Comparing structural fingerprints using a literature-based similarity benchmark.

Area of Science:

• Cheminformatics
• Medicinal Chemistry
• Drug Discovery

Background:

• Molecular similarity is a core concept in cheminformatics but lacks objective assessment.
• A practical definition is proposed: molecules are similar if a medicinal chemist would test them concurrently in a drug discovery program.
• This definition leverages medicinal chemistry literature, assuming molecules in the same activity table are considered similar.

Purpose of the Study:

• To develop a practical definition of molecular similarity for drug discovery.
• To create benchmark datasets for evaluating structural similarity measures.
• To assess the performance of various molecular fingerprints.

Main Methods:

• Defined molecular similarity based on medicinal chemists' likely testing strategies.
• Constructed benchmark datasets from ChEMBL activity data and medicinal chemistry literature.
• Evaluated 28 different molecular fingerprints using these benchmark datasets.

Main Results:

• Created two sensitive benchmark datasets for structural similarity, outperforming virtual screening results.
• Identified Extended-connectivity fingerprints (ECFP) with diameter 4 and 6, and topological torsion fingerprints as top performers for diverse structures.
• Atom pair fingerprints excelled at ranking very close analogues.

Conclusions:

• ECFP fingerprints show significantly improved performance with increased bit-vector length (1024 to 16,384) for ranking diverse structures and virtual screening.
• The developed benchmark datasets are valuable for guiding the development of improved molecular similarity measures.
• Specific fingerprints are recommended based on whether ranking diverse structures or close analogues.