Simulation of clonal expansion and latent HIV reservoir size prediction using machine learning in virally suppressed individuals in Uganda

Abstract

Background: HIV can survive as pro-viruses that do not undergo transcription mostly within resting memory CD4+ T cells despite prolonged antiretroviral therapy (ART). This Latent Viral Reservoir (LVR) has an extremely long half-life and is the major barrier to curing HIV. The LVR is mostly maintained through clonal expansion, but its dynamics are not fully understood particularly among individuals living in settings with important endemic infections like TB and Malaria. We simulated LVR dynamics and predicted LVR size in a cohort from a tropical setting with multiple endemic infections. Methods: We obtained 2 or more LVR measurements of 41 and 22 virally-suppressed men and women respectively in the Rakai Community Cohort Study, collected between 2014 and 2020. We observed trends in LVR size, measured as infectious units per million resting CD4+ T cells (IUPM); and fitted a modified version of Perelson and Ribeiro’s 2013 HIV model to estimate the clonal expansion, reactivation and decay rates, assuming 100% ART efficacy. CD4/CD8 ratio, time on ART, gender, Pol and gp41 phylogenetic distances were used to predict LVR size using different machine learning algorithms. Results: Of the 63 individuals, 43(68.3%) had LVRs with increasing trends, 20(31.7%) with declining trends, and none with a constant trend. Among those with increasing trends, the median LVR size was 0.211, 0.82, 0.495, 1.07 and 0.188 at years 1 to 5 respectively, and did not differ by sex (p=0.91); the estimated clonal expansion, activation, and decay rates were 1.517, 0.935, and 0.2 year-1 respectively; and median doubling time was 26.7 months (QR=15.2-68.1). Among those with declining trends, the median LVR size was 0.769, 0.062, 0.277, 0.373 and 0.12 at years 1 to 5 respectively, and did not differ by sex (p=0.14); the rates were 0.949, 1.184 and 0.199 year-1 respectively; and median half-life was 41.1 months (QR=20.3-62.5). The mean half-life after elimination of 3 extreme half-lives was 37.1 months, requiring up to 32 and 74 years for the LVR to fall below 10-6 and 10-14 IUPM respectively. Both trajectories did not differ by sex (p=0.64 and 0.38 respectively). The Light Gradient Boosted Machine (LBGBM) algorithm was selected and it predicted LVR size with an R2, RMSE, and MAE of 0.52, 1.39, and 0.97 respectively. Conclusion: Majority of individuals surprisingly had LVRs with increasing trends, at least in the first 5 years post-ART, which could raise the barrier for a cure in such individuals. Among those with declining trends, the dynamics were comparable to previous reports elsewhere, suggesting minimal or balanced alteration of LVR decay by immune activation if it existed. Machine learning also showed potential in predicting the LVR size which is essential in monitoring HIV infection.