Research map / Graph learning / Graph self-supervised learning
MOL-Mamba: Enhancing Molecular Representation with Structural & Electronic Insights
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-Mamba, a self-supervised molecular representation framework that fuses atom/fragment graph Mamba reasoning with electronic descriptor masked modeling via a Mamba-Transformer fuser.
Problem
Molecular representation learning often ignores the relationship between molecular structure and electronic information and the internal semantic reasoning within molecules, yielding incomplete representations for property prediction.
Approach
Combines a graph-augmented Mamba (GraphSSM with adjacency/distance-modulated state updates, fragment-aware node sorting, and positional encoding) over atom and fragment graphs with a Mamba-Transformer fuser that integrates masked electronic descriptor tokens, trained with structural distribution and e-semantic masked losses.
Key results
- MOL-Mamba reaches 75.0% ROC-AUC on BBBP and 81.3% on Tox21, outperforming the listed pretraining baselines on those datasets.
- On ESOL, MOL-Mamba achieves RMSE 0.63 versus 0.79 for Uni-Mol, and on Lipo RMSE 0.53 versus 0.60.
- Ablation shows each pretraining loss adds gains, with the full set improving BBBP from 72.5 to 75.0 over no pretraining.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MOL-Mamba instead of the strongest listed baseline | molecular property prediction · BBBP | ROC-AUC (%): 84.8 (±2.2) -> 75.0 (±0.2) | Hurt |
| Use MOL-Mamba instead of the strongest listed baseline | molecular property prediction · Tox21 | ROC-AUC (%): 79.8 (±0.7) -> 81.3 (±0.4) | Helped |
| Use MOL-Mamba instead of the strongest listed baseline | molecular property prediction · ClinTox | ROC-AUC (%): 93.2 (±1.7) -> 92.7 (±1.1) | Hurt |
| Use MOL-Mamba instead of the strongest listed baseline | molecular property prediction · HIV | ROC-AUC (%): 80.8 (±0.3) -> 81.6 (±0.5) | Helped |
| Use MOL-Mamba instead of the strongest listed baseline | molecular property prediction · BACE | ROC-AUC (%): 89.0 (±0.3) -> 86.4 (±0.3) | Hurt |
| Use MOL-Mamba instead of the strongest listed baseline | molecular property prediction · SIDER | ROC-AUC (%): 68.0 (±1.1) -> 68.3 (±0.9) | Helped |
| Use MOL-Mamba instead of the strongest listed baseline | molecular property prediction · MUV | ROC-AUC (%): 88.6 (±2.2) -> 89.0 (±1.2) | Helped |
| Use MOL-Mamba instead of the strongest listed baseline | molecular property prediction · ESOL | RMSE: 0.79 (±0.03) -> 0.63 (±0.01) | Helped |
| Remove all self-supervised losses (ablation) | MOL-Mamba · BBBP | ROC-AUC (%): 75.0 (±0.2) -> 72.5 (±0.4) | Hurt |
| Add L_d, L_s, and L_f to the training objectives (ablation) | MOL-Mamba · Tox21 | ROC-AUC (%): 78.6 (±0.6) -> 80.9 (±0.4) | Helped |
| Add L_mask to the training objectives (ablation) | MOL-Mamba · Tox21 | ROC-AUC (%): 80.9 (±0.4) -> 81.3 (±0.4) | Helped |
| Remove graph node sorting strategy (ablation) | MOL-Mamba MG module · BBBP | ROC-AUC (%): 75.0 (±0.2) -> 73.2 (±0.4) | Hurt |
Surprising findings
- Removing the fragment-level loss L_f yields small but consistent gains rather than large changes, and the paper reports L_f weighted at 20.0 versus 0.1 for others, indicating strong scale sensitivity.
Limitations
- Paper claims superiority on 8 of 11 datasets but the tables show MOL-Mamba not best on several (e.g., ClinTox, Lipo comparisons, QM9 where GEM/Uni-Mol are better), and the text contains inconsistent claims such as 78.6% Tox21 supervised-only.
- Results rely on 10-fold cross-validation with reported standard deviations; the paper does not clearly describe hyperparameter tuning of baselines.
- Several architectural descriptions are inconsistent (e.g., MT vs MF fuser naming, GNN_A described as SchNet while also as a GNN layer).
Experimental setup
Datasets: GEOM BBBP Tox21 ClinTox HIV BACE SIDER MUV FreeSolv ESOL Lipo QM9
Benchmarks: BBBP Tox21 ClinTox HIV BACE SIDER MUV FreeSolv ESOL Lipo QM9
Metrics: ROC-AUC RMSE MAE
Baselines: SchNet GIN AttentiveFP DMPNN PretrainGNN GROVER GEM GraphMVP MolCLR Uni-Mol MOLEBLEND
Builds on: Mamba PSM GIN SchNet KCL
Closest papers by meaning
- Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Model (NeurIPS 2025 · similarity 0.88)
- Enhancing Molecular Representation Learning Through the Combination of 3D and 2D Graph Machine Learning (Student Abstract) (AAAI 2025 · similarity 0.87)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.86)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.86)
- MolTailor: Tailoring Chemical Molecular Representation to Specific Tasks via Text Prompts (AAAI 2024 · similarity 0.86)
- Foundation Molecular Grammar: Multi-Modal Foundation Models Induce Interpretable Molecular Graph Languages (ICML 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.