Monomer models missing?

[jkosinski]

I have run this example from the repo:

Q8I2G6_Q8I5K4
Q8I2G6
Q8I5K4

and then got only the prediction for Q8I2G6_Q8I5K4 (/scratch/kosinski/TestAlphaAbriss/predictions/). I would expect from the format that I should get monomeric predictions for Q8I2G6 and Q8I5K4. Or that means sth else? MAybe related to the errors in issue #1 that I got in the same run?

[DimaMolod]

I got into the same issue today

[maurerv]

Seems to be related to AF internals. I ran monomers of W7JYJ3 for testing and everything appeared fine.

Would be great if someone could try to reproduce this from within an AP enviroment with different JAX version. Simply running alphafold_inference.py from the command line should suffice

[jkosinski]

Could it be that you pass the monomer to multimeric modeling pipeline? @dingquanyu will know better but I think the script needs to recognize whether the job is a monomer and pass it to the appropriate AlphaFold modeling pipeline (AlphaFold has different pipelines for monomer and multimer). Or maybe you create MultimericObject but pass it to monomer pipeline. I think it might be handled here: https://github.com/KosinskiLab/AlphaPulldown/blob/2443672127ef31e32fcf2353c6f4784a37a1d9f3/alphapulldown/run_multimer_jobs.py#L264

[maurerv]

Then it would not work on W7JYJ3 monomer, no?

The only difference is the model preset in the runner and I handle that case in alphafold_inference.py

[maurerv]

@DimaMolod, @dingquanyu would be great if one of you could test alphafold_infernce.py using your AP conda environment from the CLI to double check its not due to the jax version

[maurerv]

So I did a few experiments and Dima is cross-checking on different jax version.

Basically, the problem is that the number of recycles. If that is less than or equal to three, everything will work fine for monomers. I guess this issue is simply jax related and I will upgrade the container to a newer version.

@jkosinski, @DimaMolod, @dingquanyu Dima tried running more than 3 recycles through AP monomer mode and it also failed with the same error. So the problem is either the way AP prepares the data for AF or AF internally. I will open an issue on this in the AP github

[jkosinski]

ok, then you can close it here and we handle that there, thanks for tracking this down!