[BUG] Conformer generation struggles with organometallics like HEM

[ljarosch]

Describe the bug
Inference queries that use organometallics such as HEM groups will run into conformer timeout problems. This is because RDKit by default cannot seem to properly handle conformer generation for such compounds.

To Reproduce

from openfold3.core.data.primitives.structure.query import structure_with_ref_mol_from_ccd_code

structure_with_ref_mol_from_ccd_code("HEM", chain_id="A")

> ConformerGenerationError: Failed to generate 3D coordinates

Expected behavior
A conformer should be generated without failure.

Additional context
Our training pipeline makes use of input sanitization by pdbeccdutils, which solves these conformer generation issues. Because of that, this issue is inference-specific.

[tomgoddard]

I also have been hitting this bug, also with heme ligands. OpenFold 3 (Feb 13, 2026 github source) predictions with both CCD code and SMILES string for heme fail with RDKit unable to produce 3D coordinates. I'll attach a zip file with OpenFold 3 json input that reproduces the problem and the traceback output. AlphaFold 3 and Boltz 2 both use RDKit and handle heme fine. I compared the OpenFold 3 use of RDKit to Boltz 2 and it looked very similar. Hopefully it will be relatively easy to see how OpenFold 3 is using RDKit differently.

heme.zip

[jnwei]

Thank you for the report @tomgoddard We are working on a fix now that would incorporate the pdbeccdutils parsing into our inference pipeline.

As a temporary workaround, providing this SMILES string with dative bonds will allow the inference pipeline to generate a conformer: C=CC1=C(C)/C2=C/c3c(C)c(CCC(=O)O)c4[n]3[Fe@SP2]35<-[N]2=C1/C=c1/c(C)c(C=C)/c([n]13)=C/C1=[N]->5/C(=C\\4)C(CCC(=O)O)=C1C

[tomgoddard]

Thanks @jnwei. My comment 2 days ago was a bit misleading. AlphaFold 3 and Boltz 2 handle the Chemical Component Dictionary HEM and they use reference atom coordinates from the CCD. If you use the OpenEye canonical SMILES for heme Cc1c2n3c(c1CCC(=O)O)C=C4C(=C(C5=[N]4[Fe]36[N]7=C(C=C8N6C(=C5)C(=C8C)C=C)C(=C(C7=C2)C)C=C)C)CCC(=O)O Boltz 2 fails with the same RDKit can't compute conformer error and the prediction fails, and AlphaFold 3 also uses RDkit and fails to get a conformer, but then AlphaFold 3 uses 0 coordinates and produces a very reasonable heme structure. Here's the comment from the AF3 alphafold3/model/features.py code, line 1618

  # If no mol could be generated (can only happen when using CCD), or no
  # conformer could be generated from the mol (can happen in either case), then
  # use CCD cif instead (which will have zero coordinates for SMILES case).

But OpenFold 3 preview fails with the CCD code HEM which AF3 and Boltz 2 both handle. It may be challenging to handle the SMILES for heme. I think it is more important to make all CCD codes work. That should be easier since I think CCD cif files have reference coordinates.

Thanks for your work-around SMILES string. I have put OpenFold 3 preview in ChimeraX and my comment on this ticket was to make OpenFold more robust for the thousands of ChimeraX users.

https://www.rbvi.ucsf.edu/chimerax/data/openfold-feb2026/openfold.html

[jnwei]

Thanks for the information @tomgoddard and for bringing OpenFold to ChimeraX!

Regarding the failing CCD code for HEM, we have identified the issue as resulting from an outdated CCD file. If you rerun setup_openfold to trigger the download of an updated CCD file, or run the following to update the CCD file:

python -c "from openfold3.setup_openfold import setup_biotite_ccd; import biotite.setup_ccd; setup_biotite_ccd(ccd_path=biotite.setup_ccd.OUTPUT_CCD, force_download=False)"

This will update the local CCD file stored in biotite. With this updated CCD file, the CCD for HEM should pass.

[tomgoddard]

Updating the CCD file components.bcif (new size 63393643, previous size 59077222) made no difference for me. Same error "ConformerGenerationError: Failed to generate 3D coordinates".

[ljarosch]

Hi @tomgoddard , thanks for your report and the reproducible examples!

We have a fix in progress for this that we're hoping to add this week. There seems to be an issue with how Biotite reads some of these organometallics into RDKit objects - switching back to pdbeccdutils (which we use in training) will solve that.

[tomgoddard]

Thanks! I don't think it is a biotite problem, RDKit itself can't produce 3d coordinates the SMILES string for heme. In AlphaFold 3 the CCD reference coordinates are used during inference. AlphaFold 3 does not take the CCD SMILES string and try to get RDKit or pdbeccdutils to produce 3d coordinates. The Alphafold 3 behavior seems sensible since the CCD reference coordinates are curated by the PDB and very likely to be correct, while coordinates generated from a SMILES string can be erratic. My sense is you still prefer not to use the CCD reference coordinates because in OpenFOld training you used pdbeccdutils. It makes some sense to have inference use the same method of getting CCD coordinates as was used in training. On the other hand it should be done in the best possible way, so it might make more sense to change code now to use CCD reference coordinates as done in AlphaFold 3 so the next round of OpenFold training will use these better coordinates.