Multiple Machine Learning Comparisons of HIV Cell-based and Reverse Transcriptase Data Sets.

Area of Science:

• Computational chemistry
• Machine learning
• Virology

Background:

• Human immunodeficiency virus (HIV) remains a major global health threat, causing significant mortality and economic burden.
• Existing treatments, such as reverse transcriptase inhibitors, face challenges due to drug-resistant HIV strains, necessitating novel therapeutic strategies.

Purpose of the Study:

• To leverage Bayesian machine learning and public databases to identify novel compounds with potential anti-HIV activity.
• To compare the predictive performance of various machine learning algorithms for identifying effective HIV inhibitors.

Main Methods:

• Utilized the NIAID ChemDB HIV, Opportunistic Infection and Tuberculosis Therapeutics Database for data curation and cleaning.
• Trained and evaluated multiple machine learning models, including Support Vector Classification, Deep Neural Networks, and consensus approaches, using HIV-1 inhibition assay data.
• Employed 5-fold cross-validation and 24 training/test set combinations to assess model generalizability.

Main Results:

• Identified a correlation between compounds exhibiting high HIV-1 RT DNA polymerase inhibition and cell-based HIV-1 inhibition (Pearson r = 0.44).
• Support Vector Classification, Deep Learning, and consensus models demonstrated comparable predictive performance.
• No significant difference was observed between Support Vector Machines and Deep Neural Networks across multiple datasets.

Conclusions:

• Machine learning, particularly Support Vector Classification and Deep Learning, effectively identifies potential anti-HIV compounds from existing databases.
• The developed models provide a valuable tool for accelerating the discovery of new therapeutics against drug-resistant HIV strains.
• Findings support the utility of machine learning in drug discovery for infectious diseases.