An electron-density point-cloud framework for robust protein-ligand interaction prediction

Abstract

Abstract Accurate protein-ligand affinity prediction typically depends on precise 3D coordinates, limiting robustness when structures are low-resolution or predicted. We introduce E-CloudBind, a framework that fuses electron-density point clouds with intrinsic molecular graphs to model non-covalent and covalent interactions without relying on sub-ångström accuracy. Ligand electron densities are obtained by semi-empirical quantum calculations, whereas protein pockets are represented by van der Waals-guided Gaussian point clouds, a physically motivated proxy that preserves interaction geometry while tolerating coordinate noise. Point-cloud encoders capture local non-covalent patterns and a heterogeneous graph neural network integrates them with covalent features for affinity regression. Across PDBbind splits and out-of-distribution scenarios, E-CloudBind matches or exceeds leading sequence-, graph- and structure-based baselines, with markedly reduced sensitivity to resolution and to experimental-versus-predicted proteins. Case studies further illustrate atom-level interpretability and large-scale virtual screening. By decoupling interaction learning from exact coordinates, E-CloudBind enables robust structure-based modeling on heterogeneous conditions.