Structural Analysis and Prediction of Hematotoxicity Using Deep Learning Approaches.

Area of Science:

• Computational toxicology
• Drug discovery and development
• Machine learning in pharmacology

Background:

• Hematotoxicity is a significant, yet underappreciated, adverse effect in drug discovery.
• Existing in silico models for predicting hematotoxicity are limited.

Purpose of the Study:

• To develop and validate a robust computational model for predicting drug-induced hematotoxicity.
• To identify key structural features associated with hematotoxicity.

Main Methods:

• Construction of a high-quality dataset of 759 hematotoxic and 1623 nonhematotoxic compounds.
• Development of classification models using seven machine learning algorithms and nine molecular representations.
• Application of SHAP and atom heatmap for feature interpretation and MMPA for structural analysis.

Main Results:

• The best model, Attentive FP, achieved a balanced accuracy (BA) of 72.6% and an area under the receiver operating characteristic curve (AUC) of 76.8% on the validation set.
• The model demonstrated superior performance on an external validation set with a BA of 67.5% compared to existing methods.
• Identification of critical structural fragments and features contributing to hematotoxicity.

Conclusions:

• The developed graph-based deep learning model offers a reliable tool for assessing hematotoxicity in early drug development.
• The model's interpretability provides valuable insights into the structural basis of hematotoxicity.
• This approach can significantly contribute to the development of safer therapeutic agents.