FRAGSITE: A Fragment-Based Approach for Virtual Ligand Screening.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Virtual ligand screening (VLS) is crucial for reducing drug discovery costs.
• Traditional VLS methods have limitations in structure-based and ligand-based approaches.
• Existing methods struggle with exploring diverse chemical spaces and identifying novel binders.

Purpose of the Study:

• To develop an improved VLS approach that overcomes limitations of existing methods.
• To enhance the precision and recall of VLS for identifying potential drug candidates.
• To leverage predicted protein structures and conserved ligand-binding subpockets.

Main Methods:

• Developed FRAGSITE, a structure-based VLS approach integrating molecular fingerprints and global ligand similarity scores.
• Utilized a boosted tree regression machine learning framework.
• Benchmarked FRAGSITE on DUD-E and LIT-PCBA datasets, excluding highly similar proteins.

Main Results:

• FRAGSITE improved VLS precision by 14.3% and recall by 18.5% on the DUD-E set compared to FINDSITE^comb2.0.
• Achieved a higher mean top 1% enrichment factor (30.2 vs. 25.2).
• Outperformed state-of-the-art deep learning and docking methods on benchmark datasets and identified novel nanomolar binders for DHFR and ACVR1 in experimental validation.

Conclusions:

• FRAGSITE represents a significant advancement in virtual ligand screening methodology.
• The approach effectively identifies novel small-molecule binders with diverse scaffolds.
• FRAGSITE offers a powerful, freely available tool for academic drug discovery research.