Molecule

A molecule is a collection of atoms and bonds represented as a graph, 3D conformation, or set of measured properties.

A molecule can be represented by a string, graph, fingerprint, descriptor vector, conformer ensemble, or learned embedding:

$$r_m=\varphi (m,q)$$

where $m$ is the standardized molecular identity and $q$ is the representation protocol: SMILES tokenization, graph featurization, conformer generation, fingerprint parameters, or descriptor calculation.

Why It Matters

• It is the base object for ligands, fragments, and molecular datasets.
• Molecular models must separate topology, coordinates, stereochemistry, and assay context.
• Geometry-aware methods often treat molecules as graphs with coordinate-dependent features.
• The same raw record can map to different model inputs depending on standardization, tautomer, protonation, and stereochemistry policy.

Standardization Boundary

For public ML notes, define which molecular object is being modeled:

$$m_{\mathrm{std}}=\mathrm{standardize}(m_{\mathrm{raw}})$$

where standardization may include salt/counterion removal, charge handling, tautomer selection, stereochemistry policy, aromaticity perception, and canonicalization.

Deduplication, splitting, featurization cache keys, and label aggregation should reference $m_{\mathrm{std}}$, not the raw source string:

$$\mathrm{key}=H(m_{\mathrm{std}},\text{standardization protocol},\text{featurizer version})$$

Representation Choices

Different molecular representations preserve different information:

Representation	Keeps	Common risk
SMILES	token order and chemistry syntax	augmentation/tokenization ambiguity
Graph	atom-bond topology	lost stereochemistry or charge details
Fingerprint	substructure patterns	collision and radius dependence
Descriptor vector	curated physicochemical features	version drift and missing values
Conformer ensemble	3D geometry	conformer generation protocol bias

The note should state whether the model consumes topology, coordinates, descriptors, or a learned embedding.

Checks

• Which representation is used: SMILES, graph, fingerprint, conformer, or coordinates?
• Are stereochemistry, charges, and aromaticity normalized consistently?
• Is the molecule standardized before deduplication and splitting?
• Does the task require a single molecule, molecule pair, or molecule-protein context?
• Are generated molecules checked for validity, novelty, diversity, and usefulness?
• Does evaluation use scaffold or similarity groups when claiming ligand-side generalization?
• Is the featurizer version part of the dataset or run artifact?
• Are invalid molecules reported rather than silently dropped?
• Are tautomers, protonation states, and stereoisomers collapsed only when that matches the task?

Related

• Ligand
• Molecular modeling concepts
• Molecular standardization
• Molecular featurization contract
• SMILES
• Molecular graph
• Molecular fingerprint
• Conformer
• Stereochemistry
• Data versioning
• Preprocessing contract
• Assay
• Dataset
• Equivariant GNN