Mitochondrial toxic prediction of marine alga toxins using a predictive model based on feature coupling and ensemble learning algorithms.

Alga toxins have recently emerged as environmental risk factors to multiple human health issues. Mitochondrial toxicity is an essential element in the field of ecotoxicology, it is necessary to screen and manage mitochondrial toxicants from common alga toxins. To overcome the limitations of traditional animal and cell experiments, computational toxicology is increasingly emphasized. In this study, all the publicly available datasets were compiled to create the largest mitochondrial toxicity dataset to date, establishing a robust and high-performance QSAR screening model. The model couples and filters 12 molecular fingerprints and 318 descriptors as features, capturing more information about molecular structure and properties. By comparing 8 machine learning algorithms and using a weighted soft voting method to integrate the two optimal algorithms, we established 108 prediction models and identified the best ensemble learning model MACCS_LK for screening and defining its application domain. Additionally, the efficacy of MACCS fingerprints in representing mitochondrial toxicants was established, and a mechanistic analysis of the identified model based on the SHAP method and 11 structural alerts uncovered in this study was conducted, enhancing the interpretability of this model. This study highlights the key roles of lipophilic structures such as aromatic rings and long hydrocarbon chains and their related physicochemical properties in predicting toxicity outcomes. The mitochondrial toxicity of six algal toxins was predicted by employing this model, and the results indicating that two of them possess mitochondrial toxic effects. This model has high reliability and accuracy, making it applicable for predicting mitochondrial toxicity of more marine biotoxins.