Flex ddG: re-weighting scores

[Sh-Yana]

Dear Rosetta community,

I have another question about the flex ddG protocol – I am interested whether the scores predicted by it require an additional re-weighting/re-scaling to better fit some experimental benchmarks.

This question has arisen from a discussion I had about other Rosetta methods. For example, Rosetta Cartesian, as I understood it, outputs the scores in Rosetta Energy Units (REU), which are then multiplied by a scaling factor (1.0/1.84 in case of the talaris2014 function) to fit the actual experimental benchmarks in kcal/mol – it is described on page 20 here:
https://pmc.ncbi.nlm.nih.gov/articles/instance/5515585/bin/NIHMS867713-supplement-Suppl__Mat_.pdf.
“The energy gap between the refined mutant structure and the refined wild-type structure, multiplied by a energy-function-specific scaling factor, becomes the predicted mutational ddG. Scaling factors are introduced to fit the overall scale of estimated values to actual experimental free energies measured in kcal/mole. A least-squares fit determined a scaling factor for talaris2014 of 1.0/1.84 and for opt-nov15 of 1.0/2.94. <...> Correct assignment of stabilizing mutation implies predicting ddG < -1.0 for a mutation with experimental ddG < -1.0, and correct assignment of destabilizing ones is the opposite (ddG > 1.0).” The experimental benchmark set used by Rosetta Cartesian was mentioned here: https://onlinelibrary.wiley.com/doi/epdf/10.1002/prot.22921?src=getftr&utm_source=acs&getft_integrator=acs

Flex ddG also uses talaris2014 function, whose scores in REU are then re-weighted using a GAM nonlinear re-weighting scheme, to the best of my understanding. The objective of this re-weighting was to reduce the absolute error between flex ddG predictions and known experimental values. Therefore, as I understand it, flex ddG was benchmarked on a different experimental dataset (ZEMu) than Rosetta Cartesian, the GAM re-weighting was already implemented, and the ddG talaris2014-GAM scores, as it seems, can be used as is (?).

But still, there are many deeply technical things to consider, and I am worried I might have misunderstood or missed something important. I kindly ask more advanced users to share their knowledge and experience with these methods. Do you use the talaris2014-GAM scores from flex ddG as is, without any additional re-weighting? After GAM-reweighting, are these scores in kcal/mol, or still in REU? Do you interpret these scores using the criteria suggested by the authors: “stabilizing mutations are defined as those with ΔΔG ≤ −1.0 kcal/mol, neutral as those with −1.0 kcal/mol < ΔΔG < 1.0 kcal/mol, and destabilizing as those with ΔΔG ≥ 1.0 kcal/mol”?

Thank you so much in advance!
Yana