Reproducibility of Morgan fingerprints: The last RDKit version you can use

[flange-ipb]

Hello,

I tried to bump the version of RDKit in this project and ran into reproducibility issues for the classification outcomes. This is due to some well-known changes in RDKit's generation of Morgan fingerprints for molecular graphs with added hydrogens (see rdkit-discuss). Hydrogens are added in NPClassifier's calculate_fingerprint function.
The classification models were trained under these circumstances, thus the last version of RDKit you can use is rdkit-pypi==2021.9.4. Everything beyond that may give irreproducible classifications!

My test code:

from rdkit import Chem
from rdkit.Chem import rdMolDescriptors

smiles = "C[C@]1(C=C[C@H]2C[C@H](CC[C@@H]2[C@H]1C(=O)CCO)CO)O"

mol1 = Chem.MolFromSmiles(smiles)
mol_fp1 = rdMolDescriptors.GetMorganFingerprintAsBitVect(mol1, radius=1, bitInfo={}, nBits=2048)
# always passes
assert list(mol_fp1.GetOnBits()) == [29, 80, 142, 222, 473, 494, 622, 650, 787, 807, 848, 926, 1019, 1057, 1060,
                                     1083, 1154, 1274, 1292, 1325, 1516, 1564, 1764, 1873, 1917]

mol2 = Chem.MolFromSmiles(smiles)
mol2 = Chem.AddHs(mol2)
mol_fp2 = rdMolDescriptors.GetMorganFingerprintAsBitVect(mol2, radius=1, bitInfo={}, nBits=2048)

# passes with rdkit-pypi==2021.9.4
# fails with rdkit-pypi==2021.9.5.1 and beyond
assert list(mol_fp2.GetOnBits()) == [2, 88, 107, 114, 449, 650, 664, 695, 788, 807, 836, 866, 906, 955, 1060, 1233,
                                     1380, 1455, 1477, 1652, 1673, 1804, 1871, 1886, 1917, 2003]

NB: I use pip-based dependency management, which is why I refer to the RDKit artifacts at PyPI. If you use Conda to install RDKit like in the Docker image you'll probably run into incompatibilities with the Boost library at runtime. A working RDKit version from conda-forge with reproducible classifications is rdkit=2021.09.5.

[mwang87]

ooh thats very interesting and thanks for pushing through this. We do notice some run time issues depending up on the API from rdkit. I'm going to keep this issue open just for my own FYI. Do you think we should upgrade, its been working fine-enough for us using the 2019 versions of rdkit via conda.