ketatam / DiffDock-PP

Implementation of DiffDock-PP: Rigid Protein-Protein Docking with Diffusion Models in PyTorch (ICLR 2023 - MLDD Workshop)
https://arxiv.org/abs/2304.03889
195 stars 38 forks source link

location of confidence model score? #2

Closed HeJunhong1107 closed 1 year ago

HeJunhong1107 commented 1 year ago

Thanks for the amazing work,and I have finished the test demo of DIPS,Do there have a specific file for users to check the confidence scores of protein pairs?

HeJunhong1107 commented 1 year ago

ps:Dose there exist a cutoff for “good” confidence score?

ketatam commented 1 year ago

Hi! Thanks a lot for your interest in our work. Great to hear that you have finished the test demo of DIPS.

Currently, we do not have any such file that allows for checking the confidence scores of protein pairs. Could you elaborate more on the format of such a file and how you would use it?

To answer your second question, the confidence score is the output of the last layer of a binary classifier (the logit) that would be passed through a sigmoid to get a probability value between 0 and 1 (of whether the generated pose is within 2Å of the true one). Therefore, a value of 0 for the confidence score corresponds to a probability value of 0.5, which is usually considered a cutoff for a "good" value in classification tasks but is however not fixed and you can choose any other value with this logic, depending on how you wanna set the threshold probability. However, note that since we are ranking the generated poses by their scores, the actual values do not really matter, only their order does (that is also why we do not apply the last sigmoid activation and we just rank the outputs of that last linear layer)

Hope this helps

HeJunhong1107 commented 1 year ago

thanks for your kind reply~