Research map / Large language models / Language, safety and interpretability
Bridging Molecular Graphs and Large Language Models
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.
Graph2Token aligns a molecular graph encoder's output to LLM vocabulary embeddings via cross-attention retrieval, training only a graph tokenizer and task head on a frozen Llama3-8B for few-shot molecular classification and regression.
Problem
LLMs handle molecular graph data poorly; existing graph-language tuning methods fine-tune projectors and LLM backbones with scarce molecule-text data, which harms generalization and few-shot ability.
Approach
Represent a graph as a special token whose embedding is produced by cross multi-head attention between the graph feature (query) and a compressed set of pretrained LLM token embeddings (keys/values), so the graph token is expressed in the LLM's own vocabulary space while the LLM backbone stays frozen; IUPAC names are added to prompts.
Key results
- On 10% few-shot molecular classification, Graph2Token averages 68.7 ROC-AUC on TOX21/BBBP/BACE/HIV settings, about 3.7% above the best MolCA result (MolCA-GS 67.2).
- On few-shot PubchemQC-IUPAC regression (10%), Graph2Token reaches 0.324 eV average MAE versus 0.346 for MolCA-GS.
- Trainable parameters are 8.60M, about 0.11% of the full framework, versus 109.09M for MolCA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Graph2Token instead of MolCA-GS | Molecular classification transfer from BACE, BBBP, HIV to TOX21 · 5% few-shot | ROC-AUC: 63.6 -> 68.7 | Helped |
| use Graph2Token instead of MolCA-GS | Molecular classification transfer from TOX21 to BACE, BBBP, HIV · 5% few-shot | ROC-AUC: 65.5 -> 65.5 | No clear effect |
| use Graph2Token instead of MolCA-GS | Molecular classification transfer from BACE, BBBP, HIV to TOX21 · 10% few-shot | ROC-AUC: 67.2 -> 72.1 | Helped |
| use Graph2Token instead of MolCA-GS | Molecular classification transfer from TOX21 to BACE, BBBP, HIV · 10% few-shot | ROC-AUC: 68.5 -> 68.7 | Helped |
| use Graph2Token instead of MolCA-GS | PubchemQC-IUPAC regression · 5% few-shot | Avg MAE (eV): 0.426 -> 0.392 | Helped |
| use Graph2Token instead of MolCA-GS | PubchemQC-IUPAC regression · 10% few-shot | Avg MAE (eV): 0.346 -> 0.324 | Helped |
| add IUPAC name to the prompt (ablation) | Graph2Token · BBBP and TOX21 classification, 5% few-shot | Avg ROC-AUC: 63.5 -> 64.9 | Helped |
| add cross-association and retrieval alignment instead of linear mapping (ablation) | Graph2Token · BBBP and TOX21 classification, 5% few-shot | Avg ROC-AUC: 62.6 -> 64.9 | Helped |
| pretrain the graph encoder on molecular-text data instead of random initialization (ablation) | Graph2Token · BBBP and TOX21 classification, 5% few-shot | Avg ROC-AUC: 63.6 -> 64.9 | Helped |
| add the LLM to the pretrained GNN and linear task layer (ablation) | Graph2Token · BBBP and TOX21 classification, 5% few-shot | Avg ROC-AUC: 61.2 -> 64.9 | Helped |
| add IUPAC name to the prompt (ablation) | Graph2Token · BBBP and TOX21 classification, 10% few-shot | Avg ROC-AUC: 67.3 -> 68.7 | Helped |
| add cross-association and retrieval alignment instead of linear mapping (ablation) | Graph2Token · BBBP and TOX21 classification, 10% few-shot | Avg ROC-AUC: 65.8 -> 68.7 | Helped |
Surprising findings
- Removing the LLM entirely (GNN plus linear head) causes a sharp drop, yet the paper frames this as evidence LLM prior knowledge matters even with a frozen backbone.
- Llama3-Chat-8B with LoRA fine-tuning on the full QM9 data performs much worse than the graph-token approach on regression, which the authors attribute to LLM hallucination under label shift.
Limitations
- Few-shot evaluation covers only a limited set of molecular tasks (four classification datasets and one regression dataset).
- Regression relies on the LLM generating text that is converted to a value via a placeholder, which may limit numeric precision.
Experimental setup
Datasets: CHEBI HMDB PubChem MoleculeNet BBBP MoleculeNet BACE MoleculeNet HIV MoleculeNet TOX21 QM9 PubchemQC GIMLET Mol-Instructions
Benchmarks: BBBP BACE HIV TOX21 QM9 PubchemQC-IUPAC
Metrics: ROC-AUC MAE
Baselines: Galactica-6.7B Galactica-120B Vicuna-v1.5-13B-16k Llama-2-7B-chat Vicuna-v1.3-7B Llama3-Chat-8B Vicuna-7B Mol-Instructions MolCA-S MolCA-GS GPF-AttrMasking GPF-GCL
Builds on: CLIP GIN MolCA GPF Mol-Instructions
Closest papers by meaning
- LLaMo: Large Language Model-based Molecular Graph Assistant (NeurIPS 2024 · similarity 0.92)
- Hierarchical Graph Tokenization for Molecule-Language Alignment (ICML 2025 · similarity 0.92)
- Improving Large Molecular Language Model via Relation-aware Multimodal Collaboration (AAAI 2026 · similarity 0.88)
- Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets (ICLR 2024 · similarity 0.88)
- Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation (ICLR 2025 · similarity 0.87)
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · 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.