Incorporating Density Information for Protein-Drug Interaction

Protein machine learning primarily uses amino-acid sequences and fitted atomic coordinates, typically omitting the experimental electron-density maps used to determine crystallographic structures. Yet, a single fitted structure can omit evidence of conformational heterogeneity and local atomic support that may be relevant to protein-drug interaction. To retain this signal, we introduce CODE, a coordinate-and-density voxel encoder pre-trained by masked reconstruction on approximately 100K curated protein-ligand complexes. We transfer the frozen encoder to binding-affinity regression and reference-assisted structure-based drug design. CODE is competitive on leakage-controlled affinity benchmarks and improves VoxBind's Vina scores without increasing reference-ligand similarity. These results support experimental density as a transferable representation for protein-ligand modeling.