Research map / Graph learning / GNN architectures

Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation

ICLR 2024 · Graph learning › GNN architectures · approach: Equivariant networks for molecules · 6 citations (OpenAlex)

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 LSR-MP, a fragmentation-inspired long-short-range message-passing framework (BRICS fragments plus bipartite atom-fragment message passing) that extends EGNNs, instantiated as ViSNet-LSRM for molecular dynamics potentials on MD22 and Chignolin.

Problem

Equivariant graph neural network potentials use local radius cutoffs and thus miss long-range electrostatic and van der Waals interactions in large molecules, while increasing cutoff or depth causes over-squashing and data-efficiency loss.

Approach

Adds a long-range module on top of a short-range EGNN in which BRICS-derived fragment embeddings (scalar and vectorial) exchange equivariant messages with atoms over a bipartite atom-fragment radius graph, with distance-dependent attention; the framework generalizes EGNNs and reduces to them when the long-range cutoff equals the short-range cutoff.

Key results

• ViSNet-LSRM reduces energy/force MAE versus ViSNet on large MD22 molecules and Chignolin, with the abstract claiming up to 40% MAE reduction.
• On AT-AT-CG-CG, ViSNet-LSRM reaches force MAE 0.1063 kcal/mol/Å versus 0.1563 for ViSNet and uses 1.70M parameters versus 2.21M.
• Applying LSR-MP to PaiNN, ET, Equiformer, and ViSNet reduces force MAE by 15% to 58.8% on AT-AT-CG-CG.

Measured effects reported in the paper

Surprising findings

• Increasing the short-range cutoff beyond 4–5 Å degrades EGNN performance, which the authors attribute to over-squashing.
• A 3-layer ViSNet-LSRM outperforms an 8-layer ViSNet and PaiNN.

Limitations

• BRICS fails to fragment supramolecules in MD22, so k-means clustering is used as an alternative.
• Fragmentation is fixed rather than learned end-to-end.
• Zero-shot transfer to larger unseen molecules is poor (energy MAE 150.23 kcal/mol).

Experimental setup

Datasets: MD22 Chignolin PubChemQC Electrostatics binding energy dataset

Benchmarks: MD22 Chignolin PubChem Electrostatics binding energy

Metrics: energy MAE force MAE training time per epoch parameter count

Baselines: sGDML ViSNet PaiNN ET GemNetOC SO3krates Allegro MACE Equiformer

Builds on: ViSNet PaiNN BRICS Equiformer TorchMD-NET

Closest papers by meaning

• Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel (ICLR 2024 · similarity 0.87)
• ELoRA: Low-Rank Adaptation for Equivariant GNNs (ICML 2025 · similarity 0.87)
• Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity (ICML 2025 · similarity 0.87)
• Neural P$^3$M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs (NeurIPS 2024 · similarity 0.86)
• Learning Equivariant Non-Local Electron Density Functionals (ICLR 2025 · similarity 0.86)
• Higher-Rank Irreducible Cartesian Tensors for Equivariant Message Passing (NeurIPS 2024 · similarity 0.86)