Research question
Can an electron-density representation improve binding-affinity prediction when atomic coordinates are uncertain?
How the method works
E-CloudBind combines covalent molecular graphs with point clouds representing ligand electron density and a protein-pocket density proxy. Ligand densities are approximated with GFN2-xTB; pocket clouds use van der Waals-guided Gaussian functions. A heterogeneous graph network combines these features to predict binding affinity.
E-CloudBind: from electron-density representations to binding-affinity prediction.
Figure 2 · Yujian Liu et al., Nature Communications 17, 7424 (2026). © The Author(s) 2026. Full figure reproduced; display size and image format only are adjusted. · Published figure · CC BY 4.0
View full-size figureWhat the study shows
The reported tests span crystallographic resolution, experimental and AlphaFold2 structures, and ligand-scaffold, protein-sequence and joint holdouts. They support improved robustness to structural uncertainty. An ablation also finds similar prediction performance using a cheaper Gaussian approximation for ligand clouds.
Scope of the evidence
The model uses prepared three-dimensional structures and predefined pockets. Explicit solvent and polarization effects are omitted, and substantial geometric corruption reduces performance. These results support affinity prediction; mechanistic explanations require independent testing.
Read the publisher record for the abstract and access options.