MPNN-CWExplainer: An enhanced deep learning framework for HIV drug bioactivity prediction with class-weighted loss...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Human Immunodeficiency Virus (HIV) poses a significant global health challenge, necessitating advanced therapeutic strategies beyond current antiretroviral therapies due to drug resistance and viral mutations.
• Developing novel treatments requires improved methods for predicting molecular bioactivity and understanding the underlying structure-activity relationships.

Purpose of the Study:

• To enhance the accuracy of predicting Human Immunodeficiency Virus (HIV) bioactivity using a novel deep learning framework.
• To provide interpretable insights into the molecular determinants that influence HIV bioactivity, aiding medicinal chemists.

Main Methods:

• A graph-based deep learning framework, MPNN-CWExplainer, was developed, integrating a Message Passing Neural Network (MPNN) with a class-weighted loss function.
• GNNExplainer was employed for post-hoc interpretability, identifying critical atom and bond substructures contributing to bioactivity predictions.
• Bayesian hyperparameter optimization and multiple independent runs were utilized to ensure model robustness.

Main Results:

• MPNN-CWExplainer achieved state-of-the-art performance on the HIV dataset, with AUC-ROC of 87.63% and AUC-PRC of 86.02%.
• The class-weighted approach improved the representation of the minority class within the dataset.
• GNNExplainer successfully identified chemically relevant substructures associated with bioactivity.

Conclusions:

• The MPNN-CWExplainer framework offers improved accuracy and interpretability for HIV bioactivity prediction, crucial for computational drug discovery.
• This interpretable tool supports medicinal chemists in understanding model predictions and making informed decisions during lead optimization and molecular design.