DiffDock
DiffDock copied to clipboard
Rescoring with confidence model
I'm currently trying to rescore poses using the confidence model without running reverse diffusion. When I performed confidence model inference after setting time to zero (like in sampling()), it seemed fine without any errors. However, when I compared the confidence scores predicted this way with those obtained after reverse diffusion using DiffDock-generated poses, they showed significant differences.
I'm curious to know if anything inside the HeteroData, apart from the heavy atom coordinates, is being modified during reverse diffusion, or if I've made mistakes somewhere.