WDL-RF

WDL-RF predicts ligand bioactivities for G protein-coupled receptors (GPCRs) using weighted deep learning to generate molecular fingerprints and a random forest model to estimate values such as IC50, EC50, Ki, and Kd for virtual screening and lead identification.

Key Features:

• Two-stage pipeline: A sequential workflow that first generates molecular fingerprints and then predicts bioactivities.
• Molecular fingerprint generation: Uses a weighted deep learning method to produce data-driven molecular fingerprints from input ligands of arbitrary size.
• Bioactivity calculation: Employs a random forest model to calculate bioactivity endpoints including IC50, EC50, Ki, and Kd.
• End-to-end learning capability: Supports processing ligands without requiring predefined molecular sizes or hand-crafted feature definitions.
• Data-driven features: Leverages weighted deep learning–derived fingerprint features to improve the efficiency of short molecular fingerprints in virtual screening.
• Benchmark performance: Evaluated on twenty-six non-redundant GPCRs with 200–4000 ligand associations per receptor, achieving an average RMSE of 1.33 and r² of 0.80 against experimental measurements.

Scientific Applications:

• GPCR-focused drug discovery: Predicts ligand bioactivities for GPCR targets to support hit identification and lead optimization.
• Virtual screening: Enables screening of compound libraries when experimental interaction data are lacking.
• Affinity estimation: Provides quantitative estimates of IC50, EC50, Ki, and Kd for pharmacological assessment of ligands.

Methodology:

WDL-RF implements a two-stage computational approach: weighted deep learning to generate molecular fingerprints from input ligands, followed by a random forest model to predict bioactivity endpoints (IC50, EC50, Ki, Kd).