Research map / Probabilistic and scientific modeling / Molecules and proteins
Quadruple Attention in Many-body Systems for Accurate Molecular Property Predictions
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 MABNet, an SE(3)-equivariant graph transformer with a four-body (quadruplet) attention module for molecular energy and force prediction, reporting SOTA on MD22 and SPICE.
Problem
Existing graph neural networks and transformers approximate three- and four-body interactions via message passing, while attention models are limited to triplets, making explicit higher-order (four-body) interactions computationally expensive for molecular property prediction.
Approach
Generalizes attention from triplets to quadruplets of atoms (node-node and node-edge-node tensors with gated softmax over edge-pair keys and values), keeps equivariance by using vector-valued features, and limits cost via spatial cutoffs and sparse edges.
Key results
- On SPICE, MABNet reports average energy MAE 33.7 meV and force MAE 20.4 meV/Å, versus 50.1 and 39.0 for TensorNet.
- On MD22, MABNet attains the lowest MAE in 7 of 10 energy/force metrics on five large molecules, e.g., Ac-Ala3-NHMe energy 0.0534 kcal/mol.
- On dipeptides in SPICE, force MAE drops to 7.7 meV/Å versus 23.1 for TensorNet; ablation shows 4-body attention improves over 3-body attention.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 4-body attention instead of 3-body attention (ablation) | MABNet · SPICE Dipeptides | force MAE: 10.1 -> 7.7 | Helped |
| use 4-body attention instead of 3-body attention (ablation) | MABNet · SPICE Solvated Amino Acids | force MAE: 48.4 -> 31.9 | Helped |
| remove equivariance from 3-body attention (ablation) | MABNet · SPICE Solvated Amino Acids | force MAE: 48.4 vs. 59.4 | Hurt |
| increase 4-body attention edges from 1.5N to 6N (ablation) | MABNet · SPICE Dipeptides | force MAE: 8.4 -> 7.7 | Helped |
| increase 4-body attention edges from 1.5N to 6N (ablation) | MABNet · SPICE Solvated Amino Acids | force MAE: 36.3 -> 31.9 | Helped |
| increase 4-body attention edges from 1.5N to 6N (ablation) | MABNet · MD22 Ac-Ala3-NHMe | energy MAE: 0.058 -> 0.053 | Helped |
| increase 4-body attention edges from 1.5N to 6N (ablation) | MABNet · MD22 Ac-Ala3-NHMe | force MAE: 0.080 -> 0.077 | Helped |
| use MABNet instead of the strongest listed baseline | molecular property prediction · MD22 | MAE | Mixed |
| use MABNet instead of TensorNet | molecular property prediction · SPICE Average | energy MAE: 50.1 -> 33.7 | Helped |
| use MABNet instead of TensorNet | molecular property prediction · SPICE Average | force MAE: 39.0 -> 20.4 | Helped |
Surprising findings
- On the 370-atom carbon nanotube, MABNet's energy MAE (0.1862) is worse than ViSNet-LSRM (0.1135), which the authors attribute to pairwise-dominated periodic systems.
- MABNet's force MAE on Stachyose (0.0752) lags Equiformer-LSRM (0.0632), and the authors note force accuracy is sensitive to high-order expansions.
Limitations
- Energy MAE is worse than ViSNet-LSRM on the 370-atom carbon nanotube, and force MAE lags some baselines on several MD22 systems.
- Complexity is stated as O(|N|^2·|E|) with cutoff-based sparsity, and the paper provides no formal guarantee of the speed-accuracy trade-off.
- Results are reported on a limited set of benchmarks (MD22, SPICE) with no evaluation on other property datasets.
Experimental setup
Datasets: MD22 SPICE
Benchmarks: MD22 SPICE
Metrics: energy MAE force MAE
Baselines: SchNet sGDML PaiNN ET So3krates MACE Equiformer ViSNet LSR-MP DimeNet++ TensorNet Allegro
Builds on: VisNet Equiformer Triplet Graph Transformers GEM-2 AlphaFold triangular update
Closest papers by meaning
- Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel (ICLR 2024 · similarity 0.87)
- Learning the Electronic Hamiltonian of Large Atomic Structures (ICML 2025 · similarity 0.84)
- Neural P$^3$M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs (NeurIPS 2024 · similarity 0.84)
- The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · similarity 0.84)
- Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation (ICLR 2024 · similarity 0.84)
- Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer (ICLR 2025 · similarity 0.84)
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.