Bioinformatics Advance Access originally published online on November 15, 2007
Bioinformatics 2008 24(4):513-520; doi:10.1093/bioinformatics/btm548
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
TMBpro: secondary structure, β-contact and tertiary structure prediction of transmembrane β-barrel proteins
1School of Information and Computer Sciences, 2Institute for Genomics and Bioinformatics, University of California, Irvine, CA 92697 and 3Department of Computer Science, University of Missouri, Columbia, MO 65203, USA
*To whom correspondence should be addressed.
 |
ABSTRACT |
|---|
 TOP ABSTRACT 1 INTRODUCTION2 DATA3 METHODS4 RESULTS5 CONCLUSIONACKNOWLEDGEMENTSREFERENCES |
|---|
Motivation: Transmembrane β-barrel (TMB) proteins are embedded in the outer membranes of mitochondria, Gram-negative bacteria and chloroplasts. These proteins perform critical functions, including active ion-transport and passive nutrient intake. Therefore, there is a need for accurate prediction of secondary and tertiary structure of TMB proteins. Traditional homology modeling methods, however, fail on most TMB proteins since very few non-homologous TMB structures have been determined. Yet, because TMB structures conform to specific construction rules that restrict the conformational space drastically, it should be possible for methods that do not depend on target-template homology to be applied successfully.
Results: We develop a suite (TMBpro) of specialized predictors for predicting secondary structure (TMBpro-SS), β-contacts (TMBpro-CON) and tertiary structure (TMBpro-3D) of transmembrane β-barrel proteins. We compare our results to the recent state-of-the-art predictors transFold and PRED-TMBB using their respective benchmark datasets, and leave-one-out cross-validation. Using the transFold dataset TMBpro predicts secondary structure with per-residue accuracy (Q2) of 77.8%, a correlation coefficient of 0.54, and TMBpro predicts β-contacts with precision of 0.65 and recall of 0.67. Using the PRED-TMBB dataset, TMBpro predicts secondary structure with Q2 of 88.3% and a correlation coefficient of 0.75. All of these performance results exceed previously published results by 4% or more. Working with the PRED-TMBB dataset, TMBpro predicts the tertiary structure of transmembrane segments with RMSD <6.0 Å for 9 of 14 proteins. For 6 of 14 predictions, the RMSD is <5.0 Å, with a GDT_TS score greater than 60.0.
Availability: http://www.igb.uci.edu/servers/psss.html
Contact: pfbaldi{at}ics.uci.edu
Supplementary information: Supplementary data are available at Bioinformatics online.
|
1 INTRODUCTION |
|---|
|
TOPABSTRACT1 INTRODUCTION2 DATA3 METHODS4 RESULTS5 CONCLUSIONACKNOWLEDGEMENTSREFERENCES |
|---|
Transmembrane β-barrel (TMB) proteins are an important class of proteins embedded in the outer membrane of Gram-negative bacteria, mitochondria and chloroplasts (Schulz, 2000; Tamm et al., 2004; Wallin and Heijne, 1998). It is estimated that genomic databases currently contain thousands of TMB proteins (Wimley, 2002, 2003), and ongoing large-scale sequencing efforts promise to produce many more (Yooseph et al., 2007). These proteins carry out diverse biochemical functions including active ion transport, passive nutrient intake and defense against attack proteins (Koebnik et al., 2000; Schulz, 2000). Thus, elucidating the structure and function of TMB proteins has immediate medical relevance, as bacterial membrane proteins are potential targets of antimicrobial drugs and vaccines (Jackups and Liang, 2005). Crystallizing transmembrane (TM) proteins is especially challenging; thus, predicting the structure of TMB proteins from sequence is an interesting and important task (Casadio et al., 2003; Oberai et al., 2006).
Currently, several methods try to discriminate TMB proteins from globular and TM 
-helical proteins, or to predict their 1-dimensional (1D) secondary structure features (i.e. the positions of TM β-strands and the types of loops) (Bagos et al., 2004a, b, 2005; Bigelow and Rost, 2006; Bigelow et al., 2004; Diederichs et al., 1998; Fariselli et al., 2005; Garrow et al., 2005; Gromiha and Suwa, 2005; Gromiha et al., 1997, 2004, 2005; Jacoboni et al., 2001; Liu et al., 2003; Martelli et al., 2002; Natt et al., 2004; Park et al., 2005; Paul and Rosenbusch, 1985; Waldispühl et al., 2006b; Welte et al., 1991; Zhai and Saier, 2002).
The 1D structure predictions are very useful for constructing a coarse topology of TMB structure (Tamm et al., 2001). However, they do not provide enough information to construct a low-resolution tertiary structure for a TMB protein (Jackups and Liang, 2005). In addition, traditional homology modeling of TMB proteins is hindered by the lack of sequence similarity between the small number of TMB proteins with known structures and the thousands of TMB proteins without known structures (Jacoboni et al., 2001; Schulz, 2000).
TMB proteins adopt a common β-barrel fold and obey specific construction rules, as outlined in Schulz (2000). For instance, known TMB proteins consist of an even number of membrane spanning β-strands with an anti-parallel β-meander topology. Two recently published methods take advantage of these construction rules to predict the inter-strand β-residue pairings of TMB proteins (Jackups and Liang, 2005; Waldispühl et al., 2006a). These β-contact predictions provide strong constraints for building tertiary structure models of TMB proteins as in the reconstruction of globular protein structures using contact constraints (Skolnick et al., 1997).
Since there are fewer than 20 non-redundant (Waldispühl et al., 2006a) TMB proteins with known structures in the Protein Data Bank (PDB) (Berman et al., 2000) and membrane protein databases (Ikeda et al., 2003; Lomize et al., 2006), it is challenging to develop robust knowledge-based methods to predict inter-strand pairings in TMB proteins. To overcome the small dataset problem, the method transFOLD (Waldispühl et al., 2006a) uses pair-wise inter-strand residue statistical potentials derived from globular proteins to predict the inter-strand residue pairings of TMB proteins with moderate accuracy.
In this article, we present a three-stage pipeline to predict the tertiary structure of TMB proteins. First, we predict the two-class secondary structure with TMBpro-SS. Second, we predict β-residue contacts using TMBpro-CON (Baldi and Pollastri, 2003; Cheng and Baldi, 2005; Cheng et al., 2006a; Pollastri and Baldi, 2002). Finally, we use these feature predictions, TMB templates, and construction rules to predict tertiary structure with TMBpro-3D.
|
2 DATA |
|---|
|
TOPABSTRACT1 INTRODUCTION2 DATA3 METHODS4 RESULTS5 CONCLUSIONACKNOWLEDGEMENTSREFERENCES |
|---|
2.1 Benchmark sets
In this work ,we use two sets of TMB proteins described in the literature. The first is the dataset described in Waldispühl et al. (2006a), which consists of 14 redundancy-reduced TMB proteins. The authors divide this set into two main subsets: non-water-filled (NWF) and water-filled (WF). NWF consists of (PDB code) 1QJP [PDB] , 1QJ8, 1THQ, 1P4T, 1I78, 1K24 and 1QD6. WF consists of 1A0S, 1AF6, 1PRN, 2OMF, 1E54, 1TLY and 2POR. In our work, we treat all 14 proteins as a single set. The secondary structure assignments used for this set come from the DSSP program (Kabsch and Sander, 1983), which we condense to two classes : strand (β) and non-strand (–). These single character designations are used throughout this work when dealing with two-class representation. Following the work described in Waldispühl et al. (2006a), the group published a web-server for predicting features of TMB proteins called transFold (Waldispühl et al., 2006b). Throughout this work, we refer to this set as SetTransfold. We compare our secondary structure and β-contact prediction results to transFold using this set.
The second set is described in Bagos et al. (2004a) and also contains 14 redundancy-reduced TMBs. Nine of them overlap with SetTransfold: 1QJP, 1QJ8, 1I78, 1K24, 1A0S, 1PRN, 2OMF, 1E54 and 2POR. The five proteins that differ are: 1QD5, 2MPR, 1FEP, 2KMO and 2FCP. Rather than using the DSSP assignments, the authors manually designated TM (β) and non-TM (–) segments for each protein in this set. This approach was motivated by the observation that many of the β-strands in TMB proteins extend significantly beyond the membrane, and the authors sought to focus on the TM regions. The authors have made their method available as the web server PRED-TMBB (Bagos et al., 2004b). For the remainder of this work, we refer to this set as SetPRED-TMBB. We compare our results for secondary structure and topology prediction to PRED-TMBB using this set. We also use this set to evaluate our tertiary structure predictions.
The two datasets are created and treated independently in this work in order to make fair comparisons to previous work. For all of the proteins in SetTransFold, the secondary structure annotation comes from DSSP. For all of the proteins in SetPRED-TMBB the secondary structure, the annotation comes from manual designation. For the nine proteins common to both datasets, we keep both types of secondary structure annotation. For example, protein 1QJ8 [PDB] is present in each dataset, but with different secondary structure annotation (DSSP in SetTransFold and manual designation in SetPRED-TMBB). Results comparing our work to transFold are based only on SetTransFold annotations, and results comparing our work to PRED-TMBB are based solely on SetPRED-TMBB annotations.
We compare our results using sets SetTransFold and SetPRED-TMBB to the published results of the respective methods. To compare our β-contact predictions to those of transFold using the same predicted secondary structure, we submitted the proteins of SetTransFold to the transFold server. The transFold server predicts the secondary structure into four classes: membrane facing strand residues (M), channel facing strand residues (C), loops inside the periplasm (i) and extra-cellular loops (o). The transFold server also predicts β-residue contacts. These single character designations are used throughout this work and in the output of our server. The PRED-TMBB server predicts secondary structure into three classes: TM, periplasmic and extra-cellular. For both datasets, we expanded the two-class representation to three-class by designating ‘β’ residues as either ‘M’ or ‘C’ based on visual inspection of the structures. These representations (M, C, –) were used to train a three-class predictor.
2.2 Cross-validation
Our predictors are trained and tested using leave-one-out cross-validation (LOOCV) on SetTransFold and SetPRED-TMBB independently. A single protein is held out of the set, a model is built using the other 13, and a prediction is made on the held out protein. This process is repeated for each protein in the set to obtain the evaluation statistics in the results section. LOOCV provides the best estimate of the generalization accuracy of a predictor; however, with larger datasets LOOCV is not practical because of the training time involved in building a model for each member of the dataset. The same LOOCV procedure is applied to template usage in the tertiary structure prediction evaluation. The procedure is also commonly referred to as ‘Jackknife’.
2.3 Template construction
Our tertiary prediction evaluation is performed using SetPRED-TMBB. We created template files by extracting the backbone (N, C, C) coordinates from the monomeric PDB files. The curated (β, –) designations are used to label each residue position in the template. The set contains 2 proteins with 8 strands, 2 with 10 strands, 1 with 12 strands (1QD5), 4 with 16 strands, 2 with 18 strands and 3 with 22 strands. The strand count of the predicted secondary structure is used to select templates for modeling. If the strand count of 1QD5 is correctly predicted, no templates would be available for modeling because of the LOOCV procedure. To account for this, we built a template from one additional 12 stranded protein: 1TLY. Also, if a 14 stranded protein is predicted, no templates would be available; therefore, we built templates from two 14 stranded TMBs: 1T16 and 2F1C. The manually curated designations were not available for these three proteins, so we used the TM segment ranges published in the Orientation of Proteins in Membranes (OPM) database (Lomize et al., 2006). The template set contains no 20 stranded proteins because none are present in the PDB.
|
3 METHODS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 DATA3 METHODS4 RESULTS5 CONCLUSIONACKNOWLEDGEMENTSREFERENCES |
|---|
3.1 Secondary structure prediction
3.1.1 Neural-network implementation
The TMB secondary structure predictor uses specialized neural network architecture called a 1D Recursive Neural Network (1D-RNN). This network architecture has been used for prediction of secondary structure, SSpro (Pollastri et al., 2002), domain boundaries, DOMpro (Cheng et al., 2006b) and disordered regions, DISpro (Cheng et al., 2005). As in the prior applications, the input at each position to the neural network is the profile of the sequences in the NR database aligned to the target sequence using PSI-BLAST (Altschul et al., 1997). It has been the experience of the authors that there is little chance of over-fitting the models because of the weight sharing involved in the 1D-RNN architecture. This feature of the architecture makes it appropriate for the small datasets used in this work.
3.1.2 Two-class prediction (β, –)
For two-class prediction the 1D-RNN is trained on the two-class 1D representation: (β) and (–). When making a prediction, the output from the model is the predicted probability of class membership to each class. The initial predicted secondary structure, Sinitial, consists of the class with higher predicted probability at each position. The first row in Figure 1 contains an example of Sinitial for the TMB protein 1P4T. Since the secondary structure of TMB proteins adhere to consistent construction rules, we perform post-processing on the predicted probabilities to revise the secondary structure prediction. The lengths of β-segments and the different types of loop segments are constrained by minimum and maximum values; however, the length of N and C-terminal (–) segments are left unconstrained. Table S1 in the Supplementary Material contains a summary of the specific values used for the different segment types for each dataset. In the example in Figure 1, the initial secondary structure prediction Sinitial for protein 1P4T violates multiple constraints. To describe the post-processing strategy formally we use the additional notations: N is the number of residues in a sequence, S is any two-class secondary structure that does not violate any of the model constraints, Si is the secondary structure at position i, O is the matrix of predicted probabilities output from the 1D-RNN, Oi,β and Oi,non-β are the predicted probabilities that Si is ‘β’ or ‘–’, respectively. The post-processing objective function is the sum of predicted probabilities for each position of S as defined in Equation (1).
|
| (1) |
sum(Sinitial). To find a Smax, we developed a dynamic-programming (DP) solution that incorporates the parameters of the TMB construction rules. The search guarantees to find a Smax, but the solution may not be unique. Since, we have no objective way to discriminate between two equal scoring predictions this issue is ignored, and the single optimal path returned from the DP search is used as the final Smax.
|
We use the number of β-strands in Smax as the prediction of strand count. During the search for Smax, the DP method saves the value of sum(S) for each value of potential strand count. If the number of strands ‘
’ is provided as an additional constraint, the notation Smax, indicates an optimal S with strands. This information can be useful for assessing the confidence in the predicted secondary structure and corresponding strand count. Table S2 in the Supplementary Material contains a summary of the Smax, results for the proteins in SetPRED-TMBB. For 1QJ8 the gap between Smax,8 (130.4) and the next highest sum Smax,10 (115.2) is 11.7%, whereas for 1A0S the gap between Smax,16 (340.9) and the next highest sum Smax,18 (340.1) is only 0.2%. The larger the gap, the more confident the predictor is in its strand count. For assessing our system, this information is not useful, as the predictor will use the single best Smax; however, this information could be valuable to a user who may decide to build tertiary models from multiple strand counts.
3.1.3 Three-class prediction (M, C, –)
To predict the membrane/channel pattern within the β segments, we trained a separate neural network to predict three classes: M, C and other (–). The architecture for the three-class predictor is the same 1D-RNN architecture used for the two-class predictor. The output of the network is the probability of class membership in each of the three classes. For each β segment predicted in the final two-class prediction Smax, the membrane-channel (M/C) pattern is predicted by choosing the pattern with the higher predicted probability sum. For the example protein, 1P4T, in Figure 1, the first β segment is predicted to be from position 6 to 18. Equation (2) shows the calculation for the sum of predicted probabilities for each pattern.
|
| (2) |
3.2 β-contact prediction
Between two paired anti-parallel β-strands, only every other pair of aligned residues is hydrogen bonded. Residue pairs that are aligned, but not hydrogen bonded to one another, are still considered β-contacts. The DSSP program is used to automatically identify β-contacts in known protein structures. DSSP classifies β-contacts based on inter-residue atomic distances and angles. TMBpro-CON is trained on true β-contacts using a 2D Recursive Neural Network (2D-RNN) (Cheng and Baldi, 2005). TMBpro-CON predicts β-contacts in TMB proteins by first predicting the probability of pairing between all pairs of predicted β-strand residues. For each pair of strands the pseudo-energy (i.e. the sum of the individual predicted pairing probabilities) of all possible strand–strand alignments is calculated. Then, TMBpro-CON utilizes the following rules to restrict the search for acceptable pairings: consecutive strands must pair in anti-parallel fashion; the terminal strands must pair in anti-parallel fashion; the shear number must be between 0 and +4 with respect to the strand count; membrane facing residues must pair with other membrane facing residues and core facing residues must pair with other core facing residues. A dynamic programming method is used to find a set of contact predictions that maximizes the global pseudo-energy while conforming to the construction rules.
3.3 Tertiary structure prediction
TMBpro-3D combines de novo and template-based methods to predict tertiary structure, using a search energy composed of predicted structural feature, physical interaction and statistical terms. The conformational search is performed using simulated annealing with a move set that utilizes whole protein templates and fragment assembly.
3.3.1 Search energy
The search energy used in the conformational search is a linear combination of the following terms:
- Ebeta_pairs—favors formation of predicted β-contacts.
- Emc_pattern—favors predicted M/C pattern using template residue membrane-channel values.
- Eglobular_pairwise—rewards favorable side-chain interactions between predicted non-β positions (Zhang et al., 2003).
- Echain_break—favors close termini proximity at artificial chain break sites.
- Ecentroid_repulsion—penalizes clashes between side-chain centers of mass.
- Evdw_repulsion—penalizes steric clashes between all explicitly modeled atoms using van der Waals radii.
The details of each individual energy term and the corresponding weights are provided in the Supplementary Material.
3.3.2 Template usage
The strand count () of the predicted secondary structure is used to screen for potential templates. Each template with a strand count matching is used to generate an ensemble of models. All models are then ranking according to their energy, and the model with the best search energy is the final tertiary prediction. To allow flexible alignment of each predicted β-segment to its corresponding template segment, TMBpro creates artificial chain breaks at the center of each non-β region, dividing the model into loosely coupled sub-models. The sub-models are allowed to move independently, but their interactions are captured through the global energy function.
Four arrays of variables (
) are used to manage template utilization during the conformational search (see Figs 2 and S2). The model 
is an array containing the xyz coordinates of the backbone atoms (N, C, C), indexed by the residue number i. The template 
is a similar array built for the template protein. The template usage 
is an array of binary variables indicating whether or not 
is used to model 
at each residue position. 
i = 1 indicates that 
is used to model 
i, while 
i = 0 means 
i is modeled by fragment replacement using the fragment library (Simons et al., 1997). The alignment shifts 
is an array of length , where each position is the integer shift between model and template segment relative to center–center alignment. The centers of all model and template segments are aligned, corresponding to 
i = 0 for i = 1, ... , . From these center–center alignments, 
is set to 1 at each predicted β position that aligns to a β-residue in the template, and the rest of 
is set to 0 (Figs 2 and S2). During the search phase the values of 
and 
are modified to explore the use of 
.
|
and 
represent the model and template, respectively. 
controls where the template is used: ‘
’ indicates the position is modeled from 
(
i = 1), whereas ‘f’ means the position is modeled by fragment modeling (
i = 0). The wavy vertical lines mark the chain breaks. The center residue of each segment is boxed to help illustrate the shifts. Initially the centers of segments are aligned (all 
i = 0). In the final model, the first segment is shifted 1 position to the left (
1 = –1) and the second segment is shifted 3 positions to the right (
2 = 3).3.3.3 Move types
The following move types are used in the simulated annealing protocol to search the conformational space:
- Shift Single Segment by k:
i = 
i + k
i = segment index; k 
Z and –max k max;
max = (length of segment i)/2;
- Shift m Consecutive Segments by k:
j = 
j + k, for j = i, ... , i + m – 1
i = starting segment index; m Z and 2 m ;
k Z and –max k max;
max = (length of shortest among m segments)/2;
- Adjust Single Segment Template Usage by k:
l = 
for l = b, ... , b + k – 1
b = index of boundary residue (
b 
b+1);
= 0 (contraction) or = 1 (extension);
k Z and –max k max;
max = number of residues to next boundary;
- Replace with Fragment: use fragment to model
i, ... ,
i+k
i = index of first residue to replace;
k Z and 1 k 9;
This move is applied only to regions where the template is not used (
i, ... , 
i+k = 0).
3.3.4 Conformational search
The space of possible conformations is searched using simulated annealing with a linear cooling schedule and the move-set described above. The search is performed in two distinct phases.
Phase 1 focuses on modeling the TM-segments, while phase 2 focuses on modeling the loops. In phase 1 all move types are used and the weights for Eglobular_pairwise, Echain_break, Ecentroid_repulsion and Evdw_repulsion are set to 0 to allow the search to quickly find a conformation that satisfies the predicted strand constraints (low Ebeta_pairs and Emc_pattern). At the end of phase 1 the values of H are locked, so that the model-template alignments are no longer allowed to change. This reduces the move set in phase 2 to only Adjusting Single Segment Template Usage and Replace with Fragment. In addition, all energy terms are used in phase 2. The search is run with different random seeds to generate an ensemble of predicted models, equally utilizing the available templates. The model with the lowest final search energy is returned as the tertiary structure prediction.
|
4 RESULTS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 DATA3 METHODS4 RESULTS5 CONCLUSIONACKNOWLEDGEMENTSREFERENCES |
|---|
To assess our secondary structure prediction, we compare it to the published results of transFold (Waldispühl et al., 2006a) and PRED-TMBB (Bagos et al., 2004a). To assess our β-contact prediction we compare it to the published results of transFold, and to the server output in order to make a comparison using the same predicted secondary structure as input. To the best of our knowledge, TMBpro-3D is the first publicly available method to predict the structure of TMB proteins without relying on sequence–sequence, sequence–profile or profile–profile alignments for template usage; thus, we do not compare out tertiary prediction results to previous work.
4.1 Secondary structure prediction results
As described previously, we developed a two-class (β,–) secondary structure predictor specialized for TMB proteins. Using the two-class predictions, we predict the three-class (M, C, –) and infer four-class predictions (M, C, i, o). We developed two separate secondary structure predictors using the non-redundant datasets
SetTransfold and SetPRED-TMBB to make comparisons with the related methods.
4.1.1 Secondary structure evaluation metrics
To assess secondary structure prediction performance we use the following per-residue metrics: the two-class accuracy (Q2), three-class (M, C, –) accuracy (Q3), Mathews correlation coefficient (MCC) (Baldi et al., 2000), and segment overlap measure (SOV) (Zemla et al., 1999).We include the SOV measure for completeness, but no SOV results were provided in the studies we compare to. In addition to these common measures, we use additional measures from previous work for the sake of comparison. For comparison to transFold, we also include the per-segment recall (sensitivity) 
and precision 
, with correct prediction defined as an observed β-strand intersecting exactly one predicted β-strand, and vice versa (Waldispuhl et al., 2006a). The per-segment measures for comparison to PRED-TMBB include the number of true positives (TP), the number of false negatives (FN) and the number of false positives (FP). In addition, we include the number of correctly predicted topologies (TOP), that is when all strands and loops have been predicted correctly according to Bagos et al. (2004a).
4.1.2 Results using SetTransfold
Table 1 contains a summary of TMBpro-SS secondary structure prediction results compared to transFold. We use LOOCV on SetTransfold to assess our method and compare it to transFold. TMBpro-SS outperforms transFold significantly using the Q2 (77.84–69.91%) and MCC (0.538–0.380) measures. TMBpro-SS performs slightly better than transFold, according to the per-segment measures 
and 
.
|
4.1.3 Results using SetPRED-TMBB
Table 2 contains a summary of TMBpro-SS secondary structure prediction results compared to PRED-TMBB. We use the same LOOCV (Jackknife) procedure as the authors of the PRED-TMBB method, on the same set of proteins, to make the comparison as objective as possible. Of the 214 annotated β-strands PRED-TMBB correctly predicts 203, while TMBpro-SS correctly predicts 204. PRED-TMBB makes 13 false positive predictions (FP), while TMBpro-SS only makes 6. Using the TOP measure of correct topology prediction PRED-TMBB correctly predicts 8 topologies, while TMBpro-SS succeeds on 11. TMBpro-SS also outperforms PRED-TMBB according to the Q2 (88.3–84.2%) and MCC (0.751–0.720) measures. When comparing TMBpro-SS to itself between datasets, it has significantly higher Q2, Q3, MCC and SOV when using SetPRED-TMBB (see Tables 1 and 2). It is unclear how much of this difference is due to the five proteins that differ between the sets, and how much is due to the different types of annotation of the training data. The Q2, Q3, MCC and SOV results for individual proteins are displayed with the detailed tertiary prediction results in Table 4.
|
4.2 β-Contact prediction results
The input to TMBpro-CON is the amino acid sequence and a two-class secondary structure. Using SetTransfold, we performed β-contact prediction with three different sets of two-class secondary structure: (1) predicted by transFold server, (2) predicted by TMBpro-SS and (3) DSSP designations. We compare our results using (1) to the β-contacts predicted by the transFold server. We compare our results using (2) to the transFold published results. Using SetPRED-TMBB, we performed β-contact prediction with two sets of two-class secondary structure: predicted by TMBpro-SS and hand curated annotations from Bagos et al. (2004a). No comparison to other work is made using SetPRED-TMBB since PRED-TMBB does not predict β-contacts.
4.2.1 β-Contact evaluation metrics
For evaluation of β-contact prediction, the authors of transFold introduce the concept of a compatible pair of residues to allow contact predictions that are nearly correct to be counted. Consider a pair (i, j) to be a true β-residue pairing. The contact pairs (i, j) and (m, n) are considered to be compatible if, for a given integer , (i, j) = (m ± , n ± ). In their work they use a value of = 2 for evaluation. For our assessment we use = 2 and = 0, where only exact pairing predictions are counted. The measures we use for assessment are precision and recall. The precision is calculated by (number of correct β-contact predictions/total number of β-contact predictions) and recall by (number of correct β-contact predictions/total number of true β-contacts).
4.2.2 Results using SetTransfold
A summary of β-contact prediction results for both protein sets and all secondary structure sets is available in Table 3. Using the same secondary structure as input (the predicted secondary structure from the transFold server) TMBpro-CON performs slightly better than the transFold server by all measures. Using the predicted secondary structure from TMBpro-SS as input, the TMBpro-CON prediction results are significantly better than transFold server results and published results according to all measures. Using the DSSP assigned secondary structure as input, TMBpro-CON predicts exact β-contacts with precision 0.478 and recall 0.520. These results demonstrate the upper bound in β-contact prediction accuracy of TMBpro-CON given improvements in secondary structure prediction only.
|
4.2.3 Results using SetPRED-TMBB
Taking the predicted secondary structure from TMBpro-SS trained on SetPRED-TMBB as input, TMBpro-CON predicts exact β-contacts with precision 0.414 and recall 0.407. These values are significantly higher than the corresponding prediction using SetTransfold (see Table 3). This difference can be accounted for by the more accurate secondary structure predictions for SetPRED-TMBB. The β-contact recall results for the individual proteins are shown in the tertiary results Table 4.
|
4.3 Tertiary structure prediction results
Here, we evaluate the tertiary structure predictions of TMBpro-3D for SetPRED-TMBB using secondary structure and β-contacts predicted by TMBpro. We chose SetPRED-TMBB rather than SetTransfold for tertiary prediction experiments because of the stronger secondary structure and β-contact prediction results. Only the model with the lowest search energy is evaluated.
4.3.1 Tertiary structure evaluation metrics
The two measures we use to evaluate tertiary predictions are root-mean-square deviation (RMSD) and global distance test total score (GDT_TS) reference for GDT_TS measure (Zemla, 2003). The latter has been used as the primary numeric measure in recent critical assessment of methods of protein structure prediction (CASP) experiments (Moult et al., 2005). The TM notation is used as a subscript to indicate that the measure is calculated on only the TM segments of the true structure compared to the model.
4.3.2 Prediction results
The tertiary structure prediction results for each protein in SetPRED-TMBB are displayed in Table 4. The best prediction, in terms of the GDT_TS and RMSD on the whole structure is made on the protein with the second highest β-contact recall: 1QJP
[PDB]
. The β-contact recall is 0.65, the GDT_TS is 57.3 and RMSD is 4.3 Å. The GDT_TSTM is 68.3 and RMSDTM is 3.0 Å. The next best whole structure predictions are for proteins 1QJ8
[PDB]
(52.0, 5.5 Å), 1PRN (50.0, 7.1 Å) and 1E54 (49.3, 7.7 Å). The Supplementary Material contains a superposition file (1QJ8_pred.pdb) and an image (Fig. S1) showing the predicted structure for 1QJ8 aligned to the PDB structure. For several proteins the GDT_TSTM results are strong. For proteins 1QJ8
[PDB]
, 1QJP, 1PRN, 1I78, 1E54, 2OMF and 1FEP the GDT_TSTM is greater than 60.0. These predictions correspond to correct topology predictions and high β-contact recall when compared to the other predictions. The significantly lower GDT_TS and higher RMSD scores on the whole structures reflect the difficulty of modeling long loop regions and core domains folded inside the larger proteins.
The worst whole structure and TM segment predictions are made on proteins 1A0S [PDB] and 2MPR, both of which have true strand counts of 18, but are modeled using 16-stranded templates because of incorrect secondary structure topology predictions. Additionally, the locations of multiple strands in the 2POR prediction are incorrect resulting in an incorrect topology according to the TOP measure. The worst whole structure and TM segment prediction for a protein with correct topology prediction was made on the 10-stranded protein 1K24 [PDB] . The topology is correct using the TOP measure; however, the locations of the sixth and seventh strands are off by seven residues. Using a slightly stricter standard for topology assessment, this prediction would be considered an incorrect topology. From these results it is clear that the correct topology is necessary to build a reasonable tertiary model.
4.3.3 Self-consistency results
To evaluate the self-consistency of TMBpro, we provided the curated secondary structure and true β-contacts as input to the program. The performance was assessed both allowing and disallowing the inclusion of the native template among the available templates, and the results are displayed in the rightmost section of Table 4. When the native template is included, TMBpro always recovers the true structure (see the last column in Table 4). When the native template is not included, the RMSDTM results range from 1.5 to 4.5 Å. For 12 of 14 predictions, the RMSDTM is <2.8 Å. The only two exceptions are proteins 2FCP
[PDB]
, with an RMSDTM of 3.5 Å, and 1QD5, with an RMSDTM of 4.5 Å. At 723 residues 2FCP is one of the longest proteins in the set, so a slightly higher error is not surprising. 1QD5 is only 269 residues, but contains an irregular bulge in the first strand that is not present in its only available template (1TLY).
|
5 CONCLUSION |
|---|
|
TOPABSTRACT1 INTRODUCTION2 DATA3 METHODS4 RESULTS5 CONCLUSIONACKNOWLEDGEMENTSREFERENCES |
|---|
TMB proteins have clear biological and medical relevance. Due to their importance and the difficulty of experimentally determining their structures, accurate tertiary structure prediction of TMB proteins is an important task for the protein structure prediction community. Traditional homology modeling methods will perform well if the target protein is similar enough to a solved protein to create a quality alignment; however, for the vast majority of putative TMB proteins traditional homology modeling will fail. The construction rules TMB proteins follow provide a greatly reduced search space compared to the globular protein structure prediction problem. In this work, we demonstrated a methodology for predicting secondary structure, β-contacts and tertiary structure of TMB proteins. The tertiary structure predictor does not rely on sequence similarity between target and template. The performance of TMBpro compares favorably to other publicly available predictors. The TMBpro server, trained on all 14 proteins in SetPRED-TMBB, is publicly available at: http://www.igb.uci.edu/servers/psss.html.
|
ACKNOWLEDGEMENTS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 DATA3 METHODS4 RESULTS5 CONCLUSIONACKNOWLEDGEMENTSREFERENCES |
|---|
Work supported by NIH grant LM-07443-01, NSF grants EIA-0321390 and IIS-0513376, and a Microsoft Faculty Research Award to P.B., and a UCF faculty start-up grant to J.C.
Conflict of Interest: none declared.
|
FOOTNOTES |
|---|
Associate Editor: Alfonso Valencia
Received on July 11, 2007; revised on October 6, 2007; accepted on October 29, 2007
|
REFERENCES |
|---|
|
TOPABSTRACT1 INTRODUCTION2 DATA3 METHODS4 RESULTS5 CONCLUSIONACKNOWLEDGEMENTSREFERENCES |
|---|
Altschul S, et al. Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res (1997) 25:3389–3402.
Bagos P, et al. A hidden markov model method, capable of predicting and discriminating beta-barrel outer membrane proteins. BMC Bioinformatics (2004a) 5:29.[CrossRef][Medline]
Bagos P, et al. PRED-TMBB: a web server for predicting the topology of beta-barrel outer membrane proteins. Nucleic Acids Res (2004b) 32:W400–W404.
Bagos P, et al. Evaluation of methods for predicting the topology of beta-barrel outer membrane proteins and a consensus prediction method. BMC Bioinformatics (2005) 6:7.[CrossRef][Medline]
Baldi P, et al. Assessing the accuracy of prediction algorithms for classification: an overview. Bioinformatics (2000) 16:412–424.
Baldi P, Pollastri G. The principled design of large-scale recursive neuralnetwork architectures-DAG-RNNs and the protein structure prediction problem. J. Mach. Learn. Res (2003) 4:575–602.[CrossRef][Web of Science]
Berman H, et al. The Protein Data Bank. Nucleic Acids Res (2000) 28:235–242.
Bigelow H, Rost B. PROFtmb: a web server for predicting bacterial transmembrane beta barrel proteins. Nucleic Acids Res (2006) 34:W186–W188.
Bigelow H, et al. Predicting transmembrane beta-barrels in proteomes. Nucleic Acids Res (2004) 32:2566–2577.
Casadio R, et al. In silico prediction of the structure of membrane proteins: is it feasible? Brief. Bioinformatics (2003) 4:341–348.
Cheng J, Baldi P. Three-stage prediction of protein beta-sheets by neural networks, alignments, and graph algorithms. Bioinformatics (2005) 21(Suppl. 1):i75–i84.[Abstract]
Cheng J, et al. Accurate prediction of protein disordered regions by mining protein structure data. Data Mining Knowl. Discov (2005) 11:213–222.[CrossRef]
Cheng J, et al. Large-scale prediction of disulphide bridges using kernel methods, two-dimensional recursive neural networks, and weighted graph matching. Proteins (2006a) 62:617–629.[CrossRef][Web of Science][Medline]
Cheng J, et al. DOMpro: protein domain prediction using profiles, secondary structure, relative solvent accessibility, and recursive neural networks. Data Mining Knowl. Discov (2006b) 13:1–10.[CrossRef]
Diederichs K, et al. Prediction by a neural network of outer membrane beta-strand protein topology. Protein Sci (1998) 7:2413–2420.[Web of Science][Medline]
Fariselli P, et al. A new decoding algorithm for hidden Markov models improves the prediction of the topology of all-beta membrane proteins. BMC Bioinformatics (2005) 6:S12.
Garrow A, et al. TMB-Hunt: an amino acid composition based method to screen proteomes for beta-barrel transmembrane proteins. BMC Bioinformatics (2005) 6:56.[CrossRef][Medline]
Gromiha M, Suwa M. A simple statistical method for discriminating outer membrane proteins with better accuracy. Bioinformatics (2005) 21:961–968.
Gromiha M, et al. Identification of membrane spanning beta strands in bacterial porins. Protein Eng (1997) 10:497–500.
Gromiha M, et al. Neural network-based prediction of transmembrane beta-strand segments in outer membrane proteins. J. Comput. Chem (2004) 25:762–767.[CrossRef][Web of Science][Medline]
Gromiha M, et al. TMBETA-NET: discrimination and prediction of membrane spanning beta-strands in outer membrane proteins. Nucleic Acids Res (2005) 33:W164–W167.
Ikeda M, et al. TMPDB: a database of experimentally-characterized transmembrane topologies. Nucleic Acids Res (2003) 31:406–409.
Jackups R, Liang J. Interstrand pairing patterns in beta-barrel membrane proteins: the positive-outside rule, aromatic rescue, and strand registration prediction. J. Mol. Biol (2005) 354:979–993.[CrossRef][Web of Science][Medline]
Jacoboni I, et al. Prediction of the transmembrane regions of beta-barrel membrane proteins with a neural network based predictor. Protein Sci (2001) 10:779–787.[CrossRef][Web of Science][Medline]
Kabsch W, Sander C. Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolymers (1983) 22:2577–2637.[CrossRef][Web of Science][Medline]
Koebnik R, et al. Structure and function of bacterial outer membrane proteins: barrels in a nutshell. Mol. Microbiol (2000) 37:239–253.[CrossRef][Web of Science][Medline]
Liu Q, et al. A HMM-based method to predict the transmembrane regions of beta-barrel membrane proteins. Comput. Biol. Chem (2003) 27:69–76.[CrossRef][Web of Science][Medline]
Lomize M, et al. OPM: orientations of proteins in membrane database. Bioinformatics (2006) 22:623–625.
Martelli P, et al. A sequence-profile-based hmm for predicting and discriminating beta barrel membrane proteins. Bioinformatics (2002) 18:S46–S53.[Abstract]
Moult J, et al. Critical assessment of methods of protein structure prediction (CASP) – Round 6. Proteins (2005) 61(Suppl. 7):3–7.[CrossRef][Web of Science][Medline]
Natt N, et al. Prediction of transmembrane regions of beta-barrel proteins using ANN- and SVM-based methods. Proteins (2004) 56:11–18.[CrossRef][Web of Science][Medline]
Oberai A, et al. A limited universe of membrane protein families and folds. Protein Sci (2006) 15:1723–1734.[CrossRef][Web of Science][Medline]
Park K, et al. Discrimination of outer membrane proteins using support vector machines. Bioinformatics (2005) 21:4223–4229.
Paul C, Rosenbusch J. Folding patterns of porin and bacteriorhodopsin. EMBO J (1985) 4:1593–1597.[Web of Science][Medline]
Pollastri G, Baldi P. Prediction of contact maps by GIOHMMs and recurrent neural networks using lateral propagation from all four cardinal corners. Bioinformatics (2002) 18:S62–S70.[Abstract]
Pollastri G, et al. Improving the prediction of protein secondary structure in three and eight classes using recurrent neural networks and profiles. Proteins (2002) 47:228–235.[CrossRef][Web of Science][Medline]
Schulz G. Beta-barrel membrane proteins. Curr. Opin. Struct. Biol (2000) 10:443–447.[CrossRef][Web of Science][Medline]
Simons KT, et al. Assembly of protein tertiary structures from fragments with similar local sequences using simulated annealing and Bayesian scoring functions. J. Mol. Biol (1997) 268:209–225.[CrossRef][Web of Science][Medline]
Skolnick J, et al. Monsster: a method for folding globular proteins with a small number of distance restraints. J. Mol. Biol (1997) 265:217–241.[CrossRef][Web of Science][Medline]
Tamm L, et al. Structure and assembly of beta-barrel membrane proteins. J. Biol. Chem (2001) 276:32399–32402.
Tamm L, et al. Folding and assembly of beta barrel membrane proteins. Biochim. Biophys. Acta (2004) 1666:250–263.[Medline]
Waldispühl J, et al. Predicting transmembrane beta-barrels and interstrand residue interactions from sequence. Proteins (2006a) 65:61–74.[CrossRef][Web of Science][Medline]
Waldispühl J, et al. transFold: a web server for predicting the structure and residue contacts of transmembrane beta-barrels. Nucleic Acids Res (2006b) 34:W189–W193.
Wallin E, von Heijne G. Genome wide analysis of integral membrane proteins from eubacterial, archaean, and eukaryotic organisms. Protein Sci (1998) 7:1029–1038.[Web of Science][Medline]
Welte W, et al. Prediction of the general structure of OmpF and PhoE from the sequence and structure of porin from Rhodobacter capsulatus. Orientation of porin in the membrane. Biochim. Biophys. Acta (1991) 1080:271–274.[CrossRef][Medline]
Wimley W. Toward genomic identification of beta-barrel membrane proteins: composition and architecture of known structures. Protein Sci (2002) 11:301–312.[CrossRef][Web of Science][Medline]
Wimley W. The versatile beta-barrel membrane protein. Curr. Opin. Struct. Biol (2003) 13:404–411.[CrossRef][Web of Science][Medline]
Yooseph S, et al. The Sorcerer II Global Ocean Sampling expedition: expanding the universe of protein families. PLoS Biol (2007) 5:e16.[CrossRef][Medline]
Zemla A, et al. A modified definition of sov, a segment-based measure for protein secondary structure prediction assessment. Proteins (1999) 34:220–223.[CrossRef][Web of Science][Medline]
Zemla A. LGA: a method for finding 3D similarities in protein structures. Nucleic Acids Res (2003) 31:3370–3374.
Zhai Y, Saier M. The beta-barrel finder (BBF) program, allowing identification of outer membrane beta-barrel proteins encoded within prokaryotic genomes. Protein Sci (2002) 11:2196–2207.[CrossRef][Web of Science][Medline]
Zhang T, et al. TOUCHSTONE:II a new approach to ab initio protein structure prediction. Biophys. J (2003) 85:1145–1164.[Web of Science][Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||