FRAGSITE

FRAGSITE applies fragment-based machine learning to improve ligand virtual screening by integrating fragment scores, molecular fingerprints, and FINDSITEcomb2.0 global ligand similarity to predict protein–ligand binding using predicted low-resolution protein structures.

Key Features:

• Fragment-Based Scoring: Integrates ligand fragment scores computed from molecular fingerprints and maps interactions to stereochemically conserved subpockets, leveraging evolutionary conservation across unrelated proteins.
• Machine Learning Framework: Uses a boosted tree regression model to combine fragment-based scores with global ligand similarity scores from FINDSITEcomb2.0 to improve prediction accuracy.
• Benchmarking Performance: Benchmarked on a 102-protein DUD-E set with template proteins excluded above 30% sequence identity, reporting precision improvement of 14.3%, recall improvement of 18.5%, and a mean top 1% enrichment factor of 30.2 versus 25.2 for FINDSITEcomb2.0.
• Superiority Over Existing Methods: Outperforms deep learning and similarity/docking methods including AtomNet, ECFP4, Surflex-Dock v.3066 on LIT-PCBA, and the boosted tree regression-based vScreenML on a DEKOIS 2.0 subset.
• Experimental Validation: Experimental testing identified new nanomolar binders for DHFR and ACVR1 and revealed a kinase inhibitor binding a novel DHFR subpocket, while yielding more hits and broader chemical space exploration than FINDSITEcomb2.0.

Scientific Applications:

• Early-stage virtual ligand screening (VLS): Improves precision and recall in VLS to prioritize lead compounds in early drug discovery.
• Novel binding interaction discovery: Predicts subpocket-specific binders and novel binding interactions to expand opportunities for therapeutic agent discovery, exemplified by DHFR and ACVR1 hits.
• Comparative benchmarking: Supports comparative assessment against methods such as AtomNet, ECFP4, Surflex-Dock v.3066, and vScreenML on benchmark sets including DUD-E, LIT-PCBA, and DEKOIS 2.0.

Methodology:

Predicts low-resolution protein structures, computes molecular-fingerprint-based ligand fragment scores mapped to stereochemically conserved subpockets, and combines those fragment scores with FINDSITEcomb2.0 global ligand similarity using a boosted tree regression model.