Research map / Large language models / Instruction and fine-tuning
LLaMo: Large Language Model-based Molecular Graph Assistant
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 LLaMo, an end-to-end trained molecular graph-language assistant with a multi-level graph projector and GPT-4-generated molecular graph instruction data, evaluated on description generation, IUPAC naming, and property prediction.
Problem
Integrating molecular graphs with LLMs for instruction-following molecular understanding is underexplored; GNN-based projectors lose local/multi-scale information due to over-smoothing and instruction data for molecular graphs is scarce.
Approach
A multi-level graph projector that uses cross-attention with learnable tokens over every GNN layer's node representations plus functional-group (motif) representations, combined with GPT-4-generated multi-turn molecular graph-text instruction conversations and two-stage training (alignment then LoRA instruction tuning).
Key results
- Generalist LLaMo reaches 38.9 BLEU and 67.1 METEOR on molecule description, versus 27.0 BLEU and 52.2 METEOR for GPT-4 (ICL), and MAE 0.006 on property prediction.
- Specialist LLaMo (Galactica-1.3B) achieves 34.4 BLEU on PubChem324kV2 vs 30.3 for MolCA, and 73.4 METEOR on IUPAC prediction vs 72.1 for MolCA.
- Multi-level graph projector (MGProj) improves over MLP-with-concatenation on IUPAC prediction (49.6 vs 47.1 BLEU) and description generation (37.8 vs 34.8 BLEU).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LLaMo instead of GPT-3.5 (ICL) | Generalist molecule description generation · Molecule description generation | BLEU: 28.4 -> 38.9 | Helped |
| use LLaMo instead of GPT-4 (ICL) | Generalist IUPAC prediction · IUPAC prediction | METEOR: 62.4 -> 73.2 | Helped |
| use LLaMo instead of Mol-Instructions | Generalist property prediction · Property prediction | MAE: 0.013 -> 0.006 | Helped |
| use LLaMo instead of MolCA with Galactica-1.3B | Specialist molecule description generation · PubChem324kV2 | BLEU: 30.3 -> 34.4 | Helped |
| add motif tokens to MGProj (ablation) | MGProj · Molecule description (projector analysis) | BLEU: 36.1 -> 37.8 | Helped |
| add motif tokens to MGProj (ablation) | MGProj · IUPAC prediction (projector analysis) | BLEU: 48.8 -> 49.6 | Helped |
| add GPT-generated instruction data (ablation) | LLaMo with Stage 1 + Stage 2 training · Molecule description | BLEU: 37.2 -> 37.8 | Helped |
| add GPT-generated instruction data (ablation) | LLaMo with Stage 1 + Stage 2 training · IUPAC prediction | BLEU: 47.5 -> 49.6 | Helped |
| replace multi-task learning with instruction tuning (ablation) | LLaMo training · Molecule description | BLEU: 36.9 -> 37.8 | Helped |
| add graph input (ablation) | LLaMo · Molecule description generation | description quality | Helped |
| use MGProj instead of no MGProj (ablation) | LLaMo · Molecule description generation | description accuracy | Helped |
Surprising findings
- Without instruction tuning in stage 2, the stage-1-only model scores only 7.3 BLEU on IUPAC prediction, showing the second stage is critical for that task.
- Multi-task training with task-name identifiers yields much worse property prediction MAE (0.218) than instruction tuning (0.007).
Limitations
- Possible data leakage between LLM pretraining corpora and molecule benchmark test sets.
- LLMs are memory and compute intensive and prone to hallucination, which LLaMo inherits.
- Evaluation is on benchmark datasets rather than real-world chemical applications.
Experimental setup
Datasets: PubChem324k PubChem324kV2 Mol-Instruction ChEBI-20 QM9 (MoleculeNet) USPTO USPTO_500MT GPT-4 generated molecular graph instruction data
Benchmarks: Mol-Instruction molecule description PubChem324kV2 ChEBI-20 IUPAC prediction (PubChem324k) QM9 property QA Mol-Instruction forward reaction Mol-Instruction retrosynthesis
Metrics: BLEU-4 METEOR MAE Exact match Levenshtein RDK FTS MACCS FTS Morgan FTS Validity
Baselines: GPT-3.5 GPT-4 Galactica LLaMA2-7B Text+Chem T5 Mol-Instructions MolT5 MoMu MolCA InstructMol
Builds on: Mol-Instructions MolCA Qwen-VL GIN MoleculeSTM
Closest papers by meaning
- Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.92)
- Hierarchical Graph Tokenization for Molecule-Language Alignment (ICML 2025 · similarity 0.90)
- Improving Large Molecular Language Model via Relation-aware Multimodal Collaboration (AAAI 2026 · similarity 0.90)
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.89)
- Multimodal Large Language Models for Inverse Molecular Design with Retrosynthetic Planning (ICLR 2025 · similarity 0.89)
- Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Model (NeurIPS 2025 · similarity 0.88)
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.