XIMP: Cross Graph Inter-Message Passing for Molecular Property Prediction

ICLR 2026 Conference Submission18507 Authors

Keywords: ADMET, Drug Discovery, Expressivity, Graph Neural Networks, GNN, Hierarchical Graphs, Molecular Property Prediction

TL;DR: XIMP is a new GNN architecture that combines multiple molecular graph abstractions with higher-order message passing, enabling expressive, interpretable representations that overcome data scarcity and oversquashing, improving drug discovery tasks

Abstract: Accurate molecular property prediction is central to drug discovery, yet Graph Neural Networks (GNNs) often underperform in data-scarce regimes and can trail fixed fingerprints. We introduce XIMP (Cross Graph Inter-Message Passing), which performs message passing both within and across graph representations, integrating multiple granularities in a single model. In our chemistry setting, XIMP unifies the molecular graph with junction trees (scaffold-aware) and extended reduced graphs (pharmacophoric), enabling per-atom use of complementary views and exceeding the expressivity of the Weisfeiler-Leman isomorphism test. Unlike prior work – often limited to few abstractions, indirect exchange via the original graph, or overlooking oversquashing – XIMP supports arbitrary numbers of abstractions and both direct as well as indirect inter-abstraction communication. Across ten diverse molecular property-prediction tasks, XIMP outperforms state-of-the-art GNNs and fingerprint baselines in most cases, leveraging interpretable abstractions as an inductive bias to guide learning toward established chemical concepts and enhance generalization in data-scarce regimes.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 18507