Research map / Label-efficient and robust learning / Semi-supervised learning
Molecular Data Programming: Towards Molecule Pseudo-labeling with Systematic Weak Supervision
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 Molecular Data Programming (MDP), which builds graph-kernel, fingerprint, and topological labeling functions, aggregates them with a learnable label synchronizer into probabilistic labels, and trains a GNN molecule classifier under weak supervision.
Problem
Molecular property classification when labeled molecules are scarce; existing pseudo-labeling methods rely on a single domain knowledge source and struggle with conflicting supervision signals.
Approach
Applies data programming to molecules: multiple graph labeling functions (graph kernels, molecular fingerprints, topological features) are combined by a learnable label synchronizer (confidence vector plus MLP) that models LF confidences and correlations to produce probabilistic pseudo-labels for training a GNN.
Key results
- MDP achieves best test AUC on 7 of 8 MoleculeNet datasets under weak supervision, e.g., BACE 0.632 and Tox21 0.653, versus best baselines around 0.586 and 0.569 respectively.
- Average improvement of 9.5% over baselines on weakly supervised molecule classification.
- Ablation shows confidence learning and correlation estimation are complementary; full model reaches BACE 0.632 vs 0.484 with confidence-only-free clustering-only setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MDP instead of the strongest reported baseline | Molecular classification under weak supervision · BACE, BBBP, SIDER, ClinTox, Tox21, ToxCast, HIV, MUV | AUC: BACE: 0.586±0.016 -> 0.632±0.023; BBBP: 0.552±0.013 -> 0.618±0.011; SIDER: 0.595±0.006 -> 0.613±0.018; ClinTox: 0.545±0. | Mixed |
| Add confidence estimation and correlation estimation, without clustering (ablation) | MDP · BACE | AUC: 0.632±0.023 -> 0.529±0.096 | Hurt |
| Add confidence estimation and clustering, without correlation estimation (ablation) | MDP · BACE | AUC: 0.632±0.023 -> 0.602±0.023 | Hurt |
| Add correlation estimation and clustering, without confidence estimation (ablation) | MDP · BACE | AUC: 0.632±0.023 -> 0.619±0.025 | Hurt |
| Add confidence estimation only (ablation) | MDP · BACE | AUC: 0.632±0.023 -> 0.484±0.038 | Hurt |
| Add correlation estimation only (ablation) | MDP · BACE | AUC: 0.632±0.023 -> 0.546±0.038 | Hurt |
| Add clustering only (ablation) | MDP · BACE | AUC: 0.632±0.023 -> 0.617±0.033 | Hurt |
| Add all three components: confidence estimation, correlation estimation, and clustering (ablation) | MDP · BACE, BBBP, ClinTox, SIDER, Tox21 | AUC | Helped |
Surprising findings
- MDP underperforms all baselines on ToxCast (0.602 vs best 0.617 for edge pred), attributed to many tasks and label sparsity.
Limitations
- Performance lags on ToxCast with many binary tasks and label sparsity
- Simple MLP label synchronizer may struggle when pseudo-labels become complex
- Requires manual labeling of selected prominent examples
Experimental setup
Datasets: BACE BBBP SIDER ClinTox Tox21 ToxCast HIV MUV
Benchmarks: BACE BBBP SIDER ClinTox Tox21 ToxCast HIV MUV
Metrics: test AUC
Baselines: GCN GAT GraphSAGE GIN pseudo-labeling self-training infomax context pred attr mask edge pred GraphLoG perturb edge
Builds on: data programming Snorkel MeTaL GraphLoG pseudo-labeling
Closest papers by meaning
- Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets (ICLR 2024 · similarity 0.88)
- Empowering Active Learning for 3D Molecular Graphs with Geometric Graph Isomorphism (NeurIPS 2024 · similarity 0.86)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.85)
- Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.85)
- From Artificially Real to Real: Leveraging Pseudo Data from Large Language Models for Low-Resource Molecule Discovery (AAAI 2024 · similarity 0.85)
- Learning Molecular Representation in a Cell (ICLR 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.