Sulfoxide (tetrahedral S) stereochemistry is not conditioned — generated configuration is independent of the input

[isayev]

Summary

LoQI reproduces carbon and phosphorus stereocenters essentially perfectly, but the configuration it generates at a sulfoxide (tetrahedral sulfur) stereocenter is independent of the input SMILES chirality — it samples a fixed, ~63/37 R-biased prior regardless of whether R or S is requested.

Surfaced while building a docking ligand-prep pipeline. Many drugs are chiral sulfoxides (omeprazole/esomeprazole, modafinil/armodafinil, sulindac, oxfendazole, …), where the wrong configuration is the wrong molecule.

Evidence — enantiomer-conditioning experiment

Each stereoisomer is fed separately (30 conformers, --seed 0); the CIP label is read back from each generated 3D structure. The ETKDG column is an RDKit embedding of the same SMILES — it confirms the target is achievable and that the 3D→CIP assignment is reliable.

input	SMILES	intended	LoQI output (n=30)	ETKDG
sulfoxide, R	C[S@@](=O)c1ccccc1	R	R:18, S:12	10/10 R
sulfoxide, S	C[S@](=O)c1ccccc1	S	R:20, S:10	10/10 S
phosphine oxide, R	CC[P@](=O)(C)c1ccccc1	R	R:30	10/10 R
phosphine oxide, S	CC[P@@](=O)(C)c1ccccc1	S	S:30	10/10 S
carbon control, R	C[C@H](O)c1ccccc1	R	R:30	10/10 R

Both sulfoxide inputs collapse to the same ~63/37 R-biased output → the S-center configuration is not conditioned on the input. Phosphorus — also a tetrahedral heteroatom stereocenter — and carbon are conditioned perfectly.

At the population level, across ~730 achievable, fully-defined chiral conformers from a 1000-molecule ChEMBL run, carbon centers reproduce at ~100% while the residual stereo errors are dominated by sulfoxide centers.

Reproduction

Standalone; needs only the shim + RDKit. Run from the repo root.

import json, subprocess, sys, tempfile
from collections import Counter
from pathlib import Path
from rdkit import Chem, RDLogger
RDLogger.DisableLog("rdApp.*")

SHIM = "scripts/loqi_inference_shim.py"
CASES = [("sulfoxide R", "C[S@@](=O)c1ccccc1"), ("sulfoxide S", "C[S@](=O)c1ccccc1"),
         ("P-oxide R", "CC[P@](=O)(C)c1ccccc1"), ("P-oxide S", "CC[P@@](=O)(C)c1ccccc1")]
def cip(s):
    m = Chem.MolFromSmiles(s)
    return dict(Chem.FindMolChiralCenters(m, includeUnassigned=True, useLegacyImplementation=False))
payload, meta = [], {}
for label, smi in CASES:
    ik = Chem.MolToInchiKey(Chem.MolFromSmiles(smi))
    payload.append({"inchikey": ik, "smiles": smi}); meta[ik] = (label, cip(smi))
with tempfile.TemporaryDirectory() as t:
    t = Path(t); (t/"in.json").write_text(json.dumps(payload)); o = t/"out"; o.mkdir()
    subprocess.run([sys.executable, SHIM, "--input", str(t/"in.json"), "--output_dir", str(o),
                    "--n_confs", "30", "--seed", "0"], check=True)
    for ik, (label, intended) in meta.items():
        want = ",".join(map(str, intended.values())); obs = Counter()
        for m in Chem.SDMolSupplier(str(o/f"{ik}.conformers.sdf"), removeHs=False):
            if m is None: continue
            mm = Chem.Mol(m); Chem.AssignStereochemistryFrom3D(mm)
            obs[",".join(dict(Chem.FindMolChiralCenters(mm, includeUnassigned=True,
                useLegacyImplementation=False)).get(i, "?") for i in intended)] += 1
        print(f"{label:14s} intended {want:3s} -> {dict(obs)}")

Ruled out

• Not "chirality ignored globally": carbon and phosphorus centers condition at ~100% (input R→R, S→S).
• Not a perception artifact: the sulfoxide CIP assigns cleanly as R/S from the 3D (never ?), and ETKDG of the same SMILES returns the intended label 10/10.
• Not an impossible/degenerate geometry: ETKDG embeds both sulfoxide enantiomers cleanly; the generated S-center is a well-formed pyramid.

Likely cause

The graph featurization appears to carry chiral tags for tetrahedral carbon and phosphorus but not for the 3-coordinate sulfoxide sulfur (RDKit represents it as S(=O) / [S+][O-], a special stereo case). Worth checking whether atom.GetChiralTag() on the sulfoxide S survives graph construction / one-hot encoding.

Environment

LoQI main @ 3f1c3fe, data/loqi.ckpt, 25-step sampler, --seed 0; PyTorch 2.9.1, Lightning 2.6.1, NVIDIA L40S. Happy to share the full failing-set SMILES + per-center analysis from the ChEMBL run.