DeepCarc: Deep Learning-Powered Carcinogenicity Prediction Using Model-Level Representation.

Area of Science:

• Computational chemistry
• Toxicology
• Drug discovery

Background:

• Carcinogenicity testing is crucial but time-consuming using traditional animal studies.
• There's a need for faster, reliable methods to assess carcinogenic potential.
• Environmental chemistry and drug development require efficient carcinogen identification.

Purpose of the Study:

• To develop a deep learning model for predicting carcinogenicity of small molecules.
• To provide a robust and efficient alternative to conventional carcinogenicity testing.
• To assess the performance of the DeepCarc model against existing methods.

Main Methods:

• Developed the DeepCarc model using deep learning and model-level representations.
• Trained and evaluated the model on the National Center for Toxicological Research liver cancer database (NCTRlcdb).
• Compared DeepCarc performance against four advanced deep learning quantitative structure-activity relationship (QSAR) models.

Main Results:

• The DeepCarc model achieved a Matthews correlation coefficient (MCC) of 0.432 on the test set.
• DeepCarc demonstrated a 37% average improvement over existing DL-QSAR models.
• The model was successfully applied to screen compounds from DrugBank and Tox21.

Conclusions:

• The DeepCarc model offers a promising early detection tool for carcinogenicity assessment.
• This deep learning approach significantly improves upon traditional methods.
• DeepCarc can accelerate carcinogen identification in environmental and pharmaceutical research.