search

HannesStark / EquiBind

EquiBind: geometric deep learning for fast predictions of the 3D structure in which a small molecule binds to a protein
MIT License
469 stars 110 forks source link

Scale off for results #36

Closed jadolfbr closed 2 years ago

jadolfbr commented 2 years ago

The program runs without issue, but the scale of the SDF molecule does not match the input scale/protein. Have you seen this before? Is there any recourse here?

Screen Shot 2022-06-16 at 1 10 33 PM Screen Shot 2022-06-16 at 1 10 16 PM
HannesStark commented 2 years ago

I have not encountered this before and am not sure how the issue occurs

jadolfbr commented 2 years ago

The odd thing is, the molecule in general is in a good docked place. The SDF molecule is preserved as well, but the scale is off.

Screen Shot 2022-06-16 at 1 18 32 PM
jadolfbr commented 2 years ago

I have now tried with Mol2 format and PDB format and I still get the same strange scale, so something is wrong with the post-processing. They are consistently the same scale - IE C-C bonds are 6.1A. I will see if I can share the inputs that I am using.

HannesStark commented 2 years ago

I can not reproduce this issue and have also not encountered it before with other people using EquiBind. Sorry that I cannot be of further help here. Please let us know if you have any further insights!

jadolfbr commented 2 years ago

Sounds good. Looks like the Cuda torch version was different than system Cuda toolkit system version - 11.3 vs 11.4. I installed 11.4 and re-running on both GPU and CPU to see if anything comes out from that. Also working on being able to get you exact inputs to see if it can be reproduced, but that might take some time to get through.

jadolfbr commented 2 years ago

So digging further into this, the problem was a mismatch between cudatools version and PyTorch on our cluster. The issue is now resolved!