A comparison of machine learning techniques for classification of HIV patients with antiretroviral therapy-induced mitochondrial toxicity from those without mitochondrial toxicity

Antiretroviral therapy (ART) has significantly reduced HIV-related morbidity and mortality. However, therapeutic benefit of ART is often limited by delayed drug-associated toxicity. Nucleoside reverse transcriptase inhibitors (NRTIs) are the backbone of ART regimens. NRTIs compete with endogenous deoxyribonucleotide triphosphates (dNTPs) in incorporation into elongating DNA chain resulting in their cytotoxic or antiviral effect. Thus, the efficacy of NRTIs could be affected by direct competition with endogenous dNTPs and/or feedback inhibition of their metabolic enzymes. In this paper, we assessed whether the levels of ribonucleotides (RN) and dNTP pool sizes can be used as biomarkers in distinguishing between HIV-infected patients with ART-induced mitochondrial toxicity and HIV-infected patients without toxicity.

We used data collected through a case-control study from 50 subjects. Cases were defined as HIV-infected individuals with clinical and/or laboratory evidence of mitochondrial toxicity. Each case was age, gender, and race matched with an HIV-positive without evidence of toxicity. We used a range of machine learning procedures to distinguish between patients with and without toxicity. Using resampling methods like Monte Carlo k-fold cross validation, we compared the accuracy of several machine learning algorithms applied to our data. We used the algorithm with highest classification accuracy rate in evaluating the diagnostic performance of 12 RN and 14 dNTP pool sizes as biomarkers of mitochondrial toxicity.

We used eight classification algorithms to assess the diagnostic performance of RN and dNTP pool sizes distinguishing HIV patients with and without NRTI-associated mitochondrial toxicity. The algorithms resulted in cross-validated classification rates of 0.65-0.76 for dNTP and 0.72-0.83 for RN, following reduction of the dimensionality of the input data. The reduction of input variables improved the classification performance of the algorithms, with the most pronounced improvement for RN. Complex tree-based methods worked the best for both the deoxyribose dataset (Random Forest) and the ribose dataset (Classification Tree and AdaBoost), but it is worth noting that simple methods such as Linear Discriminant Analysis and Logistic Regression were very competitive in terms of classification performance.

Our finding of changes in RN and dNTP pools in participants with mitochondrial toxicity validates the importance of dNTP pools in mitochondrial function. Hence, levels of RN and dNTP pools can be used as biomarkers of ART-induced mitochondrial toxicity.