A Large-Scale Dataset for Molecular Structure-Language Description via a Rule-Regularized Method

Molecular function is largely determined by structure. Accurately aligning molecular structure with natural language is therefore essential for enabling large language models (LLMs) to reason about downstream chemical tasks. However, the substantial cost of human annotation makes it infeasible to construct large-scale, high-quality datasets of structure-grounded descriptions. In this work, we propose a fully automated annotation framework for generating precise molecular structure descriptions at scale. Our approach builds upon and extends a rule-based chemical nomenclature parser to interpret IUPAC names and construct enriched, structured XML metadata that explicitly encodes molecular structure. This metadata is then used to guide LLMs in producing accurate natural-language descriptions. Using this framework, we curate a large-scale dataset of approximately $163$k molecule-description pairs. A rigorous validation protocol combining LLM-based and expert human evaluation on a subset of $2,000$ molecules demonstrates a high description precision of $98.6%$. The resulting dataset provides a reliable foundation for future molecule-language alignment, and the proposed annotation method is readily extensible to larger datasets and broader chemical tasks that rely on structural descriptions.

💡 Research Summary

The paper addresses a fundamental bottleneck in chemical AI: the lack of large‑scale, high‑quality datasets that tightly couple molecular structure with natural‑language descriptions. While recent multimodal models have shown that aligning visual data with text enables powerful reasoning, analogous progress for chemistry has been hampered by the prohibitive cost of expert annotation—approximately one hour per molecule. To overcome this, the authors propose a fully automated annotation pipeline that leverages and extends the rule‑based chemical nomenclature parser OPSIN.

OPSIN can translate systematic IUPAC names into molecular graphs, but its native XML parse tree is an internal representation that discards many structural details needed for language generation, such as explicit ring‑fusion relationships, stereochemistry, and attachment positions. The authors therefore engineer a transformation layer that enriches the parse tree into a comprehensive XML metadata format. This metadata explicitly records substituent locations, ring topology (including fused, bridged, and spiro systems), stereochemical descriptors, and labeling schemes. By providing this structured “structural hint” to a large language model (LLM), the model no longer needs to infer complex topology from linear SMILES strings, dramatically reducing hallucinations.

The pipeline proceeds as follows: (1) a large collection of molecules is drawn from public databases (PubChem, ChEMBL, etc.) and their IUPAC names are extracted; (2) each name is parsed by OPSIN, and the enriched XML metadata is generated; (3) the metadata is embedded in a carefully crafted prompt and fed to GPT‑4‑Turbo, which produces a concise, unambiguous natural‑language description of the structure; (4) automatic filters remove overly long, duplicate, or chemically inconsistent outputs. The resulting corpus contains 163 085 molecule‑description pairs.

Quality assessment combines automated cross‑checking (ensuring the generated text is consistent with the metadata) and a human‑in‑the‑loop evaluation on a 2 000‑sample subset. Three chemistry experts and GPT‑4‑Turbo independently rated each description for accuracy, completeness, and ambiguity. The overall precision reached 98.6 %, and inter‑annotator agreement (Cohen’s κ) was 0.91, indicating very high consistency. Ablation studies showed that removing the enriched metadata and relying solely on SMILES reduced accuracy to 86 % and increased error types, confirming the critical role of the rule‑regularized structural information.

Key contributions of the work are: (i) a scalable, fully automated method for generating high‑fidelity structure‑grounded textual descriptions; (ii) a novel XML metadata schema that captures the full chemical topology required for reliable language generation; (iii) a publicly released dataset of over 160 k high‑quality pairs that can serve as a foundation for downstream tasks such as structure recognition, property prediction, synthesis planning, and multimodal model training. The authors discuss limitations, notably the reliance on standard IUPAC nomenclature, and outline future directions including support for non‑standard names, mixtures, and multilingual description generation. By making both the dataset and the annotation pipeline openly available, the paper aims to catalyze the development of truly structure‑aware chemical language models.