Bioaccumulation parameters of chemicals were necessary data for chemical risk assessment and management.1,2In silico methods, e.g., quantitative structure-activity relationship (QSAR) models can be valid alternatives to costly data measurement.3 In this study, QSAR models for simultaneously predicting bioconcentration factor (BCF) and biomagnification factor (BMF) values of chemicals in fish were constructed by using Dragon descriptors and 12 molecule fingerprints, as well as artificial neural network (ANN) algorithm. Single-task (ST) neural network models based on the backpropagation were established to predict BCF and BMF values, respectively. Appropriate descriptors were screened in the ST models for the development of multi-task (MT) learning with single-input-multiple-output (SIMO-MT) models and multi-input-multiple-output (MIMO-MT) models. Results showed that BCF and BMF values can be simultaneously predicted by MT models. The prediction performance of MT models was generally better than that of ST models, indicating that shared and generalized features can be learned from BCF and BMF, and jointly improve the predictive performance. Compared with SIMO-MT models, better predictive ability was observed in the MIMO-MT model. It can be inferred that the simultaneous input of multiple molecular fingerprints/descriptors can provide more learnable information, which further improves the performance of the MIMO-MT model. The application domain of the MIMO-MT model was characterized based on Tanimoto similarity of molecules, and setting an appropriate application domain range can improve the performance of QSAR models. With MT learning strategies and the case studies, better understanding of practical use of ANN in risk assessment and management of chemicals were provided, and the models developed in the present study can be valuable tools to predict the bioaccumulation parameters.
搜索
复制