Externally validated and explainable QSAR for ready biodegradability: KNIME–H2O AutoML with RDKit descriptors and Morgan fingerprints

Ready biodegradability is an important endpoint for environmental chemical assessment. We developed an externally validated QSAR workflow using KNIME, RDKit descriptors, Morgan fingerprints and H2O AutoML to classify readily biodegradable (RB) and not readily biodegradable (NRB) compounds. The benchmark contained training (n = 837), internal test (n = 218) and external validation (n = 670) subsets. Preprocessing and RF feature screening reduced 119 RDKit descriptors and 1024 Morgan bits to 311 predictors. Four model families were evaluated: GLM, Random Forest, Gradient Boosting (GB) and a neural network. Reported external ROC–AUC values were approximately 0.91 for the tree ensembles. Recalculation from the deposited GB predictions gave ROC–AUC = 0.911, accuracy = 0.866, RB recall = 0.754 and MCC = 0.668. A paired bootstrap of the external prediction table gave a 95% ROC–AUC interval of 0.886–0.935. Global and local explanation outputs highlighted surface-property descriptors, topology and selected fingerprint bits. Direct comparison of feature ranks distinguished broadly shared signals, such as smr_VSA10, from method-dependent rankings. Applicability-domain assessment and conservative interpretation of hashed fingerprint bits provide additional context for screening. The workflow supports prioritization and model auditing; it does not replace experimental biodegradability testing or establish causal degradation mechanisms.