Substructure Search

Substructure search asks whether a molecule contains a query pattern. It is a graph matching problem over atoms, bonds, aromaticity, charge, stereochemistry, and query constraints.

At a high level, the question is:

$$Q\subseteq G_m$$

where $Q$ is a query graph and $G_m$ is the molecular graph.

More explicitly, search asks whether there is an injective mapping $\varphi$ from query atoms to molecule atoms such that atom and bond constraints are preserved:

$$(i,j)\in E_Q\Rightarrow (\varphi (i),\varphi (j))\in E_m$$

Atom constraints must also be preserved:

$$c_Q(i)⪯c_m(\varphi (i))$$

where $c_Q(i)$ is the query constraint for query atom $i$ and $c_m$ is the molecule atom state. The relation $⪯$ depends on the query language, such as SMARTS.

Key Ideas

• Substructure search is exact pattern matching under a chemical query language.
• It is different from similarity search, which ranks approximate neighbors.
• Query semantics matter: aromaticity, valence, charge, stereochemistry, and atom lists can change results.
• Substructure filters are useful for motifs, alerts, scaffolds, and dataset audits.

Query Semantics

A query is not only a drawing of atoms. It is a set of constraints:

$$Q=(V_Q,E_Q,C_V,C_E)$$

where $C_V$ and $C_E$ constrain atom and bond properties.

Constraint	Example issue
atom type	element, wildcard, atom list, isotope
valence/degree	aromatic and charged atoms may differ after sanitization
aromaticity	toolkit aromaticity model changes matches
charge	protonation policy can create or remove hits
stereochemistry	unspecified stereo should not be treated as confirmed stereo
recursive SMARTS	query may encode hidden expert assumptions

For public notes, store the intention of the query in prose, not only the SMARTS string.

Search Boundary

Operation	Output	Do Not Confuse With
Substructure search	boolean match or atom mappings	molecular similarity
SMARTS alert	query-defined structural pattern	toxicity mechanism proof
Scaffold filter	retained core or matched series	scaffold split unless policy is explicit
Fragment match	occurrence of a fragment	functional equivalence
Similarity search	ranked approximate neighbors	exact subgraph containment

Dataset Use

Substructure search can be a data operation, a scientific hypothesis, or a leakage source.

Use	Good for	Caveat
alert filtering	removing known reactive or undesirable motifs	alert hit is not proof of toxicity
scaffold grouping	series analysis or split construction	query policy must be fixed before evaluation
motif enrichment	explaining model errors or clusters	post-hoc query mining can overfit test labels
candidate retrieval	finding molecules with a required group	exact match may miss bioisosteres
data audit	locating duplicated cores or shortcuts	audit should not become hidden training signal

Practical Checks

• Is the query a strict substructure, scaffold, pharmacophore-like pattern, or loose SMARTS pattern?
• Were molecules standardized before searching?
• Are tautomers, protonation states, salts, and stereochemistry handled consistently?
• Are matches used for filtering, labeling, splitting, or interpretation?
• Could the query encode a known leakage shortcut?
• Are atom mappings, match counts, or only boolean hits used downstream?
• Was the query authored before looking at test labels or benchmark failures?
• Is the toolkit aromaticity and SMARTS behavior documented?
• Are query hits treated as hypotheses rather than biological proof?

Related

• Molecular graph
• SMILES
• Molecular similarity
• Molecular standardization
• Stereochemistry
• Leakage
• Scaffold split