Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery
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 UniMatch, a dual matching few-shot molecular property predictor that combines hierarchical (atom/substructure/molecule) pooling-attention matching over GIN layers with meta-learning framed as implicit task-level matching.
Problem
Few-shot molecular property prediction in drug discovery, where scarce labels and single-scale molecular features limit generalization across properties that depend on different structural levels.
Approach
Combines explicit hierarchical molecular matching (mean pooling at each GIN layer, attention-based support-query matching per layer, and fusion of layer predictions) with implicit task-level matching via meta-learning, using a task relationship matrix to share task parameters.
Key results
- On MoleculeNet with support size 20, UniMatch reaches 82.62 AUROC on Tox21 and 77.74 on ToxCast; Pre-UniMatch reaches 80.34 on SIDER, outperforming listed baselines on those datasets.
- On FS-Mol, UniMatch improves over all baselines by 4.27%, 8.53%, 4.40%, 6.07%, and 4.26% at support sizes 16, 32, 64, 128, and 256 (in ΔAUPRC).
- On Meta-MolNet, UniMatch performs strongly on GSK3, JNK3, Tox21 and ToxCast but weaker on HIV, PCBA and MUV.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniMatch instead of the compared methods | MoleculeNet · Tox21, SIDER, ToxCast | AUROC: 2.87% | Helped |
| Use UniMatch instead of the compared methods | FS-Mol · support set size 16 | ΔAUPRC: 4.27% | Helped |
| Use UniMatch instead of the compared methods | FS-Mol · support set size 32 | ΔAUPRC: 8.53% | Helped |
| Use UniMatch instead of the compared methods | FS-Mol · support set size 64 | ΔAUPRC: 4.40% | Helped |
| Use UniMatch instead of the compared methods | FS-Mol · support set size 128 | ΔAUPRC: 6.07% | Helped |
| Use UniMatch instead of the compared methods | FS-Mol · support set size 256 | ΔAUPRC: 4.26% | Helped |
| Add pretrained parameters | UniMatch · MoleculeNet | AUROC | Helped |
| Add explicit hierarchical molecular matching and implicit task-level matching (ablation) | common GNN backbones on FS-Mol · FS-Mol | ΔAUPRC | Helped |
| Use UniMatch on Meta-MolNet | few-shot molecular property prediction · GSK3, JNK3, Tox21, ToxCast | AUROC | Mixed |
Surprising findings
- The paper reports UniMatch ranks only second on MUV, which it attributes to distribution imbalance, and it underperforms ADKF-IFT there (79.40 vs 98.18).
Limitations
- Performs poorly on the MUV dataset, likely due to distribution bias, and on HIV and PCBA in Meta-MolNet.
- Implicit task-level matching is described informally; the formulation of the task relationship matrix and its update is loosely specified.
Experimental setup
Datasets: MoleculeNet (Tox21, SIDER, MUV, ToxCast) FS-Mol Meta-MolNet (GSK3, JNK3, HIV, Tox21, ToxCast, PCBA, MUV)
Benchmarks: MoleculeNet FS-Mol Meta-MolNet
Metrics: AUROC ΔAUPRC AUPRC
Baselines: CHEF MixHop Siamese ProtoNet MAML TPN EGNN IterRefLSTM PAR ADKF-IFT MHNfs Pre-GNN
Builds on: Matching Networks MAML GIN PAR ADKF-IFT
Closest papers by meaning
- Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation (ICLR 2025 · similarity 0.87)
- M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.87)
- UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection (NeurIPS 2025 · similarity 0.87)
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.86)
- Learning Molecular Representation in a Cell (ICLR 2025 · similarity 0.86)
- Unified all-atom molecule generation with neural fields (NeurIPS 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.