Co-model for chemical toxicity prediction based on multi-task deep learning.

Area of Science:

• Medicinal Chemistry
• Pharmacology
• Computational Toxicology

Background:

• Accurate prediction of compound toxicity is vital for drug development's effectiveness and safety.
• Predicting toxicity remains a significant challenge in medicinal chemistry and pharmacology.

Purpose of the Study:

• To construct and validate single- and multi-task models for compound toxicity prediction.
• To evaluate the performance of models based on diverse molecular representations and algorithms.

Main Methods:

• Development of three model types: single-task and multi-task learning.
• Utilizing 2D/3D descriptors, molecular fingerprints, and molecular graphs as input features.
• Validation using benchmark tests on the Tox21 dataset.

Main Results:

• Multi-task learning effectively addressed data imbalance issues in the Tox21 dataset.
• Multi-task models generally showed significantly improved prediction performance over single-task models.
• An integrated Co-Model, combining diverse representations and algorithms, outperformed existing literature models.

Conclusions:

• Multi-task learning enhances compound toxicity prediction accuracy and robustness.
• The developed Co-Model demonstrates superior predictive power and robustness for toxicity assessment.
• This approach offers a promising strategy for improving drug safety and development.