bug: pdb_2022_09_28_mmcif_files.tar replacement with mmcif_files

[hegelab]

Hi,

You introduced this into run_alphafold.py, line 185:
pdb_2022_09_28_mmcif_files.tar # ~200k PDB mmCIF files in this tar.
mmcif_files/ # Directory containing ~200k PDB mmCIF files.

So run fails, since mmcif_files does not exists.

[Augustin-Zidek]

Yes, this was done to significantly speed up template search. You will have to untar your PDB database (the download script has been updated to untar it).

See https://github.com/google-deepmind/alphafold3/blob/main/fetch_databases.sh#L28 for the exact command to run on the tarfile.

[charlesbeattie]

You can pass the following flag to restore the original behaviour:

--pdb_database_path='${DB_DIR}/pdb_2022_09_28_mmcif_files.tar'

This will be considerably slower for each run of alphafold, so I would recommend untaring that file and keep the default.

[hegelab]

Thanks :-)

[hegelab]

It came into my mind:

If the structure search takes a significant amount of time then you may want to add an option not to perform it (I do not see this option; you can replace the template list with an empty list before starting inference, but this is post-processing after completed template search).

In most of my past cases with AF2 I needed to run the prediction without structural template. I suppose that AF3 works without structural templates as well as with templates (similar to AF2).

[Augustin-Zidek]

Yes, you are right, I will add an option to disable template search.

But I am also fixing the template search performance, so should be less of an issue.

[kkorotkovuky]

Yes, you are right, I will add an option to disable template search.

But I am also fixing the template search performance, so should be less of an issue.

The template-free option will be extremely useful. Also, having the option to limit the template search by the release date - as implemented in AF2 - would be great.

[hegelab]

May be important: in AF2 the filter-by-date was performed after the template search was completed; I would not perform the search on mmcif entries which can be already excluded based on the date. E.g. if I restricted the templates for later than 2050, the search was performed on all entries and finally none was used, since we were only in 2021.

[Augustin-Zidek]

Template search should now be much faster (up to ~100x in the mmCIF fetching and parsing stage after Hmmsearch) thanks to d6b06d6.

Starting work on the template-free and date filter features.

[Augustin-Zidek]

The ability to run template-free was added in 1942639.

[Augustin-Zidek]

Max template date flag added in e0cfd70.

I am going to close this issue as everything reported in here has been resolved. Thanks everyone for chiming in. Summary:

• d6b06d6: template search is now up to 100x faster in the post-Hmmsearch stage.
• 1942639: you can run template-free by setting templates to [] while keeping unpairedMsa and pairedMsa unset.
• e0cfd70: you can control max template date using the --max_template_date flag, just like in AlphaFold 2.

Happy folding -- be it with or without templates. :)

[Augustin-Zidek]

Sorry, I missed on case (MSA set to empty, templates unset). Fix in 4bebfb0.