Research map / Large language models / Instruction and fine-tuning
Multimodal Large Language Models for Inverse Molecular Design with Retrosynthetic Planning
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.
Llamole couples a base LLM with a frozen Graph Diffusion Transformer, a GNN reaction-template predictor, and A* retrosynthetic search to interleave text and molecular graph generation for inverse molecular design.
Problem
LLMs struggle to generate molecular structures that meet multi-property and synthesizability requirements because they handle graphs poorly when restricted to text representations such as SMILES, and they cannot plan multi-step retrosynthesis.
Approach
A multimodal LLM that uses trigger and query tokens to switch control to frozen graph modules (Graph DiT for molecule generation, GIN for reaction template prediction), interleaving text, molecules, and reactions, with the LLM also supplying heuristic cost estimates for A* retrosynthetic search.
Key results
- Llamole (Llama-3.1-8B base) reaches 0.913 chemical validity and 35.1% drug retrosynthesis success vs 5.5% for the best ICL baseline (Llama-3-8B).
- Llamole improves property controllability by up to 80.9% over LLM baselines across 12 metrics, outperforming 70B models.
- Material retrosynthesis success rises to 17.9% with Qwen2-7B-based Llamole, versus 4.8% for the best baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Llamole with graph-text multimodal generation instead of the strongest baseline | Molecular design · Multi-conditional molecular design | validity: 0.885 -> 0.913 | Helped |
| Use Llamole with graph-text multimodal generation instead of the strongest baseline | Molecular design · Drug design | BACE BA: 0.562 -> 0.746 | Helped |
| Use Llamole with graph-text multimodal generation instead of the strongest baseline | Molecular design · Material design | FFV MAE: 0.020 -> 0.021 | Hurt |
| Use Llamole instead of the strongest baseline | Molecular design · Multi-conditional molecular design | BLEU-4: 0.226 -> 0.262 | Helped |
| Use Llamole instead of the strongest baseline | Retrosynthesis planning · Drug retrosynthesis | success rate: 5.5 -> 35.1 | Helped |
| Use Llamole instead of the strongest baseline | Retrosynthesis planning · Material retrosynthesis | success rate: 4.8 -> 17.9 | Helped |
| Add text conditions to the Graph DiT input (ablation) | Llamole Graph DiT · Controllable molecular generation | chemical structure understanding | Helped |
| Add property conditions to the Graph DiT input (ablation) | Llamole Graph DiT · Controllable molecular generation | property controllability | Helped |
| Apply text-only SFT instead of ICL | Llama-3-8B · Drug retrosynthesis | success rate: 5.5% to below 1% | Hurt |
| Remove the planning time constraint | Llamole · Material retrosynthesis | success rate: 0.176 -> 0.312; 0.143 -> 0.273; 0.179 -> 0.273 | Helped |
| Replace default LLM heuristics with domain heuristics | Llamole A* search · Material retrosynthesis | success rate: 0.176 -> 0.176; 0.143 -> 0.147; 0.179 -> 0.181 | No clear effect |
Surprising findings
- SFT on Llama-3-8B reduces retrosynthesis success from 5.5% to below 1%.
- Model size correlates weakly with molecular design performance (average Pearson 0.366).
Limitations
- Retrosynthetic planning is limited to 30 seconds or 300 iterations, so success rate trades off against inference time.
- Training data for property labels and descriptions is generated by Llama-3-70B and GNN pseudo-labels, which may be noisy.
Experimental setup
Datasets: MolQA MolPair USPTO PubChem MoleculeNet ChEMBL ZINC PI1M Enamine building blocks
Benchmarks: MolQA
Metrics: validity similarity BLEU-4 ROUGE-L balanced accuracy mean absolute error retrosynthesis success rate
Baselines: GraphGA Llama-2-7B Mistral-7B Qwen2-7B Llama-3-8B Flan-T5-XXL Granite-13B Llama-2-13B Mistral-8x7B Llama-2-70B Llama-3-70B
Builds on: Graph DiT GraphGA CLIP LoRA A* search
Closest papers by meaning
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.89)
- LLaMo: Large Language Model-based Molecular Graph Assistant (NeurIPS 2024 · similarity 0.89)
- Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.88)
- Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.86)
- Improving Large Molecular Language Model via Relation-aware Multimodal Collaboration (AAAI 2026 · similarity 0.86)
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.86)
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.