Research map / Graph learning / GNN architectures
HAGO-Net: Hierarchical Geometric Message Passing for 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 HAGO-Graph, a hierarchical n-body geometric graph with a chirality-aware ROT feature, and HAGO-Net/DHAGO-Net message passing models for invariant molecular property prediction and equivariant force prediction.
Problem
Completely and effectively characterizing and learning 3D molecular geometry for molecular property prediction, chirality identification, and molecular dynamics (force) prediction.
Approach
Model molecular geometry hierarchically as n-body nodes (atoms, pairs, triplets, etc.) with intra-n-body geometry as node attributes and inter-n-body geometry as edge attributes, connected by horizontal and vertical edges; a message passing scheme (HAGO-MPS) aggregates across and within scales. A new chiral identity, the relative orientation of tetrahedron (ROT), distinguishes enantiomers; the der
Key results
- ROT distinguishes 602 of 624 chiral pairs on GEOM-QM9 (96.5% accuracy); HAGO-Net reaches 98.9% R/S classification accuracy on PubChem3D vs 98.5% for ChIRo.
- HAGO-Net achieves best MAE on five of twelve QM9 properties, e.g., U0 5.89 vs 6.15 for TorchMD (10.6% improvement cited).
- DHAGO-Net reduces aspirin force MAE to 11.6 vs 14.7 for PaiNN (about 25% lower), best on three of eight MD17 molecules.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HAGO-Net for QM9 molecular property prediction | QM9 property prediction · QM9 | MAE | Mixed |
| Use HAGO-Net for QM9 U0 prediction | QM9 property prediction · QM9 | MAE U0: 5.89 vs 6.59 | Helped |
| Use HAGO-Net for QM9 U prediction | QM9 property prediction · QM9 | MAE U: 6.11 vs 6.36 | Helped |
| Use HAGO-Net for QM9 G prediction | QM9 property prediction · QM9 | MAE G: 7.30 vs 7.78 | Helped |
| Use HAGO-Net for QM9 µ prediction | QM9 property prediction · QM9 | MAE µ: 0.023 vs 0.011 | Hurt |
| Use HAGO-Net for QM9 ϵHOMO prediction | QM9 property prediction · QM9 | MAE ϵHOMO: 22.6 vs 15.0 | Hurt |
| Use HAGO-Net for QM9 ϵLUMO prediction | QM9 property prediction · QM9 | MAE ϵLUMO: 19.1 vs 14.0 | Hurt |
| Use DHAGO-Net for force prediction | MD17 molecular dynamics prediction · MD17 aspirin | force MAE: 11.6 vs 14.7 | Helped |
| Use DHAGO-Net for force prediction | MD17 molecular dynamics prediction · MD17 ethanol | force MAE: 5.9 vs 9.0 | Helped |
| Use DHAGO-Net for force prediction | MD17 molecular dynamics prediction · MD17 malonaldehyde | force MAE: 9.7 vs 13.8 | Helped |
| Use DHAGO-Net for force prediction | MD17 molecular dynamics prediction · MD17 naphthalene | force MAE: 5.8 vs 3.4 | Hurt |
| Use DHAGO-Net for force prediction | MD17 molecular dynamics prediction · MD17 salicylic acid | force MAE: 8.7 vs 8.5 | Hurt |
Limitations
- Authors note HAGO-Net is only one possible implementation; more powerful architectures are possible.
- NequIP comparison uses an older reported version; the updated NequIP is better but uses more spherical harmonics.
- ROT identifies 602 of 624 chiral pairs, not all.
Experimental setup
Datasets: QM9 MD17 GEOM-QM9 PubChem3D
Benchmarks: GEOM-QM9 PubChem3D QM9 MD17
Metrics: accuracy MAE
Baselines: SchNet sGDML PhysNet DimeNet SEGNN PaiNN ComENet SphereNet TorchMD Equiformer ChIRo NequIP
Builds on: DimeNet SphereNet Equiformer Weisfeiler-Lehman test
Closest papers by meaning
- GeomGCL: Geometric Graph Contrastive Learning for Molecular Property Prediction (AAAI 2022 · similarity 0.89)
- Spherical Message Passing for 3D Molecular Graphs (ICLR 2022 · similarity 0.89)
- Empowering Active Learning for 3D Molecular Graphs with Geometric Graph Isomorphism (NeurIPS 2024 · similarity 0.88)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.88)
- Enhancing Molecular Representation Learning Through the Combination of 3D and 2D Graph Machine Learning (Student Abstract) (AAAI 2025 · similarity 0.87)
- A Plug-and-Play Quaternion Message-Passing Module for Molecular Conformation Representation (AAAI 2024 · 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.