Research map / Label-efficient and robust learning / Self-supervised and contrastive
Association Pattern-enhanced Molecular Representation Learning
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 APMP, a random-walk path-based, confidence-filtered message passing plugin that enhances atom representations of existing molecular foundation models for property prediction and builds a BBBP pattern library.
Problem
Molecular foundation models learn atom representations from local neighborhoods or predefined substructures and overlook fixed association path patterns that determine molecular properties, limiting property prediction.
Approach
An association pattern-aware message passing (APMP) module samples random-walk paths from each atom, filters them by degree-centrality confidence, mines common patterns with Transformer layers, and fuses the result as a dual-branch plugin into pretrained MFMs during finetuning, with a contribution-based scoring to build property-specific pattern libraries.
Key results
- APMP plugin improved 8 MFM variants across 11 MoleculeNet tasks, with gains up to 4.48 ROC-AUC points on classification and 19.20% relative RMSE on FreeSolv.
- On UniMol, APMP raised BBBP ROC-AUC from 89.22 to 93.70.
- A filtering ratio of about 30%-45% of sampled paths gave the most stable performance in ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add APMP to GEM | GEM · ESOL | RMSE: 0.4668±0.1277 -> 0.4606±0.0346 | Helped |
| Add APMP to N-GRAM | N-GRAM · FreeSolv | RMSE: 0.4609±0.1251 -> 0.4524±0.0897 | Helped |
| Add APMP to GEM | GEM · Lipophilicity | RMSE: 0.4685±0.0671 -> 0.4461±0.0579 | Helped |
| Add APMP to GEM | GEM · BBBP | ROC-AUC: 92.99±1.06 -> 93.76±2.65 | Helped |
| Add APMP to GEM | GEM · BACE | ROC-AUC: 88.77±1.38 -> 90.14±0.76 | Helped |
| Add APMP to GEM | GEM · ClinTox | ROC-AUC: 91.38±0.89 -> 93.19±2.28 | Helped |
| Add APMP to GEM | GEM · Tox21 | ROC-AUC: 85.58±0.72 -> 85.74±0.91 | Helped |
| Add APMP to GROVER | GROVER · ToxCast | ROC-AUC: 75.06±1.17 -> 75.51±0.68 | Helped |
| Add APMP to GEM | GEM · SIDER | ROC-AUC: 67.04±0.57 -> 67.45±2.55 | Helped |
| Add APMP to MGSSL | MGSSL · HIV | ROC-AUC: 84.32±0.40 -> 84.61±0.45 | Helped |
| Add APMP to MGSSL | MGSSL · MUV | ROC-AUC: 82.07±0.37 -> 84.65±1.62 | Helped |
| Add APMP features after the UniMol encoder and finetune the combined representation (ablation) | UniMol · BBBP | ROC-AUC: 89.22±1.58 -> 93.70±0.23 | Helped |
Surprising findings
- Gains were smaller on multi-task classification datasets than on single-task ones, which the authors attribute to common patterns interfering across tasks.
Limitations
- Complexity grows with n^2 attention over sampled paths per atom.
- Multi-task classification gains are smaller due to pattern interference across tasks.
- Pattern library validated mechanistically only for BBBP.
Experimental setup
Datasets: ESOL FreeSolv Lipophilicity BBBP BACE ClinTox Tox21 ToxCast SIDER HIV MUV
Benchmarks: ESOL FreeSolv Lipophilicity BBBP BACE ClinTox Tox21 ToxCast SIDER HIV MUV
Metrics: ROC-AUC RMSE
Baselines: N-GRAM GROVER MGSSL GEM GraphMVP MolCLR UniMol
Builds on: GROVER UniMol GEM GraphMVP MolCLR
Closest papers by meaning
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.87)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.86)
- Association Pattern-aware Fusion for Biological Entity Relationship Prediction (NeurIPS 2024 · similarity 0.86)
- DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform (NeurIPS 2025 · similarity 0.85)
- Learning Molecular Representation in a Cell (ICLR 2025 · similarity 0.85)
- MOL-Mamba: Enhancing Molecular Representation with Structural & Electronic Insights (AAAI 2025 · similarity 0.85)
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.