Closed jadolfbr closed 9 months ago
Looks like this is correlated with temp of course. At sampling temp used for binder design in RFDiffusion (.0001), this does not occur; though it sometimes occurs at .1, so mostly just increase # of samples, sort by global and go from there.
I'm wondering if the loss function is on the local or global score? I ask this because I've sometimes had instances where the local score would go down, but the global score would go up compared to the native protein.