Similar tool as NN for AF-similarity?

[ichwilldiesennamen]

The Neural Network is a helpful tool for improving the AlphaFold confidence (even though for me I try to use a designflow which typically gives >80% AF values without using NN). So it highlights regions where you should mutate.
But AF-similarity seems to be a bit difficult to control/optimize. The best way for me to improve this seems for now to load the prediction and work with that. But that screws up a lot of things (especially the score) and it takes some effort to fix it again while keeping the good AF-sim value.
Therefore do you think it's possible to have a similar tool like the NN that shows especially similarity issues? So regions that should be changed to improve similarity? Or should I try to load the prediction as guide and do modifications manually to bring them in line? Haven't tried that yet.
Anyway, if you think such a tool would be possible then I would be interested. And thx again for the NN tool which seems to show pretty reliably (as long as the design is not totally screwed up) regions that are problematic for AF-conf!

[bkoep]

Good question! I can think of some alternative ways to measure "local similarity." But actually I believe the current Neural Net Objective may still be the best option for indicating where mutations may improve similarity.

The Neural Net Objective works with AlphaFold distogram data. The distogram is a kind of abstract representation of the protein structure; it is not directly related to confidence nor similarity. But the distogram does contain some rich information about how your design "appears" to the AlphaFold network. Regions with poor local similarity are expected to show up red by the Neural Net Objective.

There are certainly limitations to the Neural Net Objective, which only considers local structure. Long range, non-local interactions contribute to the similarity score, but are not evaluated by the Neural Net Objective. If a low-similarity structure is colored all blue by the Neural Net Objective, then the low similarity probably comes from differences in the overall fold (which should be clear if you compare the AlphaFold predicted structure with your original structure).