Molecules

Molecule and ligand modeling covers small-molecule identity, representation, standardization, similarity, and property prediction. The main risk is treating strings or graphs as if they were already chemically normalized objects.

This page owns the chemical object layer. Use Molecular and Ligand Modeling when deciding the modeling route, Interaction Modeling when target or assay context is central, and Structure-Based Modeling when pose, pocket, or protein-ligand geometry is central.

$$r_L=\varphi (L,c)$$

where $L$ is a molecular object and $c$ can include pH, assay condition, conformer source, or target context.

Route Map

Question	Start	Watch
What is the modeled chemical object?	Molecule, Ligand, Molecular identity	salts, mixtures, stereo, protonation, tautomer policy
How is the molecule normalized?	Molecular standardization, Chemical state contract	deduplication before standardization
What does the model actually see?	Molecular featurization contract, RDKit, SMILES, Molecular graph	string augmentation or graph construction changing the object
Is similarity meaningful for the claim?	Molecular fingerprint, Molecular similarity	using a cheap proxy as if it were chemical equivalence
Is generation constrained by chemistry?	Fragment-SELFIES, Molecular generation	validity without utility or novelty boundary

Representation Choices

Representation	Use For	Main Risk
SMILES / string	sequence modeling, language-model style pretraining, generation	equivalent molecules can have many strings unless canonicalization or augmentation is explicit
Molecular graph	property prediction, message passing, atom/bond reasoning	stereo, aromaticity, charge, and hydrogen handling can silently change the object
Fingerprint	similarity search, scaffold analysis, cheap baselines	fixed bits hide which chemistry caused a match
Conformer / 3D coordinates	docking, geometry, shape, interaction modeling	conformer source and protonation state can dominate downstream results
Fragment representation	scaffold editing, constrained generation, substructure search	fragment vocabulary can bias generation and evaluation

Standardization Contract

Before deduplication, splitting, featurization, or model comparison, specify:

$$L_{\mathrm{raw}}\stackrel{\mathrm{standardize}}{\to }{L}_{\mathrm{std}}\stackrel{\mathrm{featurize}}{\to }{r}_L$$

where $L_{\mathrm{raw}}$ is the input molecule record, $L_{\mathrm{std}}$ is the standardized chemical object, and $r_L$ is the model representation.

RDKit Protocol Layer

RDKit is often the code layer behind this contract. The important point is not “use RDKit” but “record the exact molecular protocol.”

$$r_L=F_{\psi }(\mathrm{RDKitProtocol}(L_{\mathrm{raw}}))$$

where $\psi$ includes featurizer settings such as fingerprint radius, bit length, chirality flag, descriptor list, conformer seed, and failure policy.

RDKit Output	Typical Use	Must Record
canonical SMILES	deduplication, exact identity, sequence model input	standardization and stereo policy
Morgan fingerprint	baseline, similarity, scaffold analysis	radius, bit length, count/binary, chirality
descriptor vector	small-data baseline, interpretable model	descriptor list, NaN policy, scaling
molecular graph	GNN input	atom/bond features, hydrogens, aromaticity, charge, stereo
conformer	3D model, docking, shape	protonation, seed, force field, minimization, failure rate

Structure and Chemistry Routes

Area	Start	Use For
Conformer state	Conformer, Stereochemistry, Protonation state, Tautomer	making the chemical state explicit before 3D or assay claims
Substructure and fragments	Substructure search, Fragment-SELFIES	scaffold editing, filtering, constrained generation
Physics-inspired preparation	Force field, Energy minimization, Molecular dynamics	coordinate refinement, conformer protocols, simulation-derived features
Prediction and retrieval	Property prediction, Molecular property prediction, Similarity search	supervised endpoints and ranked candidate lists

Task Map

Task	Output	Evaluation
Property prediction	scalar, class, or calibrated probability	regression/classification metric plus scaffold or assay-aware split
Similarity search	ranked molecule list	retrieval metric and duplicate policy
Molecular generation	valid molecule samples	validity, uniqueness, novelty, diversity, constraint satisfaction
Docking / SBDD	pose, score, or candidate ranking	pose quality, enrichment, affinity, and leakage checks

Checks

• Has the molecule been standardized before deduplication and split construction?
• Are stereochemistry, protonation, tautomer, salts, and conformer source handled explicitly?
• Are force-field, minimization, or simulation protocols recorded when coordinates are modified?
• Is the split scaffold-based, temporal, assay-based, or target-aware?
• Is similarity measured on the right representation for the claim?

Related

• Computational Biology
• Interaction modeling
• Structure-based modeling
• RDKit
• Scaffold split