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

Area of Science:

• Computational chemistry
• Toxicology
• Drug development

Background:

• Predicting nephrotoxic agents is crucial for drug development due to kidney toxicity risks.
• Traditional molecular descriptors often lack the spatial and electronic details needed for accurate nephrotoxicity prediction.

Purpose of the Study:

• To develop a novel projection-based method for enhanced nephrotoxicity prediction.
• To improve feature representation and deep learning model performance using 3D to 2D molecular structure transformation.

Main Methods:

• Utilized Mollweide and Equirectangular projections to convert 3D molecular geometries into 2D maps.
• Incorporated atom-based, electrostatic, and van der Waals (vdW) projections to represent molecular properties.
• Developed a Convolutional Neural Network (CNN) model for prediction.

Main Results:

• The Mollweide projection achieved 83% predictive accuracy and an AUC of 0.86, outperforming other methods.
• Projection-based maps enhanced molecular pattern recognition by visualizing atomic positions, charge distribution, and steric potentials.
• The model's reliability was confirmed via independent testing, cross-validation, and comparison with traditional descriptor models.

Conclusions:

• Projection-based molecular representations show significant potential for effective nephrotoxicity screening.
• This approach offers advancements in toxicology prediction and contributes to improved drug safety.