DLIGAND2: an improved knowledge-based energy function for protein-ligand interactions using the distance-scaled,...

Area of Science:

• Computational Chemistry
• Structural Biology
• Drug Discovery

Background:

• Molecular docking performance relies heavily on scoring function accuracy.
• Knowledge-based potentials, derived from protein structures, are crucial scoring function types.
• Existing potentials like DLIGAND use limited atom type representations.

Purpose of the Study:

• To enhance the knowledge-based protein-ligand potential DLIGAND.
• To improve prediction accuracy for binding affinities and virtual screening.
• To develop a more effective scoring function for molecular docking.

Main Methods:

• Expanded atom type representation from 13 to 167 residue-specific types.
• Trained the updated potential (DLIGAND2) on an updated dataset of 12,450 protein chains.
• Evaluated DLIGAND2 against native complexes, docking poses, and benchmark datasets (DUD-E).

Main Results:

• DLIGAND2 demonstrated consistent improvement over DLIGAND in predicting binding affinities.
• DLIGAND2 showed a 52% increase in enrichment factors on the DUD-E decoy set compared to DLIGAND.
• DLIGAND2 outperformed Autodock Vina, other statistical potentials, and empirical/machine-learning methods in benchmark tests.

Conclusions:

• DLIGAND2 offers improved accuracy for molecular docking and virtual screening.
• Its parameter-free nature and strong performance make it valuable for pose re-assessment and scoring function development.
• DLIGAND2 represents a significant advancement in knowledge-based scoring potentials.