A deep learning based multi-model approach for predicting drug-like chemical compound's toxicity.

Area of Science:

• Computational chemistry and cheminformatics
• Pharmacology and toxicology
• Artificial intelligence in drug discovery

Background:

• Drug development is costly and time-consuming, with toxic compounds identified late in the process causing significant setbacks.
• Early and accurate prediction of compound toxicity is essential to mitigate risks and optimize resource allocation in small-molecule drug discovery.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting various types of compound toxicity.
• To integrate these models into a virtual screening pipeline for early identification of low-toxicity drug candidates.

Main Methods:

• Utilized graph convolutional networks (GCNs) for both regression (acute toxicity) and binary classification (carcinogenicity, hERG_cardiotoxicity, hepatotoxicity, mutagenicity) tasks.
• Employed diverse training strategies to handle variations in data size, label type, and distribution across different toxicity endpoints.
• Validated models using an approved drug dataset to establish prediction score thresholds.

Main Results:

• GCN regression model achieved notable performance for acute toxicity prediction (Pearson R: 0.76, 0.74, 0.65 for IP, IV, oral routes).
• GCN binary classification models demonstrated high predictive power with AUC scores ranging from 0.69 to 0.88 for carcinogenicity, hERG_cardiotoxicity, mutagenicity, and hepatotoxicity.
• Integrated models successfully identified potential low-toxicity drug candidates within a virtual screening pipeline.

Conclusions:

• Deep learning models offer a powerful approach for early and accurate prediction of compound toxicity in drug development.
• These models can significantly reduce the costs and risks associated with drug discovery by enabling faster selection of safer drug candidates.
• The developed models serve as valuable tools for virtual screening and prioritizing compounds with favorable toxicity profiles.