Conservative consensus QSAR approach for the prediction of rat acute oral toxicity

Abstract

Consensus approaches are applied in different quantitative structure–activity relationship (QSAR) modeling contexts based on the assumption that combining individual model predictions will improve prediction reliability. This study evaluated the performance of TEST, CATMoS and VEGA models for prediction of oral rat LD₅₀, both individually and in consensus, across a dataset of 6,229 organic compounds. Predicted LD₅₀ values from the models were compared for each compound, and the lowest value was assigned as the output of the conservative consensus model (CCM). Predictive accuracy was then evaluated based on the agreement of predicted LD₅₀-based GHS category assignments with those derived experimentally. The aim was to allow for the most conservative value to be identified. Results showed that CCM had the highest over-prediction rate at 37 %, compared to TEST (24 %), CATMoS (25 %) and VEGA (8 %). Meanwhile, its under-prediction rate was lowest at 2 %, relative to TEST (20 %), CATMoS (10 %) and VEGA (5 %). Due to the method applied, CCM was the most conservative across all GHS categories. Further, structural analysis demonstrated that no specific chemical classes or functional groups were consistently underpredicted or overpredicted. The utility of CCM lies in its ability to establish a foundation for contextualizing the general use of consensus modeling, in order to derive health-protective oral rat LD₅₀ estimates under conditions of uncertainty, especially where experimental data are limited or absent.