Research map / Vision-language and multimodal / Human and embodied modeling
KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge
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 KnowMol-100K, a 100K-molecule multi-level chemical annotation dataset (atoms, functional groups, structure, properties), and KnowMol, a Mol-LLM using SELFIES with a separate vocabulary and a hierarchical atom/functional-group/molecule graph tokenizer.
Problem
Existing molecular LLMs understand molecules poorly because PubChem captions are sparse and coarse, and because SMILES tokens are shared with natural language and graph encoders only yield atom-level tokens.
Approach
Builds multi-level molecule annotations (atoms, BRICS-based functional groups, structure, six physicochemical properties) using PubChem data, RDKit and GPT-4o, and encodes molecules with SELFIES plus a training-free hierarchical graph pooling (atom, functional group, molecule tokens) on top of an InstructMol-style architecture.
Key results
- On ChEBI-20 captioning, KnowMol (finetuned on ChEBI-20 and PubChem pretrain) reaches BLEU-4 0.595 vs UniMoT 0.583.
- On QM9 regression the average MAE is 0.0030 versus 0.0049 for UniMoT.
- On forward reaction prediction, exact match is 0.752 versus 0.611 for UniMoT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KnowMol-100K MRQA + DGMG instead of PubChem Captions (301K) (ablation) | KnowMol · Forward reaction prediction | Exact: 0.509 -> 0.622 | Helped |
| Add structural construction annotations to MRQA training (ablation) | KnowMol · Forward reaction prediction | Exact: 0.618 -> 0.627 | Helped |
| Add physicochemical property annotations to MRQA training (ablation) | KnowMol · Forward reaction prediction | Exact: 0.587 -> 0.627 | Helped |
| Combine MRQA and DGMG training tasks instead of either task alone (ablation) | KnowMol · Forward reaction prediction | Exact: 0.624 -> 0.622 | Helped |
| Add hierarchical graph representation (ablation) | KnowMol with MRQA + DGMG · Forward reaction prediction | Exact: 0.622 -> 0.728 | Helped |
| Add SELFIES tokenization (ablation) | KnowMol with hierarchical graph representation · Forward reaction prediction | Exact: 0.728 -> 0.752 | Helped |
| Fine-tune on PubChem pretrain-set in addition to ChEBI-20 | KnowMol · ChEBI-20 test | BLEU-4: 0.518 -> 0.595 | Helped |
| Use KnowMol rather than UniMoT | Molecule property prediction · MoleculeNet BBBP | ROC-AUC: 72.9 -> 69.2 | Hurt |
| Use KnowMol rather than Uni-Mol | Molecule property prediction · MoleculeNet MUV | ROC-AUC: 82.1 -> 61.5 | Hurt |
| Use KnowMol rather than UniMoT | Molecule generation · Caption-guided molecule generation | Exact: 0.237 -> 0.083 | Hurt |
Limitations
- No error bars reported
- No 3D molecular representation integrated
- Limited safeguards; reliance on GPT-4o generated annotations
Experimental setup
Datasets: PubChem ChEBI-20 MoleculeNet QM9 Mol-Instructions reaction datasets
Benchmarks: ChEBI-20 MoleculeNet QM9 Mol-Instructions generation tasks PubChem forward reaction
Metrics: BLEU ROUGE METEOR ROC-AUC MAE Exact match Levenshtein FTS Validity
Baselines: InstructMol HIGHT UniMoT MolCA Mol-Instructions Text+Chem T5 GIMLET Uni-Mol
Builds on: InstructMol LoRA BRICS SELFIES
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 0.91)
- Improving Large Molecular Language Model via Relation-aware Multimodal Collaboration (AAAI 2026 · similarity 0.89)
- ChemVLM: Exploring the Power of Multimodal Large Language Models in Chemistry Area (AAAI 2025 · similarity 0.89)
- Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models (ICLR 2024 · similarity 0.89)
- MolTailor: Tailoring Chemical Molecular Representation to Specific Tasks via Text Prompts (AAAI 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.