-
Notifications
You must be signed in to change notification settings - Fork 83
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Strange output from inpainting #24
Comments
@hnisonoff Thanks for the feedback! can you parse sde_func='langevin' as additional sampling args. This enables preconditioned langevin dynamics instead of reverse diffusion, which should be the default for infilling. |
@wujiewang thank you for the help! I added Should the example in the ChromaAPI.ipynb be modified to reflect this? Also, in that example I see the following line: |
Yes, let me update the code! will do a small 1.0.1 release as we gather more improvements! Thanks for bringing this to our attention! A quiet subtle point! |
Hi, thank you for chroma! I also have problems with infilling. My plan is to take an input structure and design a new domain of specified size (e.g. 100 AA) into one of its loops. This is the code that I use, which is a combination of code snippets that I have found in your examples:
It sort of works. The original structure is left untouched, but the new design is not forming a globular domain but rather very strangely wrapps around the input structure. I wonder if I do anything completely wrong. Thanks for your help! |
Do I perhaps have to combine this with a shape conditioner? |
Hi @gha2012 What you described is possible, the results can depend on the geometry of the motif. I think it is good idea to try out the shape conditioner. You can also write a conditioner that uses a restraint function that penalizes structures with large radius of gyration to make it more globular. |
I have been playing around with inpainting and after replicating the examples that you provided I tried it out on my own protein.
I found that while inpainting a region of 40 amino acids out of 157, that the output from Chroma was a protein in which it appeared to explode with a completely broken chain. I then attempted to modify the various SDE / Langevin dynamics parameters but these did not result in more stable designs. In the zipped folder attached, I have the following files:
original_protein.pdb
is the starting protein from which I am trying to inpaint a subset of the residues (indices 80-139)large_inpainting_final.cif
is the structure output from the inpainting designlarge_inpainting_traj.cif
is the trajectoryTo debug this further I tried reducing the size of the inpainting. I noticed when inpainting only 10 residues, Chroma provided reasonable looking designs. These files are:
large_inpainting_traj.cif
small_inpainting_traj.cif
Next, I increased the inpainting size from 10 to 20 residues. This time, things did not explode, however, Chroma inpainted a region that forms a knot with the rest of the protein and clashes quite significantly. These files are:
medium_inpainting_final.cif
medium_inpainting_traj.cif
The code I used to do the inpainting is:
As mentioned above, I also tried modifying
Langevin_factor
,Langevin_isothermal
, andinverse_temperature
.If you are able to provide any insight into why this may be happening or provide guidance on what I may be doing wrong, that would be great. It is also possible that this is just a case where inpainting seems to not work well for whatever reason (sampling is hard, protein is out of distribution enough). Thank you!
for_github.zip
The text was updated successfully, but these errors were encountered: