QSAR models for predicting key hormesis parameters of quinolone antibiotic mixtures on Vibrio qinghaiensis sp.-Q67

Antibiotics, as emerging pollutants, are increasingly detected in various water bodies at low-doses. The hormesis effect observed at these low-doses presents a challenge for toxicity prediction. Accurately predicting the key parameters of the hormesis effect is crucial. However, current methods for predicting the key parameters of hormesis mixtures (EC_min and ZEP) are limited. This study introduces machine learning-based QSAR (quantitative structure-activity relationship) models designed to predict these parameters. We conducted a binary mixture toxicity experiment using 10 quinolone antibiotics, with Q67 as the indicator organism, to obtain experimental data. Molecular structure descriptors of the antibiotics were calculated, and the optimal descriptors were selected. Additionally, molecular docking was used to convert the relative 3D conformation of antibiotic-protein complexes into SMILES strings. QSAR models were developed using the GA-MLR (genetic algorithms multivariate linear regression) method and the Transformer-CNN (Transformer model and convolutional neural network) method with the mixture descriptors and SMILES strings as independent variables and the toxic effect values (EC₅₀, EC_min, and ZEP) as dependent variables. The models were validated internally and externally, demonstrating reliable prediction of the toxic effect values of EC₅₀, EC_min, and ZEP at three different exposure times (4 h, 12 h, and 24 h), model quality is better with longer exposure times. The QSAR model exhibited strong internal stability and external predictive ability. A comparison of the two modelling approaches showed that the Transformer-CNN method produced QSAR models with a coefficient of determination (R²) ranging from 0.8458 to 0.9853, and a root mean square error (RMSE) ranging from 0.0409 to 0.1496, indicating higher accuracy in predicting time-dependent toxicity. This study offers a novel approach to exploring and predicting the key parameters of the hormesis effect.