Homology Modelling GLA

From Bioinformatikpedia
Revision as of 03:54, 24 June 2011 by Drexler (talk | contribs) (3D-JIGSAW)

by Benjamin Drexler and Fabian Grandke

Introduction

In this task, we performed homology modelling of the protein α-galactosidase A with the programs MODELLER, SWISS-MODEL, iTasser and 3D-JIGSAW. Homology modelling relies on the following two assumptions. First, the structure of the protein is determined by its amino acid sequence. Second, the structure of a protein is more conserved than its amino acid sequence. Usually one performs homology modelling of a protein which structure is not known. In this case, we have several PDB structures of the α-galactosidase A available and hence we are able to evaluate the resulting models of the programs afterwards.

General

Template Selection

The following table lists the best ten hits of the HHpred search of Task 1. We used 3HG3 (97% identity), 1KTB (53%) and 3CC1 (34%) as templates for the modelling process. This selection covers a wide range of sequence identity and hence we are able to evaluate how the sequence identity influence the quality of the models.

PDB-ID Name Probability E-value P-value Identity Template
> 60% sequence identity
3hg3_A Alpha-galactosidase A 1.0 0 0 97% x
> 40% sequence identity
1ktb_A Alpha-N-acetylgalactosaminidase 1.0 0 0 53% x
< 40% sequence identity
1uas_A Alpha-galactosidase 1.0 0 0 39%
3lrk_A Alpha-galactosidase 1 1.0 0 0 32%
3a5v_A Alpha-galactosidase 1.0 0 0 35%
1szn_A Alpha-galactosidase 1.0 0 0 34%
3a21_A Putative secreted alpha-galactosidase 1.0 0 0 34%
3cc1_A BH1870 protein 1.0 0 0 26% x
3a24_A Alpha-galactosidase 1.0 0 0 14%
1zy9_A Alpha-galactosidase 1.0 2.2E-37 8.8E-42 14%

Evaluation

Figure 1: The PDB structures 1R46 (red) and 1R47 (green) in cartoon representation.

The evaluation of the models consist of two parts, i.e. a visual comparison with an experimental structure and a numeric evaluation. The PDB structures 1R46 and 1R47 were used for the evaluation. 1R46 is a structure of human α-galactosidase A without galactose (apo) and the structure 1R47 contains galactose (complexed). Both, the visual comparison and the numeric evaluation, were ony done with the chain A of the structures, because the model programs also modelled one chain.

The differences between 1R46 and 1R47 are very marginal (see figure 1) and hence we did the visual comparison with one structure, i.e. 1R47.

The numeric evaluation involves the calculation of several scores.

RMSD

The root mean square deviation (RMSD) value between the model and the reference structure was calculated by the webserver of TM-align.

The calculation of the RMSD in the catalytic site was done by PyMol. We used the annotation of the UniProt entry to determine the active sites, which are Asp170 and Asp231. We applied the following workflow:

  1. Import the reference structure, e.g. 1R47
  2. Select the residues of the active site, i.e. Asp170 and Asp231
  3. Expand the selection with modify -> expand -> by 6A, residues
  4. Rename this selection to "selection_ref"
  5. Import the model
  6. Align the model to the reference structure (align -> to molecule -> 1R47)
  7. Select the residues of the active site of 1R47
  8. Expand the selection once again, but exclude residues of 1R47 (modify -> exluce -> object -> 1R47) afterwards
  9. Rename this selection to "selection_model"
  10. Align "selection_model" to "selection_ref" with align -> to selection and retrieve the RMSD

TM-Score

We also used the webserver of TM-align for the calculation of the TM-Score. The model was always used as 'Structure 1' and the reference was 'Structure 2'. This was done, because 'Structure 2' will be used for the normalization of the TM-Score and hence we always normalize with the same number of residues.

At first, we used the command line TMS to calculate the TM-Score, but the values seemed to be wrong, i.e. way too low.

QMEAN Score

SWISS-MODEL provides also two QMEAN scores (QMEANscore4 and QMEAN Z-score). QMEAN is score to describe the model quality and consist of the five following structural descriptors. For a more detailed explanation, please read the help of SWISS-MODEL.

QMEANscore4 ranges from 0 to 1 and indicates the reliability of the model. The Z-score describes the absolute quality of the model and is calculated by comparison to reference structures.

C-Score

The C-Score is calculated by I-TASSER and is a confidence score that is based on the significane of the alignments and the convergence parameters of the structure assembly simulations. The C-Score typically ranges from -5 to 2 and a high score indicates a model with a high confidence.<ref name=itasser_cscore>C-Score explanation</ref>

Calculation of Models

MODELLER

Figure 2: Representation of the resulting models of MODELLER and the reference PDB structure 1R47. The models are in superposition to the reference structure (green) and are shown in cartoon representation. (A) The model is based on the PDB structure 3HG3 (red). (B) The model is based on the PDB structure 1KTB (blue). (C) The model is based on the PDB structure 3CC1 (magenta).

MODELLER is a program to produce three-dimensional protein structures based on homology or comparative modelling. The user has to provide the sequence of the protein to be modeled and the structure and sequence of at least one related protein that is used as a template. MODELLER uses all atoms of the template protein, but the hydrogen-atoms. We used MODELLER as described in the tutorial Using Modeller for TASK 4. Therefor we had to align both sequences and convert them into pir-format. This alignment is given as input together with the template pdb-file. Unfortunately the input file has to be provided as python file. <ref name=modeller>http://salilab.org/modeller/</ref>

Pairwise Alignments

In this section, we used a pairwise alignment between the template (i.e. 3HG3, 1KTB and 3CC1) and the target as the input for MODELLER. All three models fairly match the structure of 1R47 (see figure 2). The model of 3HG3 seems to be the best (see figure 2A), closely followed by the model of 1KTB (see figure 2B). The largest deviations in respect to the reference structure can be observed in the model of 3CC1 (see figure 2C), especially in the coil regions of the protein.

The numeric evaluation confirms these observations. The differences of the RMSD values and the TM-Scores of the model by 3HG3 and 1KTB are very close in all columns. The results of 3CC1 suggest that the quality of the model is worse, but still very decent considering that a TM-Score above 0.5 indicates two structures with the same fold. In this case, it seems like that the difference of about 40% sequence identity between 3HG3 and 1KTB does not affect the quality of the model which is very interesting. In contrast the difference of 20% between 1KTB and 3CC1 leads to an quite observable loss in the quality of the model. It is also remarkable that the RMSD values of the catalytic site are lower than the overall RMSD value of 3HGH3 and 1KTB. So it seems like that their is an increase of the quality in the active site.


Apo (1R46) Complexed (1R47)
Template TM-Score RMSD RMSD catalytic site TM-Score RMSD RMSD catalytic site
3HG3 0.99473 0.53 0.326 0.99459 0.53 0.366
1KTB 0.96933 1.44 0.439 0.96977 1.43 0.437
3CC1 0.78243 3.26 3.436 0.78291 3.26 3.405


Multiple Sequence Alignments

Additionally to the pairwise approach we used a multiple alignment as template for the model. Therefor we created an alignment of the sequences, provided in the table below. Then we added the target sequence to the alignment and supervised it. The supervision showed, that the sequences aligned very well in general, but the sequences 3LRK_A and 3CC1_A. Thus, those were removed and the alignment was realigned. Both, the supervised and the unsupervised alignment have been used as input for MODELLER. The table below shows the used sequences. The green color indicates, that the sequence was contained in the MSA. The red color indicates the opposite.

Figure 3: Representation of the resulting MSA models of MODELLER and the reference PDB structure 1R47. The models are in superposition to the reference structure (green) and are shown in cartoon representation. (A) The model is based on a MSA (orange). (B) The model is based on a MSA which was modified/supervised (yellow).
PDB-ID Unsupervised Supervised Identity Comment
3LX9_A 99%
3GXP_A 99%
3H53_A 99%
3HG3_A 97%
3IGU_A 54%
1KTB_A 53%
1UAS_A 39%
3LRK_A 34% Was removed due to little sequence identity. Caused huge gaps in alignment.
3CC1_A 28% Was removed due to little sequence identity. Caused huge gaps in alignment.


Both, the model of the unsupervised and supervised MSA, perform very good. They fit the reference structure very good and there are almost no deviations (see figure 3). Considering the numeric evaluation, the MSA models are even better than the model of the 3HG3 template. It is noteworthy that the exclusion of the two sequences with a low sequence identity slightly decreased the quality of the model. So it seems like that there is a benefit in including sequences with a low identity. Another explanation could be that the unsupervised had two more sequences in the MSA and hence the quality of the model is better. After all, the results indicate that it is a viable option to do the modelling with an MSA.

Apo (1R46) Complexed (1R47)
Type TM-Score RMSD RMSD catalytic site TM-Score RMSD RMSD catalytic site
unsupervised 0.99563 0.47 0.330 0.99570 0.47 0.314
supervised 0.99521 0.50 0.335 0.99537 0.49 0.329

iTasser

Figure 4: A schematic representation of the I-TASSER protocol for protein structure and function predictions. The protein chains are colored from blue at the N-terminus to red at the C-terminus.<ref name=itasser_roy>Ambrish Roy, Alper Kucukural, Yang Zhang. I-TASSER: a unified platform for automated protein structure and function prediction. Nature Protocols, vol 5, 725-738 (2010).</ref>
Figure 5: The resulting models of I-TASSER in carton representation. The models are in superposition to the reference PDB structure 1R47 (green). (A) Model 1 (red) of the run, in which I-TASSER was restricted to the PDB-ID 3HG3. (B) Model 2 (blue) with restriction to 3HG3. (C) Model 1 (yellow) of the run, in which I-TASSER was restricted to the PDB-ID 1KTB. (D) Model 2 (silver) with restriction to 1KTB. (E) Model 1 (cyan) of the run, in which I-TASSER was restricted to the PDB-ID 3CC1. (B) Model 2 (orange) with restriction to 3CC1. (G) Model 1 (lightorange) of the run, in which I-TASSER was not restricted to any PDB ID. (B) Model 2 (lightpink) with no restriction to a PDB ID.

Figure 4 shows, that iTasser takes an amino acid sequence as input and tries to retrieve template proteins from PDB. In the next step fragments from the the templates are reassembled to a complete model. In the last step, the model is reassembled by taking energy calculations into account. Additionally biological function prediction is done, but that was not of interest of this task<ref name=itasser1>http://zhanglab.ccmb.med.umich.edu/I-TASSER/about.html</ref>. I-TASSER was developed by Zhang et. al in 2007<ref name=itasser_zhang>Yang Zhang. Template-based modeling and free modeling by I-TASSER in CASP7. Proteins, vol 69 (Suppl 8), 108-117 (2007)</ref><ref name=itasser_roy>Ambrish Roy, Alper Kucukural, Yang Zhang. I-TASSER: a unified platform for automated protein structure and function prediction. Nature Protocols, vol 5, 725-738 (2010).</ref>.

We used the iTasser-server in two different ways:

  1. Standard parameters: the protein sequence is given as input and the program searches PDB for templates. The found proteins are used to create a template to predict the structure. No further arguments are given as input.
  2. PDB-ID as input: together with the amino acid sequence a template PDB-ID is given as input. Therefore the second input field in the "Option 1" dropout menu was filled with the certain PDB ID, as explained here. The program takes all available information into account and uses them to calculate the structure.

We used 3HG3, 1KTB and 3CC1 as the PDB-ID input in run 1, 2 and 3 respectively. Run 4 was performed with the standard parameteres. It seems like that these chosen options did not influence the outcome in our case which could be an error in the usage of I-TASSER on our side.

We received in every run two models as a result. Acoording to the visual (see figure 5) and numeric evaluation, the first result (referred as model 1) of every run is almost identical and the same applies for the second result (referred as model 2). Model 2 is always worse than model 1, but is still pretty decent. A closer look at the visual representation reveals, that these models always have a problem with four beta-sheets (on the left side) that are instead modelled as a coil region.

Overall the results are very good, but this is probably due to the fact that we were not able to exclude the selfhit even though we restricted the modelling to a certain template.


Apo (1R46) Complexed (1R47) Independent
Run Model TM-Score RMSD RMSD catalytic site TM-Score RMSD RMSD catalytic site C-Score
1 1 0.99209 0.65 0.309 0.99236 0.64 0.353 1.883
1 2 0.87082 2.36 1.022 0.87069 2.35 0.617 -0.314
2 1 0.99138 0.68 0.334 0.99173 0.67 0.327 1.870
2 2 0.87165 2.34 0.640 0.87137 2.33 0.634 -0.327
3 1 0.99077 0.70 0.328 0.99101 0.69 0.331 1.883
3 2 0.87192 2.33 0.616 0.87173 2.32 0.600 -0.315
4 1 0.99072 0.70 0.333 0.99099 0.69 0.334 1.883
4 2 0.87137 2.34 0.625 0.87104 2.33 0.645 -0.314

SWISS-MODEL

Figure 6: Representation of the resulting models of SWISS-MODEL which used the PDB structure 3HG3 as a template the reference PDB structure 1R47. The models are in superposition to the reference structure (green) and are shown in cartoon representation. (A) The model (cyan) was build by using the aligned mode of SWISS-MODEL. (B) The model (yellow) was build by using the aligned mode of SWISS-MODEL.
Figure 7: Representation of the resulting models of SWISS-MODEL which used the PDB structure 1KTB as a template and the reference PDB structure 1R47. The models are in superposition to the reference structure (green) and are shown in cartoon representation. (A) The model (red) was build by using the aligned mode of SWISS-MODEL. (B) The model (blue) was build by using the aligned mode of SWISS-MODEL.
Figure 8: Representation of the resulting model of SWISS-MODEL which used the PDB structure 3CC1 as a template and the reference PDB structure 1R47. The model is in superposition to the reference structure (green) and are shown in cartoon representation. It was not possible to build a model with the automated mode of SWISS-MODEL and hence there is only a model of the aligned mode (magenta).

We used the swissmodel server with two different options:

  1. Automated Mode: The target sequence is given as input together with the PDB ID and chain name of the template protein as described in the help page. The information about the target are given in the advanced options area below the input field of the target sequence. This method should only be used, if the sequence identity between target and template is greater than 50%.
  2. Aligned Mode: A pairwise alignment of template and target sequence is given as input. We created our alignments using online ClustalW2 from EBI. There are no advanced options to specify the prediction, but several input formats are valid.

Template 3HG3

In this section, we deal with the model based on 3HG3 as the template. Both models, the aligned and the automated one, match close to perfect the reference structure (see figure 6). The numeric evaluation confirms these obversations. The TM-Score is close to 1 which indicates a very good model. Additionally, the values suggest that the two models are almost identical. Hence in this case, it does not make a difference whether the aligned or the automated mode was used.

Apo (1R46) Complexed (1R47) Independent
Mode TM-Score RMSD RMSD catalytic site TM-Score RMSD RMSD catalytic site QMEAN Z-Score QMEANscore4
Aligned 0.99504 0.51 0.278 0.99497 0.51 0.290 -0.415 0.74
Automated 0.99504 0.51 0.277 0.99497 0.51 0.291 -0.415 0.74

The additional graphics of SWISS-MODEL are provided on this page.

Template 1KTB

In contrast to the template 3HG3, there is a difference between the model of the automated and aligned mode obversable. The latter shows huge deviations in respect to the reference structure (see figure 7A), whereas the model of the automated mode matches the reference structure almost perfectly (see figure 7B). Even though there are differences to the reference structure by the model of the aligned mode, the TM-Score is very good. Nonetheless the TM-Score of the automated mode is even higher and even further the RMSD values are also better. The most remarkable thing is the extraordinary RMSD of the catlytic site in comparison to the overall result which indicates a bad quality of the model nearby the catalytic site. So in this case, the automated mode would be preferable.

Apo (1R46) Complexed (1R47) Independent
Mode TM-Score RMSD RMSD catalytic site TM-Score RMSD RMSD catalytic site QMEAN Z-Score QMEANscore4
Aligned 0.84410 2.17 5.073 0.84472 2.15 6.409 -12.996 -0.022
Automated 0.96507 1.22 0.417 0.96559 1.21 0.404 -2.742 0.599

The additional graphics of SWISS-MODEL are provided on this page.

Template 3CC1

Unfortunately SWISS-MODEL was not able to perfom the automated modelling with 3CC1 as template due to the low sequence identity. We received the following message:
building model based on 3cc1A (1-390) was not successfull go to next best template

Hence we are only able to evaluate the result of the aligned mode. In comparison to the results of the other two templates, the model of the 3CC1 template has the greatest differences to the reference structure (see figure 8). Nevertheless the TM-Score is still pretty decent. The RMSD values are overall the highest and just like in the model of the 1KTB template, the RMSD value of the catalytic site increases. It is clearly recognizable that the quality of the model suffers from the low sequence identity.

Apo (1R46) Complexed (1R47) Independent
Mode TM-Score RMSD RMSD catalytic site TM-Score RMSD RMSD catalytic site QMEAN Z-Score QMEANscore4
Aligned 0.60954 3.47 7.107 0.61051 3.53 7.357 -14.046 -0.067
Automated N/A N/A N/A N/A N/A N/A N/A N/A

The additional graphics of SWISS-MODEL are provided on this page.

3D-JIGSAW

Figure 9: Two resulting models of 3D-JIGSAW in carton representation. The models are in superposition to the reference PDB structure 1R47 (green). (A) One of the five models which was created based on 'good models' (blue). (B) One of the five models which was created based on 'mediocre models' (red).

Because we included also a template with a high sequence identity, we received models which were very accurate through out all three programs (MODELLER, I-TASSER, SWISS-MODEL). Since their TM-Score was already very close to 1 and the RMSD values close to 0, 3D-JIGSAW is probably not able to improve these models anymore. We decided that we also compile a second set of models that were mediocre and hence we will be able to evaluate whether 3D-JIGSAW is able to improve a set of models. This set is referred as the "mediocre models" one and the set of the best models is the "good models" one.

Good Models

This set contains the following models:

Program Model/Run TM-Score (1R47) RMSD (1R47)
MODELLER 3HG3 template 0.99459 0.53
MODELLER unsupervised MSA 0.99570 0.47
I-TASSER run 1, model 1 0.99236 0.64
SWISS-MODEL 3HG3, aligned mode 0.99497 0.51
SWISS-MODEL 1KTB, automated mode 0.96559 1.21


The visual comparison (see figure 9A) and the numeric evaluation suggest that the resulting models of 3D-JIGSAW are almost perfect. This is not surprisingly, since the input was a set of very accurate models.

Apo (1R46) Complexed (1R47) Independent
Model TM-Score RMSD RMSD catalytic site TM-Score RMSD RMSD catalytic site Energy
1 0.99472 0.53 0.377 0.99458 0.53 0.329 -609.82
2 0.99503 0.51 0.416 0.99496 0.51 0.289 -608.81
3 0.99503 0.51 0.416 0.99496 0.51 0.289 -608.78
4 0.99469 0.53 0.376 0.99456 0.53 0.328 -607.96
5 0.99492 0.52 0.277 0.99487 0.52 0.289 -606.74


Mediocre Models

This set contains the following models:

Program Model/Run TM-Score (1R47) RMSD (1R47)
MODELLER 3CC1 template 0.78291 3.26
I-TASSER run 1, model 2 0.87069 2.35
SWISS-MODEL 1KTB, aligned mode 0.84472 2.15
SWISS-MODEL 3CC1, aligned mode 0.61051 3.53


3D-JIGSAW provides five models, where model 1 is superior according to the TM-Score, RMSD and the energy calculation and hence we conclude that this model is the best overall. The visual representation also suggest a pretty decent result, but it seems like that the model fails escpecially with some beta-sheets at the C-terminal (see figure 9B). Surprisingly, it is outperformed by some models in the RMSD of the catalytic site. Taking model 1 as the reference for this run, 3D-JIGSAW was not able to provide an improvment in respect to the set which was given as an input.

Apo (1R46) Complexed (1R47) Independent
Model TM-Score RMSD RMSD catalytic site TM-Score RMSD RMSD catalytic site Energy
1 0.84599 2.59 1.626 0.84568 2.58 1.163 -501.82
2 0.74083 3.68 2.108 0.74099 3.77 0.851 -488.67
3 0.73951 3.61 1.531 0.74037 3.74 0.764 -486.56
4 0.74083 3.68 2.108 0.74099 3.77 0.851 -486.52
5 0.73951 3.61 0.743 0.74037 3.74 1.182 -485.56

Discussion

  • overall quality of the model is good -> the quality is even better nearby the catalytic site
  • automated mode better than alignment mode (SWISS-MODEL)

References

<references />