Virtual Screening with Gnina 1.0.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Virtual screening is crucial for drug discovery, aiming to identify compounds that bind to target molecules.
• Current virtual screening methods balance speed and accuracy, but may miss critical drug usability factors.
• Deep convolutional networks offer a novel approach to scoring protein-ligand interactions.

Purpose of the Study:

• To evaluate the virtual screening performance of the Gnina molecular docking software.
• To compare Gnina's deep learning approach against conventional empirical scoring methods.
• To investigate potential biases in benchmark datasets affecting machine learning model evaluation.

Main Methods:

• Utilized Gnina, a deep learning-based molecular docking software, for virtual screening.
• Assessed Gnina's performance on the DUD-E and LIT-PCBA virtual screening benchmarks.
• Compared Gnina's scoring function against the empirical AutoDock Vina scoring function.

Main Results:

• Gnina demonstrated superior performance compared to conventional empirical scoring on average.
• Gnina's default scoring outperformed AutoDock Vina on 89 out of 117 targets.
• A significant median 1% early enrichment factor was observed for Gnina, more than double that of Vina.
• Dataset biases were identified, potentially influencing the interpretation of machine learning model performance.

Conclusions:

• Gnina represents a significant advancement in virtual screening accuracy for drug discovery.
• The study highlights the need to address dataset biases for reliable machine learning model assessment.
• Further research is warranted to understand the impact of bias on deep learning in computational chemistry.