Canonical Representations of Molecules

• Canonical representations are unique digital encodings of molecular structures that resolve ambiguity via deterministic graph traversal, stereochemical analysis, and symmetry invariance.
• They underpin essential applications such as database deduplication, molecular similarity searches, and serving as inputs for chemical language models and property predictions.
• Recent advances extend these methods to algebraic data types and latent machine learning embeddings, ensuring reproducibility and theoretical identifiability.

Canonical representations of molecular structures are foundational to cheminformatics, computational chemistry, and molecular machine learning. They enable the unique, reproducible identification, indexing, comparison, and processing of molecules in silico. Canonicalization encompasses explicit string and graph encodings such as SMILES, algorithmic procedures over molecular graphs that ensure a unique representation, symmetry-invariant three-dimensional descriptors, and extensive stereochemistry handling—including both point and axis chirality. Furthermore, recent frameworks have extended canonicity to algebraic data types and machine learning latent spaces, aiming for representation theoretical guarantees and identifiability. This entry provides a comprehensive treatment of these approaches and exposes their theoretical underpinnings, algorithmic structures, practical implications, and outstanding challenges.

1. Formal Definitions, Purposes, and Role in Molecular Informatics

Canonical representations map each chemically distinct molecule to a unique digital object, eliminating ambiguity arising from alternative atom orderings, resonance forms, conformers, or stereochemical conventions. Formally, a canonicalization function $C : {\mathcal M} \to {\mathcal R}$ selects a single representative from each equivalence class of molecular graphs or structures, with $\mathcal M$ denoting all valid molecular inputs and $\mathcal R$ the canonicalized outputs. For example, canonical SMILES and canonical molecular graphs are constructed so that isomorphic molecules yield identical representations. These canonical encodings underpin applications such as:

• Molecular database deduplication and indexing.
• Molecular similarity search and clustering.
• Training and evaluation of chemical LLMs (CLMs), where the canonical representation mediates between chemical reality and text-processing machinery.
• Comparison of three-dimensional structures and assignment of minimal-root-mean-square deviation (RMSD).
• Providing symmetry- and invariance-respecting inputs to machine learning models for property prediction, structure generation, or quantum chemical calculation.

The absence of strict canonicalization propagates ambiguity, impairs reproducibility, and degrades data-driven model integrity (Kikuchi et al., 11 May 2025).

2. Canonical Molecular Graphs and Stereochemistry

Molassembler offers a rigorous pipeline, generalizing canonicalization across organic, inorganic, and organometallic chemistries (Sobez et al., 2020). It models molecules as undirected, connected graphs $G=(V,E)$, with atoms as vertices and labeled bonds as edges, supplemented by explicit stereopermutators for both atom- and bond-centered stereochemistry. The canonicalization proceeds through the following stages:

1. Graph Construction: Atoms and bonds, with element labels and idealized bond orders or haptic binding, instantiate $G$.
2. Substituent Ranking: IUPAC-inspired recursive ranking assigns each neighbor of atom $v$ a priority $p(v,w)$ based on atomic number, bond order, and recursively the priorities of further neighbors. Haptic substituents are handled as binding sites, ranked by sorted priority multisets and binding cardinality.
3. Shape Classification: Non-terminal atoms with degree $\geq 3$ have their neighborhood mapped onto an idealized polyhedral shape, e.g., tetrahedron, octahedron, etc., enabling systematic stereopermutator enumeration.
4. Stereodescriptor Index Assignment: Permutational indices, abstracted over all labelings and factored by the rotational symmetry group $Rot(s)$ of the shape, yield a unique integer stereodescriptor for each local configuration:

$$n_{\rm abstract} = \frac{V_s! / \prod_k m_k!}{|\mathrm{Rot}(s)|}$$

where $m_k$ is the multiplicity of each ranked character.

1. Graph Coloring and Isomorphism Testing: Vertex invariants (128-bit color hashes including element, bond order set, local shape, and stereopermutation index) are passed to the sparsenauty routine. The canonical permutation $\pi$ establishes an ordering applied to the adjacency matrix and all per-atom data.
2. Canonical Representation Output: The reordered, fully labeled graph—incorporating stereochemistry, haptic binding, and all isomeric indices—uniquely encodes the molecule, agnostic to atom ordering or symmetry-related relabelings.

The procedure is strictly deterministic: two molecular descriptions differing only in atom order, labeling, or facial rotation/reflection collapse to the same canonical object, resolving both topological and stereochemical ambiguities.

3. String-Based Canonicalization: SMILES and its Limitations

Canonical SMILES, defined as a unique string for each molecular graph, constitute an industry standard for molecular communication and computational pipelines (Kikuchi et al., 11 May 2025). Canonicalization algorithms, e.g., Daylight, RDKit, and OpenBabel, implement multi-step atom ranking and iterative invariants refinement, culminating in a canonical depth-first traversal and string serialization. Cip-based chirality, Kekulé/aromatic ring conventions, and multiset hashing of neighbor invariants play key roles.

However, two principal forms of inconsistency disrupt uniqueness:

• Grammatical inconsistency: Different tools (or even different versions) may assign distinct canonical SMILES to the same molecule. For example, disagreement rates $V$(RDKit, OpenBabel) exceed 10% on large datasets.
• Stereochemical inconsistency: Omission or nonstandard encoding of chiral centers and geometric isomers results in up to 50% of enantiomers and 27% of geometric isomers lacking explicit annotation in extensive public sets.

These inconsistencies propagate to downstream CLMs, affecting translation accuracy (reductions up to 20% in perfect reconstruction) and inducing latent space drift, although property prediction performance remains largely stable due to feature selectivity (Kikuchi et al., 11 May 2025).

Strict documentation of software, unification of canonical pipelines, and comprehensive stereochemical assignment, preferably 3D-driven, are necessary to maximize reproducibility and chemical validity.

4. Canonicalization Using Three-Dimensional Structure

Most canonicalization schemes historically operate on 2D graphs or 1D strings, failing to distinguish atom equivalence in three-dimensional context. The canonized-then-minimized RMSD approach (Li et al., 2024) introduces a stereochemically aware, 3D-consistent indexing scheme central for structure alignment, molecular docking, and conformer clustering.

The approach proceeds as follows:

1. Initial Index Assignment: Atoms are sorted and partitioned using Cahn–Ingold–Prelog (CIP)–compatible rules: atomic number, heaviest neighbor mass, coordination number, in that order.
2. Graph-Based Refinement: Neighbor canonical indices are recursively sorted (DILIN), resolving further equivalences in the partitions.
3. Stereochemical Tagging: R/S and E/Z labels are explicitly assigned using spatial vector determinants and dihedral projections from 3D coordinates.
4. Symmetry-Resolved Tiebreaking: For unresolved equivalences, tree-based search branches assign provisional canonical indices, with each candidate mapping evaluated and pruned by performing an optimal 3D alignment and minimizing RMSD.
5. Hydrogen Handling: Hydrogens are permuted only within their local parent atom, reducing factorial combinatorial explosion. This localizes symmetry resolution, keeping the method computationally tractable for molecules with up to several hundred atoms.

All steps are implemented in Python + RDKit. The method yields canonical indexings fully compatible with 3D structural alignment, guarantees stereochemical correctness, and achieves practical timings on large molecular systems. It strictly supersedes the SSL (Schneider–Sayle–Landrum) canonizer by eliminating arbitrary tiebreaks and guaranteeing minimal RMSD matches (Li et al., 2024).

5. Algebraic Data Types and Canonical Digital Molecule Constructions

Recent work advances canonical representations beyond graphs and strings, leveraging algebraic data types (ADTs) to encode not only connectivity, coordinates, and stereochemistry but also quantum mechanical structure and symmetry groups (Goldstein et al., 23 Jan 2025).

The ADT-based paradigm (in Haskell, as presented by Goldstein) formalizes a molecule as a record holding a list of Atoms and an explicit map of Bonds. Each Atom object encodes:

• Unique integer ID and element attributes.
• Optional 3D coordinate.
• Electronic configuration, shells, subshells, and orbitals.

BondType accommodates covalent, ionic, and hydrogen bonds, fully supporting delocalization, multi-center bonding, and resonance cycle labeling. Canonicalization is achieved by sorting atoms and bonds lexicographically by invariant keys (element, neighbor list, orbital data), guaranteeing a unique, reproducible digital object for each molecular graph.

Rigid rotational symmetry is moreover represented at the type level, turning the set of all rotated instances of a molecule into an algebraic group (with composition, identity, and inversion), thus enabling learning and inference methods that are equivariant or invariant under $SO(3)$ or relevant point groups.

Distinct advantages of ADT representations, compared to SMILES or even graph-based methods, include:

• Guaranteed well-formedness via type-system constraints.
• Compositional extensibility: e.g., new bond types, quantum fields, or chemical reactions without redefining syntactic rules.
• True symmetry invariance, including under coordinate rotation and permutation.
• Direct integration with probabilistic programming and Bayesian inference.

Canonicalization within the ADT approach consists of deterministic sorting and mapping steps, with key invariants (e.g., bond-map symmetry, electron conservation, group axioms) ensuring one-to-one mapping and reproducibility (Goldstein et al., 23 Jan 2025).

6. Canonical Machine Learning Embeddings and Identifiability

Canonical representations are critical not only for string and graph encodings but also for latent molecular embeddings in machine learning pipelines. Multi-task regression frameworks enable embeddings with identifiability guarantees up to permutation and scaling (Chen et al., 2023).

The theoretical setting involves two stages:

1. Multi-task Regression Network (MTRN): A shared featurizer $h_\phi(x)$ and per-task linear heads $w_t$ yield $$p(y|x,t) = \mathcal{N}(y|w_t^T h_\phi(x), \sigma^2)$$. This stage ensures linear identifiability.
2. Causal Prior Structure and Marginal Likelihood Maximization: By encoding a task- and label-dependent prior $p(z|y,t)$ over latent space—separating causal and spurious coordinates according to the task—maximization of the exact Gaussian marginal likelihood (in closed form) enforces identifiability up to permutation and scaling. This is provably optimal under sparse mechanism shift and in the presence of diverse multi-task property structure.

Empirical results demonstrate high dimensional match correlation coefficients (MCC > 0.9) between canonicalized latent embeddings and true underlying molecular structural factors, outperforming unsupervised variational autoencoder approaches. The resulting embeddings are robust, reproducible, and suitable for downstream applications such as property prediction, clustering, and structure–property disentanglement (Chen et al., 2023).

7. Generalized Symmetric Descriptors and Open Problems

Physics-informed structural representations, notably atom-centered symmetry functions (ACSF), SOAP (Smooth Overlap of Atomic Positions), Coulomb matrices, and higher-order expansion frameworks (ACE, MTP), provide continuous, symmetry-respecting descriptors for molecules and materials (Musil et al., 2021). These encode invariance under translation, rotation, and permutation, which is essential for generalizing canonicalization beyond discrete graphs and strings.

Completeness, i.e., injectivity with respect to all symmetry-distinct molecular structures, is systematically improvable but not always guaranteed at finite expansion order. 2-body and even many 3-body invariant descriptors may map distinct configurations to identical representations, indicating residual non-uniqueness. Algebraic completeness guides the construction of minimal, injective descriptors.

Open research directions include constructing finite, invertible descriptors suitable for generative molecular modeling, seamless integration of 2D graph and 3D density descriptors for QSPR or reactivity modeling, unified handling of electron density and quantum information, as well as standardized, reproducible benchmarks and open tooling (Musil et al., 2021).

---

In summary, canonical representations—across string, graph, 3D, algebraic type, and latent embedding paradigms—are central to the digital handling, analysis, and learning of molecular structure. Multiple algorithmic and mathematical frameworks ensure unambiguous, symmetry-respecting encodings, but practical limitations, especially regarding steric and stereochemical completeness, symmetry inconsistencies, and injectivity, remain active areas of development and debate.