Lopinavir Resistance Classification with Imbalanced Data Using Probabilistic Neural Networks.

Area of Science:

• Virology
• Computational Biology
• Machine Learning

Background:

• Antiretroviral drug resistance is a significant challenge in managing HIV infection.
• Predictive models for drug resistance can guide personalized therapy selection for HIV-positive individuals.

Purpose of the Study:

• To develop and evaluate probabilistic neural network (PNN) classifiers for predicting resistance to the HIV protease inhibitor lopinavir.
• To assess the performance of these classifiers against existing expert-based systems.

Main Methods:

• Feature selection using logistic regression (LR) with bootstrap and stepwise techniques identified ten input features.
• Probabilistic neural network (PNN) models were trained and optimized using bootstrap and cross-validation for the smoothing parameter.
• Four balanced PNN classifiers were developed and validated on a separate test set.

Main Results:

• The PNN classifiers demonstrated high predictive accuracy, with test set accuracies ranging from 0.89 to 0.94.
• Area Under the Curve (AUC) values for the receiver operating characteristic (ROC) curves were between 0.96 and 0.97.
• Sensitivity ranged from 0.94 to 1.00, and specificity was between 0.88 and 0.92, comparable to expert systems.

Conclusions:

• The developed PNN models effectively predict lopinavir resistance in HIV.
• These classifiers offer a valuable tool for clinical decision-making in HIV therapy.
• The performance of the PNN models is comparable to established expert-based resistance interpretation systems.