You are here

pmut and scoring

4 posts / 0 new
Last post
pmut and scoring

Hello all,

I am doing semi-rational design of my protein. So far I have one mutation that increases thermostability identified in vitro. However, Rosetta didn't predict this mutation in pmut scan. When I scored the protein variant with this mutation, the score was better than all the mutations predicted by Rosetta.

Now I'm trying to see if there is a combination of mutations (using predicted mutations and invitro predicted mutation) that will give me the best energy, but instead of the predicted mutations making the invitro predicted variant's score even better, they are actually lowering it by a lot.

Does anyone have an idea of what might be going on?

I'll appreciate any suggestions. Thanks!

Post Situation: 
Tue, 2021-09-14 00:46

Pmut has a `DDG_cutoff` setting that when set to something farcially bad will make it output all variants: I am a big fan of the application to make a heatmap of the mutational landscape of a protein... but its speed does come at the cost of backbone accuracy —proline (and a bit glycine) are no-go. I assume you scored the variant with the mutation by introducing it, re-relaxing it and scoring it with the _same_ scorefunction? Maybe there was some interesting backbone change?

One thing that is important to consider is the isoelectric point of the protein plays a part in controlling how much it aggregates regardless of how stable the protein ought to be: a pI of 7 gives funky lines in DSC or thermal shift assays. I had a protein that had exactly this issue and moving the isoelectric point away from neutral pH make it behave like a normal protein and the Gibbs free energy values started behaving.

Otherwise, the may be a technical issue with the set up. The protein template you are using for pmut_scan may not be ideal...

  • Is it a crystal or NMR structure? I-Tasser/Phyre/SwissModels/AlphaFold2 models with low scores will always return junk as per the junk-in junk-out rule. But even decent models that did not "blow up" during minimisation will return dodgy results.
  • How well did you relax it beforehand? 15 cycles is the "thorough" mode in the app. The quick mode will give odd results downstream. I relax my protein against the density map while keeping one or two key structural waters, but it does play weird tricks on pmut_scan so I have had to manually check out those residues.
Wed, 2021-09-15 02:07

Hi matteoferla,

Thank you for responding.

Yes I scored my variant with the same scorefunction after relaxing, the output .pdb file looks just the input, there are no significant changes.

I've done DSF on this protein, it has a pI of 6.8 and the results looked pretty decent, I haven't done DSC yet though.

It's a cryo-EM model and I do a single relax cycle, which takes almost 24 hours, perhaps I should try more cycles, use more mpis? I've never relaxed against the density map that could be interesting. I'd like to try that, do I just include the map as an input file in my relax script?

This is quite helpful. Thanks :)

Mon, 2021-09-20 04:33

I'd be well concerned with a pI of 6.8, so its good you've got decent SYPRO curves. And you don't get stuff like low yield or flocculation? I am code monkey, so cannot really be helpful here. But check out the Rosetta binary superchange on a better model: it might give something cool.

A day for a cycle: I know your pain. But one single cycle is a problem. There's this mover, LocalRelax, which is great for large protein and works with density maps. It runs with Rosetta XML scripts and with PyRosetta. I use the latter, but I believe the paper has an XML in the SI. If not, I can through together a colabs notebook.

It is nice, but can segfault with nasty density maps, in which case not using a density map is better. Specifically it is those map with weird densities wallpapered on the faces of the cell —I have no idea why or what these are.

However, LocalRelax runs off a modified cartesian scorefunction, which means that you either use pmut_scan with that scorefunction —not recommended as unlike dihedral ref2015 it is not 1 REU ~= kcal/mol — or run three or more cycles of FastRelax with ref2015 with a movemap preventing backbone movements —way quicker.

MPI --> the parallelisation is for multiple structures, not for a single one.

Mon, 2021-09-20 06:05