Constraint aware and uncertainty informed QSAR modeling of Tetrahymena pyriformis toxicity using physicochemical descriptors.

Conventional Quantitative Structure Activity Relationship (QSAR) models for predicting aquatic toxicity typically focus on accuracy at the expense of uncertainty awareness, physicochemical consistency, and chemical similarity leakage risks. Hence, the predictions are unstable when tested over different chemical spaces. This paper presents a machine-learning framework to predict the toxicity of Tetrahymena pyriformis based on a dataset of 1792 organic compounds whose 40-h growth inhibition (IGC50) values were experimentally determined and converted to pIGC50 for the models. The goal is to generate trustworthy and explainable results by using only eight molecular descriptors that are physicochemically relevant as model inputs. The proposed framework combines monotonicity-constrained learning, predictive uncertainty estimation, and intervention-inspired sensitivity analysis in one pipeline. The model’s reliability has been tested by cluster-aware data splitting, cross-validation, external testing, Y-scrambling, and applicability-domain assessment. The final method is capable of producing consistent, explainable, and risk-aware predictions and can be used as a reliability-centered basis for QSAR-initiated aquatic toxicity screening of chemically diverse compounds.