Foundation Molecular Grammar: Multi-Modal Foundation Models Induce Interpretable Molecular Graph Languages
Michael Sun, Weize Yuan, Gang Liu, Wojciech Matusik, Jie Chen
Abstract
Recent data-efficient molecular generation approaches exploit graph grammars to introduce interpretability into the generative models. However, grammar learning therein relies on expert annotation or unreliable heuristics for algorithmic inference. We propose Foundation Molecular Grammar (FMG), which leverages multi-modal foundation models (MMFMs) to induce an interpretable molecular language. By exploiting the chemical knowledge of an MMFM, FMG renders molecules as images, describes them as text, and aligns information across modalities using prompt learning. FMG can be used as a drop-in replacement for the prior grammar learning approaches in molecular generation and property prediction. We show that FMG not only excels in synthesizability, diversity, and data efficiency but also offers built-in chemical interpretability for automated molecular discovery workflows. Code is available at https://github.com/ shiningsunnyday/induction . Foundation Molecular Grammar: Multi-Modal Foundation Models Induce Interpretable Molecular Graph Languages Prompt I want you to think like a chemist performing a detailed analysis of the chemical composition of a toxic compound through its constituent motifs. I will highlight for you "'24"' of the substructures of a molecule. They are numbered one-by-one from Motif 0 to Motif "'23"', inclusive. I want you to explain, concisely, what each numbered motif is. Make sure to start from Motif 0 and go in order of the numbering. MAKE SURE you describe EVERY MOTIF! Response Certainly! Here is a concise description of each motif: Motif 0: A carbonyl group (C=O) attached to a carbon chain. Motif 1: A nitrile group (C ≡ N ) attached to a tertiary carbon. Motif 2: A di-substituted carbon chain with two adjacent nitrile groups (N=C-C=C-N). Motif 3: Similar to Motif 2 but with an additional carbon on one end. Motif 4: An imine group (C=N) connected to a vinylic carbon chain (C=C). Motif 5: Benzene ring (C6H6) -a six-membered aromatic carbon ring. Motif 6: A singly unsaturated carbon chain. Motif 7: A doubly unsaturated carbon chain. Motif 8: Di-substituted carbon chain with a nitrile group (N=C) on one end and another nitrile (N ≡ C) in the middle. Motif 9: Similar to Motif 8 but with different placement of carbons. Motif 10: An amino group (NH2) attached to a C=N group (imino). Motif 11: A ketone (C=O) with adjacent carbons. Motif 12: A cyano-ketone (C ≡ N and C=O) attached to adjacent carbons. Motif 13: An alkyl cyanide (R -C ≡ N ) where R is an alkyl chain. Motif 14: An amino group (NH2) connected to a nitrile (C ≡ N ). Motif 15: A benzene ring (C6H6) -slightly different representation of Motif 5. Motif 16: Another representation of a benzene ring (C6H6). Motif 17: A nitrile group (C ≡ N ) attached to a doubly unsaturated carbon chain. Motif 18: An alkyl halide (carbon chain with a chlorine, C-Cl). Motif 19: Similar to Motif 18 with a different number of carbons. Motif 20: A doubly canned ketone (C=O) group attached to an imine (C=N) group.
Ask about this paper
Your agent reads all of it.
Lune indexed this paper to the last equation, along with the top-tier papers that cite it. Ask a question and the answer quotes them.
Your agent calls
Luneget_paper_fulltext
Free to start. No credit card required.
Terminal
Install the CLIlune papers fulltext 5a4c9ad4-d68f-4f5b-9f34-622a33a7abddCited by top-tier papers3
- Graph Diffusion Transformers are In-Context Molecular DesignersGang Liu, Jie Chen, Yihan Zhu, Michael Sun et al.ICLR 2026 · 7 citations
- Post Hoc Regression Refinement via Pairwise RankingsKevin Tirta Wijaya, Michael Sun, Minghao Guo, Hans-Peter Seidel et al.NeurIPS 2025 · 1 citation
- Bayesian Post Training Enhancement of Regression Models with Calibrated RankingsKevin Tirta Wijaya, Bing Hu, Hans-Peter Seidel, Wojciech Matusik et al.ICLR 2026
Builds on13
- Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsJason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma et al.NeurIPS 2022 · 22,562 citations
- Finetuned Language Models are Zero-Shot LearnersJason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu et al.ICLR 2022 · 4,966 citations
- Hierarchical Generation of Molecular Graphs using Structural MotifsWengong Jin, Regina Barzilay, Tommi S. JaakkolaICML 2020 · 356 citations
- Debating with More Persuasive LLMs Leads to More Truthful AnswersAkbir Khan, John Hughes, Dan Valentine, Laura Ruis et al.ICML 2024 · 244 citations
- Retro*: Learning Retrosynthetic Planning with Neural Guided A* SearchBinghong Chen, Chengtao Li, Hanjun Dai, Le SongICML 2020 · 151 citations
Related papers
- Representing Molecules as Random Walks Over Interpretable GrammarsMichael Sun, Minghao Guo, Weize Yuan, Veronika Thost et al.ICML 2024 · 6 citations
- Data-Efficient Graph Grammar Learning for Molecular GenerationMinghao Guo, Veronika Thost, Beichen Li, Payel Das et al.ICLR 2022 · 46 citations
- Hierarchical Grammar-Induced Geometry for Data-Efficient Molecular Property PredictionMinghao Guo, Veronika Thost, Samuel W. Song, Adithya Balachandran et al.ICML 2023
- MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph GenerationZhaoning Yu, Hongyang GaoICLR 2025
- Barking up the right tree: an approach to search over molecule synthesis DAGsJohn Bradshaw, Brooks Paige, Matt J. Kusner, Marwin H. S. Segler et al.NeurIPS 2020 · 71 citations
