CapsCarcino: A novel sparse data deep learning tool for predicting carcinogens.

Area of Science:

• Computational toxicology
• Drug discovery
• Machine learning

Background:

• Early identification of chemical carcinogenicity is crucial in drug discovery to mitigate human health risks.
• Existing computational methods for predicting carcinogenicity have limitations in predictive power.
• There is a growing need for improved computational approaches in this field.

Purpose of the Study:

• To develop a novel deep learning architecture, CapsCarcino, for distinguishing between chemical carcinogens and noncarcinogens.
• To enhance the accuracy and efficiency of carcinogenicity prediction in early drug development.

Main Methods:

• Developed CapsCarcino, a deep learning model utilizing a dynamic routing algorithm.
• Compared CapsCarcino's performance against five other machine learning models.
• Evaluated the model's predictive and generalization abilities on an external validation dataset.

Main Results:

• CapsCarcino demonstrated significantly improved predictive and generalization abilities compared to existing models.
• The best CapsCarcino model achieved 85.0% accuracy on an external validation dataset.
• CapsCarcino maintained high performance even when trained on only 20% of the data, outperforming other methods using full datasets.

Conclusions:

• CapsCarcino offers a robust and efficient solution for carcinogen risk assessment in early drug discovery.
• The model effectively learns carcinogen characteristics, even with limited structural alert representation.
• CapsCarcino shows potential for improving the safety and efficacy of newly developed drugs.