2

I have a peptide X of length 50aa. Alphafold2 predicted the structure which takes a helix form and with high average pLDDT score (>90).

But during actual experiment (e.g. crystallography, spectroscopy), peptide X is determined to have a random coil structure.

My question is how can I resolve the difference between AF2 prediction and crystallography? To what extend we can use AF2 peptide structure prediction for downstream analysis (e.g. docking, etc)?

neversaint
  • 131
  • 4

1 Answers1

2

Possibility 1: technical error

AlphaFold2 uses MSA to infer residue co-evolution (due to epistasis) and thus proximity. A MSA can become contaminated with incorrect homologues, especially shorter protein. Generally these will have low scores and be filtered out, but if they don't they will come in like a wrecking ball.

Possibility 2: conformational switching

Even though on circular dichroism a random coil and helix could not be more opposites, several proteins adopt both conformations and switch. Amyloids are a classic example of this, switching between a beta sheet and a random coil. Synuclein is another example.

A technically problematic (vide infra) alternative option is that the helical conformation is only found when the protein is bound to another.

Solution: AlphaFold2 testing

Whereas EBI-AF2 provides a single model, the AF2 algorithm can provide multiple possible solutions, which in many cases are actually bona fide alternate conformations. On a protein I tried, the phosphorylations in PhosphoSitePlus (high-throughput) explained wonderfully a second model. The model with the best score is not the "most normal" conformer: not only such a statement makes no sense, but is the one that best obeys the evolutionary signals. I would suggest looking at https://github.com/sokrypton/ColabFold Set the target number of decoys to more than 5 and if all are helices then it's a technical issue. In particular the MSA data would need to be dissected if one really cared. But if nothing comes up... it would be a low reward area of investigation as one would need to discount other less and less likely possibilities.

If the decoys do show alternative folds, i.e. coils, then a few things spring to mind that could be tested in silico (applicable only to some cases):

  • different pH sidechain protonation
  • ion binding
  • phosphorylations
  • binding to nucleic acids
  • binding to known protein candidates —not worth exploring if there's no very likely binding partners. If different paralogues bind differ targets it's tricky
  • binding to membranes

Looking in PyMOL at the location of the high pLDDT scores and after running APBS (Adv. Poisson-Boltzmann Solver) within PyMOL may reveal something in the helical AlphaFold2 model. Once the conditions which explain the helical model (e.g. bond to DNA as a heterotetramer in the presence of zinc at pH 10 has a very low ∆G), those conditions could be tested in vitro.

Matteo Ferla
  • 4,234
  • 5
  • 19
  • thanks for your reply. 1) When you said "A problematic alternative option..." did you mean, it's very unlikely that it happens? 2) How can I tested different pH, ion binding in-silico? 3) For Adv. Poisson-Boltzmann Solver is that an option in AF2?, if not how can I get it? – neversaint Apr 08 '22 at 01:42
  • Technically problematic. Is very common, but protein protein binding requires you to know what the other protein is —string DB etc. will give you loads that don't hold true— and preferably where —XLMS datasets are sparse. ColabFold can do hetero oligomers but paralogues with different targets lose the res' co-evo signal 2. The various tests can be done in Rosetta/PyRosetta but require a lot of coding and know-how. 3. APBS is a plug in PyMOL —I spelt out the acronym just because I'd would have scrambled the letters had I not.
    • – Matteo Ferla Apr 08 '22 at 08:29