[BUG] CIF templates seem to get ignored

[claggner]

Hi, I'm struggling with CIF Direct Templates approach - it seems my templates get ignored.

(Apologies, I can't share the protein sequence or CIF templates here due to confidentiality:) My protein A has experimental structures in the PDB where it exists in a conformation 1. From close homologs, I know that these proteins also exist in a significantly different conformation 2 (think active vs. inactive state). I want to model protein A in conformation 2, using CIF templates from its homologs.

I added several CIF templates via "template_cif_paths" in the .json file. But when I run the prediction, I still only get models in conformation 1 - just like when I don't use templates at all. Adding "--use-templates=true" seems to have no effect. Likewise, "--use-msa-server=false --use-templates=true" produces a diverse set of seemingly random protein structures, without indication that the provided CIF templates contributed to the structures.

Configuration:

• OS: Fedora 44
• NVIDIA GeForce RTX 4070 16 GB
• current version via pip

[jandom]

Thanks for reporting this @claggner – is there any chance you could share more details?

Specifically

• did you run from pip or main (which commit)
• commands you ran, step-by-step
• any inputs needed

Please don't share any proprietary data but i suspect you can reproduce this on a public PDB input

[claggner]

Hi @jandom, thanks for your reply.

I installed openfold a while ago in its own conda environment (python=3.13), and updated it on July 1st with pip install -U openfold3[cuequivariance].

I managed to recreate the problem with a non-proprietary example. Let's say I want to model the kinase HCK on the PDB template 1Y57 from the related SRC kinase. Similar to before, the models I got look like the Xray structures that have been reported for HCK (e.g. 9BYJ) rather than the template. Adding/removing --use-templates=true had no noticeable effect.

Command and output:

run_openfold predict --query_json=hck_human_src_1y57_templ.json --runner_yaml=runner.yaml --use-templates=true
/home/christian/miniforge3/envs/of3_env/lib/python3.13/site-packages/torch/__init__.py:1617: UserWarning: Please use the new API settings to control TF32 behavior, such as torch.backends.cudnn.conv.fp32_precision = 'tf32' or torch.backends.cuda.matmul.fp32_precision = 'ieee'. Old settings, e.g, torch.backends.cuda.matmul.allow_tf32 = True, torch.backends.cudnn.allow_tf32 = True, allowTF32CuDNN() and allowTF32CuBLAS() will be deprecated after Pytorch 2.9. Please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:80.)
  _C._set_float32_matmul_precision(precision)
WARNING:openfold3.entry_points.experiment_runner:No version_tensor is found for this checkpoint.Assuming the user knows checkpoints are parameters are compatible, continuing...
💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
HPU available: False, using: 0 HPUs
WARNING:openfold3.core.data.tools.colabfold_msa_server:Using output directory: /tmp/of3-of-christian/colabfold_msas for ColabFold MSAs.
WARNING:openfold3.core.data.tools.colabfold_msa_server:Mapping file /tmp/of3-of-christian/colabfold_msas/mappings/seq_to_rep_id.json already exists. Appending new sequences.
WARNING:openfold3.core.data.tools.colabfold_msa_server:Mapping file /tmp/of3-of-christian/colabfold_msas/mappings/rep_id_to_seq.json already exists. Appending new sequences.
Submitting 1 sequences to the Colabfold MSA server for main MSAs...
COMPLETE: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 150/150 [elapsed: 00:05 remaining: 00:00]
/home/christian/miniforge3/envs/of3_env/lib/python3.13/site-packages/openfold3/core/data/tools/colabfold_msa_server.py:393: DeprecationWarning: Python 3.14 will, by default, filter extracted tar archives and reject files or modify their metadata. Use the filter argument to control this behavior.
  tar_gz.extractall(path)
No complexes found for paired MSA generation. Skipping...
/home/christian/miniforge3/envs/of3_env/lib/python3.13/site-packages/openfold3/core/data/tools/colabfold_msa_server.py:1229: UserWarning: Query hck_human_src_1y57_templ chain ['A'] already has template_cif_paths set. These are not overwritten with path(s) to the template CIF files from the ColabFold MSA server.
  inference_query_set = add_msa_paths_to_iqs(
/home/christian/miniforge3/envs/of3_env/lib/python3.13/multiprocessing/popen_fork.py:67: DeprecationWarning: This process (pid=1542917) is multi-threaded, use of fork() may lead to deadlocks in the child.
  self.pid = os.fork()
Preprocessing templates: 100%|███████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 217.78it/s]
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
Predicting DataLoader 0:   0%|                                                                                                        | 0/1 [00:00<?, ?it/s]Seed set to 42
W0706 12:20:21.710000 1542917 site-packages/torch/fx/_symbolic_trace.py:52] is_fx_tracing will return true for both fx.symbolic_trace and torch.export. Please use is_fx_tracing_symbolic_tracing() for specifically fx.symbolic_trace or torch.compiler.is_compiling() for specifically torch.export/compile.
Predicting DataLoader 0: 100%|████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:34<00:00,  0.03it/s]
==================================================
    PREDICTION SUMMARY (COMPLETE)    
==================================================
Total Queries Processed: 1
  - Successful Queries:  1
  - Failed Queries:      0
==================================================

Predicting DataLoader 0: 100%|████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:34<00:00,  0.03it/s]
Removing empty log directory...
Cleaning up MSA directories...

Contents of hck_human_src_1y57_templ.json:

{
  "queries":{
    "hck_human_src_1y57_templ": {
      "chains": [
        {
          "molecule_type": "protein",
          "chain_ids": ["A"],
          "sequence": "DIIVVALYDYEAIHHEDLSFQKGDQMVVLEESGEWWKARSLATRKEGYIPSNYVARVDSLETEEWFFKGISRKDAERQLLAPGNMLGSFMIRDSETTKGSYSLSVRDYDPRQGDTVKHYKIRTLDNGGFYISPRSTFSTLQELVDHYKKGNDGLCQKLSVPCMSSKPQKPWEKDAWEIPRESLKLEKKLGAGQFGEVWMATYNKHTKVAVKTMKPGSMSVEAFLAEANVMKTLQHDKLVKLHAVVTKEPIYIITEFMAKGSLLDFLKSDEGSKQPLPKLIDFSAQIAEGMAFIEQRNYIHRDLRAANILVSASLVCKIADFGLARVIEDNEYTAREGAKFPIKWTAPEAINFGSFTIKSDVWSFGILLMEIVTYGRIPYPGMSNPEVIRALERGYRMPRPENCPEELYNIMMRCWKNRPEERPTFEYIQSVLDDFYTATESQYQQQP",
          "template_cif_paths": [
            "1Y57.cif"
          ]  
        },
        {
          "molecule_type": "ligand",
          "chain_ids": "B",
          "ccd_codes": "ADP"
        },
        {
          "molecule_type": "ligand",
          "chain_ids": "C",
          "ccd_codes": "MG"
        }
      ]
    }
  }
}

Contents of runner.yaml:

model_update:
  presets:
    - predict
    - low_mem  # to use low memory settings
  custom:
    settings:
      memory:
        eval:
          use_cueq_triangle_kernels: true
          use_deepspeed_evo_attention: false
msa_computation_settings:
     cleanup_msa_dir: False
     save_mappings: True
output_writer_settings:
  structure_format: pdb

The CIF was downloaded from https://files.rcsb.org/download/1Y57.cif.

[jandom]

Thanks, very helpful – we're looking into this, I'm starting to think that this is real

[jnwei]

Hi @claggner , I'd encourage you to pull from the openfold3 repo and try your queries with cif templates again.

@rostro36 recently submitted a fix for this issue in #306 . You can see some additional discussions in related issues #234 and #235 where others were having issues with the custom cif templates and after the bug fix, see closer to expected performance.

We will aim to make an updated release with the bug fix soon.

[jandom]

@claggner i recently put in some scientific tests for this case.

Can you provide some structure / PDB code to which this will co-fold, in the absence of a template input? Like you've provided a positive (fold with this sequence + template, you get sth close to the template), I'm looking for a negative such that the test is well-calibrated.

[claggner]

Hi @claggner , I'd encourage you to pull from the openfold3 repo and try your queries with cif templates again.

Thank you @jnwei . Are there specific step I need to take to pull in those latest fixes? Installing from pip or pixi as described in the manual doesn't seem to change the behavior.

[claggner]

@claggner i recently put in some scientific tests for this case.

Can you provide some structure / PDB code to which this will co-fold, in the absence of a template input? Like you've provided a positive (fold with this sequence + template, you get sth close to the template), I'm looking for a negative such that the test is well-calibrated.

Hi @jandom. There are actually experimental structures for the example protein P08631, and the predictions seem to be driven by those. E.g. 9BYJ, 5H0B, 1AD5, etc.

[jnwei]

Thank you @jnwei . Are there specific step I need to take to pull in those latest fixes? Installing from pip or pixi as described in the manual doesn't seem to change the behavior.

You will need an up-to-date clone of the repo if you don't already have one, e.g.

# if no clone exists
$ git clone git@github.com:aqlaboratory/openfold-3.git
$ cd openfold-3

# sync the latest changes
$ git pull main

# Run up with pixi as per manual instructions
$ pixi run -e openfold3-cpu run_openfold predict ...

[claggner]

Thank you @jnwei, looks like I was already on the latest version. My problem persist.