Stereochemistry
Stereochemistry describes the 3D arrangement of atoms that can make molecules with the same graph behave differently. It matters for binding, activity, conformers, and molecular generation.
A modeling pipeline should treat stereochemistry as part of the molecular identity:
Checks
- Are chiral centers and double-bond stereochemistry preserved?
- Are unspecified stereocenters treated as unknown rather than silently flattened?
- Does featurization distinguish stereoisomers when the task requires it?
- Are generated molecules checked for valid stereochemistry?
- Is stereochemistry standardized before deduplication and split assignment?
- Does a stereoisomer pair produce distinct features when labels can differ?
Feature Contract
Stereochemistry can enter a model through several layers:
- String: isomeric SMILES or another stereo-aware string representation.
- Graph: atom chiral tags, bond direction, wedge/hash information, and stereo groups.
- 3D: conformer coordinates that respect the assigned stereochemical state.
- Fingerprint: feature generator settings that include or ignore chirality.
The featurization contract should state whether stereochemistry is:
These are different data assumptions.
Common Pitfall
If stereochemistry is dropped, two molecules with the same topology can collapse to one representation:
This may be acceptable only if the task explicitly ignores stereochemistry. For binding, docking, conformers, and many activity labels, this collapse can destroy the signal.
Failure Modes
- Enantiomers with different activity are merged during deduplication.
- A train/test split separates stereoisomers but the featurizer makes them identical.
- A generative model outputs unspecified stereochemistry and is scored as if the molecule were fully specified.
- Reflection-invariant geometric features erase chirality needed for binding.
- A benchmark reports topology validity but not stereochemical validity.