GNINA 1.0: molecular docking with deep learning.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning in drug discovery

Background:

• Molecular docking is crucial for predicting ligand-receptor interactions.
• Scoring functions are essential for evaluating molecular poses in docking.
• Current methods require optimization for accuracy and efficiency.

Purpose of the Study:

• To introduce and evaluate Gnina 1.0, a novel molecular docking software.
• To assess the performance of convolutional neural networks (CNNs) as a scoring function.
• To optimize Gnina's parameters for docking performance and computational efficiency.

Main Methods:

• Utilized an ensemble of CNNs as a scoring function within the Gnina software.
• Explored various parameter values to optimize docking performance and computational cost.
• Compared Gnina's performance against AutoDock Vina using redocking and cross-docking tasks.

Main Results:

• Gnina with CNN scoring significantly outperformed AutoDock Vina in both defined binding pocket and whole protein docking scenarios.
• Top1 accuracy increased from 58% to 73% for redocking with defined pockets and from 31% to 38% for whole protein docking.
• The CNN ensemble demonstrated generalization to unseen proteins and ligands, with scores correlating well with RMSD to known binding poses.

Conclusions:

• Gnina 1.0, powered by CNNs, offers superior performance for molecular docking compared to traditional methods.
• The software provides an effective and generalizable approach for pose prediction in drug discovery.
• Gnina is available as an open-source tool to advance computational drug design.