AEP-Net: Angle-aware Electrostatic Potential Point Network for Molecular Property Prediction

Molecular representation learning is a fundamental task in computational chemistry and drug discovery. While graph neural networks (GNNs) have significantly advanced molecular representation learning, they rely solely on conventional molecular graph, inherently failing to capture critical three-dimensional electronic structure, which are crucial determinants of molecular interactions. To address this, we propose the Angle-aware Electrostatic Potential point Network (AEP-Net), a novel paradigm that explicitly integrates high-resolution 3D Electrostatic Potential (ESP) represented as point clouds. AEP-Net represents molecules as collections of Atomic-centered Electrostatic-potential Subgraphs (AES). We design an angle-aware convolution mechanism that explicitly captures 3D spatial geometric features through learnable directional bases. Each AES processes its internal ESP points via a field-aware interaction layer, employing a triple weighting mechanism based on angular similarity, spatial proximity, and electrostatic values to capture fine-grained local electrostatic patterns. Features from different AESs are aggregated at the atomic level through global pooling, preserving the integrity of the local ESP environment while enabling a scalable architecture. Subsequently, a hierarchical hybrid network deeply integrates multi-scale features at the point cloud, atomic, and molecular levels,