AF3 cannot predict correct ligand isomer even using userCDD

[antonielgomes]

Hi,

I'm trying to predict a protein-ligand complex, but AF provides the wrong isomeric form of the ligand.
To avoid that, I generated the ligand with openbabel (which gives the correct isomer), then I generated a one-line cif file for the userCCD option in the JSON input. No success!

I feel that AF3 is not considering my input, or I'm not using the correct flags.
Here's the command :

> docker run     --volume ${input_dir}:${input_dir}\\
               --volume ${output_dir}:${output_dir}\\
               --volume /mnt/bio-8t-disk1/alphafold/:/mnt/bio-8t-disk1/alphafold/ \\
               --volume /mnt/bio-8t-disk1/alphafold/public_databases:/mnt/bio-8t-disk1/alphafold/public_databases \\
               --gpus all alphafold3 \\
  python3 run_alphafold.py --model_dir=/mnt/bio-8t-disk1/alphafold --db_dir=/mnt/bio-8t-disk1/alphafold/public_databases \\
               --json_path=${input_dir} --output_dir=${output_dir} --input_dir=${input_dir} \\
               --pdb_database_path='/mnt/bio-8t-disk1/alphafold/public_databases/mmcif_files' \\
               --norun_data_pipeline --buckets 512 --num_diffusion_samples `10

I've tested all versions (1, 2, 3, and 4) in the JSON file with no success.
I obtain the correct isomer in 1 out of 10 conformers. Checking every prediction is not feasible when performing virtual screening...
I installed AF3 in the past week. It means that I'm using the newest version.

Is there a command or flag to force AF to always generate the correct isomer from userCCD?

Best,

[Augustin-Zidek]

Hello,

could you try providing the ligand using the userCCD option and commenting out the following if-block (lines 1601--1616):

alphafold3/src/alphafold3/model/features.py

Line 1601 in a8ecdb2

if mol is not None:

This will force AlphaFold to use the provided userCCD instead of generating conformers using RDKit. This might make AlphaFold output the correct isomer. Make sure to have the ideal coordinates set in the userCCD CIF (_chem_comp_atom.pdbx_model_Cartn_{x,y,z}_ideal).

But note that even with this the model might still not do the correct thing (cf. the discussion of chirality in the AlphaFold 3 paper) and you might need to get multiple samples and select the right ones programatically.

[antonielgomes]

Thanks for the fast reply.
I'll make the suggested changes and test it again.
Yes, I'm using the ideal coordinates.
I hope it will work. I'll get back to you as soon as possible.

[YoavShamir5]

Interesting discussion! to be clear, @Augustin-Zidek, you suggestion to @antonielgomes would guarantee correct stereochemistry at the input to the model (if the user's provided CCD becomes a valid molecule and respects the desired stereochemistry), but even if this is the case, the model is not guaranteed to output the ligand with the required stereoisomer - is this true?

[antonielgomes]

Hello again.

I commented the part of the code as suggested, but I'm still getting the same results.
Should I comment the code and then recompile the docker file (docker build -t alphafold3 -f docker/Dockerfile .)?

@Augustin-Zidek I'm attaching an example containing the input json file, the command to run AF3, and the ligand cif file (including the one-line file) so you can check it...

It would be so helpful if we could solve this.

@YoavShamir5 yes, I guess. According to the "Model limitations" section of the AF3 paper, it still generates inappropriate isomers even when providing the CCD file (errors of 4,4%).