Research map / Architectures and efficiency / Attention and transformers
Generalist Equivariant Transformer Towards 3D Molecular Interaction 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 GET, an E(3)-equivariant transformer over geometric graphs of sets (blocks of atoms) that encodes proteins, small molecules and RNA/DNA with one model, and shows gains on binding affinity tasks and zero-shot RNA/DNA-ligand affinity.
Problem
Molecular interaction modeling usually encodes each molecule type (proteins, small molecules, RNA/DNA) with separate domain-specific models or single-level (atom or block) graphs, failing to learn shared interaction physics across domains.
Approach
Represent any 3D complex as a geometric graph of sets (blocks of variable-size atom sets) and use an E(3)-equivariant, permutation-invariant bilevel attention (sparse block-level plus dense atom-level), feed-forward, and layer-norm stack to process all molecule types with one model without pooling away fine-grained information.
Key results
- On PDBbind (30% sequence identity split), GET reaches RMSE 1.364 and Pearson 0.596, beating the best listed baseline (Atom3D-3DCNN RMSE 1.416; ProNet-All-Atom Pearson 0.551).
- On PPA, GET reaches Pearson 0.514 versus 0.484 for the best block-level baseline (Equiformer); GET-PS reaches LBA RMSE 1.309 versus 1.343 for the best atom-level baseline (LEFTNet).
- In zero-shot DNA/RNA-ligand affinity prediction, GET achieves Pearson 0.450 versus 0.348 for the best baseline (hierarchical ET).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GET instead of Atom3D-3DCNN | PDBbind affinity prediction · PDBbind | RMSE: 1.416\pm 0.021 -> 1.364\pm 0.009 | Helped |
| Train GET on mixed PPA and LBA datasets | GET · PPA-All | Pearson: 0.514\pm 0.011 -> 0.519\pm 0.004 | Helped |
| Train GET on mixed PPA and LBA datasets | GET · LBA | Pearson: 0.620\pm 0.004 -> 0.622\pm 0.006 | Helped |
| Define small-molecule blocks as principal subgraphs | GET on LBA · LBA | Pearson: 0.620\pm 0.004 -> 0.633\pm 0.008 | Helped |
| Remove layer normalization (ablation) | GET · PPA-All | Pearson: 0.366\pm 0.024 -> 0.514\pm 0.011 | Hurt |
| Remove equivariant coordinate normalization from layer normalization (ablation) | GET · PPA-All | Pearson: 0.368\pm 0.025 -> 0.514\pm 0.011 | Hurt |
| Remove scale information from hidden features (ablation) | GET · PPA-All | Pearson: 0.490\pm 0.027 -> 0.514\pm 0.011 | Hurt |
| Remove the equivariant feed-forward network (ablation) | GET · PPA-All | Pearson: 0.494\pm 0.010 -> 0.514\pm 0.011 | Hurt |
| Remove both layer normalization and feed-forward network (ablation) | GET · PPA-All | Pearson: 0.360\pm 0.018 -> 0.514\pm 0.011 | Hurt |
| Add layer normalization (ablation) | GET · LBA | Pearson: 0.589\pm 0.007 -> 0.620\pm 0.004 | Helped |
| Add equivariant coordinate normalization to layer normalization (ablation) | GET · LBA | Pearson: 0.591\pm 0.008 -> 0.620\pm 0.004 | Helped |
| Reflect coordinate rescaling information in hidden features (ablation) | GET · LBA | Pearson: 0.591\pm 0.002 -> 0.620\pm 0.004 | Helped |
Surprising findings
- Mixing protein-protein and protein-ligand affinity training data improves GET while most baselines degrade in most cases.
- Zero-shot transfer to RNA/DNA-ligand affinity without any in-domain training data outperforms domain-specific-style baselines.
Limitations
- Evaluation focuses on prediction tasks; generative tasks are not addressed.
- Performance drops with larger coordinate noise (5.0 Å) and protein-protein complexes are harder to predict from structure.
- No large-scale universal pretraining across molecular domains explored.
Experimental setup
Datasets: PDBbind Protein-Protein Affinity Benchmark v2 Atom3D LBA Atom3D LEP
Benchmarks: PDBbind PPA LBA LEP DNA/RNA-ligand zero-shot PDBbind
Metrics: RMSE Pearson Spearman AUROC AUPRC
Baselines: SchNet DimeNet++ EGNN TorchMD-Net (ET) GemNet MACE Equiformer LEFTNet ProNet Holoprot Atom3D-3DCNN GVP
Builds on: equivariant graph neural networks transformer attention principal subgraphs
Closest papers by meaning
- Equivariant Spherical Transformer for Efficient Molecular Modeling (NeurIPS 2026 · similarity 0.89)
- Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation (AAAI 2025 · similarity 0.88)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.88)
- Triplet Interaction Improves Graph Transformers: Accurate Molecular Graph Learning with Triplet Graph Transformers (ICML 2024 · similarity 0.87)
- GeoMFormer: A General Architecture for Geometric Molecular Representation Learning (ICML 2024 · similarity 0.87)
- Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.87)
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.