Systematic generation and analysis of counterfactuals for compound activity predictions using multi-task models.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Machine learning (ML) models often act as "black boxes," hindering trust and adoption in scientific research, particularly in drug discovery.
• The inability to understand ML predictions limits their utility in experimental design and interdisciplinary collaboration.
• Explainable Artificial Intelligence (XAI) offers methods to interpret ML models, with counterfactuals (CFs) being a promising approach.

Purpose of the Study:

• To adapt and extend a systematic method for generating molecular counterfactuals (CFs).
• To apply CF generation to multi-task predictions of protein kinase inhibitors.
• To analyze CFs in detail and explain their formation in multi-task modeling.

Main Methods:

• Systematic generation of molecular counterfactuals (CFs).
• Application of CF generation to multi-task prediction models for protein kinase inhibitors.
• Detailed analysis of generated CFs and their relationship to model predictions.

Main Results:

• Successfully generated molecular CFs for multi-task protein kinase inhibitor predictions.
• Analyzed CFs to identify structural features influencing predictions.
• Provided exemplary explanations of ML predictions using CFs, accessible to chemists.

Conclusions:

• Molecular CFs provide an intuitive and powerful tool for understanding ML predictions in medicinal chemistry.
• This approach enhances trust and facilitates the application of ML in drug discovery by rationalizing model behavior.
• The developed methods offer practical insights for chemists, bridging the gap between ML predictions and experimental validation.