Hi!
I've seen that the prediction scores may complete change depending on the long range context... For example, the same protein domain is predicted to have different interfaces if a second (native) domain is included in the calculation, even though the 2nd domain is in the opposite face and 2 nm away from the initially predicted interface...
Similarly, I've found that mutating high scoring residues can lead to the prediction of new interfaces of equally good scores at a completely different (and distant) surface in the protein...
I wander if these issues are "artifacts" of the attention layer, which may tend to prioritize and rank different interfaces, while just pointing "the best of the best" with a given confidence... if that is the case, then the computed score may not have a coherent or "absolute" scale, it will depend on how other interactions are being evaluated in the system... Hence, there seems to be room for finding even more hidden interaction interfaces...
Hi! I've seen that the prediction scores may complete change depending on the long range context... For example, the same protein domain is predicted to have different interfaces if a second (native) domain is included in the calculation, even though the 2nd domain is in the opposite face and 2 nm away from the initially predicted interface...
Similarly, I've found that mutating high scoring residues can lead to the prediction of new interfaces of equally good scores at a completely different (and distant) surface in the protein...
I wander if these issues are "artifacts" of the attention layer, which may tend to prioritize and rank different interfaces, while just pointing "the best of the best" with a given confidence... if that is the case, then the computed score may not have a coherent or "absolute" scale, it will depend on how other interactions are being evaluated in the system... Hence, there seems to be room for finding even more hidden interaction interfaces...
Best