AI-assisted interpretation of Markush structures in pharmaceutical patents: a review of emerging tools, datasets, and challenges

Automated interpretation of Markush structures widely used in pharmaceutical patents to claim large families of related compounds remains challenging due to non-machine-readable structure images, variable R-groups, dependency rules, scaffold diversity, and heterogeneous claim language. Challenges include attachment points and stereochemistry, nested/conditional dependencies, and inconsistent drafting conventions that hinder faithful enumeration. Early rule-based cheminformatics systems parsed claims and mapped Markush representations into searchable formats, but struggled with nested dependencies, cross-references, and multimodal (text + image) descriptions. More recently, artificial intelligence (AI) methods have been introduced including language-based tools, vision-based tools, and multimodal or hybrid tools. Language-based tools increasingly use large language models (LLMs) and natural language processing (NLP) capabilities to extract variable definitions, constraints, and dependency graphs from claim text; vision systems translate structure depictions into machine-readable formats (e.g., SMILES, CXSMILES); multimodal or hybrid pipelines align both for end-to-end interpretation. Emerging datasets support these efforts, though licensing, family-wise leakage, and standardized splits remain inconsistent. This narrative review synthesizes tools, datasets, and evaluation practices for AI-assisted Markush interpretation, identifies persistent failure modes, and maps open legal questions (sufficiency, enablement, enforceability). We outline priorities for the field; transparent benchmarks with family-aware splits, interpretable constraint handling, and workflows aligned with U.S. Patent Office practice, near-term use is decision support, not legal advice.