[BUG] planar sp2 C=C bond giving a 3D geometry

[jandom]

Describe the bug

This is a write up for a bug that Deanna Wood (Vizcarra lab) have reported. This is a bug that relates to bad small molecule geometry, and impacts the latest p2 release (155k checkpoint), and probably everything else.

This is in fact a common failure mode for all AF3-like models

Here is the molecule in question

To Reproduce

We tried a couple of smiles for this molecule

• O=C1NC(=S)N(c2cccc(Br)c2)C(=O)/C1=C/c1ccco1 canonical smiles
• BrC1=CC=CC(N(C(/C(C(N2)=O)=C/C3=CC=CO3)=O)C2=S)=C1 smiles
• [H]\\C(C1=CC=CO1)=C1\\C(=O)NC(=S)N(C1=O)C1=CC(Br)=CC=C1 explicitly placed hydrogen

Here are the input JSONs

{
  "queries": {
    "diaph1_smifh2": {
      "chains": [
        {
          "molecule_type": "protein",
          "chain_ids": ["A"],
          "sequence": "GLTPKKLYKPEVQLRRPNWSKLVAEDLSQDCFWTKVKEDRFENNELFAKLTLTFSAQTKTSKAKKDQEGGEEKKSVQKKKVKELKVLDSKTAQNLSIFLGSFRMPYQEIKNVILEVNEAVLTESMIQNLIKQMPEPEQLKMLSELKDEYDDLAESEQFGVVMGTVPRLRPRLNAILFKLQFSEQVENIKPEIVSVTAACEELRKSESFSNLLEITLLVGNYMNAGSRNAGAFGFNISFLCKLRDTKSTDQKMTLLHFLAELCENDYPDVLKFPDELAHVEKASRVSAENLQKNLDQMKKQISDVERDVQNFPAATDEKDKFVEKMTSFVKDAQEQYNKLRMMHSNMETLYKELGEYFLFDPKKLSVEEFFMDLHNFRNMFLQAVKENQKRRETEEKMRRAKLAKEKAEKERLEKQQKREQLIDMNAEGDETGV"
        },
        {
          "molecule_type": "protein",
          "chain_ids": ["B"],
          "sequence": "GLTPKKLYKPEVQLRRPNWSKLVAEDLSQDCFWTKVKEDRFENNELFAKLTLTFSAQTKTSKAKKDQEGGEEKKSVQKKKVKELKVLDSKTAQNLSIFLGSFRMPYQEIKNVILEVNEAVLTESMIQNLIKQMPEPEQLKMLSELKDEYDDLAESEQFGVVMGTVPRLRPRLNAILFKLQFSEQVENIKPEIVSVTAACEELRKSESFSNLLEITLLVGNYMNAGSRNAGAFGFNISFLCKLRDTKSTDQKMTLLHFLAELCENDYPDVLKFPDELAHVEKASRVSAENLQKNLDQMKKQISDVERDVQNFPAATDEKDKFVEKMTSFVKDAQEQYNKLRMMHSNMETLYKELGEYFLFDPKKLSVEEFFMDLHNFRNMFLQAVKENQKRRETEEKMRRAKLAKEKAEKERLEKQQKREQLIDMNAEGDETGV"
        },
        {
          "molecule_type": "ligand",
          "chain_ids": ["C"],
          "smiles": "O=C1NC(=S)N(c2cccc(Br)c2)C(=O)/C1=C/c1ccco1"
        },
        {
          "molecule_type": "ligand",
          "chain_ids": ["D"],
          "smiles": "O=C1NC(=S)N(c2cccc(Br)c2)C(=O)/C1=C/c1ccco1"
        }
      ]
    }
  }
}

A custom runner YAML was provided but the low-mem example YAML yields tho same result

experiment_settings:
  cache_path: /home/jandom/.openfold3

model_update:
  presets:
    - predict
    - low_mem
    - pae_enabled
    
  custom:
    settings:
      memory:
        eval:
          use_deepspeed_evo_attention: false
          use_cueq_triangle_kernels: false

pl_trainer_args:
  accelerator: gpu
  devices: 1

Expected behavior

Atoms in the plane of the C=C bond should be planar.

Actual behavior

diaph1_smifh2.pse.zip

Stack trace
N/A

Configuration (please complete the following information):

• OS: [e.g. Ubuntu 20.04], Ubuntu
• GPU used and memory available: DGX
• Python version (if relevant):
• Installation path (e.g. conda / pip / build from repo) pixi openfold3:openfold3-cuda13-pypi

Additional context

[tomgoddard]

This ligand geometry prediction error is typical of AlphaFold 3, OpenFold 3 preview 1, and Boltz 2 which all produce errors where the neural net has no notion of double versus single bonds. Also chiral centers are often flipped. I ran a test a month ago with 50 anti-viral drugs and half of them have wrong predicted geometry. This suggests that the network has no clear idea where the hydrogens are on the ligands. This probably seriously degrades both the quality of binding poses and any attempt to predict binding affinity.

I am thinking of trying to retrain or fine-tune OpenFold and Boltz using explicit hydrogens on ligands and a loss term that more strictly enforces physically reasonable ligand geometry and chirality. I think it has a decent chance of significantly improving ligand binding predictions. I'm looking for collaborators since it will take a good deal of work. I tried adding explicit ligand hydrogens in OpenFold 3 preview 1 a month ago and surprisingly there were no problems. Of course the geometry it produced (bond lengths, angles) was still completely wrong as the network has no idea what a hydrogen is. Here's a description of quantification of the ligand geometry problems for the 50 test ligands.

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

[jandom]

Adding explicit hydrogens, even on ligands, will be a major lift. Iterating on losses/metrics under the current scheme will be easier job, imo. You're right obviously that including more detail will help.

I tried adding explicit ligand hydrogens in OpenFold 3 preview 1 a month ago and surprisingly there were no problems.

Well, that's an interesting data point. I'd be keen to see where we can go with this, will shoot you a message (added you on LI)

[tomgoddard]

Getting OpenFold 3 preview 1 to make a prediction with ligand hydrogens requires modifying just two lines of code, one to add hydrogens after a conformer is generated from the SMILES string, and another to not delete the hydrogens later. The results were somewhat comical predicting cyclohexane -- the hydrogens are in the prediction but the 6-member carbon ring is still flattened like benzene, bond lengths are weird. Not surprising since OpenFold never saw a hydrogen in training.

Here are some details including the OpenFold code diff to run this test.

https://www.rbvi.ucsf.edu/chimerax/data/ligchem-feb2026/badchem.html#openfoldhydrogens

OpenFold 3 preview 1 prediction of cyclohexane including hydrogens

The predicted geometry is absurd, should look like this

I am hopeful that fine tuning OpenFold 3 to include hydrogens on ligands would not be difficult. It is easy to add hydrogens for physiological pH to the ligands in all the existing training data you released for preview 2. (I would probably do it with the addh command in the ChimeraX visualization program I develop.) I think it would be interesting to try two approaches: hydrogens only on ligands, and hydrogens on all molecules. Adding all hydrogens will double the number of atoms reducing the size of feasible predictions. I think AF3 variants implicitly understand the locations of hydrogens with amino acids so only hydrogens on the ligands is sufficient. Training with amino acids saw millions of instances of each of the 20 standard amino acids, while training with ligands saw very little of ligand chemistry, which I think is the reason it fails to understand bond-orders and physically reasonable ligand geometry.

[tomgoddard]

I updated my web page of wrong ligand hydrogens and chiralties to include OpenFold preview 2. It made fewer chirality flips, 30% of ligands vs 40% for OpenFold preview 1. Still about half the 52 ligands I tested have unphysical geometry in OpenFold 3, AlphaFold 3, and Boltz 2.

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