Error in Alpha Fold multimer

[galshuler]

Hi, I am running an Alpha fold multimer, and after a few hours of running I get the following error: ValueError: Minimization failed after 100 attempts Or : ValueError: The number of positions must match the number of atoms

The error varies depending on what I define the A chain and what I define the B chain. I would be happy to help how the error can be resolved.

Thanks in advance, Gal

[Augustin-Zidek]

Hi,

could you try setting run_relax=false? This means the relaxation step will be skipped (and there might be some clashes in the output structure), but might help with your issue.

[berjeh]

Dear @Augustin-Zidek does this happen for all structures or certain ones? What if I would want to see the relaxed versions? I am also receiving the same error

[Augustin-Zidek]

Relax has problems usually if there is large number of clashes in the predicted structure, so this should happen only occasionally.

With the newly release v2.2.0 the clashes in AlphaFold-Multimer should be much less frequent. Could you try updating your AlphaFold installation to see if that helps?

[avilella]

I tried the v2.2.0 version today, and my code won't work with run_relax true, giving the same error as the one reported by @galshuler , but it does seem to go further with run_relax=false. This is on an NVIDIA RTX A4000 with 16GB of memory.

[Rampakslue]

Hello!

Did you resolve your issue? Otherwise, it might be memory issue. Our HPC uses A100 with 40GB, and running multimer ends up maxing out the VRAM quite a lot.

[Aroids]

Would a memory issue report a value error as opposed to a OOM error?

[Htomlinson14]

Hi! In our experience the relaxation stage is less stable when use_gpu=True. This will be changed in the next release to default to False. In the meantime, this can be changed manually in the Colab.

[Aroids]

Will the next release (of the local installation of AlphaFold Multimer), attempt to correct the cause of the ValueError arising during relaxation?

For large complexes GPU are seemingly required. While turning off relaxation prevents the ValueError, it also results in structures that are less suited to applications such as molecular modelling (e.g., correct charges cannot be assigned to some atoms... perhaps due to abnormal bond distances/angles).

[avilella]

Hi Samuel and others,

Most of my attempts were on a 16Gb NVIDIA card, but we are considering a 24Gb to see if we can run the multimer predictions with a better parameter set, and thus potentially obtain a better prediction.

Is there any back of the envelope calculation that could give us any guidance as to which parameters and values we would be able to change to accommodate a full prediction with the 24Gb card? I am not sure if the A100 40Gb cards are available for bog standard workstation machines, but I can see models with 48Gb VRAM.

On Tue, Aug 23, 2022 at 3:28 PM Aroids @.***> wrote:

Will the next release (of the local installation of AlphaFold Multimer), attempt to correct the cause of the ValueError arising during relaxation?

For large complexes GPU are seemingly required. While turning off relaxation prevents the ValueError, it also results in structures that are less suited to applications such as molecular modelling (e.g., correct charges cannot be assigned to some atoms... perhaps due to abnormal bond distances/angles).

— Reply to this email directly, view it on GitHub https://github.com/deepmind/alphafold/issues/383#issuecomment-1224156673, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABGSN2QL42MATKC6PDQCG3V2TNZRANCNFSM5PQ3ETSQ . You are receiving this because you commented.Message ID: @.***>

[Augustin-Zidek]

Is this still an issue in v2.3.1? We fixed a bunch of relax issues.

[Augustin-Zidek]

Closing this issue now, feel free to re-open or open another issue if this is still a problem.