Homology based structure predictions BCKDHA
- 1 1.Calculation and evaluation of models
- 1.1 Template selection
- 1.2 General information for the evaluation
- 1.3 Modeller
- 1.4 SWISS-MODEL
- 1.5 iTasser
- 1.6 3DJigsaw
- 2 Comparison of the methods
- 3 References
1.Calculation and evaluation of models
Homology modelling is a technique to determine the secondary structure of a target protein. It is based on an alignment of the target sequence and one or more template sequences with known secondary structures. The target sequence is assigned a secondary structure based on the template structure. The better the alignment, the better the predicted secondary structure for our template. Therefore the template selection is a crucial step in homology modelling.
To find similar structures to BCKDHA we ran HHsearch using the following command:
hhsearch -i query -d database -o output
It found the following 10 hits in the pdb70 database.
|No||Hit||Prob||E-value||P-value||Score||SS||Cols||Query HMM||Template HMM||Identity|
|1||2bfd_A 2-oxoisovalerate dehydr||1.0||1||1||791.3||0.0||400||1-400||1-400 (400)||99%|
|2||1qs0_A 2-oxoisovalerate dehydr||1.0||1||1||571.5||0.0||349||32-382||52-407 (407)||39%|
|3||1w85_A Pyruvate dehydrogenase||1.0||1||1||530.8||0.0||356||8-382||6-362 (368)||34%|
|4||1umd_A E1-alpha, 2-OXO acid de||1.0||1||1||521.8||0.0||351||34-386||16-367 (367)||37%|
|5||2ozl_A PDHE1-A type I, pyruvat||1.0||1||1||482.7||0.0||331||46-380||25-356 (365)||27%|
|6||3l84_A Transketolase; TKT, str||1.0||1||1||85.4||0.0||133||161-297||113-252 (632)||21%|
|7||2r8o_A Transketolase 1, TK 1;||1.0||1||1||74.5||0.0||121||161-285||113-245 (669)||33%|
|8||2o1x_A 1-deoxy-D-xylulose-5-ph||1.0||1||1||74.2||0.0||127||161-287||122-254 (629)||18%|
|9||1gpu_A Transketolase; transfer||1.0||1||1||74.2||0.0||140||161-302||115-265 (680)||22%|
|10||3m49_A Transketolase; alpha-be||1.0||1||1||68.8||0.0||121||161-285||139-271 (690)||31%|
Before we can start working with these hits we have to check whether one of them is a PDB structure for BCKDHA. This is the case for 2bfd_A.
By looking at our results and the fact that this hit can not be used we only have structures with an identity lower than 40%. Since there are just structures available from this region we decided to take two structures out of it. One with a 39% identity and one with 18% identity so that there is still a variation in the identities.
In the following we worked with 1qs0_A (39%) and with 2o1x_A (18%).
General information for the evaluation
A detailed description of how the created models were evaluated can be found in the Evaluation Protocol. The following section presents only the modelling and evaluation results.
Three interesting score when comparing two structures for their structural similarity are the Cα RMSD, the all-atom RMSD and the TM Score. These are two measures which are usually used to measure the accuracy of modelling a structure when the native structure is known. In the following we will call the Cα RMSD only RMSD.
The RMSD is the average distance of all residue pairs in two structures. The C-alpha RMSD is the average distance between aligned alpha-carbons.
The smaller the RMSD value, the better the predicted structure. A local error (e.g. misorientation of the tail) will result in a high RMSD value, although the global structure is correct.
The all-atom RMSD is calculated of the residues which are in an area of 6 around the active sites. As in the Cα RMSD the models which have a low value are the better ones.
As the RMSD is sensitive to the local error, the TM-Score was proposed. The TM Score weights close matches stronger than distant matches and therefore the local error problem is overcome. A TM-Score <0.5 indicates a model with random structural similarity, wherease 0.5 < TM-score < 1.00 means the two compared structures are in about the same fold and therefore the predicted model has a correct topology.
MODELLER is used for homology or comparative modelling of protein three-dimensional structures. It calculates a model containing all non-hydrogen atoms. There are also many other features provided by MODELLER like de novo modelling of loops in protein structures, optimization of various models of protein structure with respect to a flexibly defined objective function, multiple alignment of protein sequences and/or structures, clustering, searching of sequence databases, comparison of protein structures, and so on.
To run modeller with more than one template we use the targets (the percentage values indicate the sequence similarity to the target):
- 3m49:A (31%)
- 2r8o:A (33%)
- 2o1x:A (18%)
- 1w85:A (34%)
- 1qs0:A (39%)
|template||molpdf||DOPE score||GA341 score|
|3M49_A, 2R8O_A, 2O1X_A, 1W85_A, 1QS0_A||123913.8||-19573.7||0.001|
The DOPE (Discrete Optimized Protein Energy) score is calculated to assess homology models. The lower the value of the DOPE score the better the model. This can be also seen in our three models. 1qs0 which has the highest sequence identity, definitely has the lowest DOPE-score which is obvious because of the high identity. The model where 2o1x was the template has a higher score which is reasonable since 2o1x has a sequence identity of 18% whereas the first model (1qs0) has a sequence identity of 39%. This shows that although both structures were in the group with the lower sequence identities there are noticeable differences between them. The last model was build with five different structures. Normally it is helpful for the programm because when more structures are included in the prediction, the model can predict more precisely than with just one template structure. In this case the problem was that all structures which are combined to build a model had no high sequence identity so the information Modeller got to build the model were not helpful. This is reflected in the scores because this model has the highest DOPE-score. So as expected the model with 1qs0 as template structure is the most homologous model.
GA341 is calculated to decide wether the result is a good model or not. A model which is quite good has a score near one. When a model has a score lower than 0.6 it is a bad model. This is also reflected by our results. The model with 1qs0 as template is avery good model since the GA341-score is 1.0. This is a bit strange since the sequence identity to our protein is not very good. Of course the DOPE score was good, too but it can not be correct that a model with a template which has only 39% sequence identity has the best possible GA341-score. The other two models have a score lower than 0.6 which shows that both of them are bad. It is interesting that the model with the 5 templates only has a score of 0.001 which seems a bit to low because the average sequence identity of the used structures is higher than the one of 2o1x which has a GA341 score of 0.419. All in all we can conclude that the 1qs0 model is the most accurate one.
To sum up the results of the two scores it is to say that although all of the structures have a low sequence identity the model with 1qs0 as template is the best one.
Comparison to experimental structure
|experimental structure||model with template||RMSD (DaliLite)||RMSD (sap)||TM-score|
|1U5B_A||3M49_A, 2R8O_A, 2O1X_A, 1W85_A, 1QS0_A||no score||11.398||0.1719|
C-alpha RMSD is a measure of the average deviation in distance between aligned alpha-carbons. The higher this distance value the worse is the model. The first model using 2o1x as template has a RMSD score of 2.3 or 0.829. In both cases the value is lower than the ones of the other two models. Since a low value indicates for a good model this model is the best of the three according to the RMSD value. The RMSD score for the 2o1x model is only a bit higher so it seems that this model is not much worse than the first one. For the third model DaliLite was not able to calculate a RMSD score at all because it could not find enough significant similarities because the structures are to dissimilar. This dissimilarity is reflected by the RMSD value which is calculated with the sap command because it is very high compared with the other two values of sap. An explanation for this bad result for the last model could be that there are to many false information used during the building process. By comparing the TM scores of the three models with each other we can see that only one model has a value higher than 0.5 which means that only one model is significant good. The 1qs0 model has a TM score of 0.829 so it is declared to be a good model whereas the other two models have a TM score of about 0.1 which is far lower than 0.5 and that indicates that both models are useless.
Additionally we calculated the all-atom RMSD scores for the three catalytic center of the three models. As in all the other scores above we can notice again that the model with 1qs0 as template is the best one. This is pointed out by the fact that at all three catalytic centers the all-atom RMSD values are the lowest ones. There is one interesting observation by comparing the values of the other two models because at the first catalytic center the model with 2o1x has a much worse score than the model with the five structures as template. At the second center the score of the 2o1x model is just a little bit lower and at the third center it is even higher. So by looking at the all-atom RMSD valus it can not be decided wether the second or the third model is the better one.
All the calculated scores above declare the model which has the structure of 1qs0 as template beeing the best model. By looking at the visulization ( Figure 1) the assertion of all these scores can be approved. As we can see especially the alpha helices are quite good aligned although there are some which are not aligned. Another thing which shwos that the two structures are not perfectly aligned is that on the left and right side of the superposition there are two structures which are completely not aligned. But all in all it seems that the model compatible with our protein. Expecially by comparing it with the two other models. The 2o1x model which is visualized in Figure 2 has no aligned structure so that it appears that there are two completely different structures superposed. This impression is supported by the calculated scores above which all show that the model which uses 2o1x as template does not fit very good. This also applies to the third model. As we can see in Figure 3 there is again no match between the two structures and so there is also no aligned structure. Again this result could be suspected because of the bac evaluation scores.
To find protein structure homology models SWISS-MODEL can be used. SWISS-MODEL is a fully automated protein structure homology-modeling server and is accessible via the ExPASy web server, or from the program DeepView (Swiss Pdb-Viewer).
It provides three different modelling modes:
- Automated Mode
- Alignment Mode
- Project Mode
The Automated Mode uses fully automated modelling and can therefore be only used when the template is very similar to the target.<ref>http://swissmodel.expasy.org/?pid=smd03&uid=&token=</ref>
As an Input for the automated mode, only an amino acid sequence (raw or FASTA format) or the Uniprot AC of the target is required as it is show in Figure 4. Optional a template PDB id can be given. Swissmodel automatically selects templates from a Blast run which are suitable due to their E-values if no template is given. The Alignment Mode has to be used for the structures with a low identity. Since we only have hits in the region < 40% we used this tool.
Global Model Quality Estimation
The QMEANscore4 is calculated to compare whole models. The score ranges between 0 and 1. The higher the value the better is the quality of the model. By comparing the score of the 1qs0 model with the score of the score of the 2o1x model it is obvious that the first one is the better one since it has a much higher QMEANscore4. But although it is better then the model with 2o1x as target it is not very good. This can be argued from the QMEANscore4 of 0.57 which is not a quarter as good as a model with the score of 1. It can be inferred from the score of only 0.18 that this model is useless.
The QMEAN Z-Score represents the absolute quality of a model. Models with a low quality have a strongly negative QMEAN Z-scores. By looking at Figure 5 and Figure 6 we can see that the QMEAN Z-score of both models is negative and both are under the black or grey graph which is shown in the figures. The fact that both scores are negative indicates that both models are not of top quality. But by comparing the scores directly we can see that the model with 1qs0 as template has a score of -3.28 and the score of the model with 2o1x as template is -9.89 so it appears that the first model is a bit better than the other one. Both values can be found in the already mentioned plots.
The plots in Figure 7 and Figure 9 show the confidence of SWISSMODEL for each residue of the built models. In the plot of the model with 1qs0 as template ( Figure 7) we can see that here the program is only unsure about the beginning and the end of the model. In the middle of the model there is also a peak indicating that those residues are not modeled with complete certainty. For the rest of the model the program predicts a low inaccuracy probability. The plot of the 2o1x model ( Figure 9) is the complete opposite. Here we can see that there are many very high peaks in the middle of the protein which suggests that the programm predicts for the middle part of the model which is more important than the ends a very high inaccuracy. When a model is wrong in the middle part it is useless and since there are so high peaks in the middle part it can be that this model is useless. The same conclusion can be found in Figure 8 which is the visualization of the 1qs0 model and in Figure 10 which is the visualization of the 201x model. In both figures we can see the coloured model. The region which is blue stands for a high assurance and when a region is red is means that this part is in all probability wrong. When we look at the model with 1qs0 as template we can see that there are only a few red parts and they are mainly in the end of the protein whereas in the center the parts are coloured blue or green which shows that these parts of the model are probably correct. In contrary to this picture the model of 2o1x is nearly completely red which supports the assertion that this model is useless because nearly the complet model is predicted to be possibly wrong.
Local Model Quality Estimation: Anolea / QMEAN
For the local model quality estimation we chose the ANOLEA potential. This program performs energy calculations on a protein chain. On the y-axis the energy of each amino acid is represented. Negative energy values (in green) represent favourable energy environment whereas positive values (in red) unfavourable energy environment for a given amino acid. By looking on both plots we can see that in both there are many red parts so both of them are perhaps not completely correct. But when we analyse the two figures seperately we can see that the energy calculation for the 1qs0 model ( Figure 11) contains a few green parts which shows that there are some favourable energy environments in the center of the protein and part with a really bad energy environment is only in the beginning of the protein. So we can deduce that this protein is perhaps correct in the important middle part. The other plot for the 2o1x model ( Figure 11) is completely red. Not only in the beginning but also in the important middle part of the model. This can denote that this model is probably not useful.
Comparison to experimental structure
|experimental structure||model with template||RMSD (DaliLite)||RMSD (sap)||TM score|
The RMSD is a measure of the average deviation of the distance between aligned alpha-carbons. The higher this distance value the worse is the model. To calculate the RMSD we used two different programms. Usually the results of both are not the same but they have the same trend. In this case it is different. By comparing the RMSD scores calculated by DaliLate which can be looked up in the table above the 1qs0 model is 0.1 higher than the score of the 2o1x model. So it appears that the model with the 2o1x structure as template is a bit better. But when we compare the scores calculated by the sap command the result is completely different. The 1qs0 has a value of 0.8771 and the 2o1x model has a value of 14.305. Following these results it is obvious that the 1qs0 model is much better which is the opposite to the other RMSD conclusion. But in this case the difference between the scores of the two models is much more significant than in the other case so it can be reasoned that the model with 1qs0 as template is the better model. To confirm this assumption we analyse the TM score. When the TM score is higher than 0.5 it can be said that a model is good. This is not the case for the 2o1x model since it has a score of 0.1686 which is really low. We can argue from this value that the model is bad. In contrary to the model with 1qs0 as it has a score of 0.8771 and so it is declared to be a good model. The conclusion of the TM score supports the one of the RMSD score so it can be said that all in all the 1qs0 model is the better one.
all atom RMSD
Additionally to the scores above we calculated the all-atom RMSD scores for the three catalytic center of the two models. The values of this score are definite. At all three catalytic centers the values for the model with 1qs0 as template are much better since low values stay for good models. The really high values for the 2o1x model indicate that this model is quite useless.
The calculated RMSD score, TM score and all-atom RMSD score indicate all that the model with 1qs0 as template is the better one and that the other model is quite useless. To check these conclusions we superposed the two models with the structure of our protein. In the visualization of the 1qs0 model superposition in Figure 13 we can see that there are only a few regions of the two structures which could be superposed completely. But the main part of the model is shifted a bit so that the secondary structure elements lay next to each other. This observation shows that the model is just a approximation of the structure but is not perfect. The visualization of the superposition of the 2o1x model in Figure 14 reflects completely the conclusion we made by analysing the scores above. The model is useless as there is no region which could be superposed perfectly and it looks like a superposition of two completely different structures.
To summerize the results of the numeric evaluation and of the comparison to experimental structure we can say that the model with 2o1x can not be used for further analysis since there is no similarity between out protein and this model. The 1qs0 model is not that bad since it has quite good scores which show that it is a good model but by looking at the visualization we see that it has indeed the same structure but it is shifted a bit. So we can work with this model but the results which base on this model won`t be completely correct.
The C-score is a measure for the quality of predicted models by I-TASSER. C-score ranges between [-5,2], where a C-score of higher value signifies a model with a high confidence. First the five models with 1qs0 were analysed. Model1 has a score of 1.174 which a high value at this chart so the quality of this model seems to be good. The only other model which also has a positive score is model4 with 0.200. This is not as high as the score of model1 but it is positive enough to say that this is also a good model. Model2 has a negative score of -0.190 but this value is still much higher than 5 so it is still high enough that it can be declared as a good model. Model3 has a C-score of -0.718. This score is nearly in the middle of the chart which indicates that this model is possibly false. The last model is quite interesting since all the other models had not that bad score but this model has the worst possible score of -5. So it is clear that this model is absolutely useless. Now the models with 2o1x as template are analysed. None of the C-scores is positive which demonstrated that these five models are obviously not very good. The best of the five models is model1 since it has a score of -0.150 which is not very negative. By looking at the scores of the other four models it has to be said that all of them can not be good models because the C-score ranges between -1.276 and -3.208. To summarize the C-scores of the ten models it has to be said that only model1 and model4 which have 1qs0 as template have positive scores indicating that it is only possible to work with them.
Comparison to experimental structure
|No||RMSD (DaliLite)||RMSD (sap)||TMscore||RMSD (DaliLite)||RMSD (sap)||TMscore|
The RMSD is a measure of the average deviation of the distance between aligned alpha-carbons. The higher this distance value the worse is the model. We calculated the RMSD score with two different programms so that we can see it if there is a strange calculation in one of the results and that we can compare the two RMSDs. The other calculated score is the TM score. When it is between 0.5 and 1.0 then the predictec model has the correct topology. In the first analysis we will just look at the models with 1qs0 as template. By comparing the scores of the five models with each other it is conspicuously that all of them have nearly the same value. It doesn't matter which RMSD score is considered. In both cases all the scores differ only minimal. When we go into more detail by looking at the DaliLite-RMSD score we recognize that model3 and model5 have a score which is a bit higher but not significant. So we can say that all five models have a well predicted structure. To get more information about the models to make a better statement we also analyzed the TM score. But here we have got the same result as with the RMSD score. All five TM scores are quite the same and are all higher than 0.5. So we can concluded considering the RMSD scores and the TM score that these five models are all very well predicted and that there is nearly no difference between them.
The next analysis is of the models which have 2o1x as template. By comparing the scores of the different models we can see that here is more divergence. Only the model2 seems to be a good model because it has low RMSD values and also the TM score is far over 0.5. The only other model which has a TM score over 0.5 is model1 but it has quite high RMSD values compared to the other models. Model3, 4 and 5 have all high RMSD scores which shows that their prediction is unconfident. Additionally all of them have a TM score which is lower than 0.5 so their topology is possibly not correct.
Out of all the results we can conclude that the five models which are build with the help of 1qs0 are all very good and useful and of the other fove models only the second one seems to be well predicted and usefull.
all atom RMSD
To calculate the RMSD of the 6A radius of the catalytic center we had to find the catalytic center first. There are three catalytic center on the positions 161, 166 and 167. We calculated the RMSD for all of them. We start with the analysis for the 1qs0 models. Here we can see that there are difference between the five different models although all of them have good values. To go into more detail it has to be said that the second model has the lowest values on each position so it is the most accurate one. Model1 also has good values but they are not as good as the ones of Model2. By looking at the other three models we can see that their values are still good but they are a bit higher than the ones of model1 and 2. By analysing the models built with 2o1x as template we can see that interestingly model2 has not only lower values than the other 2o1x models but has the lowest values of all models. So we can say that according to the all-atom RMSD model2 with 2o1x as template is the best model. This model is the only one of the models built with the help of 2o1x which is profitably. All the other models have quite high values up to 5.078 so it is not possible to work with them.
Since the above discussed results are not definite we have to look at the superpositions of the model with the structure of our protein.
As in the previous analysis we start with the models of 1qs0. In Figure 15 the superposition with model1 is visualized and we can see that the model has the same structure as our protein but it shifted a bit. This observation agrees with the assumption that the model is quite good but not perfect. In Figure 16 model2 is shown which is according to the scores a really good model. In fact there are structural elements which can be superposed perfectly but there are also parts which are shifted or can not be superposed at all. So we have to conclude that in this case the model seems to be not as good as the scores predicted.
According to the scores model3 Figure 17 is not as good as the other two already mentioned models. This can be supported by the visualization since there are many regions which are shifted or can not be superposed at all. Model4 which is shown in Figure 18 actually has a bit worse scores than model3 and this difference can also be seen articulately in the picture. Model5 is predicted to be the worst models of all because it has bad scores compared to the other models. By looking at the superposition of the structures in Figure 19 this result can be affirmed as no perfect superposed structural element can be seen.
The qualitiy of the models with 2o1x as template is very sure. The calculated scores show that model2 is a very good one. To check this it is helpful to look at the visualization of the superposition of model2 and the structure of BCKDHA in Figure 21. The overlay is not perfect but seems to be shifted in most parts of the model though there are the same structural elements which point in the same direction. By looking at the other four models ( Figure 20, Figure 22, Figure 23, Figure 24) we can see that all the models can not be superposed with the structure of our protein. This observation supports the already made assumption that all four models can not be used since they are to dissimilar to the structure of BCKDHA.
3DJigsaw is a server which builds protein models based on already predicted models for a specific target. It recombines the models and optimizes them.
Since we have only models for the low sequence-identity category we started it only once with the best models of this category. The following models were chosen to build a recombined model with 3DJigsaw because of their high TM score:
- modeller model for template 1qs0
- swissmodel model for template 1qs0
- iTasser model 1 for template 1qs0
- iTasser model 2 for template 1qs0
- iTasser model 4 for template 1qs0
3D-Jigsaw calculates the energy and the coverage for each predicted model. By comparing the coverage for each model we can see that the predicted model covers 100% of the model since the coverage for each model is 1.0. The energies are more different between the five predicted models. The lower the energy the better because a low energy indicates a stable model. Although the energies are different between the five models it is obvious that these differences are not significant big. The first one indeed has the lowest energy but only minimal lower than the second or the third model. The energy of model 4 is about 5 points higher so we can say that the first three models are better than model 4 and 5. But all in all it can be said that both the energy calculation and the coverage predict all five models to be good.
Comparison to experimental structure
|Model||RMSD (DaliLite)||RMSD (sap)||TM-score|
To find out how good the models are we calculate the RMSD twice with different tools. Due to the fact that a low RMSD value indicates a good model we can say that all five models seem to be very good. The RMSD calculated by DaliLite varies only between 1.8 and 2.1 which is a very small range. This is the same for the RMSD calculated by the sap-command. By looking at both values in more detail we can see that the first three models are again as in the energy calculation a bit better than the last two models but not much. Additionally we calculated the TM-score for each model. All of the five models have TM score higher than 0.5 which shows that all of the five models have a correct predicted topology. Again a cut can be seen between model 3 and 4 indicateing that, as the other scores suggested, the first three models are a bit better.
all atom RMSD
The calculated all atom RMSD scores affirm the assertion of the above discussed results. As we can see of the all atom RMSD the first three models are all equally good. Additionally there is again the cut after model 3. Model 4 and 5 are both a bit worse than the first three models. But by looking at the values we can see that all of them are quite low so it is obvious that all five models are good.
To find out if the so far result that the first three models are a bit better than the other two but all in all the five models are all very good is correct we look at the superposition of the model with the correct structure of BCKDHA. Figure 25 shows model 1 and it is apparent that the prediction of the structure of this model was very good as most of the two structures can be aligned perfectly. Of course there are also some parts which are not aligned but that is due to the scores which are also not perfect. These score indicate that also model 2 is very similar to the real structure. This can be supported by the superposition (Figure 26) since most of the protein is covered by the model. Again we can see that the superposition is not perfect but this was also expected. Figure 27 shows the third of the models which is predicted by the scores to be very good. This prediction can be approved since the superposition of model 3 and the structure of BCKDHA is for the most part perfect. As in the two other models there are some regions which could not be aligned but that was expected again. The superpositions of model 4 and 5 with the structure of BCKDHA ( Figure 28 and Figure 29) reflect the cut between them and the first three models. Still the model covers the real structur very good but it seems that there are more regions which can not be aligned or that there are shifts between the two structures. With the observation of the superpositions the results dicussed above are supported.
Comparison of the methods
The following tables again list the RMSD and TM-score values, which were computed before, to provide an overview of the performance of the different methods.
To compare the predicted models and the real crystallized structure of our template different scores (RMSD, TM-score) were calculated. Based on these scores it is not easy to decide which tool is the best one. For modeling with 2o1x as template which has the lowest sequence identity it is obvious that iTasser did the best job since the TMscore is much higher than the TMscore of the other two programms for models with this template. But for the models with 1qs0 as template all values are very nearby. By looking very close at the values we can see that the TMscore of Swissmodel is the best of all TMscores and additionally the RMSD score of Swissmodel is the lowest one. So we can say that Swissmodel is the best tool. It is interesting that Swissmodel is even more precise than 3D-Jigsaw although this tool worked with the best predictions of Modeller, Swissmodel and iTasser. An explanation could be that all of the models are not very good because in the beginning we had two templates which have both a low sequence identity so perhaps there are to many false information in the models so that it was very hard for 3D-Jigsaw to build a very good model out of the 5 models. But it is important to see that the difference between Modeller, Swissmodel, iTasser and 3D-Jigsaw is only minimal. We can conclude that the similarity of the template is the limiting factor for the model prediction and composes which tool is the most useful one.
back to Maple syrup urine disease main page
go to Task 5: Mapping SNPs