Large-scale prediction of activity cliffs using machine and deep learning methods of increasing complexity

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Activity cliffs (AC) are structural analogues with large potency differences against the same target.
• Previous AC prediction studies used varied methods and limited compound classes, hindering direct comparisons.
• AC prediction requires analyzing compound pairs as ACs or non-ACs, unlike typical compound classification.

Purpose of the Study:

• To conduct a large-scale comparison of machine learning methods for activity cliff prediction.
• To evaluate prediction accuracy across 100 compound activity classes.
• To identify factors influencing AC prediction performance.

Main Methods:

• Compared machine learning methods of varying complexity (nearest neighbor, decision trees, kernel methods, deep neural networks).
• Executed predictions across 100 compound activity classes.
• Analyzed the influence of methodological complexity and compound memorization on accuracy.

Main Results:

• Prediction accuracy did not correlate with methodological complexity.
• Memorization of shared compounds between ACs and non-ACs significantly influenced prediction accuracy.
• Limited training data were often sufficient for accurate models, with no clear advantage for deep learning.
• Support vector machines showed the best overall performance, with small margins over other methods.

Conclusions:

• Systematic AC prediction provides insights into achievable accuracy across diverse compound classes.
• Methodological complexity is less critical than data characteristics like compound memorization for AC prediction.
• Simpler machine learning models can be as effective as complex ones, including deep learning, for AC prediction.