Improving the Virtual Screening Ability of Target-Specific Scoring Functions Using Deep Learning Methods.

Area of Science:

• Computational chemistry
• Drug discovery and design
• Artificial intelligence in pharmacology

Background:

• Scoring functions are crucial for structure-based virtual screening in drug discovery.
• Target-specific scoring functions (TSSFs) show superior performance over universal functions in drug development.
• Developing effective methods for constructing TSSFs is vital for advancing drug design.

Purpose of the Study:

• To introduce DeepScore, a novel deep learning-based model for constructing target-specific scoring functions (TSSFs).
• To evaluate DeepScore's performance in predicting protein-ligand binding affinity.
• To demonstrate the utility of deep learning in enhancing drug design and discovery processes.

Main Methods:

• Developed DeepScore, a deep learning model utilizing a feedforward neural network to calculate protein-ligand binding affinity based on the PMF scoring function framework.
• Validated DeepScore's performance on the DUD-E dataset across 102 targets.
• Introduced DeepScoreCS, a consensus model combining DeepScore and Glide Gscore.

Main Results:

• DeepScore significantly outperformed Glide Gscore on the DUD-E validation dataset.
• Achieved an average ROC-AUC of 0.98 across 102 targets.
• DeepScore demonstrated superior performance compared to other machine learning-based TSSF construction methods.

Conclusions:

• DeepScore is a valuable tool for constructing target-specific scoring functions (TSSFs).
• The model represents a novel approach integrating deep learning with drug design principles.
• DeepScore offers a promising advancement for virtual screening and drug discovery pipelines.