RDKIT Pickling Roundtrip Error (but only on MacOS?)

[apayne97]

Thanks to @ijpulidos and @chrisiacovella for finding and troubleshooting this.

2026.02.20_test_failure.ipynb

It manifests as the loss of the names in a LigandNetwork in the edges (but not the nodes?) such that edge.componentA.name and edge.componentB.name are ''.

This warning appears: "UserWarning: RDKit does not preserve Mol properties when pickled by default, which may drop e.g. atom charges; consider setting Chem.SetDefaultPickleProperties(Chem.PropertyPickleOptions.AllProps)

When we set this property in a few different places in our code, it fixes the bug but only for ubuntu/linux. On MacOS, the warning and error still appear.

The drugforge error is here:
https://github.com/choderalab/drugforge/actions/runs/22386574178/job/64798431670?pr=132

And you can see it passes on Ubuntu here:
https://github.com/choderalab/drugforge/actions/runs/22386574178/job/64798431676?pr=132#step:8:138

On the other hand, if you run this code on MacOS, it works just fine(below)

from rdkit import Chem
from rdkit.Chem import Descriptors
import pickle
Chem.SetDefaultPickleProperties(Chem.PropertyPickleOptions.AllProps)

smiles = "CN1C=NC2=C1C(=O)N(C(=O)N2C)C"
mol = Chem.MolFromSmiles(smiles)

mol.SetProp("_Name", "Caffeine")
mol.SetProp("Source", "Coffee Beans")
mol.SetProp("MolecularWeight", f"{Descriptors.MolWt(mol):.2f}")

# Pickle the molecule
pickled_mol = pickle.dumps(mol)

print(f"Molecule '{mol.GetProp('_Name')}' has been pickled.")
print(f"Pickle size: {len(pickled_mol)} bytes")

# Unpickle to demonstrate data persistence
new_mol = pickle.loads(pickled_mol)
print(f"\nRecovered Name: {new_mol.GetProp('_Name')}")
print(f"Recovered MW: {new_mol.GetProp('MolecularWeight')}")

This suggests there's a way that Chem.SetDefaultPickleProperties(Chem.PropertyPickleOptions.AllProps) is not persisting through the code but only for MacOS.

It's not clear to us exactly where the fix needs to be made (drugforge, openfe, or gufe). I've added a jupyter notebook that shows the error in our test where it fails on MacOS + a yaml file of the environment.
2026.02.20_test_failure.ipynb

20260225_macos.yaml

Proposed solution

It seems like this property needs to be set within gufe - I wonder if changing this to set the property instead of just throwing a warning would be better?

gufe/gufe/components/explicitmoleculecomponent.py

Line 70 in c531ffe

def __getstate__(self):

[mikemhenry]

So just to be clear, setting Chem.SetDefaultPickleProperties(Chem.PropertyPickleOptions.AllProps) while working with gufe stuff works as expected on linux, but doesn't work on MacOS. But in the pure RDKit example you posted, setting Chem.SetDefaultPickleProperties(Chem.PropertyPickleOptions.AllProps) does work on MacOS.

[mikemhenry]

I haven't tested to see if setting it somewhere in gufe works to fix the issue BUT looking at this issue rdkit/rdkit#2470 (assuming setting it in gufe fixes the problem) I am inclined for it just to get set in gufe since the downsides don't seem to really apply.

[chrisiacovella]

Yeah the issue is quite odd. We now set Chem.SetDefaultPickleProperties(Chem.PropertyPickleOptions.AllProps) inside of drugforge, which corrects the issue on ubuntu, but the issue persists on MacOS.

If I check the status of `Chem.GetDefaultPickleProperties() during the test on MacOS, it returns a value of 65535 indicating all properties should be exported, but I also get the warning from GUFE, indicating that the default parameters were not set to include all properties.

test_fec_schema.py /Users/cri/miniconda3/envs/alchemy/lib/python3.12/site-packages/gufe/components/explicitmoleculecomponent.py:71: UserWarning: RDKit does not preserve Mol properties when pickled by default, which may drop e.g. atom charges; consider setting Chem.SetDefaultPickleProperties(Chem.PropertyPickleOptions.AllProps)``

So there must be some silo'ing going regarding these defaults in macOS, so the suggesting of simply setting allProps to default in GUFE should correct this issue on Mac OS.

[apayne97]

I've added a jupyter notebook + environment that doesn't include drugforge as a dependency that shows this problem on MacOS.

20260303_test_env_MacOS_from_history.yaml
20260303_test_env_MacOS.yaml
test_rdkit_roundtrip_set_chem_props.ipynb

[atravitz]

Thank you @apayne97 for the minimal reproduction! I'm able to replicate this bug locally and over on openfe: OpenFreeEnergy/openfe#1861.

However, the root(ish) cause is actually in konnektor. In konnektor v0.2, the ExplicitNetworkGenerator calls multiprocessing even when n_processes=1 (which is what openfe.ligand_network_planning.generate_network_from_names() uses).

Fortunately, OpenFreeEnergy/konnektor#210 from a couple weeks ago fixes this, so that n_process=1 doesn't use multiprocessing.

I've just released konnektor v0.3.0 and it should be up on conda-forge soon. Setting konnektor>=0.3.0 should fix the issue you're seeing - please report back to let us know!

On konnektor/gufe side we will dig into this more, and I've opened a separate issue to track that work: OpenFreeEnergy/konnektor#221

[apayne97]

The plot thickens! Nice sleuthing!

[ijpulidos]

Great job everyone! Thanks for digging into this @atravitz and @apayne97. This basically means that we should now pin konnektor to 0.3.x. Thanks!

[chrisiacovella]

I'm guessing we might need to wait until a new release of openfe-base.

Could not solve for environment specs
The following packages are incompatible
├─ konnektor =0.3.0 * is requested and can be installed;
└─ openfe >=1.8.0 * is not installable because it requires
   └─ openfe-base [==1.8.0 pyhd8ed1ab_0|==1.8.0 pyhd8ed1ab_1|...|==1.9.1 pyhd8ed1ab_2], which requires
      └─ konnektor >=0.2.0,<0.3.0 *, which conflicts with any installable versions previously reported.

[atravitz]

@chrisiacovella thanks for reporting back - I just updated openfe 1.9.1 so you can install with konnektor 0.3.0 now.