Methods to predict one connected structure from two known separate structures

Currently, I am facing such a question: how to predict one connected protein structures from two separate known structures?

It should be different from the protein-protein docking method, which is non-covalent interactions. My question is that the two proteins should "contently" link together.

Any suggestions are appreciated.

• MODELLER can do this as well. – jgreener Jul 6 '20 at 10:42

There are two approaches here.

Physiological

If domain A and B of your protein interact with each other physiologically, then you need to protein-protein docking. This is not too straightforward as protein are flexible.

Several online servers exists such as Haddock and include ClusPro, which use more traditional approaches. While the online server DEMO (I-Tasser group) does a structural homology based search and a final MD based repacking to better minimise the structures (no protein-protein docking).

Many of these work by magic as far as the user is concerned. But understanding what is going on is really important, even if it is overly laborious. When you don't know where a protein goes, it's called blind or global docking, which is computational expensive and so imprecise that the effort required to prove it is correct for peer-review is too steep for it to be useful. If you are doing local docking it is okay, which may be the case here depending on the linker. Specifically, you need to generate more 'decoys' (potential poses) the more unsure you are: tens of thousands or more. Then you have find clusters and see how each scores following more precise refinement. As an example of a tediously and manually set up run, Rosetta has several docking protocols, which explain the steps required —of note are the many judgement calls, such as in the local refinement, where one has to judge whether setting -dock_pert 3 8 (3 Å and 8° perturbation) is too strict etc.

Pretty picture

The protein-protein docking approach will result in a globular protein, which may be incorrect and majorly is not great for pictures.

If you want a pretty picture of two or more domains most likely you would want them placed like beads on a string. If you have a linker that is not solved the recommended distance to space them is $$(3.5 N_{missing})^{0.5}$$ Å —unfortunately I do not know where this lore comes from. If it helps I wrote a script that does this in automatic but manual placement works best.

If one wanted to do a better job, one tool to use is Rosetta remodel, which has a bit of a learning curve, but is very powerful for protein design. This requires the PDB to be on a single chain (pymol alter command, followed by create will do the trick: more here) and actually the amino acids to not have gaps —PDB number being the same as the pose numbering. Rosetta gives better results if the poses are energy minimised, but for this it does not matter too much. Then a "blueprint file" formatted as discussed in the documentation. Briefly, in a way that residues left alone are in the form 1 M . (where residue 1 is a methionine), while the residues before and after the insertion are like 100 D L PIKAA D (meaning D100 is changed to a Asp in loop SS; not actual numbers or residue), while between entries are added 0 X L PIKAA G etc. To model loops, -generic_aa G is a good option as this tells the default rough fitting (centroid mode) to use glycine and not valine as generic residue. The top pose will actually not have a superb score, so a 3 (regular) or more cycles (15 thorough) with Rosetta Relax application either of the whole protein or with a movemap specific to the linker, may be done.