Machine learning-based predictive models for identifying high active compounds against HIV-1 integrase.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• HIV-integrase is crucial for establishing latent HIV infection by integrating viral DNA into the host genome.
• Computer-aided drug design (CADD) is vital for discovering novel antiviral agents.
• Identifying potent inhibitors of HIV-integrase is a key strategy in HIV/AIDS treatment.

Purpose of the Study:

• To develop and validate machine learning models for predicting high-activity compounds against HIV-integrase.
• To utilize validated models for virtual screening of chemical libraries to identify potential drug leads.
• To assess the binding affinity and interaction of identified compounds with HIV-integrase.

Main Methods:

• Development of multiclass support vector machine (SVM) models.
• Feature selection using the Boruta method.
• Rigorous validation of predictive models.
• Virtual screening of the ChemBridge compound library.
• Molecular docking and binding affinity analysis of top-ranked compounds.

Main Results:

• Machine learning models, particularly multiclass SVM, demonstrated reasonable accuracy in predicting compound activity.
• Feature selection slightly improved model performance compared to using all descriptors.
• Virtual screening identified six high-activity compounds.
• Three compounds (9103124, 6642917, 9082952) exhibited favorable binding affinity and stable interactions with key HIV-integrase residues (Asp64, Glu152, Asn155).

Conclusions:

• Validated machine learning classification models are effective for accurately predicting and ranking lead compounds against HIV-integrase.
• The identified compounds show promise as potential antiviral drugs targeting HIV-integrase.
• The study underscores the utility of computational approaches in accelerating drug discovery for HIV.