Technical Summary: Molecular Representations for Large Language Models

Problem Statement

Large Language Models (LLMs) are increasingly applied to scientific discovery, including tasks requiring reasoning about molecular structures such as reaction prediction, mechanistic analysis, and structure elucidation. However, current LLMs are sequence-based models that operate on text, whereas molecules are naturally defined as graphs (atoms as nodes, bonds as edges). Consequently, molecular graphs must be serialized into text-based formats for LLM processing.

While previous studies in chemistry have predominantly utilized SMILES strings or IUPAC names as the interface between LLMs and molecular structures, the suitability of these formats for LLM reasoning has not been systematically assessed. These formats were originally designed for traditional cheminformatics workflows and human interpretation, imposing strict syntactic and semantic constraints (e.g., traversal rules for SMILES, nomenclature rules for IUPAC) that may not align with the architectural strengths of LLMs. This misalignment potentially limits the reliability of LLMs in interpreting and generating molecular graphs.

Methodology

To address this gap, the authors introduced MolJSON, a novel structured JSON schema designed specifically as an input and output format for LLMs. The study systematically compared MolJSON against five established chemical representations: SMILES, IUPAC names, SELFIES, InChI, and MOL V2000.

The MolJSON Schema

MolJSON represents molecules as JSON objects containing explicit lists of atoms and bonds, rather than relying on linear traversal or rule-based nomenclature.

• Atoms: Defined by a unique identifier and an element symbol.
• Bonds: Defined by pairs of atom identifiers and a bond order.
• Optional Fields: Includes sparse fields for non-zero formal charges and explicit hydrogens on aromatic nitrogen atoms to ensure correct valence assignment.
• Design Philosophy: Molecular identity is defined directly by the sets of atoms and bonds, independent of array ordering, eliminating the need for specific graph traversals.

Benchmark Design

The authors constructed a benchmark comprising 78,045 algorithmically generated questions to isolate the impact of representation on LLM performance, avoiding confounding factors from specialized chemistry knowledge. The evaluation utilized four models: GPT-5-nano, GPT-5-mini, GPT-5, and Claude Haiku 4.5. The tasks included:

1. Molecular Translation: Converting a molecule from one format to another (e.g., IUPAC to SMILES). Accuracy was measured by exact molecular equivalence using canonical SMILES and RDKit.
2. Shortest-Path Reasoning: Given a molecule, the model must determine the number of bonds on the shortest path between two specific halogen atoms.
3. Constrained Molecular Generation: The model must generate a molecule satisfying specific constraints (e.g., presence of specific halogens, ring counts, topologies, and path lengths).

Data was sourced from the PubChem database (molecules deposited between October 2025 and December 2025) to minimize contamination from model training data. Molecules with stereochemistry, salts, or inorganic elements were excluded.

Key Results

The study observed substantial variation in LLM performance depending on the molecular representation used, with MolJSON consistently outperforming existing formats across all tasks and models.

1. Molecular Translation

• Output Performance: When generating outputs, MolJSON significantly outperformed SMILES and IUPAC. For GPT-5, converting IUPAC names to MolJSON achieved 71.0% accuracy, compared to 43.7% for converting IUPAC to SMILES. Similarly, SMILES-to-MolJSON translation reached 92.3% accuracy, versus 43.5% for SMILES-to-IUPAC.
• Input Performance: MolJSON also served as a superior input format. When generating SMILES outputs, accuracy improved from 43.7% (IUPAC input) to 48.3% (MolJSON input).

2. Shortest-Path Reasoning

• MolJSON was the most reliable input format for graph reasoning. GPT-5 achieved 98.5% accuracy with MolJSON inputs, compared to 92.2% for SMILES and 82.7% for IUPAC.
• Efficiency: MolJSON required significantly fewer reasoning tokens. GPT-5 used an average of 1,021 tokens with MolJSON inputs versus 1,854 tokens with SMILES, suggesting that SMILES and IUPAC require the model to reconstruct the graph before reasoning, a step that is error-prone.
• Failure Modes: IUPAC performance degraded significantly with fused ring systems (dropping to 66.2% accuracy for fused rings vs. 96.9% for non-fused), whereas MolJSON remained robust (98.5% for both).

3. Constrained Generation

• MolJSON was the most reliable output format for generating molecules under constraints. GPT-5 achieved 95.3% accuracy with MolJSON, compared to 76.3% for IUPAC and 64.0% for SMILES.
• The advantage was particularly pronounced for complex topologies. For fused ring systems, GPT-5 achieved 92% accuracy with MolJSON, versus 42% for SMILES and 56% for IUPAC.

4. Error Analysis

Systematic errors were observed for SMILES and IUPAC related to molecular size and ring complexity.

• SMILES: Errors often stemmed from syntax violations and invalid molecules, suggesting the strict traversal constraints of the format are challenging for LLMs to maintain.
• IUPAC: Errors were heavily associated with fused ring systems and complex nomenclature rules.
• MolJSON: Demonstrated robustness to these failure modes. The authors note that the flexibility of MolJSON allows LLMs to use "rich" atom identifiers (e.g., functional group abbreviations) which may act as "graph tokens" in the model's reasoning process, a capability not available in rigid formats like SMILES or IUPAC.

Significance and Claims

The paper claims that the choice of molecular representation has a material impact on the ability of LLMs to interpret and generate molecular graphs. The authors argue that:

1. Explicit Graph Encodings are Superior: Formats that explicitly encode atoms and bonds (like MolJSON) are intrinsically better aligned with current LLMs than serialized formats (SMILES) or rule-based nomenclature (IUPAC).
2. Performance Without Explicit Training: Despite SMILES and IUPAC being ubiquitous in training data, MolJSON outperformed them without any explicit training on the format, suggesting the structural advantage is inherent to the representation.
3. Foundation for Agentic Systems: As LLM-based systems in chemistry move toward autonomous discovery, the reliability of molecular interaction is critical. The authors suggest that structured representations like MolJSON should be adopted for developing LLM-based chemical systems.

The authors position MolJSON as a starting point for LLM-oriented molecular representations rather than a definitive final format, noting potential improvements such as better handling of stereochemistry and atomic coordinates. The study concludes that exploring molecular representations tailored specifically for LLMs is a promising direction for the field.