Validity of Comparing Binding Energies for a Series of Mutations in a Native Peptide

[nyyount]

Hi,

I’m wondering whether it’s valid to compare the Prodigy binding energies and dissociation constants, for a series of peptide ligands where one amino acid is altered in each subsequent protein (peptide 1 - native; peptide 2 - one aa change; peptide 3 - two aa changes; etc). We are trying to introduce mutations that lead to a reduced affinity of toxic peptides for a receptor, and are hoping to pre-screen molecules in silico.

My peptide receptor pairs have been docked with Haddock.

Thanks so much!

Nannette

[amjjbonvin]

Hi Nannette

You rather want to calculate changes in binding affinity (ddG) upon mutations. We wrote a review and an article (iSee) about it you might want to check:

• C. Geng. Li J. Roel-Touris and A.M.J.J. Bonvin. Finding the ∆∆G spot: Are predictors of binding affinity changes upon mutations in protein-protein interactions ready for it? WIREs Computational Molecular Science 9, e1410 (2019).

• C. Geng, A. Vangone, G.E. Folkers, L.C. Xue and A.M.J.J. Bonvin. iSEE: Interface Structure, Evolution and Energy-based machine learning predictor of binding affinity changes upon mutations. Proteins: Struc. Funct. & Bioinformatics 87, 110-119 (2019).

Not sure PRODIGY would be the best approach (but it is fast, so don’t hurt to try).

If you interface is rathe polar, simply refining the mutants with the refinement interface of HADDOCK and comparing the scores (do also refine again your docked model), might already give quite some good results (e.g. for the famous barnase-barstar complex, this simple approach gives you already a R>0.8, but this is a highly polar interaction).

This was recently done for example in the following works:

• https://www.biorxiv.org/content/10.1101/2020.06.05.136861v1

• https://covid19.ibg.edu.tr/ibg-researchers-successfully-predicted-the-impact-of-ace2-designer-mutations/

Cheers
Alexandre

[nyyount]

Dear Dr. Bonvin,

Thank you so much for your rapid reply! It has been very helpful.

I’ve reviewed the papers, and discussed trying to incorporate some of these methods with our group.

We are currently using a machine learning system with two relatively discreet, structurally (but not functionally) related, classes of peptides. We are trying to remove the toxicty (affinity for a channel) from one group by training it to become more like the related non-toxic group. We are using Haddock, and another docking program, to score the success of our peptide transformations. Based on the references you supplied, we are considering trying to add an energetics score or function to our datasets. Do you think using general mutation based energetics data from SKEMPI 2.0, or similar, can be applied to a more specific (our two peptide groups) situation?

Thanks,

Nannette

[amjjbonvin]

I guess if you are developing some machine learning model, then you can indeed add as many scores as possible