Protein structure prediction

Question

I want to perform a molecular docking between several ligands and the transmembrane domain of a protein.

For this protein I only have the amino acid sequence, so it's necessary to do two things:

• Determine which part of the aminoacid sequence belongs to the transmembrane domain (see this)

• Model the 3D structure based on this aminoacid sequence (trRosetta software)

I would like to know if this little protocol that I have established (with the two softwares) is rigorous and the conclusions of a study based on it could be taken into account.

Answer 1

Here's what I did. Assuming that you have a part of the protein crystal structure.

In my case, I have an incomplete structure of the protein. Lets say if I have an amino acid (AA) sequence of 520 (full length), I have the pdb for certain domains which are functionally important. So, I went for homology modelling. I used two predictors Robetta and tr-Rosetta. Each predictor will give 5 models, and score accordingly. Normally the first model of robetta or tr-Rosetta is always the best. But in my experience it is not always the case.

Here's what I did to select the best out of 10 models.

As I know the crystal structure of the part of the protein. I compare the radius of gyration, secondary structure and root mean square deviations between the crystal structure (PDB) and the part of the model (Rosetta and tr-Rosetta) whose crystal structure is already known. Then I select the model based on the errors of the measured quantities. I used the priority as follows secondary structure > radius of gyration > RMSD. I did this for at least 20 proteins and found that it is not necessarily the first model given by a predictor is always the best. I sometimes find the 5th model is the best. And coming to comparison between Robetta and tr-Rosetta, I would say it depends on the protein you are modelling. For example, I modelled a HSP70 chaperone, the Radius of gyration (Rg) given by tr-rosetta is almost double the Rg given by Robetta. In this case, it is very clear that Robetta did a good job. (Because the models given by tr-Rosetta are like spaghetti, you can clearly know once you see).

One more technique you can use to know a good model is from cross linking mass spectrometry. This technique gives a good structural information of which residues are crosslinked by a specific organic molecule. For instance, the commonly used lysine specific reactive organic molecules are disuccinimidylsuberate (DSS) and bis(sulfosuccinimidyl)suberate (BS(3) ) , By using this chemical, researchers can exactly pinpoint to which lysine residues this DSS has crosslinked. If two lysine residues are crosslinked by a DSS, then it is for sure, they need to be at a certain distance between each other. The distance normally is 26-30 angstroms. Now you look at your models from Robetta and tr-Rosetta and see if this distance criteria is maintained or not. If not, discard the model.

As for docking, I used High Ambiguity Driven Docking (HADDOCK), which is a free docking server, which can dock protein-protein, protein-ligand etc..

Answer 2

I'm adding another answer because I recently find these news. The machine learning-based methods alphafold and rosettafold were recently released on github. Someone has just implemented it in Google colab as Jupiter notebook that you can simply reuse with your colab account. The only thing that you need to do is change the AA sequence. It seems that in the recent CASP edition alphafold was able to gain the first place as the best structure prediction method. You can find all these informations and code in this tweet. You can run different prediction software with different methods and the compare the results. I think this is the best way to understand if the structures you obtained are completely wrong or meaningful.

Answer 3

Your protocol is right and rigorous in the sense as if you don't have the crystal structure of your protein and want to do some predictions, them the only way is using homology modeling.

I am not a big fan of homology modeling, so I always recommend to avoid it as possible.

My addition will be that, instead using only one resource to model the 3D structure, to use several services combined with different techniques. Then, submit all the structures to services that check the quality of the structures (also using different approaches). Finally, organize the structures using as criteria the one with more successful tests.

Answer 4

In the past I had the same question and after a long search I found that every two year there is a worldwide competition to assess the quality of 3D structure prediction of proteins. In the past the winner was an online services called I-tasser that you can find here. I tested it also in the past year with the covid spike protein and after the publication of the experimental 3D structure I compared the 2 structure, finding a very good agreement between the two. I guess it is a nice starting point. You only need the AA sequence and paste it into the online for and let I-tasser do the rest. Hope this help.