Research map / Graph learning / Graph self-supervised learning
Hierarchical Graph Tokenization for Molecule-Language Alignment
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 HIGHT, a hierarchical graph tokenizer (atom, BRICS motif, molecule levels) plus a hierarchical functional-group instruction dataset (HiPubChem) to improve molecule-language alignment and reduce hallucination in LLM-based molecular models.
Problem
Existing large graph-language models tokenize molecules only at the atom/node level, ignoring hierarchical substructures (motifs, functional groups), which leads to poor molecule-language alignment and hallucinated functional-group answers.
Approach
Augment the molecular graph with BRICS motif supernodes and a whole-molecule supernode, tokenize atoms and motifs with separate VQVAEs and adapters (with Laplacian positional encodings), and augment instruction tuning with functional-group captions so the LLM learns what each motif means.
Key results
- On HBench, HIGHT-G reaches 66.8 macro F1 versus 50.0 for GIMLET and 52.6 for InstructMol, with a reported hallucination reduction of up to 40% in worst-class F1.
- On ChEBI-20 description generation, HIGHT-G achieves BLEU-4 0.405 versus 0.381 for InstructMol-G.
- On QM9 regression, HIGHT-G reaches average MAE 0.0086 versus 0.0130 for InstructMol-G.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace node-centric tokens with hierarchical atom-motif-molecule tokens | InstructMol-G · MotifHallu | F1 (neg): 9.5 -> 48.2 | Helped |
| Replace node-centric tokens with hierarchical atom-motif-molecule tokens | InstructMol-G · QM9 | Average MAE: 0.0130 -> 0.0086 | Helped |
| Replace node-centric tokens with hierarchical atom-motif-molecule tokens | InstructMol-G · CYP450 | ROC-AUC: 59.1 -> 80.5 | Helped |
| Replace node-centric tokens with hierarchical atom-motif-molecule tokens | InstructMol-G · ChEBI-20 test | BLEU-4: 0.381 -> 0.405 | Helped |
| Replace node-centric tokens with hierarchical atom-motif-molecule tokens | InstructMol-G · Forward Reaction Prediction | BLEU: 0.853 -> 0.869 | Helped |
| Add HiData motif descriptions to instruction tuning (ablation) | HIGHT-G · MotifHallu | F1 (neg): 12.5 -> 48.2 | Helped |
| Add HiData motif descriptions to instruction tuning (ablation) | HIGHT-G · MotifHallu | Accuracy: 21.6 -> 39.1 | Helped |
| Add SELFIES input to hierarchical tokenization (ablation) | HIGHT-G · MotifHallu | F1 (neg): 48.2 -> 42.7 | Hurt |
| Add HiData motif descriptions to instruction tuning (ablation) | HIGHT-GS · MotifHallu | F1 (neg): 6.5 -> 42.7 | Helped |
Surprising findings
- Adding 1D SELFIES input to node-centric LGLMs increased hallucination rather than reducing it.
- Node-centric LGLMs answered 'Yes' almost always on the functional-group hallucination benchmark (Yes ratio ~94%).
Limitations
- Focus limited to molecular graphs; generalization to other graph types (e.g., social networks) not verified.
- Does not incorporate 3D information.
- Still lags molecular foundation models such as MolT5 on molecule description generation.
Experimental setup
Datasets: PubChem HiPubChem-295k MoleculeNet ChEBI-20 Mol-Instructions USPTO RDKit functional groups
Benchmarks: HBench (motif hallucination) MoleculeNet classification (BACE, BBBP, HIV, SIDER, ClinTox, MUV, Tox21) CYP450 QM9 regression ChEBI-20 molecule description USPTO reagent/forward/retrosynthesis
Metrics: macro F1 F1 (pos/neg) accuracy ROC-AUC MAE BLEU ROUGE METEOR exact match Levenshtein
Baselines: InstructMol-G GIMLET Galactica-6.7B Mol-Instruction MolT5 MolFM Uni-Mol MoMu
Builds on: InstructMol Mole-BERT HiMol BRICS GIN
Closest papers by meaning
- Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.92)
- LLaMo: Large Language Model-based Molecular Graph Assistant (NeurIPS 2024 · similarity 0.90)
- Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation (ICLR 2025 · similarity 0.90)
- Improving Large Molecular Language Model via Relation-aware Multimodal Collaboration (AAAI 2026 · similarity 0.89)
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.88)
- 3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling (ICLR 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.