INTRODUCTION

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION CONCLUSIONS REFERENCES

The nicotinic acetylcholine receptors (nAChRs) belong to the superfamily of ligand-gated ion channels (LGIC) (Cockcroft et al., 1992; Galzi and Changeux, 1994) that are allosteric transmembrane proteins responsible for fast ionic responses to neurotransmitters. These receptors are homo- or hetero-pentamers made from a set of 16 related subunits in vertebrates (9, 4, , , and ) (review in Le Novère and Changeux, 1995). Other receptors formed of polypeptides homologous to the nAChR subunits include 5-HT3, GABA_A, GABA_C, and glycine receptors of vertebrates, as well as their invertebrate counterparts. Despite their rather different pharmacological properties (Ortells and Lunt, 1995), these receptors likely possess a common quaternary structure (Eisele et al., 1993; Langosch et al., 1988). The vertebrate GABA_A receptors are formed from a set of 16 related subunits (6, 4, 4, 1, and 1) (MacDonald and Olsen, 1994). The GABA_C receptors are homo-pentamers of 1-3 subunits (Bormann and Feigenspan, 1995). The vertebrate glycine receptors are made from a set of five related subunits (4 and 1) (Bechade and Triller, 1994).

Every mature subunit of the nAChR family is assumed to follow the same transmembrane topology (Hucho et al., 1996). A large amino-terminal portion carrying the components of the acetylcholine (ACh) binding site faces the extracellular environment. The three subsequent segments cross the membrane, followed by a large intracellular domain and a fourth segment that again crosses the membrane. The relatively short carboxy-terminal domain is extracellular.

The sequence conservation varies along the subunits. The amino-terminal signal peptide and the middle of the cytoplasmic portion are highly variable, whereas the amino-terminal moiety, as well as the membrane flanking portions of the cytoplasmic part, are well conserved. The transmembrane segments are highly conserved. In humans, the size of the subunits vary from 457 aa (1) to 627 aa (4).

The overall low resolution structure of the nAChRs was initially determined by electron microscopy on single molecules (Cartaud et al., 1973), or bi-dimensional crystals (Kistler and Stroud, 1981) of Torpedo electric organ receptor. The nAChR molecules result from the assembly of five subunits arranged around an axis of symmetry perpendicular to the membrane. The length of the molecule is ~120 Å, with an extracellular, funnel-shaped portion of 60 Å and a transmembrane portion of 30 Å. The diameter of the extracellular entry of the pore is ~25 Å wide, while the intracellular one is slightly smaller (Toyoshima and Unwin, 1988; Unwin, 1993a). A similar shape was proposed for GABA_A receptors (Nayeem et al., 1994).

Affinity labeling and site-directed mutagenesis have shown that the ligand binding sites are located at the interface of two subunits, formed by residues belonging to two components (Galzi and Changeux, 1994; Table 1). The principal component [on the subunit surface that would be reached first if following the clockwise path when the structure is viewed from the extracellular surface (Machold et al., 1995)] is carried by the  subunits and comprises at least three segments or loops (Dennis et al., 1988; Galzi et al., 1990). Facing it, the complementary component includes three (or possibly four) different segments or loops (Corringer et al., 1995; Czajkowski et al., 1993; Prince and Sine, 1996; Sine et al., 1995; reviewed in Hucho et al., 1996; Tsigelny et al., 1997). Such composite ligand binding sites appear to be conserved throughout the superfamily of LGIC. Indeed, it has been shown that the binding sites for ACh, GABA, glycine, and benzodiazepines are homologous (Schmieden et al., 1992; Vandenberg et al., 1992; reviewed in Galzi and Changeux, 1994). The ionotropic glutamate receptors constitute a separate superfamily in which agonist sites probably do not occur at subunit interfaces (Paas, 1998).

Cryoelectron microscopy of the Torpedo electric organ receptor has provided three-dimensional (3D) images of the nAChR at a resolution of 9 Å (Unwin, 1993a). Such a resolution is not sufficient to resolve the spatial position and the secondary structure assignment of any particular amino acid. Although the extracellular domain has been successfully produced in a soluble form (Wells et al., 1998), the quantities obtained are still too low to permit the production of crystals suitable for x-ray diffraction. The NMR approach has been limited to small fragments (Basus et al., 1993). Some attempts were made with other methods, such as atomic force microscopy (Lal and Yu, 1993), though with a resolution lower than that of electron microscopy on two-dimensional (2D) crystals.

It is therefore of interest to obtain information on the receptor protein organization from the data currently available, i.e., the sequences of the subunits. Accordingly, in parallel with the experimental approaches, efforts have been made to predict the structure of the individual subunit with computational techniques. Two approaches have been used. The comparative modeling techniques sought to give a structural description of a protein provided that a plausible structural model can be identified. The problem resides in the identification of a suitable template from sequence information only. However, the lack of sufficient sequence similarity between an nAChR subunit and a protein of known structure requires fold recognition methods (Gready et al., 1997; Tsigelny et al., 1997) which, as known from test cases, are only partially successful in recognizing similar folds in the absence of sequence similitude (Rost and Sander, 1996). This approach also suffers from the fact that a plausible 3D model may not exist in the currently available protein structure database (Marchler-Bauer et al., 1997). The two models proposed so far are indeed different (Gready et al., 1997; Tsigelny et al., 1997). In parallel, ab initio secondary structure predictions were performed with first-generation algorithms (single amino acid-based, 50-60% accuracy) by Finer-Moore and Stroud (1984) and Ortells (1997).

Here we present a secondary structure prediction of the nAChR subunit based on third-generation algorithms (based on multiple alignments and that are capable of achieving >70% accuracy), in order to take into account the information derived from the wealth of cloned homologous subunit sequences. The combination of several independent first- and second-generation algorithms has been shown to increase the accuracy of secondary structure predictions (Biou et al., 1988; Nishikawa and Ooi, 1986; Zhang et al., 1992). We describe a program that integrates results from several prediction algorithms and multiple homologous proteins. We applied this program to the different members of the nAChR family and LGIC superfamily to increase the signal/noise ratio. In addition, the program furnished the consensus of predicted solvent accessibility and topology. By using these data in combination with information obtained from experimental sources, we integrated the results into a 2D representation of a typical nAChR subunit.

	MATERIALS AND METHODS

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION CONCLUSIONS REFERENCES

Alignment

All sequences used in this study can be found in the ligand-gated ion channel (LGIC) subunit database at the URL (http://www.pasteur.fr/units/neubiomol/LGIC.html). For the secondary structure predictions, two multialignments were achieved with ClustalX software (Thompson et al., 1997; available at ftp-igbmc.u-strasbg.fr) (pairwise gap opening, 10; pairwise gap extension, 0.1; multiple gap opening, 5; multiple gap extension, 0.05; Blosum matrix series). One alignment was carried out with 18 subunit sequences of cationic channels (AL1). AL1 contains 5-HT3 from Mus, nicotinic 1 of Torpedo, 2-6, 9, and 2-4 of Rattus, 7-8 of Gallus, 1, , , , of Mus (one example of each paralog gene), and DEG3 of Cænorhabditis (which has still no uncovered vertebrates ortholog). Another alignment was constructed with 38 LGIC (cationic and anionic LGIC) sequences (AL2). AL2 contains AL1 subunit sequences plus GABA 1-6, 1-3, 1-3, 1-3, , glycine 1-3, and  from Rattus. The aim was to determine whether the incorporation of information from more distantly related sequences would improve the predictions. We did not use more than one sequence per group orthology because of the high similarity between orthologs (and hence the lack of additional information brought from the use of multiple orthologs). The ASSP software (Russel and Barton, 1993; available at ftp://geoff.biop.ox.ac.uk/programs/assp/) allowed us to expect a Q₃ accuracy (i.e., a percentage of three-state comparison identity) of perfect prediction in the interval [83.45-100%] for AL1 and [82.74-100%] for AL2. To study the conservation of sequence at each position along the sequence, a third multiple alignment was constructed from 152 different LGIC subunit sequences. All these sequences correspond to subunits shown to be integrated in functional receptors (thus eliminating the putative members originating from large-scale genome projects).

Secondary structure prediction by consensus average

A computer program was written in C to integrate secondary structure predictions based on different algorithms. SSPCA (for secondary structure prediction by consensus average) was designed to combine three-state predictions and probabilities from several prediction programs and several sequences (Fig. 1). The SSPCA program is also designed to treat other types of prediction such as solvent accessibility and topological arrangements for membrane proteins. The individual predictions were not weighted by sequence similitudes.

View larger version (35K): [in this window] [in a new window]   FIGURE 1   Scheme describing our approach. Only four sequences are displayed (the portion of sequence shown corresponds to E1).

As input, SSPCA takes an alignment of amino acid sequences (in a Clustal format) and a file containing the predictions. The prediction file contains for each sequence and for each method (if available) the probability for helix, , and coil [0-9], the resulting secondary structure prediction [H(elix) or E(xtended) or C(oil)], the probability of accessibility to solvent [0-9], the resulting accessibility to solvent (e(xposed) or b(uried)), and the topological state (o(utside), i(nside), T(ransmembrane)). The output of SSPCA is composed of (points 1-5 concern only the secondary structure prediction):

1. The MxS predictions P_(mi,sx) where M is the number of method, S the number of sequences, m_i the ith method, and s_x the xth sequence, projected on the alignment (insertion of gap in the predictions when present in the alignment). Each P_(mi,sx) is then a character string with the length of alignment, each character belonging to {H, E, C,`-'}).

2. The MxS(MxS  1)/2 pairwise comparisons C[P_(mi,sx), P_(mj,sy)] of the predictions P_(mi,sx) and P_(mj,sy). If is the set of positions of the alignment where neither P_(mi,sx) nor P_(mj,sy) contain a gap, that is, where both predictions are defined,

(1)

Where
and card() is the cardinal (the number of elements) of .

3. The congruence between methods µ_i,j: this parameter represents the percent identity between the consensus of predictions for all the sequences of two methods.

(2)

for every pair (i, j), i  j.

4. The congruence between sequences _x,y. This parameter represents the percent identity between the consensus of predictions for two sequences by all the methods.

(3)

for every pair (x, y), x  y. This parameter permits the comparison of the predictions for different homologous proteins.

5. The consensus predictions and the sum of probabilities: by sequences, by methods, and in toto (and the percent helix and strand for each consensus prediction). For each position, the consensus is computed as the major state. In case of identical cardinals, the arbitrary priority order is E > H > C > `-'. The percent helix and strand is given for the total nongapped consensus length.

6. The global consensus solvent accessibility. In case of identical cardinals, the arbitrary priority order is b > e > `-'.

7. The global consensus topology. In case of identical cardinals, the arbitrary priority order is T > i > o > `-'.

Secondary structure prediction programs

PHDsec (Rost and Sander, 1993a, b; 1994a) is composed of several cascading neural networks (previously trained on proteins of known structures). A first network takes as input a set of vectors representing the amino acid composition at positions of the multiple alignment in a window sliding along it. Its output is composed of a vector representing the probabilities for each of the three states of the central residue of the window. Since the secondary structure of a residue is not independent of the structure of neighboring residues, a second step takes into account these local interactions. A neural network takes as input the vectors present in a window sliding along the previous output. Its own output is a refined three-state probabilities vector. Another step consists of averaging (for each state) the outputs from independently trained networks. Finally, a "winner takes all" decision assigns the secondary structure state. No explicit rules are included in the algorithm. PHD may generate its own alignment with the submitted sequence [with the MaxHom algorithm (Sander and Schneider, 1991)]. Therefore, for every sequence of AL1 and AL2, a different alignment was generated and used for the prediction. PHDsec is accessible at the URL (http://www.embl-heidelberg.de/predictprotein/predictprotein.html).

PREDATOR (Frishman and Argos, 1996, 1997) is based on the calculated propensities of every 400 amino acid pairs to interact inside an -helix or one upon three types of -bridges. It then incorporates nonlocal interaction statistics. PREDATOR also uses propensities for -helix, -strand, and coil derived from a nearest-neighbor approach (see below). To use information obtained from homologous proteins, PREDATOR relies on local pairwise alignments. PREDATOR is able to use Clustal alignment as input. The program was employed with the option `-a,' which furnishes a prediction for every sequence of the input set. The source code is kindly distributed by the authors. PREDATOR is also accessible at the URL (http://www.embl-heidelberg.de/cgi/predator_serv.pl).

DSC (King and Sternberg, 1996) combines several explicit parameters in order to produce a "meaningful" prediction. It runs the GORIII algorithm [Gibrat et al. (1987), based on information theory applied to local interactions] on every sequence to provide mean potentials for the three states. In addition, DSC uses the presence of insertions/deletions, the distance from the end of the chain, the moment of conservation, and the moment of hydrophobicity (the two last parameters given an -helical structure and a -strand structure). A linear combination of these different attributes gives an output that is subsequently filtered. The program was used with the following options: `-a' (to turn off removal of poorly aligned sections), `-i' (to stop removal of isolated predictions), `-f1' (to apply the filtering rules once), and `-w' (Clustalw alignment). The source code is kindly distributed by the authors. DSC is also accessible at the URL (http://bonsai.lif.icnet.uk/bmm/dsc/dsc_read_align.html).

NNSSP (Salamov and Solovyev, 1995) is based on the nearest-neighbor algorithm [sometimes improperly called the "homologue" method (Levin et al., 1986; Nishikawa and Ooi, 1986)]. The basic idea of the nearest-neighbor approach is the prediction of the secondary structure state of the central residue of a test segment, based on the secondary structure of similar segments from proteins with known 3D structure. The information provided by the different templates is scored according to their similarity (according to the sequence or other properties) with the test segment. NNSSP is an enhancement of the algorithm designed by Yi and Lander (1993), which selects the neighbors by the mean an environmental score (Bowie et al., 1991) and combine by the mean of a neural network predictions made with different parameters (environmental scores, length of nearest-neighbors ... ). In addition to the latter program, it incorporates information from multiple aligned sequences (by averaging their scores for the weighting of each nearest-neighbor). An executable program was kindly provided by the authors. NNSSP is also accessible at the URL (http://dot.imgen.bcm.tmc.edu:9331/pssprediction/pssp.html).

A program was written in C to convert a Clustal alignment into NNSSP alignments, clu2nnssp. This program is available at the URL (http://www.pasteur.fr/units/neubiomol/softwares.html) or upon request.

Accessibility to solvent and topology program

PHDacc (Rost and Sander, 1994b) is able to compute the probable accessibility to solvent. It was used to refine the secondary structure predictions.

PHDhtm (Rost et al., 1995, 1996) was used to provide a more accurate prediction of the transmembrane segments position, rather than the original one, established with only a few subunits and only from hydropathy plots (Popot and Changeux, 1984).

Conservation index

A computer program, ConsIndex, was written in C to compute the sequence conservation between homologous sequences at each position of a multiple alignment. The program takes as input an alignment of a Clustal-like format and a similarity matrix. It computes first the N(N  1)/2 global similarities S_ij (identities if the identity matrix is input) of the N sequences. Then for each position of the alignment, a conservation index CI is computed as follows:

(4)

where s_ij is the relevant similarity matrix element for the sequences i and j at the considered position. In the present work, the default similarity matrix of the Wisconsin package program GAP (Devereux et al., 1984) was used. It was rescaled from [1.2,1.5] to [0-100]. The gap was added as an independent amino acid, with every matrix element involving it considered as null. ConsIndex is available at the URL (http://www.pasteur.fr/units/neubiomol/softwares.html) or upon request.

	RESULTS

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION CONCLUSIONS REFERENCES

Strategy

Previous works have shown that the accuracy of secondary structure predictions increases from the combination of several independent first- and second-generation algorithms (Biou et al., 1988; Nishikawa and Ooi, 1986; Zhang et al., 1992). Here we combined the prediction of several third-generation algorithms, using the information given by a set of aligned homologous sequences to compute the secondary structure of nAChR subunits.

The algorithms used in this study were chosen according to three criteria: 1) they analyze multiple alignments instead of single protein sequences; 2) they yield a better than 70% accuracy for three-state (H, E, C) prediction when tested on a set of proteins of known structure with sequence identities lower than 25% (Rost and Sander, 1994a) or during blind predictive situations (King, 1996; Rost, 1997); and 3) each of these algorithms is based on a different predictive approach. Each program was applied successively on every sequence of the alignments to increase the signal/noise ratio.

Two different sets of sequences were used to make the secondary structure predictions. The first (AL1) represented the entire group of cationic LGIC subunits in the acetylcholine receptor superfamily (5-HT3 and nicotinic receptors). A second set of sequences (AL2) contained the first set and, in addition, sequences covering the whole group of anionic LGIC subunits (GABA and glycine receptors).

Consistency of the predictions between methods and sequences

The congruencies between methods µ_i,j for every pair of methods are listed in Table 2. The four methods gave all µ_i,j values >67%. The use of the larger set of sequences caused a decrease of µ, which nevertheless remained >57%. The congruence between sequence consensus predictions _x,y was also examined for every pair of sequences. The predictions for the cationic LGIC subunits were found consistent, the congruencies varying from _deg3,1 = 80.8% to _3,6 = 95.3%. In the larger set, the lowest  occurred just above 64%, a value much larger than random (which is 33% for a nonbiased three-state comparison and 38% if the present bias of PDB is taken into consideration). The good congruency of the different predictions for the various members of the nAChR family is illustrated in Fig. 2 (top), where the peaks are sharp and 17 of 25 final structural elements are predicted in >90% of the cases. The sequence consensus predictions were very similar. The positions of the secondary structure were almost identical, with little variation of the assignments. The method consensus predictions were more variable, though similar. The assignment of the structures varied somewhat, as well as (but only very rarely) their occurrence.

View larger version (81K): [in this window] [in a new window]   FIGURE 2   Top: relative incidence of the three-state populations per residue position. For each residue position of AL1, (5) the sum of all the state predictions for all methods and all sequences, is represented. The diagram is cumulative, i.e., the difference between the height of the peaks at a given position is informative, not the height of the peaks itself. Bottom: conservation index per residue position. For each residue position of an alignment of 152 LGIC subunit sequences, a conservation index was computed by ConsIndex. The black squares under the graph represent the final predicted structures (-helix or -strand) in the mature subunit. Note that if an -helix is adjacent to a -strand, there is only one black square.

The resemblance of protein 3D structures (rmsd) is proportional to their sequence identity (Chothia and Lesk, 1986; Flores et al., 1991). The incorporation of distant sequence information is expected to increase the reliability of predicted structures, although decreasing the consistency of the overall prediction (Russel and Barton, 1993; Sternberg, 1996). The  values of AL2 were plotted against the global amino acid similarities determined by the conservation index program. A correlation unambiguously occurred between the sequence similarities and the structure prediction similarities (Fig. 3) (n = 703, r = 0.882, p < 0.001). Two main components emerged from the comparisons: a lower similarity population representing the comparisons of anionic/cationic (e.g., GABA_A vs. nAChR), with a higher similarity population representing the comparisons of anionic/anionic or cationic/cationic LGIC subunits. Together, these data show that the variations between the secondary structure predictions therefore were not random, as expected from algorithm imperfections. On the contrary, they relied on the variation of sequence. This reflects the fact that if the core structures are conserved between the different superfamily members, as supported by a large body of experimental evidence (Galzi and Changeux, 1994), then the structural assignation at the level of individual residue may vary (for instance at the extremities of the structures). This variation was indeed found to increase when the sequence relationship decreases. Another conclusion can be derived from Fig. 3. The more distant to our target protein are the homologs used to infer secondary structure, the less reliable is the information obtained. A trade-off is reached between the information obtained from multiple alignments (reliability of secondary structure position and assignment) and the mispredictions at the level of individual residues due to sequence divergence (Russel and Barton, 1993). There is no known method published up to now to establish the best compromise.

View larger version (20K): [in this window] [in a new window]   FIGURE 3   Correlation between sequence similarities and secondary structure prediction identities. For N sequences, there are (N  1)(N  2)/2 amino acid similarities. Here are represented the 703 dots corresponding to the 38 sequences of AL1. The amino acid similarities Sij have been computed by the program ConsIndex. They are plotted against the secondary structure identities ij, as described in the Methods section. Each point is therefore the comparison of one amino acid similarity versus 16 predictions (four prediction methods for each sequence). The bivariate regression analysis shows that the correlation is meaningful (n = 703, r = 0.882, p < 0.001). Note the two components: the bottom left concentration represents the comparisons of anionic/cationic (e.g., GABAA vs. nAChR), whereas the less dense upper right distribution represents the comparisons of anionic/anionic or cationic/cationic.

The final results obtained with the two alignments AL1 and AL2 were very similar, with only a few residues predicted to be in a different state. Every structure except one was equally predicted with both sets, and in all these cases the secondary structure assignment remained the same. Therefore, except when otherwise stated, we present the results obtained with AL1 (see Fig. 4).

View larger version (148K): [in this window] [in a new window]   FIGURE 4   Alignment of the cationic channel subunits (AL1). The first line gives the limit of the refined secondary structures and their respective designation (letters for -helix and numbers for -strand). The alignment, generated by ClustalX, follows. The colors are set according to the amino acid or the consensus at each position. For instance, a proline is always yellow, whereas a cysteine is purple if it is the consensus residue or can be blue if the consensus is a hydrophobic residue. Below the alignment and the ruler, the raw consensus secondary structure prediction is presented. The magenta and green boxes represent the refined predictions. Below the secondary structure prediction is the predicted accessibility to the solvent (b: buried, e: exposed). On the last line is reported the topology inferred from SSPCA output and experimental considerations. On this line are also reported the affinity labeling results (A-F represent the respective binding site segments) and glycosylation (N) natural or induced.

Raw secondary structure prediction

The proportions of the three-state populations in the entire set of predictions for each alignment position are presented in Fig. 2 (top). Fig. 4 shows the raw consensus prediction, in plain text just below the alignments, with the designation of the structure above. In Fig. 2 (bottom), the conservation index determined on the full superfamily of LGIC (152 subunits) is plotted, along with the predicted secondary structures (black squares below the graph). In all instances but three (E₉, H_F, and H_G) the predicted structures were located in regions of high (>50%) conservation. The region of E₉ is in fact highly conserved except for the nematode subunit unc38. The region of H_F and H_G are highly conserved in cationic channel subunits but less in anionic channels. In summary, within the cationic channel subunit family, all predicted structures were located in regions of high conservation. This fact is important since the high sequence variation between family members is likely to occur in less well-structured regions. A structure predicted in a conserved region is therefore more likely to be accurate.

Refined secondary structure predictions

PHD, DSC, and NNSSP provide prediction probabilities for the three states in addition to the predicted state. Combining these probabilities permits the correction of the threshold decisions at the level of single predictions, which may lead to false assignment, and offers the possibility of resolving some single-residue problems, such as singletons (isolated structured residue) or amino acid located at the borderline of the secondary structure motifs. The changes made in this way affect only 29 residues. The resulting refined prediction contains (without the signal peptide) nine -helices (mean length 13.9 amino acids) designated H_A to H_H, and 17 -strands, designated E₁ to E₁₇ (mean length 6.6 amino acids). Their positions and lengths are summarized in Table 3. Except for two large helices surrounding a large -strand at the amino-terminal extremity, the extracellular portion of the subunits was predicted to occur as an all- structure, formed by successive short strands.

The structure of the carboxy-terminal portion of H_A is consistent with solvent accessibility patterns (described in the following by strings of `e' for exposed and `b' for buried) i.e., "bbeebbee," its amino-terminal portion being completely exposed.

The structure at the center of E₁ is also consistent with solvent accessibility patterns "bebebe"; its two extremities being predicted as completely buried. Its carboxy-terminal portion is less consistent, since it is predicted in an -helical state in every AL1 sequence consensus (see Fig. 2, top). An -helical structure for the last four residues could be envisioned. Indeed, for AL1 these -helical residues were predicted in every sequence consensus and in three of four methods' consensus (only PHD predicted all the residues under -strand state). However, for AL2, only the PREDATOR consensus, nicotinic 8 consensus, and nicotinic  consensus presented some residues predicted as -helical. This unique -helical turn could be a specific feature of the cationic channel subunits, since it does not appear in the AL2 sequence consensus, where the extended structure is consistently predicted.

The main immunogenic region (MIR) is located from the end of H_B to the beginning of E₃ (Tzartos et al., 1990). This segment was already known to be exposed to the solvent since it is directly involved in numerous forms of the autoimmune disease myasthenia gravis (Tzartos et al., 1990). Accordingly, its central portion is predicted as totally accessible to the solvent.

The assignment of H_B appeared consistent with all predictions except those of DSC for AL1 and AL2 as well as PHD on AL2 (only some residue predicted under -strand state). The solvent accessibility pattern is more consistent with a -strand in the carboxy-terminal portion. However, the glycosylation at AL1₁₀₂ (as well as at AL1₁₉₈ and AL1₂₄₂) implies that these residues are exposed to the exterior of the receptor. Since the residue AL1₁₀₄ is labeled by tubocurarine and presumably faces the binding site, a -strand might cause steric hindrance between the ligand and the sugar, whereas an -helix would place the side chain of the two residues in opposite directions.

The structure of E₂ (three residues long) is not predicted by the analysis of AL2. It is the only structural element that differs between the two analyses. However, the assignment of E₃ is contradicted by a cross-linking experiment (Watty et al., 1998) showing that its two first residues should expose their side chain in the same direction.

E₄ is predicted to be completely buried. E₅ and E₈ are consistent with solvent accessibility "ebebebeb."

The predictions of the E_12-15 and H_C-E are probably less accurate than the extramembranous portions. Indeed, the secondary-structure prediction programs were not designed or tested with membrane proteins (see Discussion). The length of the predicted secondary structures varied considerably according to the set of sequences used. With AL2, H_D is shorter (in MII), H_E is longer, and E₁₅ shorter (in MIII). Finally, H_F and H_G are fully consistent with the solvent accessibility predictions "bbebbbebbbebbbebb" and "eebeebbebbebbbeeb," implying one face exposed, the other buried.

2D representation of the amino-terminal domain

Data may be added to the 1D structural assignments given by SSPCA. This defines an envelope of structural constraints (Fig. 5), which permits proposal of a 2D folding of the peptide chain. No data concerning the tertiary folding are included, since no - interactions are known.

View larger version (57K): [in this window] [in a new window]   FIGURE 5   2D folding of a typical subunit. The perspective is from inside the pore, perpendicular to the membrane. The -helices and the length of -strand are on the same scale. The length of the nonstructured part is approximately scaled. The gray double-arrow means that the two parts of the extracellular moiety have to be bent together. The blue segments represent the positions of the stretches (of length >4) of residues exposed to the solvent. The yellow link represents the disulfide bond. MIR: main immunogenic region. The amino acids identified by affinity labeling are noted (the numbering is that of mature chick 7). Note that this fold is a 2D representation, and has nothing in common with a 3D model. Indeed, the -strands are placed totally parallel and in a progressive order only for reasons of convenience. Note that the figure does not in any way imply any specific - interaction. The positions of the strands are chosen arbitrarily, in the order of the primary sequence.

First, on the basis of electron microscopy images, we may locate the MIR at the distal end of the receptor, respectively to the membrane (Beroukhim and Unwin, 1995). As a consequence, E₂ and E₃ are also placed at the top of the fold. E₁₁ is likely to be close to the membrane since it is adjacent to MI (see below for the position of the transmembrane domains). Then, we may assume that each stretch of at least four residues predicted to be exposed to the solvent forms a loop at the surface of the subunit. This constraint implies bending of the 1D structure between E₅ and E₆, E₈ and E₉, E₉ and E₁₀, and E₁₀ and E₁₁. The beginning of E₇ and E₈ are linked by a disulfide bond, and are thus in close proximity. This disulfide bridge forces a new bend between E₇ and E₈. This so-called "Cys-loop" is the most conserved part of the LGIC subunit amino-terminal domains. Although half of it is not predicted to be folded into a periodic structure, we may reasonably hypothesize that the entire region adopts a strongly constrained conformation. Finally, a bend is introduced between E₃ and E₅ to respect the observed size of the subunit, which protrudes 60 Å from the membrane, with a diameter of ~40 Å.

Each subunit can then be artificially subdivided into two domains. One is formed by H_A, E₁, and H_B, the other by E_2-11. On the basis of the cryoelectron microscopy images of Unwin (1993b), H_A and H_B have been disposed perpendicularly to the membrane.

This representation is fully compatible with the body of experimental data concerning the nicotinic binding site. Indeed, affinity labeling and site-directed mutagenesis led to the identification of amino acids (see Table 1) that are distributed at the interface of the subunits on six different elements, referred to as A (7W85 and 7Y92), B (7W148 and 7Y150), and C (7Y187, 7C189, 7C190, and 7Y194) for the "principal" component; and D (7W54), E (7L108, 7N110, 7Q116, and 7L118), and F (7D163 and 7E172) for the "complementary component" (note that all residues are quoted according to the mature chick 7 subunit; please see the alignment for conversion. It does not mean that these residues have been identified only in or also in this subunit). Another residue has recently been identified on the complementary component (7T34). Since it is located in E₁, its position does not add further constraints on the 2D representation, though it possibly constrains the tertiary folding.

Affinity labeling experiments with toxin derivatives assigned the principal and complementary binding components to be carried by the clockwise and counterclockwise sides of the subunits, respectively, when the receptor is seen from the extracellular compartment (Machold et al., 1995).

The H_AE₁H_B part must be folded onto the E_2-11 sheet in order to form a compact structure, contained in a 40 × 60 Å² surface, and to account for the possible contribution of residues homologous to mouse K34 to the active site. Yet, only a few data constrain the folding of the H_AE₁H_B domain.

The transmembrane and cytoplasmic domains

We used PHDhtm (Rost et al., 1995, 1996) to investigate the organization of membrane spanning segments. PHDhtm is the only program that did not predict the signal peptide as transmembrane domain, probably because of its lack of conservation. In addition, it predicted the four transmembrane domains for each LGIC members. SSPCA provided the consensus of output from PHDhtm applied to all the sequences of AL1. The results, compiled in Table 5, yielded four transmembrane segments. The length of the consensus segments are 18 for MI, 17 for MII, 19 for MIII, and 17 for MIV. For comparison, four other programs were also used on 1 and 7. All of them predicted the four transmembrane segments of AL1 sequences, although in some cases other parts of the subunit were incorrectly predicted as crossing the membrane.

View this table: [in this window] [in a new window]   TABLE 4   Experimental and predicted -helix and -strand content in an entire subunit

The length of the consensus transmembrane segments are smaller than depicted in the usual proposals (Popot and Changeux, 1984). However, this could be an artifact due to conservative prediction of PHDhtm. For comparison, four other programs were also used on 1 and 7. The results vary according to the method but also according to the sequences used. This fact supports the importance of using consensus of multiple analyses.

SSPCA predicts each of the transmembrane segments to fold in a mixed  helix/ strand fashion, with almost no coiled structures. Fig. 5 shows an attempt to represent the transmembrane portion in 2D. Yet, since the present study gives no information about the precise orientation of the structures in the membrane, the represented angles are arbitrary, except in the case of the helix present in the MII segment shown to be orientated rather perpendicular to the membrane. In addition, the length of the predicted structures in the membrane is poorly accurate.

Except for H_F and H_G, the cytoplasmic domain is predicted as totally accessible to solvent and in a nonperiodic structure. The solvent accessibility pattern of the two helices suggests that they possess one face buried and another exposed to the solvent (see Discussion).

	DISCUSSION

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION CONCLUSIONS REFERENCES

Previous works have shown that the accuracy of secondary structure predictions increases with the combination of several independent algorithms (Biou et al., 1988; Nishikawa and Ooi, 1986; Zhang et al., 1992). As reported here, in order to determine the best available prediction of the nAChR subunits' secondary structure, we integrated the results of several third-generation programs, using the information from a set of aligned homologous sequences. These programs were selected on the basis of their recognized efficiency on test sets of proteins with known secondary structure (Rost and Sander, 1994a) or during blind predictive situations (King, 1996; Rost, 1997). Moreover, each program was applied on every sequence of the alignments in order to increase the signal/noise ratio.

Two main ab initio predictions have been reported for nAChRs in the past two decades. Finer-Moore and Stroud (1984) used the algorithm of Garnier et al. (1978) for the extramembranous regions and an analysis (by Fourier transformations) of hydrophobicity periodicity for the putative transmembrane regions. Recently, Ortells (1997) presented a secondary structure prediction based on a Chou and Fasman-like algorithm (Chou and Fasman, 1978). The main difference between the initial method and the one used by Ortells resides in the definition of the secondary structure initiators. Instead of being predicted solely by the sequence (via statistical tables) as in the Chou and Fasman algorithm, these initiators were determined as follows: an initiator was defined as a residue that is constantly predicted in the same state, across different sets of LGIC subunit sequences, analyzed by first- and second-generation algorithms. Another difference resides in the fact that the propagation from the initiators was unidirectional (in Ortells, 1997), from the amino-terminal to the carboxy-terminal, while it is bi-directional in Chou and Fasman (1978). The expected prediction accuracy has already been discussed elsewhere (see Kabsch and Sander, 1983; and Nishikawa, 1983 for initial assessment, and Rost and Sander, 1994a; 1996 for recent reviews), but the difference of expected accuracy between these pioneering works and our own may reach 20%. On an identical test set, Chou and Fasman reached 49% in Q₃, whereas PHD2 reached 72.5% (Rost and Sander, 1994a).

-Helix and -strand contents

The helix and strand content of the entire nAChR was measured by several groups using different spectroscopic measurement methods (Butler and McNamee, 1993; Méthot et al., 1994; Moore et al., 1974; Yager et al., 1984). The results showed a high variability, which cannot be due solely to the differences of receptor environment. Indeed, inferred helix content varied from 18.7% (Butler and McNamee, 1993) to 48% (West et al., 1997), inferred strand content (without -turn) varied from 26% (West et al., 1997) to 42% (Butler and McNamee, 1993), and the calculated helix/strand ratio varied from 0.45 (Butler and McNamee, 1993) to 1.85 (West et al., 1997), with Méthot et al. (1994) and Yager et al. (1984) finding intermediate values of 1.11 and 1.18) (see Table 4). The corrected SSPCA consensus yielded slightly lower values of helix and strand contents, although the ratio is consistent with the mean of experimentally determined ratios. In the amino-terminal portion (as defined by West et al., 1997 and not by our transmembrane segment determination), our consensus gives an equivalent predicted helix content (13.7% vs. 12%) and less strand (31.7% vs. 51%) compared to the experimentally observed one in the unique study of West et al. (1997).

Comparison with other predictions of the amino-terminal domain

At the level of the extracellular amino-terminal domain, all approaches predicted a structure mainly folded in -strand (Fig. 6 A). However, the position of the structures, as well as their number, differ considerably among the different studies. The high  content is also consistent with the cryoelectron microscopy images (although three helices were proposed in these latter investigations; Unwin, 1993b, 1996). The structures predicted by Ortells (1997) are longer than those presented in the present work and longer than the value observed in the PDB. Notably, the two large -helices predicted in the amino-terminal half of the extracellular portion are 20 aa long, whereas we predict 12 and 14 aa and the PDB average is 9. Also, the mean length of -strand is 7.2 in Ortells (1997), 5.8 in the present work, and 5.1 in the PDB. These discrepancies are likely due to the method Ortells used to propagate the structural elements. When initiated, each element is extended forward until a different initiator or a proline or a glycine is reached.

View larger version (29K): [in this window] [in a new window]   FIGURE 6   (A) Comparison of our secondary structure prediction with those of Gready et al. (1997), Tsigelny et al. (1997), and Ortells (1997). All the predictions are projected on the amino-terminal part of Torpedo 1. Numbering of the residues is that of mature Torpedo 1 and that of AL1. (B) Comparison of our secondary structure prediction with that of Ortells and Lunt (1996) based on analogy with enterotoxin. The predictions are projected on the transmembrane part of Gallus 7. Numbering of the residues is that of mature Gallus 7 and that of AL1.

Fig. 5 A also provides a comparison with the secondary structures derived from threading methods (Gready et al., 1997; Tsigelny et al., 1997). In this case, not only the lengths of the motifs, but also their positions in the sequence are very different.

Comparison with other predictions of the membrane-spanning segments

The location of the four putative transmembrane segments was originally performed by hydropathy plot analysis. This method, though of great interest and easy to use, does not apply satisfactorily in the case of membrane channels. Indeed, the residues lining the pore in the open state are not anticipated to be hydrophobic. Moreover, in a protein with multiple membrane crossings, such as the nAChR, the internal transmembrane segments may be isolated from the lipid environment. In addition, some hydrophobic stretches can be external to the membrane (in close proximity to, or embedded in, the core protein). As a consequence, some transmembrane segments were not correctly predicted. For instance, for the rat glycine 1 subunit, the program SOSUI (Hirokawa et al., 1998), based on amino acid physical properties, did not predict the MII and MIV segments as transmembrane units, nor did the program TMpred (Hofman and Stoffel, 1993), based on the comparison with a database of known transmembrane segments.

The original predictions vary from one author to another (Fig. 7). The membrane-spanning position is set by PHDhtm with 95% accuracy. Such a precision is superior to the original variations.

View larger version (11K): [in this window] [in a new window]   FIGURE 7   Comparison of the transmembrane segments as originally suggested by Claudio et al. (1983) (actual analysis is that of Torpedo ), Devillers-Thiéry et al. (1983), Noda et al. (1983), and Nef et al. (1988) (actual analysis is that of chick  and ).

Some structural prediction have already been made for the transmembrane domain based on analogy arguments. Unwin (1993b) suggested, on the basis of his images from electron microscopy, that the transmembrane region of the nAChR could have a folding similar to that of some pentameric enterotoxin domains. Ortells and Lunt (1996) further exploited this idea to model part of the LGIC transmembrane region based on the crystallographic structure of the Escherichia coli heat-labile enterotoxin domain B (Sixma et al., 1993). The resulting model presents a mixed / secondary structure, where MII is all-, MI is all-, and MIII is / (Fig. 6 B), the MIII -helical region being a posteriori added to the template. Several remarks can be make about this study apart from the fact that the template was never found, up to now, by any threading algorithm. First of all, the enterotoxin is not an integral membrane protein, and thus may not be an adequate template for the nAChR transmembrane domain. Ortells and Lunt (1996) removed the first strand that interacts with the fifth. The resulting template might then be less stable, one sheet being composed of only two antiparallel strands. As stated by the authors, the further addition of MIII and MIV, modeled as helices (partially for MIII) may result in a segregation of the enterotoxin-modeled moiety from the lipids. The secondary structure predictions presented here do not agree with those proposed by Ortells and Lunt (1996; Fig. 6 B). A three-state comparison between this study and the present prediction gives only 33% of identical residues.

The nicotinic ligand-binding site

The 2D representation accounts for the basic information concerning the ligand binding site for ACh and competitive antagonists. Secondary structure predictions suggest that whereas binding elements B, F, and C are carried by segments without regular structural patterns, binding elements A, E, and D are at least partially carried by structured segments.

At the level of the complementary component, the affinity-labeled W57 (AL1₁₀₄) is located near the center of helix B. Mutations at position AL1₁₀₆ have been shown to modulate agonist and antagonist pharmacology (Chiara and Cohen, 1997; Corringer et al., 1995; Harvey and Luetje, 1996). The side chains of residues AL1₁₀₂ and AL1₁₀₄, however, point outward from opposite faces of the helix, implying that AL1₁₀₄ mediates its effects indirectly, possibly through local alteration of the structure. At the level of element E, two successive -strands are predicted, E₅ carrying the identified mouse S111 (AL1₁₆₁) (Sine et al., 1995), and E₆ carrying mouse Y117 (AL1₁₇₅). One possibility could be that these -strands interact in an antiparallel -sheet, which would direct the side chains of these binding residues in the same direction and in close proximity. Finally, the segment carrying element F has been shown to contain the calcium binding site involved in agonist potentiation. The predicted arrangement of this region without a regular structure is consistent with the notion that the segment AL1₂₃₄-AL1₂₄₅ folds into a specific pocket that constitutes a calcium binding site, as observed for the corresponding synthetic peptide (Galzi et al., 1996).

At the level of the principal component, mutagenesis experiments have shown that several mutations, located at the vicinity of labeled residues from elements B and C, profoundly altered the pharmacological properties (regions AL1₂₁₀-AL1₂₁₄ and AL1₂₅₆-AL1₂₅₉) (Corringer et al., 1998). Since the entire corresponding regions are predicted to lack a regular secondary structure, they may fold into loops, such that the mutations could possibly alter agonist binding indirectly, through structural reorganization of these putatively flexible segments.

Transmembrane segments as an / structure

Each transmembrane segment of the receptor is predicted to fold in a mixed / structure. This prediction should be taken with extreme caution, since, as noted above, the programs used were not designed to work on membrane proteins. Prediction methods based on analyses of globular proteins could incorrectly predict strands in helical transmembrane regions.

Direct transitions are seen at the end of MI, MIII, and MIV. Such transitions are impossible following a helix of more than four residues. Due to the low reliability of the predictions in these regions, a small hinge could in fact link the -helices and the following -strands.

Also, affinity labeling experiments with a radioactive hydrophobic probe support an organization of the MIII and MIV transmembrane segments in -helix (Blanton and Cohen, 1994). MIII was predicted to be -helical until AL1₃₆₂ (7F283), while the H_E was predicted to reach only AL1₃₅₅ (7S276), and MIV was predicted to be -helical until AL1₆₆₈ (7I463), whereas H_H is predicted to reach only AL1₆₅₇ (7F452).

At the level of the MII segment, known to face the lumen of the ion channel, our predictions could lead to a reconsideration of the currently accepted architecture of the ion pathway. MII is predicted here to start at the level of amino acid AL1₃₂₃, four residues after the standard model. In addition, the MII helix is predicted to be slightly shorter. Much of the data coming from affinity labeling and site-directed mutagenesis experiments are readily represented by a helical structure (Akabas et al., 1994; Revah et al., 1990). However, recent results (Wilson and Karlin, 1998) support an elongated strand for the short segment-spanning residues AL1₃₁₀ (7S335) to AL1₃₁₉ (7S240). Moreover, it is thought that MI and MII are in close proximity (Akabas and Karlin, 1995). Consequently, the cytoplasmic portion linking MI and MII is predicted to be longer, and could fold into a -hairpin (E₁₃-E₁₄), the length of loop linking the strands being variable according to the subunit. Recent mutagenesis experiments from this laboratory point to a major contribution of the center of this cytoplasmic portion to the selectivity filter of the ion channel. Furthermore, it was found that its conformation, rather than its precise amino acid sequence, had a critical effect on the selectivity properties of the ion channel (Corringer et al., 1999). This large cytoplasmic region could thus fold in such a way that some carbonyl of the peptide backbone would be exposed in the correct geometry for dehydration of specific ions, as observed in the case of a bacterial potassium channel (Doyle et al., 1998).

The cytoplasmic portion and the oligomerization

H_F and H_G are predicted to be amphipathic, with one face exposed to the solvent and the other buried. The maximum hydrophobic moment (as determined with the program MOMENT of the Wisconsin Package (Devereux et al., 1984) with a window of eight residues) is 0.19 for H_F (low) and 0.57 for H_G (high). In addition, both helices present a leucine-zipper signature (on 79 sequences: at AL1₃₉₃, 61L; AL1₄₀₀, 62L, 14M; AL1₆₁₁, 30L, 6M; AL1₆₁₈, only 2L, but 21I and a conserved hydrophobic position in AL1₆₁₅; AL1₆₂₅, 17L, 30M, and a conserved hydrophobic position in AL1₆₂₂). These two cytoplasmic helices could interact in a coiled-coil arrangement, within the subunit or even between subunits. This motif could be critical for the oligomerization process. Indeed, Yu and Hall (1994) have demonstrated that two deletions of amino acids belonging to H_F and H_G imply their intervention in the formation of the pentamer.