Critical comparison of virtual screening methods against the MUV data set

Area of Science:

• Computational chemistry
• Drug discovery
• cheminformatics

Background:

• Ligand-based virtual screening (LBVS) is crucial for identifying potential drug candidates.
• Evaluating the performance of LBVS tools is essential for optimizing drug discovery pipelines.

Purpose of the Study:

• To assess the performance of five similarity search and two pharmacophore elucidation tools using the MUV dataset.
• To investigate the impact of template selection strategies (chemical diversity vs. random) on tool performance.
• To evaluate the effect of data fusion on virtual screening outcomes.

Main Methods:

• Tested seven LBVS tools against the MUV dataset.
• Employed single active molecules and diverse active compound sets as templates for similarity searches.
• Utilized random and chemically diverse active molecules for pharmacophore model building.
• Calculated performance metrics such as enrichment factors and AUC values.
• Analyzed the influence of data fusion on screening results.

Main Results:

• Chemically diverse template sets yielded superior results for similarity search methods.
• Randomly selected training sets were optimal for pharmacophore elucidators, highlighting automation limitations.
• Data fusion strategies showed varied effects on overall performance.

Conclusions:

• The choice of template selection strategy significantly impacts LBVS tool performance.
• Pharmacophore modeling requires careful, non-automated training set selection.
• Activity cliffs pose a challenge in LBVS and warrant further investigation for benchmark set improvement.