MUIGSL: Molecular Uncertainty-Aware Iterative Graph Structure Learning for Molecular Property Prediction

Abstract

Molecular Property Prediction (MPP) is a core task in computer-aided drug discovery. Traditional methods primarily focus on intramolecular information propagation while neglecting intermolecular relationships. While recent graph structure learning methods attempt to construct intermolecular graphs, they often ignore differences in node information quality, and the constructed graph structures are typically restricted to symmetry. To address these limitations, we propose a Molecular Uncertainty-Aware Iterative Graph Structure Learning (MUIGSL) for MPP. It adaptively reduces the influence of nodes with high uncertainty by estimating the uncertainty of node information and using it to adjust the strength of directional connections in the molecular similarity graph. In this way, the model can suppress the interference of low-quality nodes while integrating intramolecular and intermolecular information. MUIGSL is systematically evaluated on eight MoleculeNet benchmark datasets. Extensive experiments demonstrate that MUIGSL achieves state-of-the-art or near state-of-the-art performance across most molecular property prediction tasks. Furthermore, by integrating intramolecular and intermolecular information, MUIGSL provides an effective solution for accurate molecular property prediction and facilitates the discovery of potential drug candidates.