Improving Predictive Efficacy for Drug Resistance in Novel HIV-1 Protease Inhibitors through Transfer Learning...

Area of Science:

• Virology
• Computational Biology
• Drug Discovery

Background:

• Human immunodeficiency virus (HIV) poses a global health threat due to rapid mutation and drug resistance.
• Machine learning (ML) and deep learning (DL) show promise in predicting drug resistance but struggle with generalizing inhibitor representations.
• Predicting HIV drug resistance requires models that learn from both viral isolates and drug characteristics.

Purpose of the Study:

• To develop a novel drug-isolate-fold change (DIF) model framework for predicting HIV drug resistance scores.
• To enhance the molecular learning capacity of ML/DL models by incorporating inhibitor representations.
• To evaluate the effectiveness of DIF models using realistic validation strategies and diverse datasets.

Main Methods:

• Proposed a novel drug-isolate-fold change (DIF) model framework integrating protein sequence and inhibitor representations.
• Employed transfer learning by pretraining a graph neural network (GNN) for activity prediction on 4855 HIV-1 protease inhibitors (PIs).
• Analyzed various ML/DL models, inhibitor, and protein representations using internal and external genotype-phenotype datasets.

Main Results:

• DIF models demonstrated improved predictive ability by leveraging learned inhibitor representations.
• Achieved high performance metrics: 0.802 accuracy, 0.874 AUROC, and 0.727 r for unseen external PIs.
• Statistically significant improvements were observed compared to isolate-fold change (IF) models, confirming the benefit of molecular representations.

Conclusions:

• The DIF model framework effectively predicts drug resistance scores for HIV-1 protease inhibitors.
• Incorporating molecular representations significantly enhances the predictive power of ML/DL models for HIV drug resistance.
• The DIF models offer a robust approach for evaluating novel PIs against drug-resistant HIV strains.