PharmRF

PharmRF predicts and prioritizes protein–ligand complexes by using a random forest regressor that correlates protein pocket descriptors with ligand pharmacophoric elements to produce pharmacophore scores and predicted binding affinities for structure-based virtual screening.

Key Features:

• Machine-learning scoring function: Implements a random forest regressor to correlate protein pocket descriptors with ligand pharmacophoric elements and predict binding affinities.
• Descriptor usage: Uses descriptors of the protein pocket and ligand pharmacophoric elements as input features for the model.
• Training dataset: Trained and tested on the PDBbind v2018 dataset.
• Template prioritization: Scores and prunes Protein Data Bank (PDB) entries to prioritize templates for structure-based pharmacophore screening.
• Predicted affinity scores: Produces scores representing predicted binding affinities used to identify and prioritize high-affinity ligands.
• Benchmark performance (DUD-E): On 10 protein–ligand systems from DUD-E achieved an average success rate of 77.61% and a median success rate of 87.16%, outperforming Vina docking scores.
• Benchmark performance (CASF-2016): On the CASF-2016 benchmark set observed a correlation of 0.591 between PharmRF-predicted and experimental binding affinities compared against 25 other scoring functions.
• Benchmark performance (LIT-PCBA): On LIT-PCBA reported average and median success rates of 57.55% and 74.72%, respectively, with four targets exceeding 90% success.

Scientific Applications:

• Structure-based pharmacophore screening: Prioritizing PDB-derived protein–ligand complexes as templates for structure-based pharmacophore model generation.
• Virtual screening enrichment: Improving enrichment rates in pharmacophore-based virtual screening by selecting high-scoring templates.
• Binding affinity prediction: Predicting relative binding affinities of protein–ligand complexes using machine-learning-derived scores.
• Benchmarking and method comparison: Evaluating and comparing scoring functions using CASF-2016, DUD-E, and LIT-PCBA benchmarks.

Methodology:

Uses a random forest regressor to correlate protein pocket descriptors with ligand pharmacophoric elements; trained and tested on PDBbind v2018; evaluated on DUD-E (10 systems), CASF-2016, and LIT-PCBA; compared performance against Vina docking scores.