Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development.

Area of Science:

• Computational chemistry
• Pharmacology
• Data science

Background:

• Artificial Intelligence (AI) and Machine Learning (ML) are increasingly used in drug development.
• Interpreting complex AI/ML model predictions remains a significant challenge, hindering clinical adoption.
• Lack of transparency in AI/ML models limits trust and effective decision-making.

Purpose of the Study:

• To provide a practical guide to SHapley Additive exPlanations (SHAP) for interpreting AI/ML models.
• To enhance the transparency and trustworthiness of AI/ML models in drug development.
• To facilitate deeper understanding and clinical application of AI/ML predictions.

Main Methods:

• Focus on SHAP, a feature-based interpretability method for supervised ML models.
• Tutorial covers application to regression and classification problems.
• Demonstrates SHAP analysis on standard ML black-box and inherently explainable models.

Main Results:

• Overview of SHAP visualization plots and their interpretation.
• Discussion of available software for SHAP implementation.
• Highlights best practices and considerations for binary endpoints and time-series models.

Conclusions:

• SHAP analysis offers a practical approach to interpreting AI/ML models in drug development.
• Enhanced model interpretability through SHAP can improve clinical decision-making.
• Ongoing advancements aim to address current limitations of SHAP.