IPSAE lower than ipTM?

[xtzhang0216]

Hi, thanks for the great work — this tool is very useful!

I have a question: when evaluating my structures, I found that the IPSAE scores are consistently lower than the AF3 ipTM, even when I set the PAE cutoff >32. My understanding is that ipSAE, by excluding disordered regions, should give higher scores in such cases.

Thanks!

[dunbrack]

I'm glad you're finding it useful. When you use PAE cutoff of >32, you are not excluding disordered regions anymore.

The difference is in how d0 is calculated. AF2/AF3 use the total length of both proteins (assuming two proteins in the model) to calculate d0. ipSAE uses the length of only the "scored protein." That is, for every residue in chain A, all of chain B is scored and d0 is a function of len(chain B). For every residue in chain B, all of chain A is scored and d0 is a function of len(chain A). AF2/AF3 uses d0 as a function of len(chainA) + len(chainB) and therefore d0 is larger (it's proportional to the cube root of the length).

Because AF2/AF3's value of d0 is always larger than ipSAE's, the denominator of the pTM_ij expression, (1+ (dj/d0)**2 ) will always be larger than ipSAE's. Therefore the pTM_ij expression will be smaller:

pTM_ij = 1/ (1+(dj/d0)**2).

When the PAE cutoff is a useful value (10 or 15), then ipSAE may be larger or smaller than ipTM. It's larger when there is lots of disorder in both proteins (or non-interacting domains), because their poor pTM_ij's are not included in the averages. It's smaller if one of the two proteins is completely ordered, for the d0 reason above.

Thanks for asking this question.

Roland