Research map / Large language models / Retrieval and knowledge
Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Model
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 Mol-LLaMA, a molecular LLM trained on a 284k-sample GPT-4o-generated instruction dataset (structural descriptions, structure-to-feature explanations, multi-turn conversations) with a 2D-3D blending module feeding a Q-Former into Llama-2-7B-chat or Llama-3.1-8B-Instruct.
Problem
Existing molecular LLMs lack general molecular knowledge and explainable reasoning about structures and properties, and rely on single-type molecular encoders that misinterpret structures.
Approach
Builds a reasoning-oriented instruction dataset from IUPAC names and PubChem descriptions via GPT-4o across three data types, and blends 2D (MoleculeSTM) and 3D (Uni-Mol) encoder features with a cross-attention blending module before a Q-Former projector.
Key results
- Mol-LLaMA (Llama3.1-8B) reaches overall relative score 1.125 (structural), 1.251 (chemical), 1.744 (biological) against GPT-4o as reference, above all baselines.
- On zero-shot PAMPA with Llama3.1-8B, Mol-LLaMA attains 63.55% accuracy (default) versus GPT-4o 48.65%, and 72.48% with task info.
- On MoleculeQA, Mol-LLaMA on Llama-3.1-8B scores 70.76 total, above 3D-MoLM (69.10) and Mol-Instructions (68.45).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use comprehensive conversations only instead of the full instruction dataset (ablation) | Mol-LLaMA · Molecular understanding (chemical) | overall relative score: 1.251 -> 0.689 | Hurt |
| Use comprehensive conversations only instead of the full instruction dataset (ablation) | Mol-LLaMA · PAMPA (w/ Task Info.) | accuracy: 72.48 -> 83.05 | Helped |
| Use detailed structural descriptions only instead of the full instruction dataset (ablation) | Mol-LLaMA · PAMPA (default) | accuracy: 63.55 -> 37.35 | Hurt |
| Use structural descriptions and structure-to-feature explanations only instead of the full instruction dataset (ablation) | Mol-LLaMA · Molecular understanding (biological) | overall relative score: 1.744 -> 1.754 | Helped |
| Use a single 2D molecular encoder instead of blended 2D+3D representations (ablation) | Mol-LLaMA · Molecular understanding (structural) | overall relative score: 1.125 -> 0.907 | Hurt |
| Use a single 3D molecular encoder instead of blended 2D+3D representations (ablation) | Mol-LLaMA · Molecular understanding (structural) | overall relative score: 1.125 -> 1.071 | Hurt |
| Concatenate 2D and 3D representations without the blending module (ablation) | Mol-LLaMA · Molecular understanding (structural) | overall relative score: 1.125 -> 1.037 | Hurt |
| Use CoT prompting instead of default prompting | Mol-LLaMA with Llama3.1-8B-Instruct · PAMPA | accuracy: 63.55 -> 64.37 | Helped |
| Provide task-specific information instead of default prompting | Mol-LLaMA with Llama3.1-8B-Instruct · PAMPA | accuracy: 63.55 -> 72.48 | Helped |
| Increase in-context examples from 0-shot to 5-shot | Mol-LLaMA · PAMPA | accuracy: 63.55 -> 80.10 | Helped |
| Use diverse conformations rather than a fixed conformation | Mol-LLaMA · Molecular understanding (biological) | overall relative score: 1.744 -> 1.841 | No clear effect |
Surprising findings
- Concatenating 2D and 3D embeddings without the blending module degrades structural understanding relative to single encoders.
- Baseline molecular LLMs such as LLaMo and 3D-MoLM lose helpful explanation ability in zero-shot PAMPA, with fidelity and helpfulness below 0.3 for some models.
Limitations
- Relies on GPT-4o-generated instruction data, whose factual accuracy is filtered only by GPT-4o judgment.
- Evaluation of response quality uses GPT-4o as judge with relative scores against GPT-4o itself.
- Gains on BBBP are limited; GPT-4o outperforms all baselines on BBBP.
Experimental setup
Datasets: Mol-LLaMA-Instruct (284k, from PubChem 324k) PubChem MoleculeQA Mol-Instructions PAMPA BBBP
Benchmarks: PAMPA BBBP MoleculeQA Mol-Instructions 3D-MoIT
Metrics: relative score (helpfulness, relevance, accuracy, details, o accuracy fidelity helpfulness few-shot accuracy
Baselines: GPT-4o Llama2-7B-Chat Llama3.1-8B Mol-Instructions LlasMol 3D-MoLM LLaMo BioMedGPT-LM MolT5 BioT5 MolCA
Builds on: MolCA 3D-MoLM LLaMo Q-Former MoleculeSTM
Closest papers by meaning
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.91)
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.91)
- Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models (ICLR 2024 · similarity 0.90)
- Can LLMs Solve Molecule Puzzles? A Multimodal Benchmark for Molecular Structure Elucidation (NeurIPS 2024 · similarity 0.89)
- MolTailor: Tailoring Chemical Molecular Representation to Specific Tasks via Text Prompts (AAAI 2024 · similarity 0.89)
- Towards 3D Molecule-Text Interpretation in Language Models (ICLR 2024 · 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.