Classification of HIV-1 Protease Inhibitors by Machine Learning Methods

Area of Science:

• Medicinal Chemistry
• Computational Biology
• Drug Discovery

Background:

• HIV-1 protease is crucial for viral replication and a key therapeutic target.
• Developing effective HIV-1 protease inhibitors (PIs) is vital for combating HIV-1 infection.

Purpose of the Study:

• To build and validate machine learning models for predicting the activity of HIV-1 protease inhibitors.
• To identify key molecular descriptors contributing to the bioactivity of PIs.

Main Methods:

• Utilized a dataset of 4855 HIV-1 PIs from ChEMBL.
• Developed 15 classification models using k-nearest neighbors, decision tree, random forest, support vector machine, and deep neural network algorithms.
• Characterized molecular structures using MACCS and PubChem fingerprints, and physicochemical descriptors from CORINA Symphony.

Main Results:

• Individual models achieved prediction accuracies over 70%, with the best SVM model reaching 83.07%.
• Consensus models demonstrated improved performance, with model C3a (MACCS fingerprints) achieving 83.15% accuracy on the test set.
• External validation using DUD database and recent literature yielded a top accuracy of 98.37% with a random forest model (CORINA Symphony descriptors).

Conclusions:

• Machine learning models, particularly those using SVM and Random Forest, are effective for predicting HIV-1 protease inhibitor activity.
• Aromatic systems and hydrogen-bonding atoms are significant contributors to the bioactivity of PIs.
• The developed models and identified descriptors can aid in the design of novel and potent HIV-1 PIs.