Reducing false positive rate of docking-based virtual screening by active learning.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Machine learning-based scoring functions (MLSFs) offer superior performance over classical methods in virtual screening.
• Existing MLSFs often suffer from biased or poorly characterized negative data, impacting their accuracy.
• The quality of inactive molecule datasets is crucial for developing robust MLSFs.

Purpose of the Study:

• To develop an easy-to-use method (AMLSF) that iteratively enhances MLSFs by improving the quality of inactive sets.
• To reduce the false positive rate in virtual screening through active learning and negative molecular selection.
• To validate the efficacy of AMLSF in identifying active molecules across multiple targets.

Main Methods:

• Proposed an active learning method for MLSFs (AMLSF) incorporating negative molecular selection strategies.
• Utilized energy auxiliary terms learning as the MLSF.
• Validated AMLSF on eight targets from the DUD-E dataset, screening the IterBioScreen database and comparing with control models.

Main Results:

• AMLSF identified a significantly higher number of active molecules in the top 1000 screened compounds compared to control models.
• Free energy calculations for top-ranked molecules confirmed AMLSF's ability to identify more active compounds.
• The proposed method demonstrated a reduction in the false positive rate for virtual screening.

Conclusions:

• AMLSF effectively improves the quality of inactive datasets, leading to enhanced MLSF performance.
• The method offers a practical approach to reduce false positives and increase the hit rate in virtual screening.
• AMLSF represents a valuable advancement for efficient drug discovery and lead identification.