Explainable Machine Learning for Acute Toxicity QSAR Modeling: Balancing Predictive Accuracy, Descriptor Efficiency, and Regulatory Transparency

Acute toxicity assessment remains a central pillar of chemical safety evaluation, yet conventional reliance on in vivo testing raises well-documented ethical, financial, and temporal concerns that have driven sustained interest in computational alternatives. Quantitative structure-activity relationship (QSAR) modeling, particularly when combined with machine learning, has emerged as a leading strategy for predicting acute oral toxicity (LD ۵۰) without additional animal experimentation; however, the field has increasingly favored high-dimensional descriptor sets and architecturally opaque models that achieve strong predictive performance at the cost of interpretability and regulatory transparency. This review synthesizes the published literature on machine learning-based QSAR modeling of acute toxicity, with particular attention to descriptor generation strategies, feature selection methodologies, model architectures, and the integration of explainable artificial intelligence (XAI) techniques such as SHapley Additive explanations (SHAP). The synthesis reveals a persistent and largely unresolved tension between predictive accuracy and model transparency: architectures capable of capturing complex nonlinear structure-toxicity relationships typically require explainability to be appended post hoc to an already-optimized model, whereas frameworks designed for interpretability from inception tend to report comparatively lower or more variable external predictive performance. Comparative analysis of reported performance metrics, feature selection strategies, and applicability domain characterization across the reviewed studies indicates that descriptor efficiency, explainability, and external generalizability have thus far been pursued largely as independent objectives rather than co-optimized within a single validated framework. This gap suggests that the field's central challenge is not an intrinsic incompatibility between accuracy and interpretability, but the absence of design frameworks that treat descriptor parsimony and explainability as primary, jointly optimized modeling criteria.

کلیدواژه ها:

Acute Oral Toxicity ، LD۵۰ ، QSAR ، Explainable Artificial Intelligence ، SHAP ، Molecular Descriptors ، Feature Selection ، Machine Learning ، Computational Toxicology ، Applicability Domain