Research map / Label-efficient and robust learning / Recognition and long-tail
Instructor-inspired Machine Learning for Robust Molecular Property Prediction
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 InstructMol, a semi-supervised proxy-labeling method where an instructor model predicts pseudo-label reliability to reweight the target molecular model's loss, improving molecular property prediction under scarce labels and OOD shift.
Problem
Molecular property prediction suffers from scarce labeled data and distribution shift; existing proxy-labeling methods rely on classification probabilities and fail for regression, and pretraining benefits can vanish under distribution gaps.
Approach
An instructor network g judges whether each (input, label) pair is a true label or a pseudo-label, and its confidence p_j reweights the target model's loss via (2p_j - 1) weighting over all unlabeled data, enabling regression-compatible proxy labeling without multi-domain knowledge transfer.
Key results
- InstructMol on GIN improves ROC-AUC by an average of 8.6% across six classification MoleculeNet tasks and cuts RMSE by 9.3% on three regression tasks.
- GEM+InstructMol achieves SOTA on all MoleculeNet tasks with an average 4.35% improvement and about 9.98% RMSE decrease on regression tasks.
- On GOOD-HIV scaffold OOD split, InstructMol reaches 72.10 vs 69.58 for ERM, outperforming all listed OOD baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add InstructMol to GEM | GEM · MoleculeNet BBBP | ROC-AUC: 72.4(0.4) -> 73.3(0.8) | Helped |
| Add InstructMol to GEM | GEM · MoleculeNet FreeSolv | RMSE: 1.877(0.094) -> 1.604(0.043) | Helped |
| Use InstructMol instead of ERM | GOOD-HIV scaffold OOD · GOOD-HIV scaffold OOD | performance: 69.58 -> 72.10 | Helped |
| Use InstructMol instead of ERM | GOOD-PCBA size OOD · GOOD-PCBA size OOD | performance: 17.86 -> 19.20 | Helped |
| Add unlabeled molecular data (ablation) | InstructMol · MoleculeNet downstream tasks | performance | Helped |
| Use InstructMol rather than supervised-only prediction | LogP prediction · ZINC15 | RMSE: 5.63 -> 0.63 | Helped |
| Use InstructMol for drug Ki prediction | InstructMol · ZA202303678A 5-HT1A compounds | Ki prediction error | Helped |
Surprising findings
- Pseudo-label-driven training on ZINC15 reduces LogP prediction RMSE from 5.63 to 0.63 even with only 14 labeled molecules, per the authors.
Limitations
- Combination with self-supervised pretraining relies on existing methods (GEM, Uni-Mol, GROVER) rather than a tailored SSL approach.
- Code release is deferred until acceptance; reproducibility depends on appendix details.
- Real-world patent validation is limited to 9 compounds for one target (5-HT1A).
Experimental setup
Datasets: MoleculeNet ZINC15 GOOD-HIV GOOD-PCBA CHEMBL214_Ki Kaggle LogP dataset
Benchmarks: MoleculeNet GOOD-HIV GOOD-PCBA
Metrics: ROC-AUC RMSE
Baselines: GIN GAT GCN Semi-GAN pi-model UPS GROVER Uni-Mol GEM MolCLR GraphMVP IRM
Builds on: UPS pi-model Mean Teacher GEM Semi-GAN
Closest papers by meaning
- TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence (NeurIPS 2025 · similarity 0.86)
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.85)
- Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models (ICLR 2024 · similarity 0.84)
- Pin-Tuning: Parameter-Efficient In-Context Tuning for Few-Shot Molecular Property Prediction (NeurIPS 2024 · similarity 0.84)
- From Molecules to Materials: Pre-training Large Generalizable Models for Atomic Property Prediction (ICLR 2024 · similarity 0.84)
- BOOM: Benchmarking Out-Of-distribution Molecular Property Predictions of Machine Learning Models (NeurIPS 2025 · similarity 0.84)
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.