Explaining Black Box Drug Target Prediction Through Model Agnostic Counterfactual Samples

Area of Science:

• Computational Biology
• Artificial Intelligence
• Drug Discovery

Background:

• High-performance drug-target affinity (DTA) deep learning models are often black boxes, limiting interpretability and trust.
• Explainable AI (XAI) techniques are crucial for understanding DTA model behavior and extracting biological knowledge.
• Counterfactual explanations offer insights by showing how model predictions change with input variations.

Purpose of the Study:

• To develop a novel multi-agent reinforcement learning framework, MACDA, for generating counterfactual explanations in DTA prediction.
• To provide human-interpretable counterfactual instances for drug-protein complexes.
• To simultaneously optimize both drug and target inputs for effective counterfactual generation.

Main Methods:

• Proposed a Multi-Agent Counterfactual Drug-target binding Affinity (MACDA) framework utilizing multi-agent reinforcement learning.
• Generated counterfactual explanations for drug-protein complexes.
• Benchmarked MACDA on the Davis and PDBBind datasets.

Main Results:

• MACDA produced more parsimonious explanations compared to existing methods, with no loss in explanation validity (measured by encoding similarity).
• The framework successfully generated human-interpretable counterfactual instances.
• A case study with ABL1 and Nilotinib demonstrated MACDA's ability to explain DTA model predictions and reveal underlying substructure interactions.

Conclusions:

• MACDA offers a powerful approach to enhance the interpretability and trustworthiness of DTA deep learning models.
• The framework can distill valuable biological knowledge by revealing input-prediction relationships.
• MACDA's counterfactual explanations align with and support existing domain knowledge in drug discovery.