Deffini: A family-specific deep neural network model for structure-based virtual screening

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Traditional docking-based virtual screening methods have limitations in accuracy.
• Deep learning models offer improved accuracy for virtual screening.
• Structure-based virtual screening requires accurate prediction of protein-ligand interactions.

Purpose of the Study:

• To develop and evaluate Deffini, a deep neural network for structure-based virtual screening.
• To assess the impact of dataset quality and training strategy on model performance.
• To improve the prediction accuracy of drug-target interactions using machine learning.

Main Methods:

• Developed Deffini, a structure-based neural network model for virtual screening.
• Trained and validated Deffini on benchmark datasets (DUD-E, MUV).
• Employed a family-specific training approach and introduced the Kernie dataset for protein kinase screening.

Main Results:

• Deffini outperformed Smina on the DUD-E dataset (AUC ROC 0.92).
• Initial performance on the MUV dataset was poor (AUC ROC 0.517).
• Family-specific models and the Kernie dataset significantly improved performance on MUV kinases (AUC ROC 0.745).

Conclusions:

• Deep learning virtual screening models require careful dataset selection and training strategies.
• Family-specific models outperform pan-family models for improved predictive power.
• High-quality, curated datasets are crucial for enhancing deep neural network performance in drug discovery.