Extended Activity Cliffs-Driven Approaches on Data Splitting for the Study of Bioactivity Machine Learning...

Area of Science:

• Computational Chemistry
• Cheminformatics
• Quantitative Structure-Activity Relationship (QSAR) Modeling

Background:

• Activity Cliffs (ACs) are known to pose significant challenges in Quantitative Structure-Activity Relationship (QSAR) modeling.
• Machine Learning (ML) QSAR models are highly data-dependent and sensitive to the distribution of chemical data, including ACs.

Purpose of the Study:

• To investigate the impact of Activity Cliff distribution within training and test sets on ML QSAR model performance.
• To propose and evaluate extended similarity and extended Structure-Activity Relationship (SALI) methods for analyzing ACs.

Main Methods:

• Development and application of extended similarity metrics.
• Development and application of extended SALI (Structure-Activity Relationship) methods.
• Analysis of the influence of non-uniform versus uniform AC distribution on model errors.

Main Results:

• Non-uniform distribution of Activity Cliffs and chemical space leads to poorer model performance compared to uniform distributions.
• Machine Learning modeling on datasets rich in Activity Cliffs requires careful, case-by-case analysis.
• The proposed methods serve as valuable tools for dataset analysis.

Conclusions:

• Dataset splitting strategies significantly impact ML QSAR model performance, especially in the presence of Activity Cliffs.
• Uniform distribution of ACs and chemical space generally results in improved model quality.
• Random data splitting demonstrated superior generalization performance overall compared to other tested splitting alternatives.