Environment-Aware Multiscale Geometric Interaction for Equivariant Molecular Spectral Prediction

Haoran Li; Weiran Cui; Minghui Li

IJCAI 20260 citations

Environment-Aware Multiscale Geometric Interaction for Equivariant Molecular Spectral Prediction

Abstract

Predicting molecular spectra requires modeling 3D conformation and solvent modulation. However, E(3)-equivariant networks based on local message passing exhibit limited sensitivity to long-range geometric dependencies, affecting the discrimination of globally distinct conformers. We introduce the Multiscale Geometric Interaction Layer (MGIL), which integrates global context by augmenting local features with centroid-referenced anchors, geometric moments, and virtual nodes. This design explicitly encodes global anisotropy while maintaining equivariance. Furthermore, we propose a Solvent Field Modulator (SFM) to encode solvent topology for conditional feature adaptation. Experiments demonstrate that MGIL enhances the capture of global structural variations, yielding consistent performance gains across spectral prediction benchmarks while maintaining linear computational efficiency.

Machine Learning: Deep learning architecturesMachine Learning: Geometric learningMachine Learning: Representation learningMachine Learning: Sequence and graph learning

BibTeX

@inproceedings{ijcai2026_environmentaware,
  title = {Environment-Aware Multiscale Geometric Interaction for Equivariant Molecular Spectral Prediction},
  author = {Haoran Li and Weiran Cui and Minghui Li},
  booktitle = {IJCAI 2026},
  year = {2026}
}