Ties in proximity and clustering compounds.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Data Science

Background:

• Hierarchical clustering is widely used for compound selection and diversity analysis in cheminformatics.
• Common applications employ binary chemical structure representations (e.g., MACCS keys, Daylight fingerprints) and dissimilarity measures (e.g., Euclidean, Soergel).

Purpose of the Study:

• To investigate the 'ties in proximity' problem in hierarchical clustering of chemical compounds.
• To understand how this ambiguity affects compound selection and diversity analysis.
• To establish conditions under which ambiguous ties can be avoided.

Main Methods:

• Probabilistic argument to determine the theoretical limit for avoiding ambiguous ties.
• Analysis of common similarity measures and clustering algorithms on diverse chemical collections.
• Evaluation of the relationship between bit vector length, similarity measure distribution, and tie frequency.

Main Results:

• Hierarchical clustering algorithms are susceptible to ambiguous ties, where compounds or clusters are equidistant.
• The 'ties in proximity' problem is significant even for small compound sets (hundreds of compounds).
• Common similarity measures often produce statistically preferred proximity values, increasing the likelihood of ties.

Conclusions:

• The 'ties in proximity' problem poses a significant challenge for reliable hierarchical clustering in cheminformatics.
• Ensuring unambiguous clustering requires careful consideration of dataset size and the properties of similarity measures.
• The findings highlight the need for robust methods to mitigate ambiguity in compound selection and diversity analysis.