Feed

To address limitations of conventional Quantitative Structure-Activity Relationship (QSAR) descriptors in capturing molecular electronic and spatial complexity, we developed a high-dimensional framework using three-dimensional electron density features. Electron densities were computed via density functional theory (DFT), converted to 3D point clouds, and encoded into multi-scale descriptors including radial distribution functions, spherical harmonic expansions, point feature histograms, and persistent homology. This design enabled molecular characterization across statistical, geometric, and topological dimensions. The proposed descriptors consistently improved performance across multiple machine learning models; for instance, Area Under the Curve (AUC) increased from 0.88 to 0.96 with Light Gradient Boosting Machine (LightGBM). Benchmarking demonstrated superior performance versus industry-standard ECFP4 fingerprints. Control experiments using purely geometric (CPK) point clouds yielded substantially lower performance, confirming that predictive gains stem from electronic structure information rather than high-dimensional geometry alone. Feature attribution analysis revealed that local geometric descriptors and intensity-based electronic features were primary contributors, while integration with conventional 1D/2D descriptors further enhanced accuracy, indicating strong complementarity. Model robustness was validated through DeLong and permutation tests, calibration assessments, and applicability domain analysis. This study provides proof-of-concept evidence that DFT-derived electron density features can be systematically integrated into QSAR modeling. Despite computational cost limitations and reduced chemical interpretability, results demonstrate that electronic-structure-based descriptors offer valuable complementarity to established approaches, opening new avenues for molecular representation in drug discovery.