Research map / Probabilistic and scientific modeling / Molecules and proteins
SaVeNet: A Scalable Vector Network for Enhanced Molecular Representation Learning
Sarp Aykent, Tian Xia
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add direction noise (DN) during training (ablation) | SAVENET (6 layers, no VA) · QM9 | std. MAE: 0.79 -> 0.77 | Improved |
| scale from 6 to 8 layers (ablation) | SAVENET without VA · QM9 | std. MAE: 0.77 -> 0.71 | Improved |
| add vector activations (VA) to 8-layer model (ablation) | SAVENET · QM9 | std. MAE: 0.71 -> 0.69 | Improved |
| expand model from 8 to 12 layers (ablation) | SAVENET · QM9 | std. MAE: 0.69 -> 0.63 | Improved |
| add vector activations (VA) to 6-layer model, with direction noise (ablation) | SAVENET · QM9 | log MAE: -5.85 -> -5.95 | Improved |
| scale from 8 to 12 layers (ablation) | SAVENET · QM9 | log MAE: -6.04 -> -6.14 | Improved |
| use SAVENET-B instead of Equiformer | SAVENET · QM9 efficiency | latency (ms): 337.3 -> 23.1 | Improved |
| use SAVENET-B instead of Equiformer | SAVENET · QM9 | std. MAE: 0.70 -> 0.69 | Improved |
| scale SAVENET-B to SAVENET-L | SAVENET · QM9 | std. MAE: 0.69 -> 0.63 | Improved |
| scale SAVENET-B to SAVENET-L | SAVENET · QM9 efficiency | latency (ms): 23.1 -> 50.6 | Hurt |
| scale SAVENET-B to SAVENET-L | SAVENET · Molecule3D | std. MAE: 0.0156 -> 0.0128 | Improved |
| SAVENET vs Equiformer on frontier orbital targets | SAVENET-L · QM9 | epsilon_HOMO MAE (meV): 15.4 -> 16.6 | Hurt |
Closest papers by meaning
- HAGO-Net: Hierarchical Geometric Message Passing for Molecular Representation Learning (AAAI 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)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.87)
- De novo Protein Design Using Geometric Vector Field Networks (ICLR 2024 · similarity 0.87)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.86)
- Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation (AAAI 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.
