Enhancing toxicity prediction for natural products in herbal medicine and dietary supplements: Integrating (Q)STR models and in vitro assays.

New approach methods (NAMs) are required to predict human toxicity effectively, particularly due to limitations in conducting in vivo studies. While NAMs have been established for various industries, such as cosmetics, pesticides, and drugs, their applications in natural products (NPs) are lacking. NPs' complexity (multiple ingredients and structural differences from synthetic compounds) complicates NAM development. In this study, we devised NAMs for NPs using (quantitative) structure-toxicity relationship (Q)STR models and in vitro assays. Validation involved testing each method with single compounds isolated from NPs. A linear regression model was developed for (Q)STR prediction (R on test set: 0.52), with an applicability domain analysis demonstrating its reliability across NPs. This model was applied to predict the LD range of species, aiding in the development of herbal medicine and dietary supplements. In vitro screening employed three reporter cell lines (AP-1, P53, and Nrf2), with Tox scores derived by integrating in silico and in vitro data. Nimbolide exhibited the highest Tox score, with experimental studies corroborating the accuracy and reliability of the predictions made via Tox score analysis. The findings of the study align well with the purpose, as the suggested NAMs, utilizing (Q)STR models and in vitro assays, provide a Tox score to efficiently prioritize NPs for herbal medicine and dietary supplements.