Interpretation of Compound Activity Predictions from Complex Machine Learning Models Using Local Approximations and...

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Machine learning (ML) models are crucial for identifying structural patterns in structure-activity relationship (SAR) studies.
• Understanding the decision-making process of complex ML models, especially deep learning (DL) architectures, is challenging but vital for drug design.
• Interpreting ML results offers an additional layer of model validation through expert knowledge.

Purpose of the Study:

• To introduce Shapley additive explanations (SHAP) as a method for rationalizing activity predictions from any ML algorithm.
• To demonstrate the application of SHAP in interpreting Random Forest (RF), Support Vector Machine (SVM), and Deep Neural Network (DNN) models.
• To identify and map structural patterns that influence predicted compound activity.

Main Methods:

• Application of Shapley additive explanations (SHAP) to interpret ML models.
• Training and evaluation of RF, nonlinear SVM, and DNN models for SAR prediction.
• Mapping identified structural patterns onto test compounds to visualize their contribution to activity predictions.

Main Results:

• SHAP successfully rationalized predictions from diverse ML algorithms, including RF, SVM, and DNN.
• Key structural features influencing compound activity predictions were identified and visualized.
• The interpretability provided by SHAP enhances the understanding of ML model behavior in SAR studies.

Conclusions:

• SHAP is a powerful and versatile tool for explaining the predictions of complex ML models in SAR research.
• The method facilitates a deeper understanding of structure-activity relationships, aiding in rational compound design.
• SHAP offers significant potential for validating and refining ML models used in drug discovery.