Title

Transmembrane helices predicted at 95% accuracy

Burkhard Rost ¹, Rita Casadio ², Piero Fariselli ² and Chris Sander ³

1 Corresponding author: then: EMBL
now: Dept. Biochemistry & Mol. Biophys., Columbia Univ., 650 West 168^th Street BB217, New York, NY 10032, USA, rost@columbia.edu, http://cubic.bioc.columbia.edu/

2 then and now: Bologna University

3 then: EMBL; now: MIT

QUOTE: Protein Science 1995 Vol. 4, pp. 521-533

Abstract

We describe a neural network system that predicts the locations of transmembrane helices in integral membrane proteins. By using evolutionary information as input to the network system, the method significantly improved on a previously published neural network prediction method that had been based on single sequence information. The input data was derived from multiple alignments for each position in a window of 13 adjacent residues: amino acid frequency, conservation weights, number of insertions and deletions, and position of the window with respect to the ends of the protein chain. Additional input was the amino acid composition and length of the whole protein. A rigorous cross-validation test on 69 proteins with experimentally determined locations of transmembrane segments yielded an overall two-state per-residue accuracy of 95%. About 94% of all segments were predicted correctly. When applied to known globular proteins as a negative control, the network system incorrectly predicted fewer than 5% of globular proteins as having transmembrane helices. The method was applied to all 269 open reading frames from the complete yeast VIII chromosome. For 59 of these at least two transmembrane helices were predicted. Thus, the prediction is that about one fourth of all proteins from yeast VIII contain one transmembrane helix, and some 20% more than one.

Introduction

Given the rapid advance of large scale gene-sequencing projects (Oliver et al., 1992, Johnston et al., 1994), most protein sequences of key organisms will be known in about five years' time. Experimental structure determination is becoming more of a routine (Lattman, 1994); and the number of proteins with known sequence for which the 3D structure can be predicted rather accurately by homology modelling is constantly increasing (today more than 25% of all sequences in the SWISSPROT sequence data base (Bairoch & Boeckmann, 1994) can be modelled with reasonable accuracy by homology (Sander & Schneider, 1994)). Even in such an optimistic scenario, experimental knowledge about membrane proteins is likely to be sparse. However, membrane proteins represent a very important class of protein structures. To what extent can structural aspects for membrane proteins be predicted from sequence information?

Two types of membrane proteins. So far, the 3D structures of two types of membrane proteins have been determined, so far. The first type are helical proteins: photosynthetic reaction centre (Deisenhofer et al., 1985), bacteriorhodopsin (Henderson et al., 1990) and the light harvesting complex II (Wang et al., 1993, Kühlbrandt et al., 1994); these proteins consist of typically apolar helices of some 20 residues that traverse the membrane perpendicular to its surface (Fig. 1). The second type is represented by the structure of porin (Weiss & Schulz, 1992, Cowan & Rosenbusch, 1994), a 16-stranded b-barrel.

Membrane proteins easier to predict than globular ones. Typical methods for the prediction of transmembrane segments focus on helical transmembrane proteins (von Heijne, 1981, Argos et al., 1982, Eisenberg et al., 1984a, Engelman et al., 1986, von Heijne, 1986, von Heijne & Gavel, 1988). It is commonly believed that the prediction of structure is simpler for membrane proteins than for globular ones as the lipid bilayer imposes strong constraints on the degrees of freedom of structure (Taylor et al., 1994).

Fig. 1: Prediction of the location of transmembrane helices. In one class of membrane proteins, typically apolar helical segments are embedded in the lipid bilayer oriented perpendicular to the surface of the membrane. The helical segments can be regarded as more or less rigid cylinders. Thus, the 3D structure of the membrane spanning protein region can be determined by: the location of segments with respect to sequence; the orientation of helical axes; the inclination of helical axes with respect to lipid bilayer; and the phase of helices with respect to each other (orientation of helical wheel). Here, we simplify extremely by projecting 3D structure onto a 1D string describing which residues of the protein are part of a transmembrane helices. Input to the prediction tool (neural network system) is a protein sequence (in general a sequence alignment), output is a prediction of the location of transmembrane segments.

The example shown (sequence of cytochrome O ubiquinol oxidase subunit I, cyob_eco in SWISSPROT , Bairoch & Boeckmann, 1994) contained one of the few segments that were under-predicted (missed). The numbers give the reliability of the prediction for each residue on a scale of 0 to 9 (Fig. 2). Non-transmembrane regions, when predicted correctly, usually reached the highest reliability (9). Thus, the unusually low reliability values for the under-predicted segment might have enabled the expert user to improve the automatic prediction by interpreting this region as non-loop.

Prediction of transmembrane segments. Methods for prediction of transmembrane helices are usually based on (i) hydrophobicity analyses (Argos et al., 1982, Kyte & Doolittle, 1982, Engelman et al., 1986, Cornette et al., 1987, Degli Esposti et al., 1990); (ii) the preponderance of positively charged residues on the cytoplasmic side of the transmembrane segment (interior), established as the "positive inside rule" (von Heijne, 1981, von Heijne, 1986, von Heijne & Gavel, 1988, von Heijne, 1991, von Heijne, 1992, Sipos & von Heijne, 1993); or (iii) statistical procedures which perform significantly better when combined with multiple alignments (Persson & Argos, 1994). In general, prediction of transmembrane segments is relatively straightforward. But, can detailed aspects of 3D structure be predicted from sequence for helical transmembrane proteins?

Prediction of 3D structure for helical transmembrane proteins. Cytoplasmic and extra-cytoplasmic regions have different amino acid compositions (von Heijne & Gavel, 1988, Nakashima & Nishikawa, 1992). This difference allows for a successful prediction of not only the location of helices but, as well, of their orientation with respect to the cell (pointing inside or outside the cell) (Landolt-Marticorena et al., 1992, Sipos & von Heijne, 1993, Jones et al., 1994). Going further, Taylor and colleagues enumerate all possible models for packing seven-helix transmembrane proteins and select the "better models" (Taylor et al., 1994). The selection criterion for "better models'' is the crucial point of the method. The authors report that the native conformation is found in "most cases" tested. However, the N-, and C-terminal ends of the transmembrane helices have to be predicted very accurately for a successful automatic prediction of 3D structure from sequence (Taylor et al., 1994).

Fig. 2: Two level system of neural networks for helical transmembrane prediction. For each position in the alignment the amino-acid frequencies were compiled, the numbers of insertions and deletions counted, and a conservation weight computed (defined in Rost & Sander, 1993b). Furthermore, 'global information' (beyond the window of 13 adjacent residues) about the search sequence was compiled: amino acid composition, length, and the position of the current window with respect to the N-, and C-terminal end of the protein. All this information was fed into the neural network input for w = 13 adjacent residues (shown w = 7 ). The input layer was fully connected to a layer with three hidden units, and from there to the two output units coding for the central residues in the window (here 'LII') to be in a helical transmembrane region (HTM ) or not. The output of the first level was fed into a second level of structure-to-structure network, which additionally used the global information and the conservation weight as input. For this network 15 hidden units were used. The two output units code again for the secondary structure state of the central residues (here 'LII').

Abbreviations for 1st level input units: the local information is coded by w ¥ (21+3) units, 20 for each amino acid, one for a spacer (for allowing windows to extent beyond protein ends, such that the first and last w-1 residues in a protein can be used as central residue), and three for conservation weights, numbers of insertions and numbers of deletions. The global information is coded by 32 additional units; 20 for the frequency of each amino acid in the protein, four for the length of the protein, and four for the distance of the central residue to the N- and four for the distance to the C-term of the protein.

Abbreviations for 2nd level input units: the local information is coded by w ¥ (3+1) units, two for each output unit of the first level (HTM , not HTM ), one for a spacer, and one for the conservation weight of that residue. The global information is used as in the 1st level input.

Fig. 3: Generating multiple alignments for the network input. First, for each protein the SWISSPROT data base of protein sequences (Bairoch & Boeckmann, 1994) was searched for putative homologues with a fast alignment method (FASTA , Pearson & Lipman, 1988, Pearson & Miller , 1992). Second, the list of putative homologues was re-examined with a more sensitive profile-based multiple alignment method (MaxHom, Sander & Schneider, 1991). Third, a length dependent cut-off for the sequence identity between the search sequence and the aligned ones was applied to distinguish correct hits for homologues from false positives (for more than 80 residues aligned, the cut-off was chosen 25%+5%; where the '+5%' reflects a safety margin above the line observed to separate correct and false homologues (Sander & Schneider, 1991)). Fourth, a window of 13 adjacent residues was shifted along the protein sequence. Each such window constituted one training or testing example for the neural network.

Table 1 Prediction accuracy cross-validated on helical transmembrane proteins.

Abbreviations for performance accuracy: N , number of proteins used for prediction; Q₂ , percentage of correctly predicted residues; Info , information or entropy of prediction (Rost & Sander, 1993b); Q_TM, accuracy of predicting transmembrane helices (HTM); %obs Q_TM, correctly predicted residues in HTM as percentage of residues observed in HTM; %prd Q_TM , correctly predicted residues in HTM as percentage of residues predicted as HTM; Corr , Matthew correlation (Matthews, 1975) for residues in HTM; <L> , average length of predicted HTM (the observed average is <L> = 22); %obs Sov , segment overlap for HTM computed as percentage of observed segments (Rost et al., 1994); %prd Sov , segment overlap for HTM computed as percentage of predicted segments (Rost et al., 1994); Nseg over , number of segments predicted but not observed as HTM; Nseg under , number of segments observed but not predicted as HTM.

Abbreviations for methods: no profiles , two-level network system using single sequences as input (Casadio et al., 1994); PHDhtm , three-level network system + filter using all information from multiple alignments as input (Fig. 2); Jones+,1994 = (Jones et al., 1994); Edelman, 1993 = (Edelman, 1993); Persson+,1994 = (Persson & Argos, 1994).

Data sets used: set 1: set of 69 proteins with experimentally well determined transmembrane helices (see Methods); set 2: set of 37 transmembrane proteins used by (Edelman, 1993); set 3: set 1 without glra_rat and 2mlt; set 4: set of 28 transmembrane proteins used by (Persson & Argos, 1994).

‡ : discrepancy in assigning transmembrane helices for atpi_pea; both methods compared predict five transmembrane helices, in SWISSPROT only four are annotated, thus we initially counted our prediction as wrong, whereas Persson & Argos (Persson & Argos, 1994) based their evaluation on the hypothesis that the protein contains five and not four transmembrane helices.

Note 1, Nseg : whenever predicted and observed segments overlapped by at least three residues the segment was counted as correct (Rost et al., 1993, Rost et al., 1994). A similar measure seems to have been used by others. A more reasonable score is the segment overlap Sov (Rost et al., 1994).

Note 2, cross-validation: all results except for those in the last row were based on cross-validation tests. Persson & Argos (Persson & Argos, 1994) report that for their method the results with or without cross-validation analysis are similar and only give the 'non-cross-validated results on proteins in their training set.

Can the accuracy of predicting not just the location of transmembrane helices but, as well, of the N-, and C-terminal ends be improved?

Better prediction of transmembrane helix location. Prediction accuracy has recently been improved significantly (Sipos & von Heijne, 1993, Jones et al., 1994, Persson & Argos, 1994). A system of neural networks using single sequences as input (Fariselli et al., 1993, Casadio et al., 1994) appears to be slightly inferior to these methods. However, using information from multiple sequence alignments as input, neural networks have been shown to yield the most accurate prediction of secondary structure for globular proteins (Rost & Sander, 1993c, Rost & Sander, 1993a, Rost & Sander, 1994a). Here, we used a similar system of neural networks to predict transmembrane helices based on evolutionary information (Fig. 1, Fig. 2). The goal was to predict the location of transmembrane helices (defined as helix caps given in Swissprot (Bairoch & Boeckmann, 1994)) more accurately than alternative methods (Sipos & von Heijne, 1993, Casadio et al., 1994, Jones et al., 1994, Persson & Argos, 1994). The neural network system was tested in five-fold cross-validation on 69 proteins with experimentally well determined transmembrane helices (Methods). Network input was the information derived for successive windows of 13 adjacent residues from a multiple sequence alignment (Fig. 3). Output were two units, one for each state of the central residue (in membrane helix / not in membrane helix; Fig. 2).

Results and Discussion

Evolutionary information improves prediction accuracy significantly

Better prediction in terms of per-residue and segment based scores. Compared to a simple neural network, the per-residue accuracy of the full three-level system using explicitly various aspects of evolutionary information increased by some five percentage points (Table 1). The improvement in prediction accuracy was even more significant in terms of segment-based scores: from some 75% correctly predicted segments to 94%.

Fig. 4a Fig. 4b

Fig. 4: Reliability of prediction. The reliability index for the prediction was defined as proportional to the difference between the two output units: RI = INTEGER (10 * out(HTM) — out (not-HTM).

The factor 10 scales the reliability index to values 0-9. (a) Overall two-state per-residue accuracy versus the cumulative percentage of residues with a reliability index RI „ n, n = 0, ..., 9. Note that with RI „ 0 is the rightmost point representing 100% of the predicted residues. Results were averaged over the residues in all 69 transmembrane proteins used for the cross-validation test. A network system that used multiple alignments as input was compared to a network using single sequence information only. For example, 90% of all residues were predicted with RI „ 6. For these the prediction accuracy for the network using multiple alignment information reached a value of QSDO3(2) > 97%. (b) Percentage of residues correctly predicted in transmembrane helices versus cumulative percentage of residues predicted in transmembrane helices with a reliability index RI „ n. Results are given as percentages of the number of residues observed in transmembrane helices (open triangles) and as percentages of the number of residues predicted in transmembrane helices (filled circles). For example, about 70% of all residues predicted in transmembrane segments had a reliability index RI „ 7. 95% of these were predicted correctly.

Table 2 Observed and predicted transmembrane helices for 69 proteins.

protein observed predicted

HTM HTM

1brd (bacr_halha)

23-42 24-43

57-76 55-87

95-114 92-116

121-140 121-143

148-167 145-169

191-210 185-211

217-236 213-239

1prc_H

12-35 12-31

1prc_M

52-76 43-59

- 63-78

111-137 110-130

143-166 143-170

198-223 198-223

260-284 262-292

1prc_L

33-53 21-38

- 42-58

84-111 81-103

116-139 115-146

171-198 173-196

226-249 223-255

2mlt

2-10 -

12-25 -

protein observed predicted

HTM HTM

4f2_human

82-104 82-104

5ht3_mouse

246-272 238-270

278-296 282-301

306-324 307-331

465-484 457-484

a1aa_human

54-79 56-79

92-117 92-116

128-150 128-150

172-196 173-189

210-233 213-235

307-331 309-329

339-363 -

a2aa_human

34-59 32-60

71-96 69-100

107-129 106-133

150-173 151-169

193-217 196-221

375-399 375-399

407-430 405-429

a4_human

700-723 702-722

protein observed predicted

HTM HTM

aa1r_canfa

11-33 12-35

47-69 39-53

- 61-74

81-102 80-110

124-146 125-144

177-201 176-206

236-259 235-261

268-292 266-291

aa2a_canfa

8-30 10-32

44-66 40-71

78-100 77-105

121-143 122-141

174-198 174-203

235-258 234-260

267-290 266-290

adt_ricpr

34-54 31-46

68-88 60-87

93-113 92-115

148-168 134-148

- 156-170

185-205 185-206

219-239 217-239

280-300 271-298

321-341 322-342

protein observed predicted

HTM HTM

adt_ricpr (continued)

349-369 348-371

380-400 377-400

439-459 444-461

466-486 469-485

bach_halhm

23-42 24-43

57-76 55-87

95-114 92-116

121-140 121-143

148-167 145-169

191-210 185-211

217-236 213-239

cb21_pea

62-81 69-75

114-134 115-134

182-198 184-196

cek2_chick

365-389 371-389

cyoa_ecoli

- 12-24

51-69 44-66

93-111 90-109

protein observed predicted

HTM HTM

cyob_ecoli

17-35 -

58-76 61-77

102-121 101-131

144-162 146-158

195-213 191-212

232-250 227-252

277-296 286-302

320-339 315-335

348-366 349-368

382-401 380-401

410-429 415-440

457-476 457-470

494-513 498-519

588-607 592-608

614-634 612-626

cyoc_ecoli

32-50 29-50

67-85 67-85

102-120 101-116

143-161 138-162

185-203 178-202

cyod_ecoli

18-36 20-39

46-64 45-64

81-99 80-101

protein observed predicted

HTM HTM

cyoe_ecoli

10-28 12-24

38-56 44-66

79-97 90-109

108-126 109-127

- 142-158

- 166-181

198-216 198-222

229-247 228-252

269-287 265-287

edg1_human

47-71 45-72

79-107 80-107

122-140 116-145

160-185 160-180

202-222 201-227

256-277 254-282

294-314 288-312

egfr_human

646-668 648-666

fce2_human

22-47 27-47

glp_pig

63-85 63-84

glpa_human

92-114 91-114

protein observed predicted

HTM HTM

glpc_human 58-81 57-81

glra_rat

539-558 536-557

585-603 -

614-632 615-636

806-826 807-826

gmcr_human

321-346 326-351

gp1b_human

148-172 147-171

gpt_crilo

7-32 12-38

58-79 59-83

95-114 96-115

126-145 127-150

165-184 157-181

195-211 187-210

222-240 224-242

253-269 249-269

275-294 277-292

379-397 379-402

hema_cdvo

35-55 37-58

hema_measi

35-55 37-58

hprotein observed predicted

HTM HTM

ema_pi4ha

35-59 37-59

hg2a_human

46-72 50-67

iggb_strsp

- 18-32

- 91-103

423-443 425-439

il2a_human

241-259 235-258

il2b_human

241-265 236-267

ita5_mouse

356-381 355-383

lacy_ecoli

11-33 11-36

47-67 46-67

75-99 75-98

103-125 104-126

145-163 148-161

168-187 169-187

212-234 219-238

260-281 265-288

291-310 294-314

315-334 320-337

347-366 343-371

380-399 377-400

lech_human

40-60 40-59

leci_mouse

40-60 40-59

protein observed predicted

HTM HTM

lep_ecoli

4-22 4-23

58-76 63-82

magl_mouse

517-536 515-534

malf_ecoli

17-35 21-35

40-58 43-58

73-91 71-93

277-295 278-306

319-337 318-339

371-389 370-390

418-436 418-444

486-504 486-505

motb_ecoli

28-49 30-51

mprd_human

186-210 185-211

myp0_human

- 14-31

154-179 155-183

ngfr_human

251-272 253-272

protein observed predicted

HTM HTM

nep_human

28-50 30-49

oppb_salty

10-30 10-29

100-121 96-120

138-158 130-162

173-190 168-193

227-250 228-259

272-293 273-298

oppc_salty

38-59 39-59

102-122 98-126

140-160 141-158

164-180 166-182

216-236 210-225

- 232-248

268-290 268-289

ops1_calvi

48-72 47-75

85-110 85-110

125-144 116-145

164-187 162-187

212-239 212-239

275-298 275-298

306-330 306-329

ops2_drome

57-81 55-84

94-119 94-118

134-153 124-153

173-196 171-196

221-248 221-248

284-307 284-307

315-339 315-338

protein observed predicted

HTM HTM

ops3_drome

58-82 57-85

95-119 95-119

134-152 125-153

172-196 169-194

221-248 221-248

285-308 285-308

317-341 317-340

ops4_drome

54-78 53-81

91-113 91-115

130-149 121-150

168-192 166-191

217-244 217-244

281-304 281-304

313-337 313-336

opsb_human

34-58 33-59

71-96 71-100

111-130 112-135

150-173 149-173

200-227 200-227

250-272 251-275

282-306 281-306

opsd_bovin

37-61 36-62

74-99 74-104

114-133 115-139

153-176 152-176

203-230 203-230

252-276 253-279

285-309 285-309

opsg_human

53-77 52-78

90-115 90-120

130-149 131-155

169-192 168-192

219-246 219-245

269-292 269-295

301-325 301-325

protein observed predicted

HTM HTM

opsr_human

53-77 52-78

90-115 90-119

130-149 131-155

169-192 168-192

219-246 219-245

269-292 270-295

301-325 301-325

pigr_human

621-643 624-643

protein observed predicted

HTM HTM

pt2m_ecoli

25-44 20-42

51-69 54-65

135-154 133-156

166-184 167-181

- 249-262

274-291 270-283

314-333 312-332

protein observed predicted

HTM HTM

sece_ecoli

19-36 20-34

45-63 42-62

93-111 93-123

suis_human

13-32 12-33

tcb1_rabit

292-313 285-312

protein observed predicted

HTM HTM

trbm_human

516-539 515-536

trsr_human

63-88 67-86

vmt2_iaann

25-42 27-51

vnb_inbbe

19-40 19-42

For the 69 transmembrane proteins used for cross-validation the following data is listed: (i) the protein name, given by the SWISSPROT identifier (Bairoch & Boeckmann, 1994); if the 3D structure is known then the PDB code plus chain identifier is used (Bernstein et al., 1977, Kabsch & Sander, 1983)) (ii) the positions for the transmembrane helices observed (= SWISSPROT documentation, or DSSP (Kabsch & Sander, 1983), counted from the first residue in SWISSPROT or DSSP ), and (iii) the cross-validated prediction by the network system PHDhtm . Except for 2mlt and glra_rat the list comprises a subset of the proteins used by David Jones (Jones et al., 1994) and Gunnar von Heijne (von Heijne & Gavel, 1988, von Heijne, 1992, Sipos & von Heijne, 1993).

Reliability index of practical use to refine prediction accuracy. For some 70% of all proteins 100% of all segments were predicted correctly (data not shown). The reliability of the prediction (reliability index defined in Fig. 4) can help to estimate whether or not a protein is likely to belong to the majority of proteins for which all segments are predicted correctly (Fig. 4). Furthermore, the reliability index was used to control the filtering procedure (Fig. 5).

Performance similar to that of the best alternative methods

Recently, two groups reported significant improvements in predicting transmembrane helices. Jones et al. (1994) use a new method with five output states (HTM-inside/middle/outside and not-HTM inside/outside, where inside/outside refers to inside/outside the cell). Persson and Argos (1994) use four output states (HTM-begin/middle/end and not-HTM) plus multiple alignment information. The system described here resulted in an accuracy in predicting the transmembrane helices similar to these two methods although we used only two output states. An exact comparison of the performance accuracy is made difficult as for both methods neither per-residue scores are published nor the segment measures used are defined (caption Table 1). Surprisingly, the errors made by the network system are often different from those made by the two statistical methods (Table 2 in comparison to (Jones et al., 1994, Persson & Argos, 1994)).

High reliability in discriminating between proteins with and without transmembrane helices

Does the prediction method distinguish transmembrane from non-transmembrane proteins? Two questions are of interest. First, did the network system correctly predict all transmembrane proteins used for the cross-validation analysis as transmembrane proteins? And second, were some globular proteins falsely predicted to contain transmembrane segments?

Transmembrane proteins correctly identified. Both the network system using single sequences as input and the network using only profiles did identify all but two proteins in the test set as transmembrane proteins: melittin (2mlt) and the immunoglobulin G-binding protein precursor (iggb_strsp). Melittin is a special case, as the DSSP (Kabsch & Sander, 1983) assignment of secondary structure splits the long helix of the 26-residue molecule into two that were so short that the filtering procedure would miss this protein even on the basis of the known 3D structure. The ultimate network system PHDhtm missed only melittin, all other membrane proteins were correctly identified.

Fewer than 5% false positives. To test whether globular proteins were falsely predicted to contain transmembrane helices we chose a set of 278 unique globular proteins. (No network predicted a transmembrane helix in the b-barrel porin.) PHDhtm mispredicted fewer than 5% of the globular proteins (Table 3). False positives were often globular water-soluble proteins with highly hydrophobic beta strands in the core. An exception was the only globular protein predicted to contain more than three segments: photo-synthetic reaction centre (4rcr) for which 11 segments with an average length of 21 residues were predicted as transmembrane helices (mandelate race mace (2mnr) was predicted with three long helices). The network using only profiles as input predicted transmembrane helices for less than 2% of the globular proteins.

Table 3 Prediction accuracy on globular proteins (negative control).

method	number of globular proteins used	number of proteins predicted with HTM	number of HTM segments longer than 16 residues	% of false classifications

no profiles	278	18	7	6.5%
profiles only	278	5	4	1.8%
PHDhtm	278	12	7	4.3%

Jones+, 1994	155	5	-	3.2%
Edelman, 1993	14	3	-	21.4%

Abbreviations for methods as in Table 1 and Table 4 . We considered a globular protein to be mispredicted if either at least two transmembrane segments are predicted with more than 10 residues, or at least one with more than 17 residues. Results from Edelman, 1993 and Jones et al., 1994 (Jones+, 1994 ) were taken from the literature.

Multi-level system improves significantly over simple neural network

Alignment information improves performance. The most significant improvement in prediction accuracy (compared to a simpler neural network prediction) stemmed from including the information contained in multiple alignments. Roughly one half of the improvement attributed to simply using residue substitution frequencies (Table 4), and one half to using additionally more details contained in the alignments (conservation weight, number of insertions and deletions) and information about the whole protein (Table 4).

Balanced versus unbalanced training. The balanced training procedure (equally often presenting residues in transmembrane and residues not in transmembrane segments, Methods) tended to over-predict transmembrane helices, while an unbalanced training procedure (presentation of examples according to the distribution in the training set, Methods) tended to under-predict transmembrane segments.

Jury decision finds a compromise between balanced and unbalanced training. Both, balanced and unbalanced training had advantages and disadvantages. Which of the two methods should be used for prediction? A reasonable compromise (effectively between over- and under-prediction) was found by the jury decision, i.e. the arithmetic average over the output values of balanced and unbalanced networks.

Second level elongates helices. The effect of the second level (structure-to-structure) network was to elongate or delete short helical segments. The effect was an increase in the average length of a predicted helical segment from 15 residues for the first level, to 27 residues for the second level (Table 4). In other words, the first level networks (Fig. 2) yielded an average length for transmembrane segments 5-7 residues shorter than observed, the second level networks (Fig. 2) resulted in segments up to 13 residues longer than observed. Thus, the second level networks tended to elongate helices (Table 4).

Final filtering procedure. Short loop regions were often missed by the second network that tended to elongate helices too much (note that the input window is too narrow to learn a maximal length for transmembrane segments). This drawback was compensated by a relatively straightforward filtering procedure (Methods). Filtering improved the prediction accuracy both in terms of per-residue and segment-based measures for prediction accuracy (Table 4).

Table 4 Analysis of the performance for each element of the network system

The column 'system levels' describes which levels of the network system were used (Fig. 2): 1, only 1st level, 2, 1st and 2nd level; 3, jury average over different 2nd level networks (Methods); filter, application of the filtering procedure (Fig. 5). Set 1 contains 69 transmembrane proteins (Methods), set 5 , is the subset of set 1 without the PDB proteins 2mlt, 1prc (chains H, L, M) and 1brd. Abbreviations of measures as in Table 1 .

Abbreviations of methods: PHDhtm , three-level network system + filter using all information from multiple alignments as input (Fig. 2); no profiles , two-level network system using single sequences as input (Casadio et al., 1994); profiles only , same as before, but using evolutionary profiles (and no further information derived from the multiple alignment) as input; 1st unbalanced , first level network with unbalanced training (Methods); 1st balanced , first level network with balanced training (Methods); 1st x - 2nd y , a second level network with y (balanced or unbalanced) training which uses as input the prediction from a first level network with x (balanced or unbalanced) training; jury over 4 networks , arithmetic average over the four different second level networks given above.

Conclusions

Selection of data set. The 3D structure is experimentally known for only five (1prc_H, 1prc_L, 1prc_M, 1brd, 2mlt) of the 69 protein chains used for the cross-validation analysis. This implies that the results ought to be taken with caution. To increase confidence in the results, we deliberately chose proteins for which there is 'reliable' experimental evidence about the locations of the transmembrane regions (list taken from Jones et al., 1994), rather than working with a larger data set including less well known segments.

Improved prediction of transmembrane helices. Using various aspects of evolutionary information improved the overall per-residue accuracy of predicting residues in transmembrane helices by some five percentage points. This improvement could be significant enough to warrant use of the predictions as a starting point for a complete ab initio prediction of 3D structure for transmembrane regions (Baldwin, 1993, Taylor et al., 1994). Our best network system (called PHDhtm) correctly predicted some 94% of all segments and the correct location of some 90% of all residues observed in transmembrane helices. For only four out of 15 incorrectly predicted (either under-, or over-predicted) segments the defined reliability index would have lead the user to have suspect a wrong prediction (Fig. 1).

Prediction for globular proteins sufficiently accurate. The two-level network system using only profiles as input mispredicted less than 2% of globular proteins as containing transmembrane helices (Table 3). An unsatisfactory disadvantage of the most accurate network system PHDhtm was that this error rate was clearly higher (<5%). However, for most practical purposes this rate of false positives is sufficiently low. All transmembrane proteins were predicted to contain at least one transmembrane helix, except for melittin which would not have been recognised as transmembrane helix even from the crystal structure: the strongly bent helix is split into two short helices by the program assigning the secondary structure automatically from 3D structures (DSSP, Kabsch & Sander, 1983).

Weak point. A rather inconvenient aspect of the method described here is the necessity to apply a filter procedure (Fig. 5) at the end of the prediction. This disadvantage will be one of the details which still have to be improved in a more general tool.

Possible improvements of the prediction. There are methods which predict whether or not a loop region is located inside or outside the cell (von Heijne & Gavel, 1988, Nakashima & Nishikawa, 1992, von Heijne, 1992, Sipos & von Heijne, 1993, Jones et al., 1994). Such tools could be used to either complement the network prediction, or directly to train a network to predict transmembrane topology (direction of transmembrane helices with respect to cell).

b -strand membrane proteins. How can transmembrane segments for b-barrel proteins such as porin be predicted from sequence? Interestingly, the network system trained on water-soluble globular proteins (PHDsec ), predicts the b-strands of the membrane protein porin more accurately than the helices of the photo-reaction centre, bacteriorhodopsin or the light harvesting complex. The reason may be that the pore of porin is exposed to solvent and thus resembles globular proteins in some respects. The prediction of b-strands, combined with hydrophobicity scales (Eisenberg et al., 1984b) and/or predictions of solvent accessibility (Rost & Sander, 1994b) has been used to infer which of the porin strands may be in contact with lipids. Unfortunately, however, the structure of very few b-strand membrane proteins are known. Thus, training of neural networks, as well as the application of statistical methods is premature.

3D structure prediction. How can one come closer to the goal of 3D prediction for helical membrane proteins? One way to go from accurate predictions of helical transmembrane locations to 3D structure has been indicated by Taylor et al. (Taylor et al., 1994). Whether or not the network predictions described here in combination with a prediction of segment orientation relative to the membrane surface will be useful remains to be shown.

Keeping up with the flow of genome data. All results reported here, refer to completely automatic usage of PHDhtm . In some cases, prediction accuracy can certainly be improved by expert knowledge, e.g. by fine tuning the alignment. However, fully automatic use permits the analysis of many proteins, e.g. all open reading frames of complete chromosomes. For example, less than an hour of CPU time (on a SUN SPARC10 workstation) was required for the transmembrane helix prediction of all proteins of yeast chromosome VIII (Johnston et al., 1994), given the multiple sequence alignments. For 59 of the 269 proteins at least two transmembrane helices were predicted (Table 5), for another 27 of the proteins one transmembrane helix was predicted. Given an error rate of 5%, this implies that 20-25% of all yeast VIII proteins were predicted to contain transmembrane helices.

Availability of the network prediction. Predictions of transmembrane helices (as well as secondary structure and solvent accessibility for globular proteins) using the method presented here are provided via an automatic electronic mail server. If you send the sequence of your protein, the server will return a multiple sequence alignment and a prediction of the location of transmembrane helices. For further information, send the word help to the internet address PredictProtein@EMBL-Heidelberg.DE by electronic mail, or use the World Wide Web (WWW) site http://www.embl-heidelberg.de/predictprotein/predictprotein.html. (NOTE: new www site http://cubic.bioc.columbia.edu/predictprotein).

Methods

Database

Selection of proteins

We based our analyses on a set of 69 proteins for which experimental information about the location of transmembrane helices is annotated in the SWISSPROT data base (Manoil & Beckwith, 1986, von Heijne & Gavel, 1988, von Heijne, 1992, Sipos & von Heijne, 1993, Jones et al., 1994). This set in particular was chosen to meet three criteria: (i) the experimental information should be as reliable as possible (Manoil & Beckwith, 1986, von Heijne, 1992); (ii) comparability: to enable a comparison to similar methods, the data set should be similar to those used by others; (iii) availability: the list (Table 2) was the subset of those proteins used by Jones et al. (1994) that were available in SWISSPROT when we had started the project (melittin (2mlt) and the glutamic receptor (glra_rat, O'Hara et al., 1993) were added). For the few known 3D structures, the location of the transmembrane regions was taken from DSSP (Kabsch & Sander, 1983). The exact locations of the transmembrane helices are often controversial. To enable a straightforward comparison to future methods and for making our results easily reproducible for others, we decided to always use the definitions found in Swissprot (Bairoch & Boeckmann, 1994).

Table 5 Prediction of transmembrane helices for yeast chromosome VIII.

CAPTION: Table 5
As typical example for the application of the method and as an independent test of the predictive power of the method, we predicted the transmembrane helices for all proteins from the complete yeast chromosome VIII (Johnston et al., 1994). For 59 proteins (out of 269) two or more transmembrane helices were predicted. The proteins are labelled by the identifier used in (Johnston et al., 1994). Shown are the predictions only for those proteins for which sufficient alignment information was available (Bork, Ouzounis, Sander, manuscript in preparation) or which were predicted to have more than six transmembrane segments. Abbreviations: Nres : length of protein; Nali : number of sequences in the multiple alignment ('1' means that the prediction is based on a single sequence only); Nhtm : predicted number of transmembrane segments.

In some cases confirmation of the correctness of the prediction comes from detailed sequence analysis (Johnston et al., 1994, Bork, Ouzounis and Sander, unpublished): the likely function identified on the basis of sequence similarity to proteins of known function is consistent with the presence of helical transmembrane regions. Examples are: YHR026w, an ATPase; YHR048w, a resistance protein, probably works by pumping substances out of the cell through a membrane pore; YHR050w /92c /94c /96c, potential transporters; YHR190w, farnesyltransferase; YHR123w, phosphor transferase; YHR005c, G-protein alpha subunit; YHR183w /39c, dehydrogenase.

Generation of multiple alignments

For each of the initial 69 proteins a multiple sequence alignment was generated using the program MaxHom (Sander & Schneider, 1991; Fig. 3). All sequences from Swissprot with a sequence identity above a length-dependent cut-off were included in the alignment (Sander & Schneider, 1991), assuming that this is valid not only for globular but also for membrane proteins.
Cross-validation test

The set of 69 transmembrane proteins (Table 2) was divided into 52 proteins used for training and 17 used for testing the method. This was repeated five times (five-fold cross-validation), until each protein had been in a test set once. The sets were chosen such that no protein at in the multiple alignments used for testing had more than 25% sequence identity to any protein in the multiple alignments of the training set. All results reported are averages over proteins in various test sets.
Neural network system

First level: sequence-to-structure

The principles of neural networks for secondary structure prediction (Fariselli et al., 1993, Rost & Sander, 1993a) and of coding multiple sequence information (Rost & Sander, 1993b, Rost & Sander, 1994a, Rost & Sander, 1994b) are described in detail elsewhere. Here, only some basic concepts will be recapitulated and details regarding the application to transmembrane helices will be introduced.

Input to the first level network consisted of two contributions, (i) one local in sequence, i.e. taken from a window of 13 adjacent residues; and (ii) another global in sequence, i.e. compiled from the whole protein (Fig. 2). (i) The local information computed for each residue in the window was the frequency of occurrence of each amino acid at that position in the multiple alignment, the number of insertions and deletions in the alignment for that residue, and a position specific conservation weight (Fig. 2). (ii) As global information we used the amino acid composition and length of the protein and, furthermore, the distance (number of residues) of the first residue in the window of 13 adjacent residues from the protein begin (N-term), and the distance of the last residue in the window to the protein end (C-term).

Output of the first level network were two units, one representing examples with the central residue of the window being in a transmembrane helix, the other representing examples with the central residue not being in transmembrane helices (Fig. 2).
Balanced and unbalanced training

Training was performed with the usual gradient descent (also known as back-propagation Rumelhart et al., 1986)):

(1)

where t is the algorithmic time step (i.e. change of all connections for one pattern), E is the error, given by the difference between actual network output and the desired output (i.e. the value observed for the central residue); J_ij is the connection from unit j to unit i on the next layer (input to hidden, hidden to output); e is the learning speed, chosen here to be 0.01; and a the momentum term (permitting up-hill moves), chosen here to be 0.2. Two modes were used. First, unbalanced training: at each time step of the error minimisation one pattern was chosen at random from the training set, and all connections of the network were changed. Second, balanced training: at each time step of the error minimisation (eqn. 1) one pattern from the class 'transmembrane helix' and one from the class 'not transmembrane helix' was used to change all connections.
Network parameters

All units were connected to all those on the next layer (input to hidden, hidden to output). The network parameters such as criterion to terminate the training procedure, number of hidden units, training speed (e in eqn. 1), and momentum term (a in eqn. 1) were chosen arbitrarily based on our experience with secondary structure prediction for globular proteins. In other words, these parameters were not influenced by the test set. Training was stopped when the training set had been learned to an accuracy of 93% for the first and of 95% for the second level network. As for the number of hidden units, we started arbitrarily with 3 hidden units for the first level of network, and increased the number for the second level network to 15 since training too often ended in local minima.

Fig. 5: Filtering the prediction. The output of the third level (jury prediction) was filtered to delete too short and to split too long predicted transmembrane helices. The splitting of too long segments was usually done exactly in the middle of the segment by flipping the prediction for one residue from HTM (helical transmembrane) to not-HTM . Two exceptions were: (i) if there was a residue in a three-residue neighbourhood of the central residue with a lower reliability index than that of the central one, than splitting was performed at that residue; (ii) if the two residues on both sides of the central residue were predicted with a reliability index (RI) < 3, than up to five residues in total are flipped from the state HTM to not-HTM .

Second level: structure-to-structure

The input to the second level network consisted - as to the first level - of a contribution local in sequence and a contribution global in sequence (Fig. 2). (i) For each residue in the input window the local input were the values of the two output units of the first level network and the conservation weight. (ii) The global input information was the same as for the first level network. The output of the second level network - as for the first - consisted of two units for the central residue either being in a transmembrane helix or not.
Third level: jury decision

To find a compromise between networks with balanced and those with unbalanced training a final jury decision was performed (effectively a compromise between over- and under-prediction, Results). The jury decision was a simple arithmetic average over four differently trained networks: all combinations (2 ¥ 2) of first level network with balanced and unbalanced training, and with balanced or unbalanced training of second level network. The final prediction was assigned to the unit with maximal output value ('winner takes all').
Fourth level: filtering the prediction

In contrast to earlier prediction methods (Jones et al., 1992, von Heijne, 1992, Persson & Argos, 1994), which explicitly fix the length of predicted transmembrane segments to typically 17-25 residues, the second level network occasionally resulted in transmembrane helices that were either too short or too long. This was corrected by a non optimised filter that was guided by the experiences of previous work (von Heijne, 1986, von Heijne & Gavel, 1988, von Heijne, 1992, Sipos & von Heijne, 1993, Casadio et al., 1994, Jones et al., 1994).

Too long helices were either split in the middle into two shorter helices or were shortened (Fig. 5). Too short helices were either elongated or deleted. All these decisions (split or shorten; elongate or delete) were based both on the strength of the prediction (reliability index, Fig. 2) and on the length of the predicted transmembrane helix (Fig. 5).

Acknowledgements

We should like to express our gratitude to: Reinhard Schneider (EMBL, Heidelberg) for providing the latest version of the alignment program MaxHom ; Chiara Taroni (Bologna) and Mario Compiani (Camerino) for helpful discussions; David Jones (London) for help with the data set; Gunnar von Heijne (Huddinge) for motivating discussions; and Christos Ouzounis (EMBL) for providing the multiple alignments for yeast VIII. Furthermore, thanks to the two referees, who helped improving the text by their detailed criticism. Last not least, thanks to all those who deposit experimental results in public data bases.
References

Argos P, Rao JKM and Hargrave PA. 1982. Structural prediction of Membrane-bound proteins. Eur. J. Biochem. 128:565-575.

Bairoch A & Boeckmann B. 1994. The SWISS-PROT protein sequence data bank: current status. Nucl. Acids Res. 22:3578-3580.

Baldwin JM. 1993. The probable arrangement of the helices in G protein-coupled receptors. EMBO J. 12:1693-1703?

Bernstein FC, Koetzle TF, Williams GJB, Meyer EF, Brice MD, Rodgers JR, Kennard O, Shimanouchi T and Tasumi M. 1977. The Protein Data Bank: a computer based archival file for macromolecular structures. J. Mol. Biol. 112:535-542.

Casadio R, Fariselli P, Taroni C and Compiani M. 1994. A predictor of transmembrane a-helix domains of proteins based on neural networks. European Journal of Biophysics submitted, 8/94.

Cornette JL, Cease KB, Margalit H, Spouge JL, Berzofsky JA and DeLisi C. 1987. Hydrophobicity Scales and Computational Techniques for Detecting Amphipathic Structures in Proteins. J. Mol. Biol. 195:659-685.

Cowan SW & Rosenbusch JP. 1994. Folding pattern diversity of integral membrane proteins. Science 264:914-916.

Degli Esposti M, Crimi M and Venturoli G. 1990. A critical evaluation of the hydropathy profile of membrane proteins. Eur. J. Biochem. 190:207-219.

Deisenhofer J, Epp O, Mii K, Huber R and Michel H. 1985. Structure of the protein subunits in the photosynthetic reaction centre of Rhodopseudomonas viridis at 3Å resolution. Nature 318:618-624.

Edelman J. 1993. Quadratic Minimization of Predictors for Protein Secondary Structure: Application to Transmembrane a-Helices. J. Mol. Biol. 232:165-191.

Eisenberg D, Schwartz E, Komaromy M and Wall R. 1984a. Analysis of membrane and surface protein sequences with the hydrophobic moment plot. J. Mol. Biol. 179:125-142.

Eisenberg D, Weiss RM and Terwilliger TC. 1984b. The hydrophobic moment detects periodicity in protein hydrophobicity. Proc. Natl. Acad. Sc. U.S.A. 81:140-144.

Engelman DM, Steitz TA and Goldman A. 1986. Identifying nonpolar transbilayer helices in amino acid sequences of membrane proteins. Annual Review of Biophysics and Biophysical Chemistry 15:321-353.

Fariselli P, Compiani M and Casadio R. 1993. Predicting secondary structures of membrane proteins with neural networks. Eur. Biophys. J. 22:41-51.

Henderson R, Baldwin JM, Ceska TA, Zemlin F, Beckmann E and Downing KH. 1990. Model for the structure of bacteriorhodopsin based on high-resolution electron cryo-microscopy. J. Mol. Biol. 213:899-929.

Johnston M, Andrews S, Brinkman R, Cooper J, Ding H, Dover J, Du Z, Favello A, Fulton L, Gattung S, Geisel C, Kirsten J, Kucaba T, Hillier L, Jier M, Johnston L, Langston Y, Latreille P, Louis EJ, Macri C, Mardis E, Menezes S, Mouser L, Nhan M, Rifkin L, Riles L, Peter HS, Trevaskis E, Vaughan K, Vignati D, Wilcox L, Wohldman P, Waterston R, Wilson R and Vaudin M. 1994. Complete Nucleotide Sequence of Saccaromyces cerevisiae Chromosome VIII. Science 265:2077-2082.

Jones DT, Taylor WR and Thornton JM. 1992. The rapid generation of mutation data matrices from protein sequences. CABIOS 8:275-282.

Jones DT, Taylor WR and Thornton JM. 1994. A model recognition approach to the prediction of all-helical membrane protein structure and topology. Biochem. 33:3038-3049.

Kabsch W & Sander C. 1983. Dictionary of Protein secondary structure: Pattern recognition of hydrogen bonded and geometrical features. Biopolymers 22:2577-2637.

Kühlbrandt W, Wang DN and Fujiyoshi Y. 1994. Atomic model of plant light-harvesting complex by electron crystallography. Nature 367:614-621.

Kyte J & Doolittle RF. 1982. A Simple Method for Displaying the Hydrophathic Character of a Protein. J. Mol. Biol. 157:105-132.

Landolt-Marticorena C, Williams KA, Deber CM and Reithmeier RAF. 1992. Non-random distribution of amino acids in the transmembrane segments of human type I single span membrane proteins. J. Mol. Biol. 229:602-608.

Lattman EE. 1994. Protein Crystallography for All. Proteins 18:103-106.

Manoil C & Beckwith J. 1986. A genetic approach to analyzing membrane protein topology. Science 233:1403-1408.

Matthews BW. 1975. Comparison of the predicted and observed secondary structure of T4 phage lysozyme. Biochim. Biophys. Ac. 405:442-451.

Nakashima H & Nishikawa K. 1992. The amino acid composition is different between the cytoplasmic and extracellular sides in membrane proteins. FEBS Lett. 303:141-146.

O'Hara PJ, Sheppard PO, Thøgersen H, Venezia D, Haldeman BA, McGrane V, Houamed KM, Thomsen C, Gilbert TL and Mulvihill ER. 1993. The ligand-binding domain in metabotropic glutamate receptors is related to bacterial periplasmic binding proteins. Neuron 11:41-52.

Oliver S, M. vdAQJ, Agostioni-Carbone ML, Aigle M, Alberghina L, Alexandraki D, Antoine G, Anwar R, Ballesta JPG, Benit P, Berben G, Bergantino E, Biteau N, Bolle PA, Bolotin-Fukuhara M, Brown A, Brown AJP, Buhler JM, Carcano C, Carignani G, Cederberg H, Chanet R, Contreras R, Crouzet M, Daignan-Fornier B, Defoor E, Delgado M, Demolder J, Doira C, Dubois E, Dujon B, Dusterhoft A, Erdmann D, Esteban M, Fabre F, Fairhead C, Faye G, Feldmann H, Fiers W, Francingues-Gaillard MC, Franco L, Frantali L, Fukuhara H, Fuller LJ, Galland P, Gent ME, Gigot D, Gilliquet V, Glansdorff N, Goffeau A, Grenson M, Grisanti P, Grivell LA, Haan Md, Haasemann M, Hatat D, Hoenicka J, Hegeemann J, Herbert CJ, Hilger F, Hohmann S, Hollenberg CP, Huse K, Iborra F, Indge KJ, Isono K, Jacq C, Jacquet M, James CM, Jauniaux JC, Jia Y, Jimenez A, Kelly A, Kleinhans U, Kreisl P, Lanfranchi G, Lewis C, van der Linden CG, Lucchini G, Lutzenkirchen K, Maat MJ, Mallet L, Mannhaupt G, Martegani E, Mathieu A, Maurer CTC, McDonnell D, McKee RA, Messenguy F, Mewes HW, Molemans F, Montague MA, Muzi Falconi M, Navas L, Newlon CS, Noone D, Pallier C, Panzeri L, Pearson BM, Perea J, Philippsen P, Pierard A, Planta RJ, Plevani P, Poetsch B, Pohl F, Purnelle B, Ramezani Rad M, Rasmussen SW, Raynal A, Remacha M, Richertich P, Roberts AB, Rodriquez F, Sanz E, Schaaff-Gerstenschlager I, Scherens B, Schweitzer B, Shu Y, Skala J, Slonimski PP, Sor F, Soustelle C, Spiegelberg R, Stateva LI, Steensma HY, Steiner S, Thierry A, Thireos G, Tzermia M, Urrestarazu LA, Valle G, Vetter I, van Vliet-Reedijk JC, Voet M, Volckaert G, Vreken P, Wang H, Warmington JR, Wettstein Dv, Wicksteed BL, Wilson C, Wurst H, Xu G, Yoshikawa A, Zimmermann FK and Sgouros JG. 1992. The complete DNA sequence of yeast chromosome III. Nature 357:38-46.

Pearson WR & Lipman DJ. 1988. Improved tools for biological sequence comparison. Proc. Natl. Acad. Sc. U.S.A. 85:2444-2448.

Pearson WR & Miller W. 1992. Dynamic programming algorithms for biological sequence comparison. Meth. Enzymol. 210:575-601.

Persson B & Argos P. 1994. Prediction of Transmembrane Segments in Proteins Utilising Multiple Sequence Alignments. J. Mol. Biol. 237:182-192.

Rost B & Sander C. 1993a. Improved prediction of protein secondary structure by use of sequence profiles and neural networks. Proceedings National Academy of Science 90:7558-7562.

Rost B & Sander C. 1993b. Prediction of protein secondary structure at better than 70% accuracy. J. Mol. Biol. 232:584-599.

Rost B & Sander C. 1993c. Secondary structure prediction of all-helical proteins in two states. Prot. Engin. 6:831-836.

Rost B & Sander C. 1994a. Combining evolutionary information and neural networks to predict protein secondary structure. Proteins 19:55-72.

Rost B & Sander C. 1994b. Conservation and prediction of solvent accessibility in protein families. Proteins 20:216-226.

Rost B, Sander C and Schneider R. 1993. Progress in protein structure prediction? TIBS 18:120-123.

Rost B, Sander C and Schneider R. 1994. Redefining the goals of protein secondary structure prediction. J. Mol. Biol. 235:13-26.

Rumelhart DE, Hinton GE and Williams RJ. 1986. Learning representations by back-propagating error. Nature 323:533-536.

Sander C & Schneider R. 1991. Database of Homology-Derived Structures and the Structurally Meaning of Sequence Alignment. Proteins 9:56-68.

Sander C & Schneider R. 1994. The HSSP database of protein structure-sequence alignments. Nucl. Acids Res. 22:3597-3599.

Sipos L & von Heijne G. 1993. Predicting the topology of eukaryotic membrane proteins. Eur. J. Biochem. 213:1333-1340.

Taylor WR, Jones DT and Green NM. 1994. A Method for a-Helical Integral Membrane Protein Fold Prediction. Proteins 18:281-294.

von Heijne G. 1981. Membrane proteins-the amino acid composition of membrane-penetrating segments. Eur. J. Biochem. 120:275-278.

von Heijne G. 1986. A new method for predicting signal sequence cleavage sites. Nucl. Acids Res. 14:4683-4690.

von Heijne G. 1991. Computer analysis of DNA and protein sequences. Eur. J. Biochem. 199:253-256.

von Heijne G. 1992. Membrane Protein Structure Prediction. J. Mol. Biol. 225:487-494.

von Heijne G & Gavel Y. 1988. Topogenic signals in integral membrane proteins. Eur. J. Biochem. 174:671-678.

Wang DN, Kühlbrandt W, Sarabiah V and Reithmeier RAF. 1993. Two-dimensional structure of the membrane domain of human Band 3, the anion transport protein of erythrocyte membrane. EMBO J. 12:2233-2239.

Weiss MS & Schulz GE. 1992. Structure of porin refined at 1.8 Å resolution. J. Mol. Biol. 227:493-509.