Multitask Learning for Six-Pack Toxicity Prediction

The assessment of the six-pack toxicity, the crucial six systems and organ toxicities, is vital for ensuring the safe use of chemicals.Computational models capable of providing reliable predictions are acceptable for regulatory use to replace animal testing.However, data scarcity issues hindered the development of prediction models.This study proposed the first application of multitask learning to the six-pack toxicity for addressing data scarcity issues.Five algorithms were implemented and compared.Results showed that the distinct chemical space of tasks impedes the learning of shared representation of conventional algorithms, with performance worse than baseline models.In contrast, the MTForestNet algorithm built on a biological readacross concept performed best, with 3.1% and 3.3% improvement on AUC and accuracy, respectively.These findings demonstrate that biologically informed multitask learning can effectively overcome data scarcity and enhance toxicity prediction.