Higher-Order Molecular Grammars for Generative and Foundation Models in Chemistry
The paper turns molecular topology into grammar-rule sequences that ordinary sequence models can use.
HGR lifts molecules into combinatorial complexes, then parses them with a context-free higher-order grammar to preserve structures like rings and recurring motifs. The authors say this avoids the cost of explicit higher-order encodings while keeping validity decodable by construction. They also introduce RingDiv, a 1.18 million-molecule benchmark, plus an RDI measure for ring-system coverage. In their reported tests, HGR-based models hit 100% validity and led FCD across five generation benchmarks, while HGR-FM topped mean AUC across seven MoleculeNet benchmarks. ArXiv · AI/CL/LG's note
HGR lifts molecules into combinatorial complexes, then parses them with a context-free higher-order grammar to preserve structures like rings and recurring motifs. The authors say this avoids the cost of explicit higher-order encodings while keeping validity decodable by construction. They also introduce RingDiv, a 1.18 million-molecule benchmark, plus an RDI measure for ring-system coverage. In their reported tests, HGR-based models hit 100% validity and led FCD across five generation benchmarks, while HGR-FM topped mean AUC across seven MoleculeNet benchmarks. ArXiv · AI/CL/LG's note
score 4