Deep reinforcement learning enables better bias control in benchmark for virtual screening.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Artificial intelligence in medicine

Background:

• Virtual screening (VS) is integral to modern drug discovery.
• Machine learning (ML) is revolutionizing VS, but faces challenges with existing datasets.
• Current benchmarks suffer from limited data volume, narrow applicability domains, and inherent biases.

Purpose of the Study:

• Introduce MUBD^syn, a novel benchmark dataset for VS.
• Address limitations of existing benchmarks by incorporating synthetic decoys.
• Provide a robust platform for evaluating ML models in drug discovery.

Main Methods:

• Developed MUBD^syn utilizing synthetic decoys (presumed inactives).
• Employed deep reinforcement learning for bias control during decoy generation.
• Conducted extensive validation comparing MUBD^syn with classical benchmarks.

Main Results:

• MUBD^syn demonstrated superiority in controlling domain bias, artificial enrichment bias, and analogue bias.
• Assessment of ML models using MUBD^syn revealed reduced bias, including asymmetric validation embedding bias.
• MUBD^syn provides a more challenging and effective benchmark for deep learning models compared to NRLiSt-BDB.

Conclusions:

• MUBD^syn is a near-ideal benchmark for virtual screening.
• The benchmark effectively mitigates biases present in traditional datasets.
• MUBD^syn facilitates more accurate and reliable evaluation of ML models in drug discovery.
• The computational tool for MUBD^syn is publicly available.