Research map / Graph learning / GNN architectures
Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel
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 Neural Atoms, a plug-in attention module that projects atoms into a few learnable super-nodes, exchanges information among them, and projects back, so GNNs can capture long-range interactions in molecular graphs.
Problem
GNNs capture short-range interactions well but struggle with long-range interactions (LRI) in molecular graphs, due to over-squashing and over-smoothing as multi-hop information must pass through many intermediate atoms.
Approach
Learnable attention-based grouping projects all atoms into a small set of neural atoms, which communicate via self-attention so any distant atom pair interacts in a single hop; results are projected back to enhance atom features. The paper links this to Ewald summation's long-range term.
Key results
- Neural Atoms improve GINE on Peptides-struct MAE from 0.3547 to 0.2553 (up to 27.32% relative gain) and GCN Peptides-func AP from 0.5930 to 0.6220.
- On OE62 with PaiNN, + Neural Atoms reduces energy MAE from 0.6049 to 0.0558 with 6.05M vs 12.52M parameters.
- GatedGCN+Neural Atoms reaches 0.6562 AP on Peptides-func, exceeding the Transformer+LapPE baseline (0.6326).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Neural Atoms | GatedGCN · Peptides-func test | AP: 0.5864 $\pm$ 0.0077 -> 0.6562 $\pm$ 0.0075 | Helped |
| add Neural Atoms | GINE · Peptides-struct test | MAE: 0.3547 $\pm$ 0.0045 -> 0.2553 $\pm$ 0.0005 | Helped |
| add Neural Atoms | GatedGCN+RWSE · Peptides-struct test | MAE: 0.3357 $\pm$ 0.0006 -> 0.2568 $\pm$ 0.0005 | Helped |
| add Neural Atoms | GatedGCN+RWSE · PCQM-Contact test | MRR: 0.3242 $\pm$ 0.0008 -> 0.3262 $\pm$ 0.0010 | Helped |
| add Neural Atoms | SchNet · OE62 validation | Energy MAE: 0.1351 -> 0.0834 | Helped |
| add Neural Atoms | PaiNN · OE62 validation | Energy MAE: 0.6049 -> 0.0558 | Helped |
| add Neural Atoms | PaiNN · OE62 validation | Energy MSE: 0.0133 -> 0.0122 | Helped |
| add Neural Atoms | DimeNet++ · OE62 validation | Energy MAE: 0.0501 -> 0.0551 | Hurt |
| use incremental rather than fixed grouping (ablation) | GCN + Neural Atoms · Peptides-struct test | MAE: 0.2582 $\pm$ 0.0011 -> 0.3239 $\pm$ 0.0014 | Hurt |
| use decremental rather than fixed grouping (ablation) | GCN + Neural Atoms · Peptides-struct test | MAE: 0.2582 $\pm$ 0.0011 -> 0.2606$\pm$ 0.0003 | No clear effect |
| set neural-atom proportion to 0.9 rather than 0.1 (ablation) | GCN + Neural Atoms · Peptides-func test | AP: 0.5859 $\pm$ 0.0073 -> 0.6220 $\pm$ 0.0046 | Helped |
Surprising findings
- Multiple virtual nodes with pairwise connections perform nearly identically regardless of count, whereas neural atoms improve with more atoms.
- Incremental neural atom counts across layers hurt performance (e.g., GCNII MAE rises to 0.4084 on Peptides-struct).
Limitations
- Focuses on 2D molecular graphs without 3D coordinates; coordinate-aware extension left to future work.
- Emphasizes intra-molecular interactions rather than inter-molecular ones such as protein docking.
- Grouping strategy is learned without expert knowledge, limiting interpretability of groups.
Experimental setup
Datasets: Peptides-func Peptides-struct PCQM-Contact OE62 Mutagenicity ogbn-arXiv Amazon Product-Computer
Benchmarks: Peptides-func Peptides-struct PCQM-Contact OE62
Metrics: AP MAE MRR energy MAE energy MSE
Baselines: GCN GINE GCNII GatedGCN GatedGCN+RWSE Transformer+LapPE SAN+LapPE GraphGPS SchNet PaiNN DimeNet++ Ewald Block
Builds on: GraphGPS Ewald-based message passing virtual node graph transformer
Closest papers by meaning
- Neural P$^3$M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs (NeurIPS 2024 · similarity 0.90)
- Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer (ICLR 2025 · similarity 0.88)
- Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation (ICLR 2024 · similarity 0.87)
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.87)
- Pre-Training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2025 · similarity 0.87)
- The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · 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.