, we were able to identify families that contain polytopic membrane protein domains. We went through a relatively conservative procedure to identify potential families of polytopic membrane domains. First, a Pfam family needed to have a significant number of proteins containing no fewer than two TM regions to be identified as a polytopic membrane domain family. Second, all families in Pfam-A and some in Pfam-B that have more than seven members are analyzed, as the Pfam-B database is under development and contains thousands of small protein families. Finally, we identified 183 Pfam-A and 152 Pfam-B families. Proteins in these families contain 36,878 TM regions, representing approximately 70% of the total TM regions extracted from SWISS-PROT. We analyzed sizes of the loops between all the TM regions, as shown in the inner chart of Figure 1. By Pfam's protein domain classification, most loops (> 95%) are short peptides, containing less than 80 amino acids.

Proteins from 26 genomes were submitted to TMHMM server for TM-helix prediction [6]. Predicted membrane proteins were searched for polytopic membrane domains, using a rule, generated from the above result, that the intramembrane-domain loop sizes must be less than 80 amino acids. To identify domains that are included in the Pfam families that have been identified, we searched the defined polytopic membrane domains for SWISS-PROT ID matches and regional matches. Unmatched domains are further classified on the basis of Pfam's classification, and additional 48 Pfam-A and 166 Pfam-B families are identified (small size Pfam-B families with no less than four members and no less than three matches are selected). In total, we identified 231 Pfam-A and 318 Pfam-B families as polytopic membrane domains. As not all proteins from the 26 genomes are included in Pfam, we then tried to assign the unclassified polytopic membrane domains to the identified Pfam families by sequence similarity matching to proteins in these families. We used the FASTA program [18] to search for matches, and matches with E-values less than 0.01 were considered positive. Obviously, one can assign Pfam-A domains using the HMMer software [29], which they are closely associated with. However, we chose to take a somewhat simpler tack, using FASTA. This is a somewhat more conservative approach (finding fewer homologs) which has the advantage of using consistent thresholds that can be applied to all the searches. Query domains were assigned to Pfam families that their best matches belong to.

As for those that have not been classified into Pfam families by either ID match or by sequence-similarity match, we tried to cluster these into families on the basis of their sequence similarities. This procedure was done by an all-against-all sequence similarity search (E-value < 0.01) using FASTA, and polytopic membrane domains were clustered by applying a multiple linkage clustering method [30] to the FASTA results. N family members must have more than 0.9N (N-1) links to other members, with tolerance of 10% missing links among members. We selected 121 clustered families that contain no fewer than four members, and aligned protein sequences in each family using the CLUSTAL W program [31]. For a complete list of assigned polytopic membrane domains see Additional data files and [32].

We assume that all protein domains in a classified family have a defined number of TM-helices. To identify the number of TM-helices, we made a hydrophobic plot for each family of polytopic membrane domain. We took the aligned sequences in Pfam's families and in clustered families, and calculated the averaged GES hydrophobic values [8] of all the residues at each aligned position (Deleted and inserted residues, represented by '-' and '.' respectively, are given 0 individual values.) The plot for each family was generated by the averaged GES values along their corresponding aligned positions. Most hydrophobic regions were clearly defined, as most TM-helices aligned well in each family. By identifying hydrophobic regions in the plots, we assigned numbers of TM-helices to classified families of polytopic membrane proteins. We also eliminated 3 Pfam-A and 20 Pfam-B families, as they did not contain multiple hydrophobic regions in their hydrophobicity plots. Therefore, we have 228 Pfam-A, 298 Pfam-B and 121 clustered families for further analysis.

We analyzed 168 Pfam-A families with more than 20 members and generated consensus sequences with their sequence logos of all aligned sequences in these families using the Alpro sequence logo program [19]. The selected family size threshold of 20 members is somewhat arbitrary. We chose it because: first, a significant portion (~75%) of the 228 classified Pfam-A families had more than 20 members; and second, the potential bias from small families could be reduced as they tend to have more conserved residues than big families. However, we can show that our results remain unaffected by changing this threshold. In particular, we analyzed Pfam-A families containing more than 25, 30, 35, or 40 members, and got essentially the same results. Amino acids with sequence conservation values (R_sequence) of no less than 3.0 (top 15% of all values) were considered as conserved residues. For all the families, we counted the occurrences of amino acids in the consensus sequences and in all aligned sequences in hydrophobic regions, which are defined to have no fewer than 10 continuous amino acids with GES hydrophobicity value greater than 0.

We used the pair definition from a previous study [12]. For example, a pair XYn (X and Y represent amino acids and n a number) corresponds to amino acids X and Y separated by (n-1) residues. We analyzed occurrences of pair motifs of all combinations of amino acids separated by 1 to 10 residues. This result was compared with a previous study of the 200 most significant over-represented pairs [12,33].

Using simple cross-referencing based on the above procedure, proteomic entries in each genome were searched for matches of polytopic membrane domains of classified families. Numbers of membrane domains in classified families were counted and analyzed in all genomes studied.