Interpreting predictions_with_good_interpae.csv

[calebmallery]

Wanted to say thank you for the development of this great tool, it has been extremely useful for generating hundreds of structures in a high-throughput way!

My question would be if you have any recommendations on how to confidently call protein-protein interactions. For example, I have ran a homo-dimer as bait, with many candidate proteins and would like to know which protein-protein interactions between my bait and candidate might be "real:. My hesitation with using AF2 iptm score is that an experimentally confirmed protein-protein interaction receives an iptm score of 0.44, we also believe this is not indicative of a poor model because the predicted structure superimposes with our cryo-EM structure very well. Additionally, the iptm score seems a little bit biased to small proteins, which makes sense but are often not well characterized. When looking at the pi score, this does not seem to be very discriminatory, and looking at the int_area seems to be a bit biased towards membrane proteins whose interactions would not be feasible for my protein bait of interest.

Do you know what the best approach might be to help somewhat confidently call potential protein-protein interactions? Are there any scores that your group or any collaborators have used that might have led to experimentally validated interactions?

Any advice or suggestions would be greatly appreciated, thank you!