DeepDOX1: A Dual-Drive Framework Integrating Deep Learning and First-Principles Quantum Chemistry for Drug-Protein...

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Drug Discovery

Background:

• The field of drug-protein affinity (DPA) prediction has seen a rise in AI-based tools, driven by deep learning.
• These tools often require extensive experimental data for training, posing a challenge due to data limitations.

Purpose of the Study:

• To develop DeepDOX1, a novel DPA prediction tool.
• To integrate a concise AI architecture with interpretable, quantum chemistry-based representations for improved prediction accuracy and generalization.

Main Methods:

• Developed DeepDOX1, a dual-drive DPA prediction tool.
• Utilized first-principles quantum chemistry-generated features for protein-ligand interactions.
• Employed a convolutional neural network (CNN) model trained on a limited dataset (9,938 binders).

Main Results:

• DeepDOX1 demonstrated exceptional generalization capabilities on an independent test set (1,281 binders).
• The tool outperformed existing AI-based DPA prediction methods in hit-to-lead optimization simulations and challenging test cases.
• DeepDOX1 successfully guided the design of novel covalent inhibitors with validated strengthened activity.

Conclusions:

• The combination of deep learning and first-principles quantum chemistry represents a promising breakthrough in DPA prediction.
• DeepDOX1 showcases practical utility and effectiveness in real-world drug design applications.
• The developed tool can accelerate drug discovery by accurately predicting drug-protein interactions.