In lab experiment, binding affinity is pretty strong, but pae_interaction is only around 28

[Huilin-Li]

So, I have an example protein A that is very similar to my protein B I aim to study. The protein A has a binder A-binder, and I use A-binder as a scaffold to design my binder B-binder to my protein B.

As I learned here, at the last step, I will use af2_initial_guess/predict.py to assess designed binders and pick binders whose pae_interaction<10 .

However, pae_interaction~=28 between protein A and A-binder. I feel confused about it a lot, because its binding affinity is very strong in lab experiment.

I worked very hard to think of how to improve pae_interaction below 10. Unfortunately, over 95% pae_interaction is around 26~28. Should I still work hard to look for pae_interaction<10 ?

I don't understand these two scenarios: 1) why pae_interaction is bad, but binding affinity is very strong in lab experiment? 2) Is pae_interaction<10 very important? Is there any other criteria I should to consider when I assess my designed binders?

Thanks a lot!!!

[Jiny-Wang]

hi, I have also encountered this problem and not seen a reasonable explanation so far.

I think the calculation of pae_interaction is related to the size of binder (mine is about 70) and the interface area. How many aa does your B-binder have?

At the moment it looks like: "pae_interaction < 10 is a good predictor of a binder working experimentally", but if your lab experiment are good, other evaluation scores may be needed.

[nsh87]

PAE is definitely one proposed metric to correlate with binding, but 3D interactions are complex and other metrics like from certain PLM likelihoods show better correlation (definitely so for certain molecules). I would identify a correlative metric, then optimize for that rather than picking a model off the shelf. We are here to calculate and test PAE correlation vs other metrics in the lab, as well. PAE shows almost no correlation to our other metrics themselves. Feel free to DM if you want some advice

[uioQPS]

PAE is definitely one proposed metric to correlate with binding, but 3D interactions are complex and other metrics like from certain PLM likelihoods show better correlation (definitely so for certain molecules). I would identify a correlative metric, then optimize for that rather than picking a model off the shelf. We are here to calculate and test PAE correlation vs other metrics in the lab, as well. PAE shows almost no correlation to our other metrics themselves. Feel free to DM if you want some advice

what feature did you choose to get the correlative metric from the software? I want to find some filter parameters but it seems can't work out sadly..Thx :>