Questions about accuracy: average of DDGun3D and DDGunSeq; multiple mutations

[tuckerburgin]

In the use case where a 3D protein structure is available, it should always be possible to also obtain a sequence. Is there any reason to ever prefer the results from DDGun3D alone, rather than running both 3D and seq and taking the average result? You would generally expect the result from the 3D code to be a little more accurate on average, but the average of the two methods should have substantially lower mean squared error (that is, it should be more precise).

Also, I have a question about the "T_DDG" values reported by the software. In my understanding, usually in the literature the stabilizing/destabilizing effects are assumed to be additive, but according to your paper, DDGun uses the maximum of the single mutation energies, plus the minimum, minus the mean. It is explained in the paper as such:

In case of more than two mutations, the most relevant points that may affect the total ΔΔG are the minimum and the maximum values, so that we decided to take their sum and centre them in the mean...

Can you provide any further explanation of this? It produces some very unexpected results, such as for example, if you apply dozens of mutations that all individually have an expected ∆∆G of 1 kcal/mol, versus only a single mutation with that ∆∆G, DDGun would report the same "T_DDG" of just 1 kcal/mol in either case.

Thanks!