Responsible AI for Toxicology
RISE is developing Responsible AI systems that integrate scientific logic, causal reasoning, and interpretability to ensure trustworthy toxicological hazard screening, moving beyond simple statistical correlations.
Traditional AI/ML models in toxicology often rely on statistical correlations, which can overlook the underlying mechanisms that drive chemical behavior. At RISE, we are developing Responsible AI systems that integrate scientific logic, causal reasoning, and interpretability to ensure trustworthy toxicological hazard screening.
Our Approach: Three Pillars
Our approach combines three pillars:
Logic-based Reasoning
We embed expert-defined domain-specific rules into the model’s loss function. These rules penalize biologically implausible predictions, grounding AI in established toxicological knowledge.
Causality-aware Modeling
Correlation is not causation. By integrating causal constraints, the model focuses on relationships that reflect real cause-and-effect behavior rather than spurious correlations.
Interpretability
Using Integrated Gradients and rule-alignment checks, the model reveals how molecular features and mechanistic triggers influence predictions, ensuring domain rules are respected and causal behavior is transparent.
The Responsible AI Framework
Together, these components form a Responsible AI framework that is both accurate and scientifically aligned.
The model achieves competitive performance with AUC scores of 0.85 (training), 0.82 (validation), and 0.87 (external test), as demonstrated in a biodegradation case study, while still maintaining full mechanistic interpretability.
Impact
This work forms part of the Mistra SafeChem program and reflects RISE commitment to AI that enhances innovation safely, transparently, and responsibly.
