From Chemical Scaffold to Predicted Toxicity: A Scaffold-Aware Pipeline for Reliable In Silico Prioritization of...

In silico methods generate numerous toxicity predictions for natural compounds, yet their mutual consistency and transferability to new structures are rarely verified, and the predictions are often taken at face value. Using furanocoumarins of the genus Heracleum as a model class of natural products, we developed a reproducible computational pipeline that evaluates not the predictions themselves but their reliability. Two datasets were analyzed: an extended reference set (DS1, 2008 compounds from PubChem) and a taxonomically verified natural set (DS2, 102 Heracleum compounds). The pipeline combined scaffold analysis, molecular docking against 44 off-target proteins, redocking, prediction of ten phenotypic toxicity endpoints, scaffold-resolved separability testing, and scaffold-aware machine learning. The chemical space was concentrated around two scaffolds, and the descriptor space was non-randomly organized with respect to them along all 13 axes (false discovery rate (FDR)-adjusted p < 0.01), indicating that the predicted toxicological profile is determined predominantly by molecular structure. The mean docking score corresponded to only the first principal component of the multidimensional binding profile, and the ten endpoints were largely independent of one another. Critically, scaffold separability did not predict transferability: structurally determined signals were only partially reproduced on unseen scaffolds, and to differing degrees across endpoints. A multi-criteria prioritization identified 21 top-priority candidates for experimental follow-up. The reliability of in silico assessments for this class cannot be assumed without scaffold-aware validation.