I want to get a series of enzyme-substrate binding pose in one step of my research, but I met some questions:
- My enzyme strutures lack of experimental structure information, and some of them can form into homo-n-mer, so which one should I choose? Monomer from Alphafold database or homo-complex strutcure from SWISS-MODEL?
Of course, it depends on the active pocket information, but what I only know is some substrate binding sites and active sites from UniProt(based on similarity)
So I used p2rank tool to predict pockets, and tried to determine active pocket based on these two information.
- So the monomer is better if the pocket is not in subunit interface, and homo-complex is better when the pocket is in subunit interface, right? But in some cases, the rigid local enzyme structure will mask the "real" pocket, since some key residue were not in the final result...
- And there is still a question, the protocol I followed said before docking, the predicted enzyme struture should be relaxed, but when I checked the structure quality by MolProbity, the relaxed struture showed higher molProbity score(lower is better) and much higher clash score, so which one should I trust? The relax protocol made struture better or worse?
AF2 vs Swissmodel. It depends if the Swissmodel is threaded against something with at least more than 50% homology to the template and if the metrics are good.
I have a web app, https://venus.cmd.ox.ac.uk that has to face this exact issue and prefers SwissModel even if everyone is snooty these days against it!
So... If the scores are bad for the SwissModel threaded model I would strongly recommend using a notebook from https://github.com/sokrypton/ColabFold which applies the AF2 algorithm to homo-oligomers really well. As a bonus I have written a colab notebook that can analyse the interface Gibbs of the complex via PyRosetta and one to migrate ligand over from a Blast hit. However, there is the caveat that, some proteins, mammalian ones in particular, have spaghetti loops which may dock other protein etc, but are disordered alone: these from AF2 need to be trimmed for any meaningful analysis.
active pocket information Active as in there is an inactive state? Say phosphorylation turns it on? If so look at all the models from an AF2 notebook as these may hide alt conformations. If you know that an interaction is lost in the active state, you can force it according to the twittersphere, but it is not trivial.
Regular enzymes vs. domain swapping ones. Spot on.
Quality. SwissModel threaded models do "blow up" at times for me— Is the RMSD of the unrelaxed vs relaxed crazy? I would suggest a FastRelax with a movemap with fixed backbone and then one with flexible backbone. More cycles is always a good choice —the fastrelax binary has the thorough argument, which is 15 cycles when scripted. Obviously, relaxing against the template's electron density even with a weight well below the 30 from the examples is a stupid idea, but I tried it anyway —Science!— and can confirm it makes it worse. But in recap, trust the relaxed drop in score more than MolProbility's if you are sure you did a good job.