Proteome-wide functional classification and identification of prokaryotic transmembrane proteins by transmembrane topology similarity comparison

Abstract

We propose a new method for classifying and identifying transmembrane (TM) protein functions in proteome-scale by applying a single-linkage clustering method based on TM topology similarity, which is calculated simply from comparing the lengths of loop regions. In this study, we focused on 87 prokaryotic TM proteomes consisting of 31 proteobacteria, 22 gram-positive bacteria, 19 other bacteria, and 15 archaea. Prior to performing the clustering, we first categorized individual TM protein sequences as “known,” “putative” (similar to “known” sequences), or “unknown” by using the homology search and the sequence similarity comparison against SWISS-PROT to assess the current status of the functional annotation of the TM proteomes based on sequence similarity only. More than three-quarters, that is, 75.7% of the TM protein sequences are functionally “unknown,” with only 3.8% and 20.5% of them being classified as “known” and “putative,” respectively. Using our clustering approach based on TM topology similarity, we succeeded in increasing the rate of TM protein sequences functionally classified and identified from 24.3% to 60.9%. Obtained clusters correspond well to functional superfamilies or families, and the functional classification and identification are successfully achieved by this approach. For example, in an obtained cluster of TM proteins with six TM segments, 109 sequences out of 119 sequences annotated as “ATP-binding cassette transporter” are properly included and 122 “unknown” sequences are also contained.

Genome projects have provided an enormous number of potential protein sequences, of which functions tried to be identified by using computer-based methods. Many of these proteins, however, have not yet been annotated, with about half of all proteome sequences being classified as functionally “unknown” or “putative” at best (Serres et al. 2001). Such is the case, in particular, for transmembrane (TM) proteins, which account for as much as 20%~30% of the total number of proteins in individual proteomes (Boyd et al. 1998; Jones 1998; Wallin and von Heijne 1998; Mitaku et al. 1999; Pasquier and Hamodrakas 1999; Stevens and Arkin 2000; Krogh et al. 2001; Liu and Rost 2001; Arai et al. 2003). As will be described later, functionally “unknown” sequences make up more than three-quarters of all TM proteomes (see later). Furthermore, in 70,228 full-length protein sequences with a function annotated as “known” in SWISS-PROT release 41 (containing 122,564 sequences in total; Boeckmann et al. 2003), the number of TM protein sequences is only 10,796, compared with 59,432 soluble protein sequences (details described in the section of Materials and Methods). This shortage of “known” TM protein sequences in the SWISS-PROT database, as a matter of course, would cause a serious delay in the classification and identification of TM protein functions if sequence similarity is used as the only criteria.

On the other hand, recent studies have revealed that TM protein functions are closely related to TM topology (the number of TM segments [TMSs], positions of the TMSs and N-tail location), and can be classified and identified with high accuracies using TM topology information as the primary basis even without using sequence similarity itself directly (Sugiyama et al. 2003; Inoue et al. 2004). Individual functional groups have their own specific TM topologies, that is, characteristic combination patterns of loop lengths. The similarity of TM topologies between two TM protein sequences can be evaluated rather easily from comparing the lengths of corresponding loop regions between the two sequences, as described in detail in the Materials and Methods section. It is generally true that a pair of TM protein sequences with a higher sequence identity usually shows a higher TM topology similarity. In some cases, however, the TM topology similarity is kept at a high level between two sequences belonging to the same functional groups (at the superfamily level) even if the sequence similarity is below the twilight zone. For example, we have a pair of TM protein sequences, mouse GABA receptor α 6 (GAA6_MOUSE) and human neuronal acetylcholine receptor α 5 (ACH5_HUMAN), between which sequence identity is only 15.8%, while the TM topology similarity is as high as 96.9%. Thus, it is expected that the classification and identification of TM protein functions on proteome scale should be improved to a large degree by making good use of TM topology information in addition to sequence similarity.

One example of the approaches for obtaining reliable and more accurate TM topology prediction data is the ConPred program (Ikeda et al. 2002; Arai et al. 2004; Xia et al. 2004), which is based on a consensus strategy by combining several proposed prediction methods, and achieves an accuracy increase of as much as 10%, for example, predicting the entire TM topology of prokaryotic TM protein sequences, from 56.5% (by MEMSAT 1.8 [Jones et al. 1994] and HMMTOP 2.0 [Tusnády and Simon 1998] and HMMTOP 2.0 [Tusnády and Simon 2001]) to 68.1% (Arai et al. 2004).

In this study, we propose a new approach for classifying and identifying TM proteome functions by using a clustering method based on TM topology similarity. We focused on predicted TM proteins from 87 completed prokaryotic (72 bacterial and 15 archaean) genome sequences. In this approach, in the case when sequences of unknown function are segregated into a cluster together with sequences of known function, not only the functional classification but also the functional identification are achieved. Prior to carrying out the clustering, we first identified functions of the predicted TM protein sequences and classified them into three categories by using homology search and sequence comparison on the basis of sequence similarity, that is, “known,” “putative” (similar to “known” sequences), and “unknown.”

Results and Discussion

Table 1 summarizes the 87 prokaryotic (31 proteobacterial, 22 gram-positive bacterial, 19 other bacterial, and 15 archaean) proteomes used in this study. Out of 239,359 protein sequences in the 87 proteomes, 53,053 TM protein sequences (22.2% of the 87 proteomes) were obtained together with their TM topologies following the procedure described in the Materials and Methods section. We focused on the TM proteins with between 1 and 12 TMSs (1~12-tms), because only 3.8% of all of the TM proteins in the proteomes have more than 12 TMSs. The number and the fraction of predicted 1~12-tms TM proteins in each proteome are also listed in Table 1. Most of the proteomes fall in a narrow range around 21% over the four categories of prokaryotic species, with a few extremes, for example, 13.2% for Buchnera aphidicola and 29.1% for Tropheryma whipplei. The average fraction of TM proteins per proteome was calculated as 21.3% over all 87 species. The distribution of the number of TMSs in the 51,044 TM protein sequences is given in the second column of Table 2.

Table 1.

The number of ORFs and predicted 1–12-tms TM proteins for 87 prokaryotic genomes

Category/species	Abbreviationa	ORFs	Predicted 1–12-tms proteins	Fraction, %
Proteobacteria (31)
Escherichia coli K-12 MG1655	eco	4289	915	21.3
Salmonella typhi CT18	sty	4395	925	21.0
Salmonella typhimurium LT2	stm	4451	968	21.7
Yersinia pestis CO92	ype	3885	852	21.9
Shigella flexneri 301	sfl	4180	897	21.5
Buchnera sp. APS	buc	564	77	13.7
Buchnera aphidicola Sg	bas	545	72	13.2
Wigglesworthia brevipalpis	wbr	611	116	19.0
Haemophilus influenzae Rd	hin	1709	320	18.7
Pasteurella multocida PM70	pmu	2014	453	22.5
Xylella fastidiosa 9a5c	xfa	2766	551	19.9
Xanthomonas campestris pv. campestris ATCC 33913	xcc	4181	870	20.8
Xanthomonas axonopodis pv. citri 306	xac	4312	890	20.6
Vibrio cholerae El Tor N16961	vch	3828	844	22.0
Vibrio vulnificus CMCP6	vvu	4537	1003	22.1
Pseudomonas aeruginosa PA01	pae	5565	1206	21.7
Pseudomonas putida KT2440	ppu	5350	1119	20.9
Pseudomonas syringae pv. tomato DC3000	pst	5471	1073	19.6
Shewanella oneidensis MR-1	son	4630	957	20.7
Neisseria meningitidis MC58	nme	2025	364	18.0
Ralstonia solanacearum GMI1000	rso	3440	713	20.7
Helicobacter pylori 26695	hpy	1566	293	18.7
Campylobacter jejuni NCTC11168	cje	1634	336	20.6
Rickettsia prowazekii Madrid E	rpr	834	217	26.0
Rickettsia conorii Malish 7	rco	1374	302	22.0
Sinorhizobium meliloti 1021	sme	3341	656	19.6
Agrobacterium tumefaciens C58 (Cereon)	atc	4556	788	17.3
Brucella melitensis 16M	bme	3198	632	19.8
Brucella suis 1330	bms	3264	651	19.9
Bradyrhizobium japonicum USDA110	bja	8317	1736	20.9
Caulobacter crescentus	ccr	3737	730	19.5
Total		104,569	21,526	20.6
gram-positive bacteria (22)
Bacillus subtilis 168	bsu	4100	979	23.9
Bacillus halodurans C-125	bha	4066	921	22.7
Oceanobacillus iheyensis HTE831	oih	3496	877	25.1
Staphylococcus aureus N315	sau	2593	586	22.6
Staphylococcus epidermidis ATCC 12228	sep	2419	546	22.6
Listeria monocytogenes EGD-e	lmo	2846	673	23.6
Listeria innocua CLIP 11262	lin	2968	661	22.3
Lactococcus lactis IL1403	lla	2266	516	22.8
Streptococcus pyogenes SF370	spy	1696	341	20.1
Streptococcus pneumoniae TIGR4	spn	2094	459	21.9
Streptococcus agalactiae 2603	sag	2124	478	22.5
Streptococcus mutans UA159	smu	1960	468	23.9
Lactobacillus plantarum WCFSl	lpl	3009	719	23.9
Clostridium acetobutylicum ATCC824	cac	3672	830	22.6
Clostridium perfringens 13	cpe	2660	685	25.8
Clostridium tetani E88	ctc	2373	592	24.9
Thermoanaerobacter tengcongensis MB4T	tte	2588	583	22.5
Mycoplasma genitalium G-37	mge	480	102	21.3
Mycoplasma pneumoniae M129	mpn	688	129	18.8
Mycoplasma pulmonis	mpu	782	190	24.3
Mycoplasma penetrans HF-2	mpe	1037	248	23.9
Ureaplasma urealyticum	uur	611	133	21.8
Total		50,528	11,716	23.2
Other bacteria (19)
Mycobacterium tuberculosis H37Rv	mtu	3869	763	19.7
Mycobacterium leprae TN	mle	1605	365	22.7
Corynebacterium efficiens YS-314	cef	2950	673	22.8
Streptomyces coelicolor A3(2)	sco	7512	1531	20.4
Bifidobacterium longum NCC2705	blo	1729	401	23.2
Tropheryma whipplei Twist	twh	808	235	29.1
Fusobacterium nucleatum ATCC 25586	fnu	2068	428	20.7
Chlaymdia muridarum	cmu	909	213	23.4
Chlamydophila pneumoniae CWL029	cpn	1052	289	27.5
Borrelia burgdorferi B31	bbu	850	221	26.0
Treponema pallidum Nichols	tpa	1031	216	21.0
Leptospira interrogans 56601	lil	4727	912	19.3
Synechocystis sp. PCC6803	syn	3169	724	22.8
Thermosynechococcus elongatus BP-1	tel	2475	569	23.0
Anabaena sp. PCC7120	ana	5366	1183	22.0
Chlorobium tepidum TLS	cte	2252	407	18.1
Deinococcus radiodurans R1	dra	2936	523	17.8
Aquifex aeolicus VF5	aae	1522	270	17.7
Thermotoga maritima MSB8	tma	1846	374	20.3
Total		48,676	10,297	21.2
Archaea (15)
Methanococcus jannaschii DSM2661	mja	1715	308	18.0
Methanosarcina acetivorans C2A	mac	4540	1055	23.2
Methanosarcina mazei Goel	mma	3371	695	20.6
Methanobacterium thermoautotrophicum deltaH	mth	1869	389	20.8
Methanopyrus kandleri AV19	mka	1691	301	17.8
Archaeoglobus fulgidus DSM4304	afu	2407	470	19.5
Halobacterium sp. NRC-1	hal	2058	419	20.4
Thermoplasma acidophilum	tac	1478	295	20.0
Thermoplasma volcanium GSSl	tvo	1526	306	20.1
Pyrococcus horikoshii OT3	pho	2064	506	24.5
Pyrococcus abyssi	pab	1765	368	20.8
Aeropyrum pernix K1	ape	2694	564	20.9
Sulfolobus solfataricus P2	sso	2977	622	20.9
Sulfolobus tokodaii strain7	sto	2826	638	22.6
Pyrobaculum aerophilum IM2	pai	2605	569	21.8
Total		35,586	7505	21.1
Grand total		239,359	51,044	21.3

^a Three character codes are used as abbreviations following the KEGG database (Kanehisa et al. 2004).

Table 2.

The current status of the functional identification of the 1–12-tms TM proteins from the 87 prokaryotic genomes based on a homology search and sequence similarity search

		Functionally annotated sequences
tms	Total sequences	“Known” (%)	“Putative” (%)	“Unknown” (%)	Identifieda (%)
1	14,590	584 (4.0)	2191 (15.0)	11,815 (81.0)	19.0
2	6928	229 (3.3)	785 (11.3)	5914 (85.4)	14.6
3	4059	105 (2.6)	602 (14.8)	3352 (82.6)	17.4
4	4493	130 (2.9)	813 (18.1)	3550 (79.0)	21.0
5	3643	131 (3.6)	923 (25.3)	2589 (71.1)	28.9
6	4628	180 (3.9)	1411 (30.5)	3037 (65.6)	34.4
7	2076	82 (3.9)	515 (24.8)	1479 (71.2)	28.8
8	1965	82 (4.2)	572 (29.1)	1311 (66.7)	33.3
9	2015	100 (5.0)	704 (34.9)	1211 (60.1)	39.9
10	2061	89 (4.3)	525 (25.5)	1447 (70.2)	29.8
11	2045	94 (4.6)	625 (30.6)	1326 (64.8)	35.2
12	2541	132 (5.2)	794 (31.2)	1615 (63.6)	36.4
Total	51,044	1938 (3.8)	10,460 (20.5)	38,646 (75.7)	24.3

^a “Known” and “Putative” sequences are counted.

Current status of the proteome-wide functional identification of TM protein sequences based on sequence similarity only

The current level of functional identification of 1~12-tms TM proteins obtained by sequence homology searches (and similarity comparisons) is shown in Table 2. The fractions of TM protein sequences identified as “known” by our approach, which are defined as almost identical to or exactly the same as one of the sequences registered in the SWISS-PROT database with an unambiguous function, are extremely low: 5.2% for 12-tms TM proteins and 5.0% for 9-tms TM proteins at the highest, and only 3.8% as an average over all 1~12-tms TM proteins.

The fractions of “putative” sequences, the functions of which are inferable from the functionally known sequences in SWISS-PROT, range widely from the minimum, 11.3% for 2-tms, to the maximum, 34.9% for 9-tms TM proteins, with an overall average of 20.5%. The “known” and “putative” sequences added together amount to only 24.3%, that is, about one-quarter of the TM proteomes, indicating the majority (i.e., more than three-quarters) of TM proteomes are still classified as functionally unknown.

The results listed in Table 2 are illustrated in detail separately for each species in Figure 1 ▶. As expected, Escherichia coli has the highest percentage of known sequences, with over 40% of its TM proteome sequences classified as “known.” The high rate of “known” sequences for Shigella flexneri as much as for E. coli is due to the close phylogenic relationship between them that about three-quarters of TM protein sequences of S. flexneri are almost identical to those of E. coli. The “known” fractions for the archaean genomes are extremely small, that is, just 1.1% as an average over the 15 genomes.

Figure 1.

Current status of the functional identification of the 1~12-tms TM proteins from the 87 prokaryotic genomes based on a homology search and sequence similarity comparison together with the results of the functional classification and identification based on TM topology similarity: black bar, “known” sequences; dark gray bar, “putative” sequences; gray bar, newly classified and identified “unknown” sequences; white bar, still “unknown” sequences. Abbreviations of all 87 species are the same as in Table 1.

As with the fraction of “known” and “putative” sequences put together, 10 species belonging to γproteobacteria in proteobacteria (from E. coli to P. multocida in the list) stand out among the other species. This is again the contribution from the large number of “known” E. coli sequences in SWISS-PROT. Overall, the proteobacteria genomes far exceed the other three species categories in the fractions of “known” and “putative” sequences. The archaean TM proteomes have the smallest fractions of “known” plus “putative” sequences, 8.4% as an average over the 15 species. Interestingly, 65.1% of the “putative” sequences over all the archaean genomes are annotated after the proteobacterial “known” sequences, while only 23.3% of them are directly after the archaean “known” sequences.

Threshold TM topology similarities and the minimum cluster size

We assumed the proteome-scale functional classification using the clustering approach was successful when more than 50% of all the sequences were included in the clusters of at least 10 sequences (the minimum cluster size). The threshold TM topology similarities as the criteria for clustering were determined based on this assumption. The conditions (the 50% coverage and the minimum cluster size of 10) adopted in our approach are not based on any scientific data, but rather are purely empirical ones. This assumption is, however, supported by the relationships between the threshold TM topology similarities versus the minimum cluster size, where with increasing minimum cluster size, the threshold TM topology similarities decrease rapidly at first and then reach saturated levels at a minimum cluster size of around 10 for most numbers of TMSs (see Supplemental Fig. 1 ▶). Hereafter, we refer to clusters whose size is larger than nine as “large clusters,” and all others including orphan clusters as “small clusters.”

Threshold TM topology similarities thus determined are, for example, 98%, 85%, and 82% for 1-tms, 6-tms, and 12-tms TM protein sequences, respectively, as shown in the third column of Table 3. As expected, stricter threshold similarity values are obtained for the smaller numbers of TMSs.

Table 3.

The results of the functional classification and identification of the 1–12-tms TM proteins from the 87 prokaryotic genomes based on TM topology similarity

			Large cluster (with ≥10 sequences)
					Functionally annotated sequences
tms	Total sequences	Threshold TM topology similarity (%)	Clusters	Sequences (%)	“Known” (%)	“Putative” (%)	“Unknown” (%)	Classified and identified (%)a
1	14,590	98	74	7337 (50.3)	332 (4.5)	1295 (17.7)	5710 (77.8)	58.2
2	6928	92	46	3660 (52.8)	157 (4.3)	534 (14.6)	2969 (81.1)	57.5
3	4059	85	32	2281 (56.2)	75 (3.3)	426 (18.7)	1780 (78.0)	61.3
4	4493	84	41	2515 (56.0)	97 (3.9)	561 (22.3)	1857 (73.8)	62.3
5	3643	81	33	1923 (52.8)	76 (4.0)	625 (32.5)	1222 (63.5)	62.5
6	4628	85	27	2464 (53.2)	108 (4.4)	1024 (41.6)	1332 (54.1)	63.2
7	2076	75	25	1075 (51.8)	44 (4.1)	330 (30.7)	701 (65.2)	62.5
8	1965	73	26	1037 (52.8)	52 (5.0)	398 (38.4)	587 (56.6)	63.2
9	2015	74	30	1033 (51.3)	67 (6.5)	501 (48.5)	465 (45.0)	63.0
10	2061	74	31	1090 (52.9)	42 (3.9)	293 (26.9)	755 (69.3)	66.4
11	2045	75	23	1087 (53.2)	62 (5.7)	400 (36.8)	625 (57.5)	65.7
12	2541	82	22	1286 (50.6)	80 (6.2)	499 (38.8)	707 (55.0)	64.3
Total	51,044		410	26,788 (52.5)	1192 (4.4)	6886 (25.7)	18,710 (69.8)	60.9

^a Sequences in the large clusters and “known” and “putative” sequences in the small clusters (with 1–9 sequences) are included.

Comprehensive functional classification and identification of TM protein sequences based on TM topology similarity

The results of the functional classification and identification of TM proteomes using the single-linkage clustering method based on TM topology similarity are summarized in Table 3 for 1~12-tms TM proteins. The numbers of large clusters generated range from 22~74, with more clusters generated for the smaller numbers of TMSs and less for larger, in general. In these large clusters, more than half of all of the TM proteome sequences are included, a large majority of which (69.8%) are “unknown” sequences together with “known” and “putative” sequences, indicating a large amount of “unknown” sequences have been functionally classified and identified by this approach. Taking into account the “known” plus “putative” sequences included in the small clusters all together, the number of functionally annotated TM protein sequences runs up to 60.9% of the TM proteome sequences, a significant improvement over the 24.3% obtained from the sequence homology search plus similarity comparison.

The percentages of newly classified and identified sequences using this approach are displayed in Figure 1 ▶ (gray bars) individually for the respective species. Averaged over the 31 proteobacteria, 22 gram-positive bacteria, 19 other bacteria, and 15 archaea, 32.3%, 40.3%, 38.9%, and 36.7% of all the sequences are newly classified and identified, respectively. These correspond to more than half of all the “unknown” sequences in the four individual species categories. Several γproteobacteria belonging to proteobacteria, that is, E. coli, Salmonella typhi, Salmonella typhimurium, Yersinia pestis, S. flexneri, Buchnera sp., B. aphidicola, Wigglesworthia brevipalpis, Haemorphilus influenzae and Pasteurella multocida have smaller numbers of newly classified and identified TM proteins when compared with the other species, although the total levels of functional annotation achieved were remarkably high, as much as around 80%. It is also noted that the number of classified and identified archaean sequences that were originally “unknown” significantly increased from 8.4% to 45.9%.

The following describes the details of the functional classification and identification attained by this approach, exemplifying 6-tms TM proteins.

Table 4 provides the list of the 27 large clusters generated by single-linkage clustering based on TM topology similarity (threshold similarity 85%) for 6-tms TM proteins enumerated in order of cluster size. The largest cluster, Cluster 1, includes 1085 sequences, nearly one-fourth of all of the 6-tms TM protein sequences, with the “known” plus “putative” sequences (679 in total) annotated as “transport system permease protein” except for one sequence (as photosystem II chlorophyll-binding protein). This implies the 406 “unknown” sequences (37.4% of the 1085 sequences) included in the cluster also could be annotated as transport system permease proteins. By further clustering based on sequence similarity (threshold sequence identity 30%) within Cluster 1, we obtained 46 subclusters that correspond to functional subgroups that are, for example, “dipeptide transport system permease dppB” (in total 228 sequences including “unknown” sequences), “maltose transport system permease malD” (212 sequences), “lactose transport system permeases lacF” (181 sequences), “sulfate transport system permease cysT” (118 sequences), etc., suggesting that the TM topology-based clustering may correspond to a superfamily-or family-level classification, whereas the sequence similarity-based clustering to a family- or subfamily-level one in this case.

Table 4.

A list of 27 large clusters (with ≥10 sequences) based on a threshold TM topology similarity of 85% for 6-tms TM proteins

Cluster number	Sequences (“known” or “putative”)	Signal peptide	N-tail location	Speciesa	Subclustersb	Major functional descriptions for“known” or “putative” sequences
1	1085 (679)	—	in	P(25)/G(22)/O(18)/A(15)	46	Dipeptide transport system permease protein dppB (116)
2	288 (26)	—	in	P(26)/G(18)/O(17)/A(13)	123	Daunorubicin resistance transmembrane protein (5)
						Membrane protein nosY precursor (5)
						Teichoic acid translocation permease protein tag (5)
3	231 (109)	—	in	P(26)/G(22)/O(14)/A(3)	23	Multidrug resistance-like ATP-binding protein mdlA (17)
						Transport ATP-binding protein cydD (17)
4	111 (14)	—	in	P(22)/G(8)/O(10)/A(4)	74	Sensor protein pilS (4)
						Membrane-protein yhjW (3)
						Sensor histidine kinase regB (3)
5	101 (1)	—	in	P(17)/G(17)/O(4)/A(12)	67	Protein natB (1)
6	95 (46)	—	out	P(27)/G(14)/O(7)/A(5)	26	Electron transport complex protein rnfA (17)
						Cytochrome c-type biogenesis protein ccdA (16)
7	78 (36)	—	in	P(20)/G(13)/O(7)/A(8)	12	Zinc transporter zitB (13)
						Protein P34 (12)
8	78 (34)	—	in	P(27)/G(14)/O(7)/A(3)	19	Sec-independent protein translocase protein tatC (29)
9	62 (0)	—	in	P(27)/G(0)/O(11)/A(0)	26	Unknown
10	52 (30)	—	in	P(15)/G(7)/O(12)/A(10)	16	Protein-export membrane protein secD (26)
11	37 (32)	—	in	P(20)/G(4)/O(8)/A(4)	2	Protein-export membrane protein secF (32)
12	27 (25)	—	in	P(20)/G(1)/O(1)/A(0)	3	Heme exporter protein C (25)
13	24 (17)	—	in	P(16)/G(2)/O(2)/A(3)	5	Heme exporter protein B (17)
14	24 (4)	—	in	P(0)/G(12)/O(2)/A(0)	9	Phage infection protein (4)
15	19 (1)	—	out	P(7)/G(10)/O(1)/A(0)	6	Capsular polysaccharide biosynthesis protein rkpI (1)
16	19 (0)	—	out	P(9)/G(7)/O(0)/A(2)	1	Unknown
17	15 (15)	—	in	P(11)/G(2)/O(2)/A(0)	1	Aquaporin Z (13)
18	15 (12)	—	out	P(9)/G(5)/O(0)/A(1)	1	Lead, cadmium, zinc and mercury transporting ATPase (9)
19	14 (9)	—	out	P(8)/G(3)/O(1)/A(0)	6	Flagellar biosynthetic protein fliR (5)
						Secretion system apparatus protein ssaT (3)
20	14 (1)	—	in	P(11)/G(2)/O(0)/A(0)	5	Intracellular septation protein (1)
21	13 (9)	—	in	P(2)/G(5)/O(1)/A(0)	3	Phosphonates transport system permease protein phnE (7)
22	11 (7)	—	in	P(1)/G(9)/O(0)/A(1)	3	Spermidine/putrescine transport system permease protein potB (6)
23	11 (0)	—	in	P(0)/G(0)/O(0)/A(11)	4	Unknown
24	10 (10)	—	in	P(9)/G(0)/O(0)/A(0)	1	Glycerol uptake facilitator protein (10)
25	10 (8)	—	in	P(9)/G(0)/O(0)/A(0)	1	Protein glpG (8)
26	10 (7)	—	in	P(2)/G(5)/O(2)/A(1)	1	Lead, cadmium, zinc and mercury transporting ATPase (4)
27	10 (0)	—	in	P(10)/G(0)/O(0)/A(0)	2	Unknown
Total (27)	2464 (1132)				486
Others (1667)	2164 (459)

^a P(i), G(j), O(k), and A(l) represent the number of proteobacterial (i), gram-positive bacterial (j), other bacterial (k), and archaean (l) species included in the cluster, respectively.

^b These were generated by further clustering within the cluster based on sequence similarity (threshold sequence identity 30%).

The top 13 clusters, except for Clusters 9 and 12, contain sequences that distribute over all the species categories, indicating the TM proteins of these functional groups are essential for the life of prokaryotic species. By comparison, Cluster 14 (phage infection protein) contains sequences from only gram-positive bacterial and other bacterial species, and the sequences in Cluster 20 (intracellular separation protein) exist only in proteobacterial and gram-positive bacterial genomes.

In Table 4, we have four clusters composed of only “unknown” sequences, Clusters 9, 16, 23, and 27. Of these, Clusters 23 and 27 comprise the sequences from only archaean and proteobacterial species, respectively. These “unknown” protein sequences must be not only novel but also biologically important functional groups. We expect further experimental studies would characterize these sequences and elucidate their functions in detail.

Cluster 3 (231 sequences, of which 109 are “known” or “putative” assigned as “ATP-binding cassette [ABC] transporters”) clearly illustrates how well the TM topology-based clustering works in the functional classification and identification of TM proteins. Out of 119 6-tms sequences annotated as “ABC transporter,” 109 sequences (91.6%) are captured properly in this cluster, and the remaining 10 sequences are spread across nine small clusters: one sequence in a cluster with the size of nine (including nine sequences in total, N-in topology), one in a size-four cluster (N-out), two in a size-two (N-in), one in a size-two (N-out, +SP), and five orphan sequences (all N-in). The other 122 sequences are all “unknown,” and no sequences with other functions are included in this cluster at all.

TM topology models of the 231 sequences are illustrated in Figure 2 ▶, where TMS, cytoplasmic, and noncytoplasmic loop regions are represented with black, gray, and dark gray bars, respectively. The sequences in this cluster have the following characteristics with regard to TM topology: (1) No signal peptide (SP) is present; (2) the N-tail loop is located on the cytoplasmic side; (3) the cytoplasmic loops (including the N- and C-tails) are long, in particular, the C-tail loop is extremely long; and (4) the noncytoplasmic loops are short with connecting adjacent TMSs to form three typical “helical-hairpin” domains in the TM topology architecture (Gafvelin and von Heijne 1994; Gafvelin et al. 1997). The sequence-similarity based clustering within Cluster 3 segregates these 231 sequences into 23 subclusters (including 13 orphan clusters), which are shown in Figure 2 ▶ divided by the space lines. In the largest subcluster (182 sequences), 105 sequences are “known” or “putative” as ABC transporters, and the remaining 77 are “unknown.” The other 22 subclusters are composed of only “unknown” sequences, with the exception of a few subclusters that contain ABC transporter sequences.

Figure 2.

TM topology models of the 231 sequences in 6-tms Cluster 3 (“ABC transporter”; see Table 4), with the 23 subclusters (including orphan clusters) separated by spaces and enumerated in order of subcluster size: black bar, TMS region; gray bar, cytoplasmic loop region; dark gray bar, noncytoplasmic loop region. These TM topologies have traits such as very long cytoplasmic and short noncytoplasmic loops, and seem to be formed from three typical “helical-hairpin” domains in the TM topology architecture.

We would like to show another typical example, that is, Cluster 10 (52 sequences are included) of which TM topology models are presented in Figure 3 ▶. Among the 30 “known” or “putative” sequences included in this cluster, 26 sequences are actually secD proteins and four sequences are of other functions. We have 33 6-tms sequences that were annotated as “secD” (both “known” and “putative”) by the homology search plus sequence similarity comparison procedure. This means 79% (26 of 33) of 6-tms secD protein sequences were correctly classified into this cluster by this approach. The remaining seven secD protein sequences are dissipated into two small clusters: six sequences in a size-six cluster, and one orphan sequence, both of which have N-out topology predicted in error. Among the four sequences other than secD in Cluster 10, one sequence is “known” as “protein-export membrane protein secF,” of which the sequence is similar to that of secD and mostly classified together with secD into a unified secD_secF family in the family databases such as Pfam (Bateman et al. 2004). The other three are “putative” sequences annotated as “inner membrane protein creD” with rather low sequence identities, that is, 42.9%, 36.8%, and 30.6% to a hit sequence in SWISS-PROT. There are two other small 6-tms clusters holding two “known” creD sequences each: a size-three cluster of which another member is “unknown” (N-out), and a size-two cluster (N-out, +SP). Because the exact TM topology model of creD protein has not been determined experimentally yet, it is not ascertained currently which predicted TM topology model is the true one: N-out or N-in. At the same time, however, our results indicate accidentally that both TM proteins have similar predicted TM topologies except for the N-tail location, that is, N-in for secD and N-out for creD. Although the function of creD seems different apparently from that of secD (suggested an enhancing effect on the transcription of phoA; Drury and Buxton 1988), both functions might be somewhat related possibly.

Figure 3.

TM topology models of the 52 sequences in 6-tms Cluster 10 (“secD protein”; see Table 4), with the 16 subclusters separated by the space lines: black bar, TMS region; gray bar, cytoplasmic loop region; dark gray bar, noncytoplasmic loop region. The leftmost three-letter codes are abbreviations for the species name (underline, functionally “unknown” sequence) and are the same as in Table 1. The Arabic numerals next to the three-letter codes indicate species-categories (1, proteobacteria; 2, gram-positive bacteria; 3, other bacteria; 4, archaea).

The TM proteins contained in Cluster 10 have the following characteristics with the TM topology, as seen in Figure 3 ▶: (1) The N-tail loop is short; (2) the noncytoplasmic loop connecting the first and second TMSs is extremely long; (3) the second cytoplasmic loop is short; (4) the other two noncytoplasmic loops are short, with connecting adjacent TMSs to form two “helical-hairpin” domains; and (5) the remaining two cytoplasmic loops are relatively long. It should be pointed out that the secD and secF families (the TM topology models of “secF” cluster, that is, Cluster 11, are shown in Supplemental Fig. 2 ▶) are defined as one unified family (i.e., SecD_SecF) in the Pfam database as mentioned previously, while these two families are correctly split into two clusters by our approach. It is interesting that the differences in TM topology between secD and secF proteins are enough to bring about this result, but the rather high sequence similarities between these protein families resulted in being classified together into one family. Further sequence-similarity based clustering within the cluster (threshold identity 30%) yields 16 subclusters (including nine orphan subclusters). The largest subcluster consists of 22 “known” sequences distributed over three of the four species groups, the second largest consists of seven “unknown” sequences specific to only archaean species, and the third consists of four “known” sequences from only “other bacteria.” This is a good example of how our approach is effective in classifying and identifying the functions of TM proteins even in species that are distantly related. This example also demonstrates that TM topologies are more conservative than amino acid sequences themselves for preserving the TM protein functions.

Materials and methods

Data source

We used 239,359 open reading frames (ORFs) from 87 sequenced prokaryotic genomes registered in GenBank (Benson et al. 2004) for this study, as listed in Table 1. The ORFs were downloaded from ftp://ncbi.nlm.nih.gov/genbank/genomes/ on March 6, 2003. The 87 genomes included 31 proteobacteria, 22 gram-positive bacteria, 19 other bacteria, and 15 archaea according to the classification in GenBank.

Prediction of TM protein sequences and their TM topologies from the proteomes

Out of the protein sequences translated from the ORFs, we segregated TM protein sequences and predicted their TM topologies according to the following procedure: (1) prediction of TM protein sequence candidates using SOSUI (≥98% accuracy; Hirokawa et al. 1998); (2) removal of predicted SP regions using DetecSig (88% accuracy; Lao and Shimizu 2001^; Lao et al. 2002); and (3) prediction of TM topology by ConPred (68.1% accuracy; Arai et al. 2004). A more detailed description of this procedure is given in our previous article (Arai et al. 2003).

Functional identification of TM protein sequences based on sequence similarity

We first categorized the 114,965 full-length protein sequences in SWISS-PROT release 41 into “known,” “putative,” or “unknown” groups according to the level of functional annotation. For this categorization, we adopted the simple but rational criteria given in the GTOP database (http://spock.genes.nig.ac.jp/?genome/func.html; Kawabata et al. 2002). The criterion for discriminating sequences with a “known” function requires at least one of the following: (1) more than five letters with functional information in the DE line, (2) at least one informative word in the KW line, or (3) both “-!- FUNCTION” and “-!- CATALYTIC ACTIVITY” in the CC line. Sequence entries were classified as “putative” if the entry contains one of the following descriptions: (1) “HOMO-LOG,” “HOMOLOGY,” “HYPOTHETICAL,” “POTENTIAL,” “POSSIBLE,” “PROBABLE,” or “PUTATIVE” in the DE line; (2) “BY SIMILARITY,” “HYPOTHETICAL,” “POTENTIAL,” “POSSIBLE,” “PROBABLE,” or “PUTATIVE” in the “CC -!FUNCTION” or “CC -!- CATALYTIC ACTIVITY” line; and (3) “HYPOTHETICAL PROTEIN” in the KW line. When only the “known” criterion is satisfied, the sequence is regarded as “known.” In cases when both “known” and “putative” criteria are true, the sequence is classified as “putative.” The sequences to which the “known” criterion does not fit are categorized as “unknown,” even if the “putative” criterion fits. Through this procedure, we obtained 70,228 “known” (10,796 TM protein sequences), 39,296 “putative” (6643), and 5441 “unknown” (754) sequences from SWISS-PROT release 41.

Next, we classified the 51,044 predicted TM protein sequences from the 87 prokaryotic genomes into three categories in agreement with the functional description levels in SWISS-PROT using a BLAST homology search (Altschul et al. 1990, 1997) and an ALIGN (Myers and Miller 1988) sequence comparison, as illustrated in Figure 4 ▶. The BLAST search was carried out with the default settings (first gap penalty, -11; additional gap penalty, -1; substitution matrix, BLOSUM 62; Henikoff and Henikoff 1992) against the grouped full-length sequences from the SWISS-PROT database. If a query sequence matches one of the SWISS-PROT sequences of the “known” group with an E-value less than 10⁻⁵, it was treated as a candidate for the “known” or “putative” category; otherwise, it was classified as “unknown.”

Figure 4.

Illustration of the procedure to categorize TM protein sequences into “known”, “putative”, and “unknown” groups according to the level of functional annotation in SWISS-PROT as ascertained by a homology search and sequence similarity comparison using BLAST and ALIGN, respectively. BLAST is used only for detecting candidates for “known” and “putative” sequences against SWISS-PROT; the level of functional annotation is determined in compliance with the value of the sequence identity calculated using ALIGN.

Next, the “known” or “putative” candidate sequence from the BLAST search process was aligned with the matched SWISS-PROT sequences to calculate the global sequence identities between them using the ALIGN program with the default settings, except for the substitution matrix (BLOSUM 62 was used). The matched SWISS-PROT sequence with the highest identity was characterized as the most similar one to the candidate sequence, and the candidate sequence was finally classified into one of the three categories according to the value of the highest identity: “known” (with a highest identity of ≥95%), “putative” (30%~95%), or “unknown” (<30%). When a query sequence is categorized into “known” or “putative,” it is considered to be a functionally identified TM protein and the function of the matched SWISS-PROT sequence is given to the query sequence as its function.

Functional classification and identification of TM protein sequences based on TM topology similarity

The procedure for classifying and identifying TM protein functions based on TM topology similarity is illustrated in Figure 5 ▶. The 51,044 TM protein sequences annotated using a BLAST homology search and ALIGN sequence comparison were divided into 36 data sets according to the number of TMSs, the presence or absence of a signal peptide, and N-tail location. The sequences are clustered within individual data sets by a single-linkage method based on TM topology similarity. In this single-linkage clustering, the TM topology similarity is used as the determining factor defined as:

Figure 5.

Illustration of the procedure used to classify and identify TM protein functions by single-linkage clustering based on TM topology similarity and sequence similarity using ALIGN (threshold sequence identity 30%). TM topology similarities used in the clustering are determined in the subsection, “Threshold TM topology similarities and the minimum cluster size.” Prior to clustering, the predicted 1~12-tms TM protein sequences were divided into 36 data sets according to the number of TMSs, the existence of a signal peptide and N-tail location.

(1)

where, n, l_{1, i} and l_{2, i} are the number of TMSs and the length of the i-th loop in sequences 1 and 2, respectively, and min (l_{1, i}, l_{2, i}) and max (l₁, _i, l2, _i) are the lengths of the shorter and longer loops in l_{1, i} and l_{2, i}, respectively.

Within the individual TM-topology based clusters, the sequences are further clustered by a single-linkage method based on sequence similarity (threshold sequence identity 30%) using the ALIGN program with the default settings, except for the substitution matrix (BLOSUM 62 was used), to generate subclusters that must correspond to functional subgroups in the TM-topology based clusters, as illustrated in Figure 5 ▶.

Electronic supplementary material

Supplemental materials are (1) lists of the obtained large clusters based on TM topology similarity for 1~12-tms TM proteins (named “Supple_Table1.doc”), (2) Supplemental Figure legends (“Supple_Fig_legends.doc”), (3) Supplemental Figure 1 ▶ (“Supple_ Fig1.xls”), (4) Supplemental Figure 2 ▶ (“Supple_Fig2.ppt”), (5) data sets of 87 prokaryotic TM proteome sequences functionally annotated by homology search plus sequence similarity comparison (e.g., “eco.db” for E. coli), and (6) ID lists of the sequences in the obtained clusters (both the large and small clusters) for individual 1~12-tms TM proteins (e.g., “tms06.db” for 6-tms TM proteins). These files are also available at ftp://bioinfo.si.hirosaki-u.ac.jp/TopClust.

Abbreviations

• ABC, ATP-binding cassette

• n-tms, with n transmembrane segment(s)

• ORF, open reading frame

• SP, signal peptide

• TM, transmembrane

• TMS, transmembrane segment