Learning from Docked Ligands: Ligand-Based Features Rescue Structure-Based Scoring Functions When Trained on Docked...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Machine learning (ML) scoring functions excel with crystal structures for protein-ligand binding affinity.
• Their performance with computationally docked poses is less understood.
• Accurate binding affinity prediction is crucial for drug discovery.

Purpose of the Study:

• To evaluate ML scoring functions using docked versus crystallographic poses.
• To develop and assess a hybrid scoring function for improved docked pose prediction.
• To introduce a new validation dataset for binding affinity prediction.

Main Methods:

• Trained and tested ML scoring functions on both crystal and docked poses using PDBbind Core Sets.
• Developed a hybrid scoring function integrating structure-based and ligand-based features.
• Created and utilized the Updated DUD-E Diverse Subset for validation.

Main Results:

• ML scoring functions trained/tested on docked poses showed reduced performance compared to crystal poses.
• The hybrid scoring function achieved performance comparable to structure-based methods on crystal poses when using docked poses.
• The hybrid model sometimes generalized poorly to new protein targets.

Conclusions:

• Docked poses decrease ML scoring function performance for binding affinity prediction.
• Hybrid scoring functions offer a promising avenue for improving predictions with docked poses.
• Further development and validation benchmarks are needed for robust ML scoring functions.