Homology-based structure prediction (PKU)
- 1 Short Task Description
- 2 Reference
- 3 Model Construction
- 4 Model evaluation
- 5 References
Short Task Description
After the sequence based predictions of function and secondary structure for our protein we will determine the 3D structure of the wild type protein and observe the influence one or several SNPs have on this structure. Of the variety of methods to be used for tertiary structure prediction, we choose homology modeling as a first approach to our goal. Read the complete task description here. The protocol of commands and scripts can be found in our journal
Due to our prior knowledge of the protein responsible for PKU, the evaluation of the methods applied, is easier than for a completely unknown sequence. In <xr id="fig:1pahstruct" /> one can see the monomer and the active site of Phenylalaninehydroxylase. On the other side ( <xr id="fig:2pahstruct" />) one can see the polymere in its active form which can be found in the human body.
Here we will show the steps we took building the models we then use and evaluate. In order to start the sheer model-building we first have to construct some datasets, which will be the founding of our models.
These datasets were derived from several sources. They all consist of PDB-entries, but we ensured to no include the already known structure of our protein, so we have a better insight in the topic of homology modeling with a completely unknown sequence.
For this set of datasets we used the webservice of sequence similarity search provided by the pdb called PDBeXplore, which can be accessed here. In the used dataset (see <xr id="tab:datasetpdbe" /> we restricted the received data from pdb, such as we did not use the structure of both the monomer and the dimer etc. We also did not use the structure with different ligands in order to keep the variability high.
In the dataset of sequences above 80% we only found one significant hit, which is the structure for Phenylalanine Hydroxilase dephosphorylated. This is a marginal case for the noninclusion of the protein itself, but we decided, since its from another organism, that we include it.
The dataset with sequence identity from 40% to 80% sequence identity only contain structures in connection with aromatic hydroxylation namely Tryptophan and Tyrosin from chicken and rat though the structure gained from the rat also contains the tetramerisation domain we also find in our reference structure. But we also found Tryptophan and Tyrosin hydroxylase structures in the pdb derived from human.
As for the lower than 30% dataset, we can not really expect to find useful output here, because the best E-value we could find is 6.7.
The dataset with the highest sequence similarity in <xr id="tab:datasetHHPred"/> contains two structures with a very high similarity, with is due to the fact, that the structure is that of the original protein in different states. One is the protein in complex with Tetrahydrobiopterin (BH4), which is a co-factor for the PheOH-activity. The other is the phosphorylated proteinstructure.
In the second dataset (40%-80%) we find two of the structures which were alread explained above.
The last dataset from HHPred contains five structures which only decend from bacteria with only one of the structures has a direct connection with PheOH as this one binds L-Phe. The others all are connected or part of the ACT-domain which is known to be controlled by amino-acid concentration, which relates to our target protein.
In the above-80%-identity-dataset we find again our structure from above.
Unfortunately we did not receive any result for our second dataset.
But the choice for our third dataset was great. We chose the PheOH-counterpart from the CHROMOBACTERIUM VIOLACEUM namely 1ltu and one (2v27) from COLWELLIA PSYCHRERYTHRAEA 34H which is a version of the protein, that works in a much colder environment. 3luy as well as  plus the new 2qmx and 2qmw fall into the ACT-domain-group as mentioned at the HHPreddataset
Comparison of datasets
In summary one can say, that the approaches all find similar results for the dataset with 80% sequence similarity or above. Those are all structures of Phenylalaninehydroxylase with different modifications or from different organisms. But only Coma and HHPred find the same (1phz) structure, whereas PDBeXplore finds a completely different identifier, which is the protein with its co-substrate.
Most differences occur in the second dataset, in which PDBeXplore finds a lot of possible candidates with a very good e-Value range. But the other two do only find two or even no result at all.
But in the third datset (<30% identity) PDBeXplore finds some candidates, but the e-Values are to high too be considered good hits. In contrary to HHPred and Coma which both found good hits with a low e-Value as well as identity in the desired range.
What also might have been observed by an interested reader, are the differences in identity and e-Value throughout the supposed to be identical hits, like 3luy. We are not entirely sure where these might arise, but since the difference is not that significant we expect them to descend from different alignment scores and or weighting.
We created homology models with different methods, in order to examine the structure of our unknown protein.
Here we will show you the models one can gain from Modeller <ref name="modeller">A. Šali and T. L. Blundell. Comparative protein modelling by satisfaction of spatial restraints. J. Mol. Biol. 234, 779-815, 1993. </ref>. We used a local version at our private machines.
Modeller offers mainly two possibilities:
- single template modelling
- multi-template modelling
We are going to show you the differences and possibilities this offers.
</figure>In order to show the quality of the models we will show the reference protein in blue for all the folloing images, these were aligned using the pymol alignment algorithm. You might see some modelled domains, which do not appear in the 1PAH structure shown in <xr id="fig:1pahstruct" />. This stems from the pdb entry, which only contains the catalytic domain structure of the protein. So we decided to show in <xr id="fig:compare1and2pah" /> the two 'different' structures so you know the difference when we refer to the binding domain or the catalytic domain of the protein.
In this part you choose one sequence which you believe to be the closest relative to your target sequence and model it after the alignment of those two. As we did some dataset creation before, the choice of sequences is already done. We now only have to use single template to first align and then actually model this pairwise alignment. Then it will be assessed with DOPE-score and GA341<ref name="GA341">Francisco Melo, Roberto Sánchez, Andrej Sali; Protein science : a publication of the Protein Society, Vol. 11, No. 2. (February 2002), pp. 430-448, doi:10.1002/pro.110430</ref><ref name="DOPE">Min-yi Shen and Andrej Sali Protein Sci. 2006 November; 15(11): 2507–2524.doi: 10.1110/ps.062416606</ref>.
In the following we show you three exemplary models from each of the sequence identity parts.
To see the optimal quality one can gain from a modeling we used 1phz which we expected to have the highest similarity to 1pah. In <xr id="fig:1phzmodelling" /> one sees the simliarity to <xr id="fig:1pahstruct" />
The structures almost match perfectly in the catalytic domain, but the coiled ends and beginnings. The binding domain is missing completely.
With the lower sequence identity we used the one with the highest e-Value in order to see the optimal performance here. The colorcoding is the same as above.
Again we see the coiled ends, which in this part are even longer than before, which is due to the fact, that the 1TOH related sequence is about 100 residues shorter than the 1pah. But we also see some differences in the catalytic domain as well. For example the helix left of the binding domain is a bit more kinked and therefore overlaps the helix of the reference. A bit on top of this, there is a sheet, which only exists in the model, which might be a result of the difference of substrates the two domains have. Over all the differences are really minor and the catalytic domain is almost peferctly matched. In contrast to the modelling with 1phz, one has to point out the bindingdomain which is modeled here.
With a sequence identity as low as in this model one would expect very bad results. Our results can bee seen in <xr id="fig:2v27modelling" />
They are really of low quality, as almost no helix or sheet from the model matches one from the reference and even the binding domain is just coiled and seperated. This time the catalytic domain is not matched very well and the binding domain is not matched at all.
We created three models using multiple templates. The procedure follows the tutorial from previous students
Used templates are:
- 1VKJ and 3HV0
<xr id="tab:modeller_multi"/> on the left shows the overlay of a model created with two templates of low (<30%) sequence similarity in red with the true structure in blue. As can be seen, much of the model is rather randomly folded, especially the loop formed by residues approximately 320 to 380 is badly misshaped.
- 1VKJ and 2PHM
<xr id="tab:modeller_multi"/> in the middle shows the overlay of a model created with one templates of low (<30%) and one with high (<80%) sequence similarity in red with the true structure in blue. Here we see a big improvment to the model before. The catalytic domain is almost perfectly matched and the coiled part which could be found at the left picture has been reduced.
- 2PHM and 1TOH
<xr id="tab:modeller_multi"/> on the right shows the overlay of a model created with two templates of high (one >80% and one >60%) sequence similarity in red with the true structure in blue. We almost got the same structure as in <xr id="tab:modeller_multi"/> in the middle. But this time there is even a well shaped bindingdomain modelled. Here you can see that modeller is able to get better results with a multi template version than with one single. In <xr id="fig:1phzmodelling" /> the catalytic domain was matched equally well, but the binding domain was missing, which is now completely shaped.
<figtable id="tab:modeller_multi"> Modeller predictions with multiple templates
We submitted several templates to the Swiss Model server, the resulting models are shown in <xr id="tab:Swiss_Model"/>. As Swiss-Model results were mostly dimeric with identical chains, we restricted the results to only using one chain.
- Fully automatic prediction
The server choose the best template on its own and used the PDB structure 1J8U:A. Here again we find a perfect match as we could see in Modeller before.
Which is not extremely surprising, since the identity of 1J8U and 1PAH is 100%, but due to the co-substrate binding at 1J8U some differences might be expected.
Feeding 1TOH manually as template results in the models seen in <xr id="tab:Swiss_Model"/> middle and right. Again we see a very good matching of the structures, apart from the length of some helices of sheets, but the positioning of the domains and even the bindingdomain is matched very exactly. The results are even closer to the reference than they were with modeller. Aligning the template with T-Coffee did not change the result significantly as the automatic alignment appears already very accurate.
We also tried a number of other structures with lower sequence similarity as templates, both with and without a prealigned sequence. In each case, the server aborted the prediction because the alignment quality was too low to be used as template properly.
<figtable id="tab:Swiss_Model"> Swiss Model server predictions with and without prealigned templates
We submitted the following templates to the I-Tasser web server:
- fullly automatic (chosen by server: 2PHM:A)
Here again we find a very nice matching of reference and model, which is even better than the modeller approach, because here the binding domain is nicely formed.
In this modelling approach the catalytic domain is matched quite similar to the one with 2phm, but the binding domain is not matched at all. Though it is much better than the related modeller model, which predicted this as one long coiled region.
Surprisingly with this template I-Tasser matches the reference better than with 1toh, but in the model evaluation we will show you possible explanations for that. The result is almost as good as the one with template 2phm, but for a few small differences in helix and sheet positioning.
<figtable id="tab:I-Tasser"> I-Tasser server predictions
We show the TM scores, GDT scores and RMSD value for all presented models in <xr id="tab:modelling_scores"/> and comment on the individual methods or models in the appropriate subsections.
<figtable id="tab:modelling_scores"> Prediction Scores for the prediction methods
|template ID||TM-Score||GDT-TS Score||GDT-HA Score||RMSD value|
|Swiss Model server|
|Modeller single templates|
|Modeller multiple templates|
Modeling with a high similarity is considered easy, but especially in those cases errors can have more weight than in other. For this model, the dopescore per residue of template and target can be found in <xr id="fig:1phzdope" />. One clearly sees the dopescore is in all cases higher for the model, but the difference is never high. Like mentioned above, both ends of the model are only existent in the model due to smaller sequenceparts in the pdb.
1vkj and 3hv0
1vkj and 2phm
2phm and 1toh
I-Tasser provides five different models which can be used and their C-Score to evaluate them, but in real cases there wouldn't be a reference to compare them with to choose the best model. For the pictures, we choose simply the (preferred) models with the best scores, but large differences in the models built from one template might make it difficult to evaluate the accuracy of the prediction. To get a feeling about this, we imposed all the models we derived from I-Tasser, to see the conserved and the unsure regions.
We always show the five models colored automaticly by pymol, since the models do not need to be adressed separately, but rather as a set of models. But the reference will be colored in the same blue color as before.
So we see in <xr id="fig:automaticmodels" /> that the catalytic domain is well conserved and the differences between the models are minor, the only big difference between them is the positioning of the bindingdomain, which can be seen at the righ side of the picture.
In <xr id="fig:1tohmodels" /> we also see a rather conserved catalytic domain, with more differences than in <xr id="fig:automaticmodels"/>. Additionally we also have some coiled region, which do not match the reference.
In the modelling part we were surprised about the extremly well results I-Tasser produced with its modell, but when we now look at <xr id="fig:1vkjmodels" /> we see that this was just due to the choice of the one model with the best score of the five I-Tasser provided.
We see just a fuzzy ball, with no conserved regions throughout the models. Most of them barely match the reference structure. So the extremly good result even with a low quality template has to be dampened, but human choice allows for a decent model building.