Reassessing SHAP-based interpretations in QSAR: Model-centric limits and unsupervised alternatives for fluorocarbon...

Ye et al. (2025) report strong QSAR performance for fluorocarbon inhalation toxicity using 'SVM-RBF' and 'XGBoost', complemented by SHAP analyses to identify influential molecular descriptors. While predictive accuracy and generalization are commendable, the interpretability claims warrant caution. Supervised models possess two distinct accuracies-target prediction and feature-importance reliability-the latter lacking ground truth validation. Consequently, SHAP, as a model-dependent explainer, can faithfully reproduce and even amplify model biases, is sensitive to model specification, struggles with correlated descriptors, and does not infer causality. High accuracy does not guarantee reliable importances. We recommend augmenting the pipeline with unsupervised, label-agnostic descriptor prioritization (e.g., 'feature agglomeration', 'highly variable feature selection') followed by non-targeted association screening (e.g., Spearman correlation with p-values) to improve stability and mitigate model-induced interpretative errors.