ECHO: A Multi-Scale Graph Learning Framework for Enhanced Molecular Representation

ICLR 2026 Conference Submission784 Authors

Keywords: Graph Neural Network; Multi-scale Representation Learning; Drug Discovery; Cheminformatics; Molecular Property Prediction

TL;DR: ECHO is a novel multi-scale graph learning framework that comprehensively represents molecules by integrating diverse chemical information, leading to superior molecular property prediction and enhanced explainability.

Abstract: Accurate molecular property prediction is a cornerstone of drug discovery. While graph-based models excel at capturing molecular topology, they operate on a single atomic scale, overlooking crucial higher-order information from functional groups and spatial geometry like bond angles. To address this, we propose ECHO, a multi-scale graph learning framework that hierarchically integrates information from atom-level topology, functional-group semantics, and bond-angle geometry. ECHO introduces a hierarchical cross-scale attention mechanism for bidirectional information flow between fine-grained and coarse-grained graph representations, enabling mutual refinement. To further synthesize these diverse features, a novel hypergraph fusion module is designed to capture high-order interactions. Extensive experiments show that ECHO consistently outperforms state-of-the-art baselines, demonstrating the significant advantage of its multi-scale approach and offering new insights into the interplay between different structural scales in molecular representation learning. Code is now available at \url{https://anonymous.4open.science/r/ECHO-3C40}.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 784