Confidence Module Output

[ludwigwinkler]

Dear authors,

first of all thank you for your work and for releasing everything in an accessible manner for others to incorporate into their research. :)

We're working on top of the confidence predictions.

1. Is my understanding correct that the confidence is modeled by a single logit which we can pass through a sigmoid to obtain a probability of confidence?
   It seems to be used as such in the inference code:
   

   DiffDock/inference.py

   Line 278 in b4704d9

   if confidence is not None:

   
   Upon some searching I found
   

   DiffDock/confidence/dataset.py

   Line 193 in b4704d9

   complex_graph.y = torch.tensor(rmsds[sample] < self.rmsd_classification_cutoff).float().unsqueeze(0)

   
   where the labels are scalar quantities per sample and the confidence module predicts a single scalar as well to be trained against here

   DiffDock/confidence/confidence_train.py

   Line 130 in b4704d9

   confidence_loss = F.binary_cross_entropy_with_logits(pred, labels)

2. We evaluated the confidence logits for the provided examples and got the following values:
   Ligand 1a46: roughly -2.11 -> ~10%
   Ligand 6w70: roughly -1.10 -> ~24%
   Ligand 6moa: roughly +0.38 -> ~60%
   Do these values seem reasonable?

Thanks again for all the work you've put into this.

Cheers.