STAMP as a topology-aware molecular language for fragment-level learning and editing
Topology-aware molecular language, fragment-level learning, and executable editing
Yuxuan Ren, Jiatong Li, Ziqiao Meng, and Yatao Bian. (2026). STAMP as a topology-aware molecular language for fragment-level learning and editing. arXiv:2610.18438.
About this work
STAMP—Sequential Tree Attachment-based Motif-Preserving representation—is a topology-aware molecular language that explicitly encodes motif identity, attachment sites, and fragment-tree topology. This makes chemically meaningful fragments reusable across contexts and gives models directly addressable subtrees for structural editing.
Across molecular generation, text-to-molecule translation, property prediction, activity-cliff prediction, optimization, and instruction-guided editing, STAMP connects learned fragment representations with executable operations. Its deterministic compiler also allows frozen language models to act on indexed molecular fragments instead of rewriting complete molecular strings.