Benchmarking Data Sets from PubChem BioAssay Data: Current Scenario and Room for Improvement

Area of Science:

• Cheminformatics
• Computational Chemistry
• Drug Discovery

Background:

• Traditional virtual screening datasets like DUD, DUD-E, and DEKOIS have limitations, notably the lack of experimental validation for inactive compounds, potentially causing false negatives.
• The PubChem BioAssay database offers a valuable resource for constructing realistic screening datasets due to its repository of experimental bioactivity data.
• However, direct utilization of PubChem BioAssay data presents specific challenges that require careful consideration and methodological adjustments.

Purpose of the Study:

• To provide an overview of existing benchmarking datasets derived from PubChem BioAssay experimental data.
• To discuss critical issues and considerations for designing new ligand sets using PubChem BioAssay.
• To guide future efforts in developing improved evaluation tools for virtual screening methods.

Main Methods:

• Literature review of existing benchmarking datasets built on PubChem BioAssay.
• Analysis and discussion of challenges associated with using PubChem BioAssay for dataset construction.
• Identification of key considerations for future dataset design.

Main Results:

• Existing datasets derived from PubChem BioAssay are reviewed, highlighting their strengths and weaknesses.
• Several critical issues in data curation and interpretation from PubChem BioAssay are identified.
• Recommendations for addressing these issues are proposed to enhance dataset quality.

Conclusions:

• PubChem BioAssay is a promising source for realistic virtual screening datasets, but requires careful handling.
• Addressing the discussed challenges will lead to more robust datasets for evaluating virtual screening methods.
• This work aims to facilitate the development of superior in silico screening evaluation tools.