Machine Learning Bolsters Evidence That D1, Nef, and Tat Influence HIV Reservoir Dynamics

Abstract

Background: The primary hurdle to curing HIV is due to the establishment of a reservoir early in infection. In an effort to find new treatment strategies, we and others have focused on understanding the selection pressures exerted on the reservoir by studying how proviral sequences change over time.

Methods: To gain insights into the dynamics of the HIV reservoir we analyzed longitudinal near full-length sequences from 7 people living with HIV between 1 and 20 years following the initiation of antiretroviral treatment. We used this data to employ Bayesian mixed effects models to characterize the decay of the reservoir using single-phase and multiphasic decay models based on near full-length sequencing. In addition, we developed a machine-learning approach utilizing logistic regression to identify elements within the HIV genome most associated with proviral decay and persistence. By systematically analyzing proviruses that are deleted for a specific element, we gain insights into their role in reservoir contraction and expansion.

Results: Our analyses indicate that biphasic decay models of intact reservoir dynamics were better than single-phase models with a stronger statistical fit. Based on the biphasic decay pattern of the intact reservoir, we estimated the half-lives of the first and second phases of decay to be 18.2 (17.3 to 19.2, 95%CI) and 433 (227 to 6400, 95%CI) months, respectively.In contrast, the dynamics of defective proviruses differed favoring neither model definitively, with an estimated half-life of 87.3 (78.1 to 98.8, 95% CI) months during the first phase of the biphasic model. Machine-learning analysis of HIV genomes at the nucleotide level revealed that the presence of the splice donor site D1 was the principal genomic element associated with contraction. This role of D1 was then validated in an in vitro system. Using the same approach, we additionally found supporting evidence that HIV nef may confer a protective advantage for latently infected T cells while tat was associated with clonal expansion.

Conclusions: The nature of intact reservoir decay suggests that the long-lived HIV reservoir contains at least 2 distinct compartments. The first compartment decays faster than the second compartment. Our machine-learning analysis of HIV proviral sequences reveals specific genomic elements are associated with contraction while others are associated with persistence and expansion. Together, these opposing forces shape the reservoir over time.

Keywords: HIV Reservoir; Machine Learning; NFL Sequencing.

Figure 1.

Decay of intact HIV. Concentration of intact HIV provirus per million CD4 plotted at all time points for the 7 individuals in the study. Thick black line represents the best fit to a single-phase decay model. Dotted black line represents best fit to a biphasic decay model. The dotted red line represents the inflection point of the biphasic model. A) Best fit of the 2 models with all data included in the analysis. B) Best fit of the 2 models with clones removed. The fit of the 2 models is drastically different both with all data included and with clones removed, with the biphasic model providing a much better fit to the data. CT stands for chronically treated. In other words, an individual who was treated during the chronic phase of infection.

Figure 2.

Decay of defective HIV. Concentration of defective HIV proviruses per million CD4 plotted at all time points for the 7 individuals in the study. Thick black line represents the best fit to the single-phase model. Dotted black line represents the best fit to the biphasic decay model. The dotted red line represents the inflection point of the biphasic model. A) Best fit of the 2 models with all data included in the analysis. B) Best fit of the 2 models with clones removed. With all of the data included, the biphasic model provides a slightly better fit to the data; however, with the clones removed, the model fits are more similar between the single-phase and biphasic models.

Figure 3.

Computational expression of HIV elements and their association with proviral persistence vs decay. A) Frequency of HIV elements with all proviral data included. B) Frequency of HIV elements with all definitive clonal population removed. C) Frequency of HIV element using proviral sequences from definitive clonal populations. To present the data more clearly, Force Factors for only 28 of the 33 elements considered in our analysis are shown. Env loops V1-V5 were excluded as none of them had a large individual contribution to either persistence or decay. Notably, the upper and lower 5% of sequences associated with the most decay and most persistence are different in A, B, and C. Thus, A is not an average of B and C.

Figure 4.

Proviruses that contract (D1+) appear to favor protein expression. A) The construct labeled as red represents the common 3’ deleted proviruses that contract over time while the common 5’ deleted proviruses are labeled as blue. We generated these categories by restriction enzyme digestion and regulation after filling in with Klenow Fragment. Both clones were engineered to be able to express GFP after D1 spliced to A7 (red) or after D4 spliced to A7 (blue). B) HIV RU5 RNA levels were measured by RT-PCR and (C) percentage HIV GFP expression was measured by flow cytometry. The 95% confidence intervals for (B) and (C) were derived from the t distribution. Each proviral category appear to express distinct levels of HIV RNA. Proviruses that lack D1 express significantly less HIV GFP.