TOP: A deep mixture representation learning method for boosting molecular toxicity prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Molecule toxicity prediction is vital in early drug discovery to prevent clinical trial failures.
• Existing methods often rely on limited molecular representations (SMILES or physiochemical properties).

Purpose of the Study:

• To develop a novel deep learning framework, TOP (TOxicity Prediction), for enhanced molecular toxicity prediction.
• To create a new molecular representation method that combines structural and property information.

Main Methods:

• Developed TOP, a deep learning framework integrating data preprocessing, a Bidirectional Gated Recurrent Unit (BiGRU) RNN, and fully connected neural networks.
• TOP learns a mixed molecular representation from SMILES contextual information and physiochemical properties.
• Utilized end-to-end learning for toxicity prediction.

Main Results:

• TOP demonstrated significantly improved toxicity prediction accuracy across 14 tasks and 3 benchmark datasets.
• Outperformed baseline machine learning and five state-of-the-art methods.
• Effectively handled both balanced and imbalanced datasets.

Conclusions:

• The novel molecular representation method integrated into TOP enhances toxicity prediction accuracy.
• TOP offers a robust and effective deep learning solution for early-stage drug candidate evaluation.
• This approach can accelerate drug discovery by improving the reliability of toxicity assessments.