FFENCODER-PL: Pair Wise Energy Descriptors for Protein-Ligand Pose Selection.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and molecular modeling
• Machine learning in bioinformatics

Background:

• Accurate scoring functions are crucial for molecular docking to identify correct ligand poses.
• Previous work showed random forest (RF) algorithms enhance knowledge-based potentials for protein-ligand decoy detection.
• Force field (FF) potentials, like Amber, are vital but extracting pair wise energies for RF models is challenging.

Purpose of the Study:

• To develop a method for calculating pair wise energies from Amber force fields suitable for RF modeling.
• To create and evaluate RF-based scoring functions using these energies for improved protein-ligand docking.
• To assess the contribution of RF algorithms and force field potentials in scoring function performance.

Main Methods:

• Developed FFENCODER-PL to compute pair wise energies using FF14SB and GAFF2 force fields from Amber.
• Validated FFENCODER-PL with 275 ligand and 21 protein-ligand structures.
• Built RF models using FFENCODER-PL energies and tested them using the CASF-2016 benchmark.

Main Results:

• FFENCODER-PL successfully extracted force field-based pair wise energies.
• The developed RF models outperformed 33 existing scoring functions in accuracy and native ranking tests.
• RF models achieved a best decoy RMSD of approximately 2 Å from the native pose.
• Both RF algorithms and force field potentials were found to be critical for high accuracy.

Conclusions:

• FFENCODER-PL enables the use of force field-based pair wise energies in machine learning scoring functions.
• The combination of FFENCODER-PL and RF algorithms represents a significant advancement in protein-ligand docking accuracy.
• This approach facilitates the development of more effective scoring functions for drug discovery.