Multitask CapsNet: An Imbalanced Data Deep Learning Method for Predicting Toxicants.

Area of Science:

• Computational toxicology
• Machine learning in drug discovery
• cheminformatics

Background:

• Drug development faces high failure rates, with compound safety and toxicity prediction being major challenges.
• Existing in silico toxicity prediction models struggle with imbalanced datasets, leading to poor accuracy.
• Multitask learning and capsule neural networks have shown promise in improving predictive performance.

Purpose of the Study:

• To develop a novel multitask framework based on capsule neural networks (multitask CapsNet) for simultaneous prediction of 12 toxic effects.
• To evaluate the performance of multitask CapsNet in overcoming data imbalance issues in toxicity prediction.
• To compare the efficacy of multitask CapsNet against other computational approaches.

Main Methods:

• Development of a multitask capsule neural network (multitask CapsNet) architecture.
• Training and evaluation of the model on the Tox21 Data Challenge dataset, known for its imbalanced sample ratios.
• Comparative analysis of multitask CapsNet performance against single-task and other multitask computational methods.

Main Results:

• Multitask CapsNet demonstrated superior performance in predicting 12 different toxic effects compared to other computational methods.
• The model achieved significantly improved prediction accuracy on imbalanced datasets, a common issue in toxicity prediction.
• Multitask CapsNet achieved the highest accuracy ratio (8/12) on the Tox21 Data Challenge and predicted NR.PPAR.gamma toxicity with 96.6% accuracy, despite a 36:1 negative-to-positive sample ratio.

Conclusions:

• Multitask CapsNet effectively overcomes data bias challenges in toxicity prediction.
• The proposed model offers a novel, accurate, and efficient computational approach for predicting compound toxicities.
• This method holds significant potential for reducing risks and improving the success rate in drug development.