Learning-to-rank technique based on ignoring meaningless ranking orders between compounds.

Area of Science:

• Drug Discovery and Development
• Computational Chemistry
• Bioinformatics

Background:

• Ligand-based virtual screening is crucial for reducing drug development costs.
• The accuracy of virtual screening models relies heavily on the quality of experimental compound activity data.
• Large errors in activity values can introduce meaningless order relations, hindering effective screening.

Purpose of the Study:

• To develop a novel learning-to-rank method for virtual screening that addresses the issue of meaningless ranking orders.
• To improve the accuracy of rank prediction models in ligand-based virtual screening.
• To enhance the efficiency and reliability of early-stage drug discovery processes.

Main Methods:

• Developed a new learning-to-rank algorithm specifically designed to ignore irrelevant ranking orders.
• The method identifies and excludes rankings between compounds with highly similar activity values.
• It also excludes rankings between compounds identified as inactive.
• Evaluated the method using five high-throughput screening assay datasets from the PubChem BioAssay database.

Main Results:

• The proposed learning-to-rank method significantly improved the accuracy of virtual screening predictions.
• Ignoring meaningless ranking orders effectively addressed a key challenge in virtual screening.
• The method demonstrated enhanced predictive performance across multiple datasets.
• Validation on PubChem BioAssay datasets confirmed the method's effectiveness.

Conclusions:

• The novel learning-to-rank approach successfully enhances virtual screening accuracy by filtering out noise.
• This simple yet effective method facilitates more reliable and accurate drug candidate identification.
• The developed technique offers a valuable tool for optimizing the drug discovery pipeline.
• The source code is publicly available for further research and application.