Research map / Vision-language and multimodal / Human and embodied modeling
Foundation Molecular Grammar: Multi-Modal Foundation Models Induce Interpretable Molecular Graph Languages
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Foundation Molecular Grammar (FMG), which uses GPT-4o on rendered molecule images and text to guide clique-tree decomposition and induce a hyperedge replacement grammar for data-efficient molecular generation.
Problem
Graph grammar learning for molecular generation relies on expert annotation or unreliable heuristics, limiting interpretability and quality of the induced substructure vocabulary and rules.
Approach
Replace expert annotation and heuristic choices in clique tree decomposition with multi-modal foundation model selections over rendered substructure images and self-generated textual descriptions, then rank runs by an LLM tournament over their chain-of-thought "design stories" and pool the top rule sets into a hyperedge replacement grammar.
Key results
- On small datasets FMG achieves 100% validity and uniqueness with 99.6-99.8% membership, and 93.0% RS on Acrylates versus 84-98% for baselines.
- On HOPV and PTC, FMG reaches 100% uniqueness and RS of 70% and 78% for HOPV, and is one of only two methods with 100% uniqueness.
- LLM judge agrees with expert in 77/108 (71%) pairwise comparisons, p=1.1e-5 against a coin-flip null.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FMG instead of RW (expert) | Molecular generation on HOPV · HOPV | Retro* Score: 58% -> 70% | Helped |
| Use FMG instead of RW (expert) | Molecular generation on PTC · PTC | Retro* Score: 60% -> 78% | Helped |
| Replace molecular images with atom-numbered text encoding (ablation) | FMG · Small datasets | Diversity | Hurt |
| Replace molecular images with atom-numbered text encoding (ablation) | FMG · Chain Extenders | Generation performance | No clear effect |
| Use Top-k tournament-ranked runs instead of the first k runs (ablation) | FMG · Small datasets | Diversity | No clear effect |
| Use Top-k tournament-ranked runs instead of the first k runs (ablation) | FMG · Small datasets | Membership | Helped |
| Increase the number of pooled rule-set candidates k (ablation) | FMG · Small datasets | Diversity | Helped |
| Increase the number of pooled rule-set candidates k (ablation) | FMG · Small datasets | Membership | Hurt |
| Add MMFM-guided motif merging (ablation) | FMG · Isocyanates | Retro* Score: 32.6±5.7 -> 44.3±3.4 | Helped |
| Add MMFM-guided motif merging (ablation) | FMG · Acrylates | Membership: 16.61±0.78 -> 35.48±4.02 | Helped |
| Add MMFM-guided spanning-tree edge selection instead of the maximal spanning tree heuristic (ablation) | FMG · Acrylates | Retro* Score: 93.5±1.8 -> 87.4±1.5 | Mixed |
| Select the root clique with MMFM instead of at random (ablation) | FMG · Acrylates | Membership: 22.90±2.53 -> 35.48±4.02 | Helped |
Surprising findings
- FMG performs poorly on membership for HOPV (thiophenes) despite strong results on PTC.
- Pooling grammars from more runs improves diversity and RS but reduces class membership, showing an intrinsic trade-off.
Limitations
- Low coverage on small datasets: grammar rules with complex RHS are less applicable, and diversity is low for very small datasets.
- Relies on GPT-4o API and stochastic multi-run ensembles, which are costly and non-deterministic.
- Membership is not strictly defined for HOPV and PTC, and performance on HOPV membership is poor.
Experimental setup
Datasets: Isocyanates Acrylates Chain Extenders HOPV PTC MOSES
Benchmarks: Isocyanates Acrylates Chain Extenders HOPV PTC MOSES
Metrics: Valid Unique Novelty Diversity Retro* score Membership
Baselines: MHG DEG Random Walk (RW) STONED JT-VAE Hier-VAE MoLeR MolGPT MolT5 Text+Chem T5 GPT-4o ICL
Builds on: hyperedge replacement grammar minimal hyperedge replacement grammar junction tree algorithm Bradley-Terry model
Closest papers by meaning
- Representing Molecules as Random Walks Over Interpretable Grammars (ICML 2024 · similarity 0.88)
- Domain-Agnostic Molecular Generation with Chemical Feedback (ICLR 2024 · similarity 0.88)
- MORE: Molecule Pretraining with Multi-Level Pretext Task (AAAI 2025 · similarity 0.87)
- From Artificially Real to Real: Leveraging Pseudo Data from Large Language Models for Low-Resource Molecule Discovery (AAAI 2024 · similarity 0.87)
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.87)
- Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.