A CRISP approach to QSP: XAI enabling fit-for-purpose models

Simpler models can be more trustworthy: The CRISP workflow reduces overly complex drug development models by removing irrelevant parameters, making it easier for scientists and regulators to understand how predictions are made and why—a key principle of responsible AI.

Explainability improves drug safety: By clarifying the causal structure of biological models through the Manifold Boundary Approximation Method (MBAM), the approach helps identify which factors actually matter for predicting clinical outcomes, reducing the risk of hidden assumptions leading to unsafe treatments.

Fit-for-purpose design prevents misuse: Rather than creating one-size-fits-all models, CRISP tailors model complexity to specific questions (personalized medicine, toxicology, target discovery), preventing the misapplication of overly general models to contexts where they weren't validated.

Computational efficiency enables broader access: The resulting streamlined models are computationally efficient, potentially making advanced drug development tools more accessible to researchers and organizations with limited computational resources, democratizing responsible innovation.

Validated on real biological systems: The approach was tested on complex real-world cases (coagulation cascade and SHIV infection), demonstrating that explainability and model reduction don't sacrifice scientific rigor—a critical assurance for stakeholders evaluating AI in healthcare.