Research map / Large language models / Instruction and fine-tuning
Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for 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.
Introduces Mol-Instructions, a 2M-instruction biomolecular dataset (molecule, protein, and biomolecular text tasks) and instruction-tunes LLaMA-7B variants to show gains on biomolecular tasks.
Problem
LLMs lack specialized instruction data for biomolecular domains (small molecules, proteins, biomedical text), and no standardized instruction dataset exists for this field.
Approach
A large, licensed instruction dataset built by combining GPT-3.5-generated task descriptions (human-reviewed), database mining (PubChem, USPTO, UniProtKB, PubMed), and template-based conversion of protein annotations into textual design instructions, with SELFIES for molecules and MMseqs-based redundancy filtering.
Key results
- On molecular property prediction, the tuned LLaMA reports MAE 0.013, versus 0.568 for Galactica and 5.553 for the base LLaMA.
- On description-guided molecule design, the tuned model reaches validity 1.000 but lower exact match (0.002) and BLEU (0.345) than MolT5 (BLEU 0.546).
- The paper reports that the tuned LLM outperforms baselines on biomolecular text NLP tasks but lags specialized small models on molecule generation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Instruction-tune LLaMA-7B with Mol-Instructions | LLaMA-7B · Molecular property prediction | MAE: 0.568 -> 0.013 | Helped |
| Instruction-tune LLaMA-7B with Mol-Instructions | LLaMA-7B · Description-guided molecule design | Validity: 0.992 -> 1.000 | Helped |
| Instruction-tune LLaMA-7B with Mol-Instructions | LLaMA-7B · Reagent prediction | BLEU: 0.225 -> 0.224 | No clear effect |
| Instruction-tune LLaMA-7B with Mol-Instructions | LLaMA-7B · Reagent prediction | Validity: 0.995 -> 1.000 | Helped |
| Instruction-tune LLaMA-7B with Mol-Instructions | LLaMA-7B · Forward reaction prediction | BLEU: 0.782 -> 0.654 | Hurt |
| Instruction-tune LLaMA-7B with Mol-Instructions | LLaMA-7B · Forward reaction prediction | Validity: 0.946 -> 1.000 | Helped |
| Instruction-tune LLaMA-7B with Mol-Instructions | LLaMA-7B · Retrosynthesis | BLEU: 0.765 -> 0.705 | Hurt |
| Instruction-tune LLaMA-7B with Mol-Instructions | LLaMA-7B · Retrosynthesis | Validity: 0.986 -> 1.000 | Helped |
Surprising findings
- On molecule generation, the tuned LLM trails specialized small models such as MolT5 and Text+Chem T5 on exact match and similarity metrics, despite reaching perfect validity.
Limitations
- LLMs still lag specialized small models on molecule generation tasks.
- Current LoRA-based tuning with a text vocabulary may not fully capture biomolecular language.
- Evaluation relies on metrics commonly used in broader domains, which capture only part of output quality; no expert validation.
Experimental setup
Datasets: PubChem USPTO USPTO_500MT UniProtKB/Swiss-Prot UniProtKB/TrEMBL PubMed abstracts QM9 (MoleculeNet) BC5CDR ChemProt BC4CHEMD PubMedQA MedMCQA
Benchmarks: Mol-Instructions test split
Metrics: MAE BLEU Exact match Levenshtein RDK FTS MACCS FTS Morgan FTS Validity
Baselines: Alpaca Baize-7B ChatGLM-6B Vicuna Galactica Text+Chem T5 MolT5 PMC-LLaMA-13B
Builds on: Self-Instruct Alpaca-LoRA LLaMA
Closest papers by meaning
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.91)
- Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Model (NeurIPS 2025 · similarity 0.90)
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.89)
- Towards 3D Molecule-Text Interpretation in Language Models (ICLR 2024 · similarity 0.88)
- 3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling (ICLR 2025 · similarity 0.88)
- Training Compute-Optimal Protein Language Models (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.