In Silico Protein Models: How To Trust?

Question

[K]

I have been involved in the study of a protein family for which not much information was previously available.

As a part of our efforts to try and get a clearer insight on what those proteins do and how they could be acting, we have performed some in-silico models. The models actually look pretty, but traditionally in-silico 3-d models have generally been regarded as proving nothing by themselves. So, to try to add confidence on the likeliness of the models obtained being close to truth, or to find out if those models contain some orientative information on how these proteins actually are, what can I do? I have tried the QMEAN server, which gives some scores related to different protein parameters. This could be used to argue that the model is "possible", if the scores are good. Other servers exist, as far as I know.

However: if there were no chance for protein crystallization (i.e. no experimental checking), what could I do, and how, to check the plausibility of the model only by in-silico means? In other words: what is the strongest argumentation in favour of the model that I could possibly get by only in-silico methods? What is the strongest case I could possibly make to defend these models? (of course, assuming that they are actually close to true!)

Thanks a lot in advance.

Answer 1

[João Rodrigues]

The best thing you can get is experimental data of any kind. You don't need a crystal structure to validate your model, not necessarily at least. If your protein does something, you can check the model and try to think of residues to mutate to abolish that function based on the 3D structure. You can also use other lower resolution methods (CD) to investigate the secondary structure for example and see if it matches the model.

Computationally, your best bet is I guess homology. If your protein is part of a family and another member of that family has a structure determined, then go ahead and compare them. Do they resemble? This is also obviously linked to how you modelled your protein. If you got it through homology modelling, this is a non-sense comparison. But the sequence identity for example gives you a confidence range.. if your identity is over 40% you can be roughly sure of the model. If it is over 70%, you can be very sure. Under 40%... it depends a lot..

All in all, depends a lot on the function of the protein.

EDIT:

Most validation servers will look at the packing, clashes, etc in your model. If you used Rosetta for example, these issues are usually taken care of and your model will score beautifully although it can be pretty crap from a biological point of view.

[K]

Hi, João, and thanks for the reply.

Actually, I obtained the models by ab-initio methods, in part because no strong homology exists to any other protein families (thus no homology modelling possible). I have indeed found some good scoring models, precisely with the program you mentioned. Even, the models look beautiful in the eye of the beholder...

I guess I could try to somehow link any published functional data with the structure, as you suggest. I will chew on that for a while to see if I can find an appropriate support in that way. Anyway, if further ideas happen to pop up, you are very welcome back!

Again, Thanks a lot in any case, João.

Kindest regards from Spain,

Karel

P.S. True: one time I remember getting a model which seemed to pass the scoring/filters with flying colours. However, looking at it, the structure did just seem to be a big mess...