Higher-Order Molecular Grammars for Generative and Foundation Models in Chemistry

Abstract

Higher-order Grammar Representation (HGR) encodes molecular topology as compact sequences of grammar rules. Molecules are lifted to combinatorial complexes and parsed with a context-free higher-order grammar, allowing standard sequence models to learn from ring systems and recurring motifs. The framework supports molecular generation and transferable representation learning through generative models and HGR-FM. The work also introduces RingDiv, a ring-enriched molecular benchmark, its curated RingDiv300k subset, and the ring diversity index (RDI) for measuring ring-system coverage.

Yiming Huang

PhD Candidate

My research focuses on topological deep learning for complex and higher-order graph structures, as well as generative models, graph representation learning, and foundation models.