Human Immunodeficiency Virus (HIV) Genetic Diversity Informs Stage of HIV-1 Infection Among Patients Receiving Antiretroviral Therapy in Botswana

Abstract

Background: Human immunodeficiency virus (HIV)-1 genetic diversity increases during infection and can help infer the time elapsed since infection. However, the effect of antiretroviral treatment (ART) on the inference remains unknown.

Methods: Participants with estimated duration of HIV-1 infection based on repeated testing were sourced from cohorts in Botswana (n = 1944). Full-length HIV genome sequencing was performed from proviral deoxyribonucleic acid. We optimized a machine learning model to classify infections as < or >1 year based on viral genetic diversity, demographic, and clinical data.

Results: The best predictive model included variables for genetic diversity of HIV-1 gag, pol, and env, viral load, age, sex, and ART status. Most participants were on ART. Balanced accuracy was 90.6% (95% confidence interval, 86.7%-94.1%). We tested the algorithm among newly diagnosed participants with or without documented negative HIV tests. Among those without records, those who self-reported a negative HIV test within <1 year were more frequently classified as recent than those who reported a test >1 year previously. There was no difference in classification between those self-reporting a negative HIV test <1 year, whether or not they had a record.

Conclusions: These results indicate that recency of HIV-1 infection can be inferred from viral sequence diversity even among patients on suppressive ART.

Keywords: ART; HIV; HIV treatment; NGS; early HIV infection.

Figure 1.

Viaplot of log mean entropy for participants based on stage of infection (chronic and recent) and antiretroviral treatment (ART) status (naive or treated). Log mean entropy for recent infections (−4.45; −5.33 to −2.70) was significantly below that of chronic infections (−3.57; −5.34 to −2.34). Averaged across gag, pol, and env.

Figure 2.

(A) Model accuracy, (B) balanced accuracy, (C) sensitivity, and (D) specificity with cross-validation for 4 models with different sets of predictors (1) demographic/clinical predictors only (age, sex, viral load, and antiretroviral treatment [ART] status), (2) diversity (in each of the 3 genes) only, (3) diversity and demographics, and (4) diversity and ART status. Each model was fitted and evaluated 1000 times, splitting the complete data into training (70%) and test (30%) data each time. The no information rate for accuracy is the proportion of the dominant class (here, 89%). The equivalent no information rate for balanced accuracy would be 50%.

Figure 3.

(A) Sensitivity, (B) specificity, (C) balanced accuracy, and (D) percentage of missing predictions for the logistic regression and machine learning models. Statistics are calculated by fitting the model each time to a training dataset, then evaluating it in a test dataset. Note that the xgboost model was always able to predict recency even in the absence of some predictors (D).

Figure 4.

Balanced accuracy of the predicted stage of infection for participants based on antiretroviral treatment (ART) status. In the joint model, the model was fit to all participants regardless of ART status, and ART status was included as a predictor. In the split model, the model was fit separately to ART-treated and ART-naive participants. The split model improved balanced accuracy for both ART-treated and ART-naive participants (P < 10–16).