The Structural Bottleneck of Linear Molecular Representations

52 concepts46 papers

Deep learning has transformed chemistry, largely by treating molecules as text. By flattening three-dimensional chemical structures into one-dimensional strings like SMILES or SELFIES, researchers have unlocked the power of large language models (LLMs) and sequential neural networks for molecular design, synthesis planning, and property prediction.

However, this convenience comes at a steep conceptual cost. A molecule is not a sentence; it is a multi-dimensional quantum graph. Forcing this complex reality into a linear sequence introduces a fundamental representation bottleneck that limits the robustness, scalability, and domain applicability of chemical AI.

The Fragility of String Synonyms

To a language model, a string is a sequence of discrete tokens. To a chemist, a string like CCO is an instruction set to build ethanol. The core issue is that the mapping from molecular graphs to linear strings is rarely one-to-one. A single molecule can be written in dozens of equally valid ways depending on where one chooses to start numbering atoms or how hydrogens are explicitly drawn.

This lack of semantic invariance severely trips up modern chemical language models. In Lost in Translation: Chemical Language Models and the Misunderstanding of Molecule Structures, Veronika Ganeeva et al. demonstrate that models fail to recognize these simple "identity transformations." When faced with structural synonyms, models exhibit massive performance drops in retrieval and property prediction tasks.

The root of this vulnerability lies in how models process text. As Kuzma Khrabrov et al. systematically evaluated in Measuring Chemical LLM robustness to molecular representations: a SMILES variation-based framework, state-of-the-art models fail to maintain consistent internal representations when faced with simple structural variations of the same chemical entity. Because these models treat strings as arbitrary characters rather than structured chemical information—a phenomenon highlighted by Khiem Le et al. in MolX: Enhancing Large Language Models for Molecular Understanding With A Multi-Modal Extension—they lack the inductive bias required to understand that different token sequences map to the exact same physical object.

Graph Representation (Invariant)
       OH
       /
  CH2—CH3  

Valid SMILES Strings (Variant)
  1. CCO
  2. OCC
  3. C(O)C

Flattening Topology and Losing Long-Range Context

When a 2D or 3D molecular graph is flattened into a 1D string, spatial proximity is shattered. Two atoms that are physically adjacent or electronically coupled in 3D space might end up on opposite ends of a text string, separated by dozens of characters. For small, rigid molecules, deep neural networks can sometimes memorize these sequential patterns, but the strategy breaks down entirely when scaling to complex architectures.

Consider macromolecules and polymers. In Transferring a Molecular Foundation Model for Polymer Property Predictions, Zachary R Fox et al. show that even when linear notations are adapted for polymers, they are insufficient for predicting properties that depend on long-range structural features and inter-unit interactions. The sequential nature of strings obscures the vital topological and connectivity information required to understand how a macromolecule folds or interacts with its environment.

This sentiment is echoed by Ao Shen et al. in Complementary multi-modality molecular self-supervised learning via non-overlapping masking for property prediction, who note that SMILES strings inherently struggle to capture complex topological information. When models are forced to rely purely on character embeddings, as argued by Jinxian Wang et al. in Molecular property prediction by contrastive learning with attention-guided positive sample selection, they overlook the rich functional and structural relationships that dictate real-world chemical behavior. This makes reasoning about complex phenomena, such as drug-drug interactions or organ-specific toxicities, incredibly difficult for text-only models (Zhaoyue Sun et al.).

The Organic Bias and the Organometallic Gap

The limitations of linear strings are not just syntactic; they are deeply historical. Formats like SMILES were built primarily for organic chemistry—systems dominated by well-behaved, localized covalent bonds between carbon, nitrogen, oxygen, and hydrogen.

When AI models attempt to venture into inorganic chemistry or computational catalysis, they hit a wall. Transition metal complexes, organometallic catalysts, and crystalline lattices do not conform to simple valence rules. In SELFIES and the future of molecular string representations, Qianxiang Ai et al. highlight that current molecular representations fail to capture the complex bonding and structural features of transition metal compounds. As Konstantinos D. Vogiatzis et al. explain in Boosting Computational Catalysis and Chemical Reactivity with Artificial Intelligence, standard representations cannot accurately describe the intricate electronic structures and d-orbital bonding characteristics unique to these elements.

This representation gap directly impairs generative and predictive models in materials science:

A transition metal catalyst's behavior depends on complex 3D coordination geometries and electronic states that cannot be captured by an adjacency token in a text string.

Grammatical Constraints on Stereochemistry and Scale

Even when researchers attempt to patch linear strings by introducing stricter grammars or fragment-based tokens, they run into fundamental expressivity ceilings. For example, fragment-based extensions designed for mesoscopic simulations often restrict nested monomers or fail to directly encode intrinsic attributes like electric charges and chirality without creating entirely separate, cumbersome notations (Andreas Truszkowski et al. in A molecular fragment cheminformatics roadmap for mesoscopic simulation).

Similarly, advanced frameworks like Group SELFIES struggle to natively support complex chiral arrangements and overlapping functional groups without explicit, manual definitions (Austin Henry Cheng et al. in Group SELFIES: A Robust Fragment-Based Molecular String Representation). Furthermore, early semantic web specifications and older cheminformatics toolkits frequently struggle to scale to larger molecular weights or handle complex repeating structures like polymers with arbitrary units (Leonid L. Chepelev in Chemical Entity Semantic Specification: Knowledge representation for efficient semantic cheminformatics and facile data integration).

Ultimately, treating chemistry as a language has provided a powerful starting point for molecular AI. However, the field is rapidly outgrowing the constraints of the one-dimensional string. Overcoming the structural bottleneck requires a shift toward representations—such as multi-modal extensions, 3D molecular graphs, and electronic structure embeddings—that natively respect the topological, geometric, and quantum-mechanical realities of molecules.

Key Foundational Work

Chemical Entity Semantic Specification: Knowledge representation for efficient semantic cheminformatics and facile data integration (2011) — contributed 1 concept elements to this direction.
Structure-based classification and ontology in chemistry (2012) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_limitations_of_smiles_representation_4c73b3be,
  title = {limitations of SMILES representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/limitations-of-smiles-representation/rdr_4c73b3be53c44ba08a01d8012c095361},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_limitations_of_smiles_representation_4c73b3be_v1,
  title = {limitations of SMILES representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/limitations-of-smiles-representation/rdr_4c73b3be53c44ba08a01d8012c095361/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake