Interpretation of machine learning models using shapley values: application to compound potency and multi-target...

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Machine learning (ML) interpretability is crucial for pharmaceutical research, but complex models like deep neural networks (DNNs) pose challenges.
• Existing methods often lack universality, limiting their application across diverse ML architectures and ensembles.

Purpose of the Study:

• To evaluate and extend the SHapley Additive exPlanations (SHAP) methodology for ML model interpretability in pharmaceutical research.
• To compare a variant for exact Shapley value calculation in decision trees against the model-independent SHAP approach.
• To demonstrate new applications of SHAP for interpreting DNNs and ensemble models.

Main Methods:

• Application and extension of the SHapley Additive exPlanations (SHAP) methodology.
• Investigation of a variant for exact Shapley value computation in decision tree models.
• Systematic comparison of SHAP variants in compound activity and potency predictions.
• Interpretation of deep neural network (DNN) models for multi-target activity profiling.
• Analysis of ensemble regression models for potency prediction.

Main Results:

• SHAP provides a unified approach to identify and prioritize features driving ML predictions in compound activity and classification.
• The evaluated SHAP variant offers accurate Shapley value computation for decision tree methods.
• SHAP successfully interprets complex DNNs for generating multi-target activity profiles and ensemble models for potency prediction.

Conclusions:

• SHAP methodology significantly enhances the interpretability and trustworthiness of ML models in pharmaceutical research.
• The SHAP approach is versatile, applicable to various ML models including DNNs and ensembles.
• SHAP facilitates a deeper understanding of structure-activity relationships, aiding in drug discovery and development.