Research map / Large language models / Instruction and fine-tuning
Omni-Mol: Multitask Molecular Model for Any-to-any Modalities
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.
Omni-Mol builds a 1.4M-sample molecular instruction dataset across 16 small-molecule tasks and trains a LLaMA-3.2-1B-based multimodal LLM with Gradient Adaptive LoRA (GAL) and a Mixture-of-GAL-Experts (MoGE) to unify understanding and generation.
Problem
Building a single general-purpose molecular multimodal LLM is hard due to small, narrow task datasets, distributional shifts and competition among molecular tasks, and differing intrinsic dimensionality of tasks and representations.
Approach
Categorizes small-molecule tasks into Mol2Mol, Mol2Text, Mol2Num, and Text2Mol; introduces GAL, a LoRA variant with a learnable scaling factor gamma = alpha/r^p + beta to adapt to task intrinsic dimension; and combines GAL adapters with a shared-plus-routed MoE (MoGE) upcycled from the FFN layers.
Key results
- On forward reaction prediction Omni-Mol reaches 0.73 exact match vs 0.69 for re-implemented PRESTO and 0.54 for InstructMol, with 2.2B parameters.
- Omni-Mol attains Description QA BLEU-4/ROUGE-2 gains over prior baselines and state-of-the-art on 13 of 16 tasks as claimed by the authors.
- Omni-Mol is reported to improve Mol2Num R^2 by 21.4% and reduce MAE by 25.1%, 25.8%, 39.3% on Weight, LogP, and TPSA regression.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Omni-Mol instead of PRESTO* | forward reaction prediction · Forward Reaction Prediction | Exa: 0.69 -> 0.73 | Helped |
| Use Omni-Mol instead of PRESTO* | retrosynthesis · Retrosynthesis | Exa: 0.53 -> 0.57 | Helped |
| Use Omni-Mol instead of PRESTO* | reagent prediction · Reagent Prediction | Exa: 0.21 -> 0.23 | Helped |
| Use Omni-Mol instead of PRESTO | catalyst prediction · Catalyst Prediction | Exa: 0.77 -> 0.72 | Hurt |
| Use Omni-Mol instead of InstructMol | quantum mechanics property prediction · Quantum Mechanics Property Prediction | Avg.: 0.0050 -> 0.0044 | Helped |
| Use Omni-Mol instead of GPT-4-0314 | molecular captioning · Molecular Captioning | B-4: 0.525 -> 0.440 | Hurt |
| Use Omni-Mol instead of DeepSeekV3 | description Q&A · Description Q&A | B-4: 0.31 -> 0.44 | Helped |
| Use Omni-Mol instead of PRESTO | yield prediction · Yield Prediction | B-H: 0.94 -> 0.94 | No clear effect |
| Use Omni-Mol instead of Llama3.2-1B* | IUPAC-to-SELFIES generation · IUPAC2SELFIES | Exa: 0.31 -> 0.39 | Helped |
| Use Omni-Mol instead of Llama3.2-1B* | text-guided molecule generation · Text Guided Molecule Generation | Exa: 0.04 -> 0.12 | Helped |
| Add MoGE to joint training (ablation) | InstructMol on the Omni-Mol dataset · Reagent Prediction, Molecular Captioning, Yield Prediction, MolDesign | average performance | Helped |
| Use joint rather than separate training (ablation) | InstructMol on the Omni-Mol dataset · Mol2Mol, Mol2Text, Mol2Num, Text2Mol tasks | performance | Helped |
Surprising findings
- InstructMol's representations become less similar as more tasks are added, whereas Omni-Mol's become more similar, interpreted as convergence toward a universal representation.
- Catalyst prediction: the PRESTO baseline (6.7B) slightly beats Omni-Mol on exact match (0.77 vs 0.72), a case where the model is not best.
Limitations
- Computational resources limited scaling exploration beyond 8B backbone.
- Tasks focus on small molecules; proteins and protein-small molecule interactions are not covered.
Experimental setup
Datasets: PubChem USPTO USPTO_500_MT USPTO_500K Mol-Instruction 3D-MOIT
Benchmarks: Omni-Mol dataset tasks (forward reaction, retrosynthesis, reagent, catalyst, solvent prediction, QM property, molecular
Metrics: exact match BLEU Levenshtein RDK similarity MACCS similarity Morgan similarity validity MAE R^2 ROUGE
Baselines: InstructMol HIGHT PRESTO 3D-MoLM Mol-Instruction DeepSeekV3 Llama2 Vicuna v1.5 T5Chem nach0-base GPT-4-0314 BioMedGPT
Builds on: LoRA InstructMol PRESTO LLaMA 3 MoleculeSTM
Closest papers by meaning
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.91)
- Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models (ICLR 2024 · similarity 0.91)
- Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Model (NeurIPS 2025 · similarity 0.91)
- Improving Large Molecular Language Model via Relation-aware Multimodal Collaboration (AAAI 2026 · similarity 0.89)
- MolTailor: Tailoring Chemical Molecular Representation to Specific Tasks via Text Prompts (AAAI 2024 · similarity 0.89)
- ChemVLM: Exploring the Power of Multimodal Large Language Models in Chemistry Area (AAAI 2025 · similarity 0.89)
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.