Research map / 3D vision / 3D generation
UniCorn: A Unified Contrastive Learning Approach for Multi-view 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.
Unifies 2D graph masking, 2D-3D contrastive learning, and 3D denoising as contrastive clustering of molecular views, and proposes UniCorn, a pre-training framework combining fragment masking, torsion-augmented denoising, and cross-modal 2D-to-3D distillation.
Problem
No universal molecular pre-training model works well across quantum, physicochemical, and biological property prediction, and the relationships among existing SSL methods (2D masking, 2D-3D contrastive learning, 3D denoising) are poorly understood.
Approach
Views the three SSL families as contrastive learning with different augmentations that cluster molecular views at different granularity, then combines fragment masking (BRICS), torsion-augmented 3D denoising, and cross-modal contrastive distillation from 2D graph to 3D conformers into one 3D-encoder pre-training objective.
Key results
- On QM9, UniCorn achieves best MAE on 10 of 12 tasks (e.g., mu 0.0085 D, U0 3.99 meV), outperforming denoising baselines.
- On MoleculeNet biological classification, UniCorn averages 78.4 ROC-AUC vs 76.2 for the second-best MoleBLEND, best on 7 of 8 tasks.
- UniCorn sets new state-of-the-art on 6 of 8 MD17 molecules and all 7 MD22 tasks; overall best on 33 of 38 tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine fragment masking, torsion-augmented denoising, and cross-modal distillation | UniCorn · MoleculeNet biological classification | ROC-AUC average: 76.2 -> 78.4 | Helped |
| Remove the denoising loss (ablation) | UniCorn · QM9 | εHOMO (meV) MAE: 13.0 -> 16.8 | Hurt |
| Add the denoising loss (ablation) | UniCorn · QM9 | εHOMO (meV) MAE: 16.8 -> 13.0 | Helped |
| Remove fragment masking and cross-modal distillation (ablation) | UniCorn · MoleculeNet BBBP | ROC-AUC: 74.2 (1.1) -> 69.3 (0.9) | Hurt |
| Add fragment masking and cross-modal distillation (ablation) | UniCorn · MoleculeNet BBBP | ROC-AUC: 69.3 (0.9) -> 74.2 (1.1) | Helped |
| Remove fragment masking and cross-modal distillation (ablation) | UniCorn · MoleculeNet ESOL | RMSE: 0.817(0.034) -> 0.825(0.036) | Hurt |
| Add RDKit pretraining data to DFT data (ablation) | UniCorn · QM9 | average MAE: 16.7 -> 16.6 | Helped |
Surprising findings
- Pre-training with DFT-only conformers achieves QM9 performance comparable to using additional RDKit conformers, suggesting denoising prefers accurate conformations.
- Ablation shows removing the denoising loss degrades QM9 performance notably, e.g., Delta-epsilon from 24.9 to 31.0 meV.
Limitations
- Unifying theory relies on conditions (non-zero eigenvalue bounds, full column rank decoder) that are only discussed, not verified in practice.
- The hierarchical clustering claims are supported mainly by a self-constructed dataset and qualitative t-SNE visualization.
- Some baseline numbers are taken from prior papers and MD17 SE(3)-DDM uses a different Benzene dataset version.
Experimental setup
Datasets: 15M molecules pre-training corpus (PubChemQC, Uni-Mol/RDKit conformers) QM9 MD17 MD22 MoleculeNet (BBBP, Tox21, MUV, BACE, ToxCast, SIDER, ClinTox, HIV, ESOL, FreeSolv, Lipophilicity)
Benchmarks: QM9 MD17 MD22 MoleculeNet
Metrics: MAE ROC-AUC RMSE Davies-Bouldin Index
Baselines: 3D InfoMax GraphMVP MoleculeSDE MoleBLEND MoleculeJAE Transformer-M SE(3)-DDM 3D-EMGP Coord Frad Coord(NonEq) AttrMask
Builds on: TorchMD-NET Graphormer BRICS SimSiam Noisy Node
Closest papers by meaning
- A probability contrastive learning framework for 3D molecular representation learning (NeurIPS 2024 · similarity 0.88)
- MORE: Molecule Pretraining with Multi-Level Pretext Task (AAAI 2025 · similarity 0.88)
- Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation (ICLR 2025 · similarity 0.87)
- Multimodal Molecular Pretraining via Modality Blending (ICLR 2024 · similarity 0.87)
- 3D Denoisers Are Good 2D Teachers: Molecular Pretraining via Denoising and Cross-Modal Distillation (AAAI 2025 · similarity 0.86)
- TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence (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.