Drug Resistance Prediction Using Deep Learning Techniques on HIV-1 Sequence Data.

Area of Science:

• Virology
• Computational Biology
• Genetics

Background:

• Human immunodeficiency virus (HIV) rapidly mutates due to its replication rate and lack of repair mechanisms.
• Some mutations lead to resistance against antiretroviral therapies (ART), compromising treatment efficacy.
• Understanding HIV drug resistance is crucial for the sustained effectiveness of ART.

Purpose of the Study:

• To investigate the link between deep learning models and evolutionary principles in HIV drug resistance.
• To evaluate the performance of different deep learning architectures for predicting HIV drug resistance.
• To analyze the biological relevance of features identified by deep learning models.

Main Methods:

• Utilized publicly available HIV-1 sequence data and drug resistance assay results for 18 ART drugs.
• Evaluated three deep learning architectures: multilayer perceptron, bidirectional recurrent neural network, and convolutional neural network.
• Performed joint biological analysis alongside deep learning model evaluation.

Main Results:

• Convolutional neural networks (CNNs) demonstrated the best performance in predicting drug resistance.
• Model performance correlated with the importance of biologically relevant features.
• Deep learning models identified drug resistance mutations (DRMs) as key predictors.
• Models also weighted non-DRM features, highlighting the need for interpretability.

Conclusions:

• Deep learning models, particularly CNNs, are effective tools for predicting HIV drug resistance.
• Model interpretability is essential to understand the causal relationships between viral genotype and drug resistance phenotype.
• This approach aids in real-time evaluation of viral evolution and informs ART strategies.