RDKit

RDKit is a cheminformatics toolkit commonly used to parse molecules, standardize chemical records, compute descriptors and fingerprints, build molecular graphs, search substructures, and generate conformers. In this wiki, RDKit belongs under reusable molecular modeling concepts, not under paper notes.

The main role of RDKit in ML workflows is to define a deterministic transformation:

$$m_{\mathrm{raw}}\stackrel{\mathrm{RDKit}\mathrm{protocol}}{\to }(\tilde{m},z,c)$$

where $m_{\mathrm{raw}}$ is a raw record such as SMILES or SDF, $\tilde{m}$ is a standardized molecule, $z$ is a feature such as a graph, fingerprint, descriptor vector, or conformer, and $c$ is metadata about the protocol.

Common Uses

Use	Output	Main Risk
parsing and sanitization	molecule object	silent drop or repair policy changes dataset
standardization	canonical identity	over-collapsing salts, tautomers, stereo, or charge
fingerprinting	bit/count vector	radius, bit length, chirality, and collisions change similarity
descriptors	numeric feature vector	missing/NaN handling and version drift
substructure search	match set	query semantics and aromaticity policy
conformer generation	3D coordinates	seed, protonation, force field, and failure policy dominate

Standardization Boundary

RDKit can clean up and canonicalize molecular records, but the chemical policy is still a modeling decision. A typical pipeline is:

$$m\stackrel{\mathrm{sanitize}}{\to }{m}_1\stackrel{fragment/parent}{\to }{m}_2\stackrel{\mathrm{uncharged}\mathrm{or}\mathrm{chosen}\mathrm{charge}}{\to }{m}_3\stackrel{\mathrm{tautomer}\mathrm{policy}}{\to }\tilde{m}$$

This pipeline should run before deduplication and split assignment. If split keys are built from raw records but features are built from standardized records, equivalent molecules can appear on both sides of train/test.

Featurization Contract

For every RDKit-derived feature, record:

Field	Example
input identity	canonical SMILES or standardized molecule hash
standardization protocol	salt, tautomer, charge, stereo policy
feature family	Morgan fingerprint, descriptor vector, graph, conformer
feature parameters	radius, bit length, count/binary, chirality flag
software version	RDKit version and wrapper code version
failure policy	parse failure, sanitization failure, conformer failure
cache key	hash of standardized molecule plus featurizer version

The cache key should change whenever either molecule identity or featurizer parameters change:

$$k=H(\tilde{m},{\psi }_{\mathrm{rdkit}})$$

ML Checks

• Standardize before deduplication and split construction.
• Compute scaffold or cluster split on the same standardized molecule used by the model.
• Do not silently drop invalid molecules without recording denominator changes.
• State whether chirality and bond stereo are preserved.
• State fingerprint radius, bit length, count/binary mode, and chirality flag.
• State whether conformer generation is deterministic, seeded, minimized, and available at inference.
• Re-run cheap RDKit baselines when the split, standardization, or featurizer changes.

Where It Does Not Belong

RDKit itself is not a paper result. Put reusable protocol notes here. Put a paper under Papers only when the page is about a specific method, benchmark, or empirical claim.

Related

• Molecular standardization
• Molecular featurization contract
• Molecular fingerprint
• Molecular graph
• Conformer
• Molecules
• Data and evaluation