Projection-based molecular feature maps for CNN-driven nephrotoxicity prediction.

The development of reliable predictive models for nephrotoxic agents remains a critical challenge in drug development, given safety concerns associated with kidney toxicity. Conventional molecular descriptors generally fail to capture essential spatial and electronic features necessary for accurate nephrotoxicity prediction, underscoring the need for novel descriptor approaches. This study presents a novel projection-based method for nephrotoxicity prediction by converting chemical structures into 2D maps for deep learning via 3D spatial transformation to enhance both feature representation and model performance. Both Mollweide and Equirectangular projections were utilized to transform 3D molecular geometries into optimized 2D representations. The 2D molecular maps incorporated three key molecular properties to display the information of atom-based projections showing atomic positions and identities, electrostatic projections visualizing charge distribution, and vdW projections illustrating molecular steric potentials. The Mollweide projection based on atom color demonstrated superior predictive performance, achieving 83% predictive accuracy with an AUC of 0.86, establishing it as the most effective CNN model. The electrostatic and vdW projections transformed atomic spatial data into electrostatically and sterically informative maps, enabling more nuanced molecular pattern recognition and enhanced representation. The reliability of the model was validated through different methods, including independent verification on test set combined with five-fold cross validation as well as comparisons with traditional descriptor-based models using the benchmarking method. Our findings demonstrate that projection-based molecular representations show strong potential for nephrotoxicity screening, opening new possibilities for toxicology prediction and drug safety advancement.