Dominant CD4⁺ T cell receptors remain stable throughout antiretroviral therapy-mediated immune restoration in people with HIV

Summary

In people with HIV (PWH), the post-antiretroviral therapy (ART) window is critical for immune restoration and HIV reservoir stabilization. We employ deep immune profiling and T cell receptor (TCR) sequencing and examine proliferation to assess how ART impacts T cell homeostasis. In PWH on long-term ART, lymphocyte frequencies and phenotypes are mostly stable. By contrast, broad phenotypic changes in natural killer (NK) cells, γδ T cells, B cells, and CD4⁺ and CD8⁺ T cells are observed in the post-ART window. Whereas CD8⁺ T cells mostly restore, memory CD4⁺ T subsets and cytolytic NK cells show incomplete restoration 1.4 years post ART. Surprisingly, the hierarchies and frequencies of dominant CD4 TCR clonotypes (0.1%–11% of all CD4⁺ T cells) remain stable post ART, suggesting that clonal homeostasis can be independent of homeostatic processes regulating CD4⁺ T cell absolute number, phenotypes, and function. The slow restoration of host immunity post ART also has implications for the design of ART interruption studies.

Graphical abstract

Highlights

• •Almost all immune lineages are dysregulated in people with HIV (PWH)
• •Antiretroviral therapy (ART) rapidly controls HIV and initiates immune restoration
• •Restoration of frequency, function, and phenotype of immune lineages takes years
• •Remarkably, dominant T cell receptor clonotypes are stable in the post-ART window

In this study, we demonstrate the widespread immune dysregulation wrought by HIV infection and show that, while drugs can control HIV levels within months, many facets of our immune system take years to restore. This study has direct implications for how we design studies aimed at curing HIV.

Introduction

Immune homeostasis is a broad term that describes processes that maintain equilibrium within and between immune compartments. This can include maintenance of immune cell numbers, phenotype, function, and immune memory.

Human immunodeficiency virus-1 (HIV), which mostly infects CD4⁺ T cells, profoundly impacts immune homeostasis. HIV viremia increases levels of pro-inflammatory cytokines¹ and cell cycling and drives chronic cellular activation across multiple immune lineages (e.g., CD38, Human Leukocyte Antigen [HLA]-DR, PD-1),^2,3 which contributes to dysregulation of CD4⁺ and CD8⁺ T cells,^4,5,6,7 γδ T cells,^8,9 natural killer (NK) cells,^10,11,12 B cells,^13,14,15 and monocytes.^16,17

Dysregulation of T cell homeostasis in HIV infection is associated with altered levels of the homeostatic cytokines interleukin-2 (IL-2), IL-7, and IL-15.^18,19,20 IL-7/IL-7 receptor signaling, which increases expression of anti-apoptotic proteins that include Bcl-2, is particularly important for the generation of long-lived virus-specific CD4⁺ and CD8⁺ T cells in humans and animal models.^{7,19,21,22,23,24} The IL-7 receptor is comprised of the common γ chain and an α chain, CD127. In HIV infection, long-lived T cells expressing CD127 are lost,^19,25,26 contributing to loss of immunological memory, evidenced by poor responses to vaccination.⁷

Modern combination antiretroviral therapy (ART) blocks multiple stages of HIV replication and rapidly decreases HIV viremia within weeks to months of treatment whether individuals are in acute or chronic stages of infection.^27,28 People with HIV (PWH) on ART now enjoy near-normal life spans.^29,30 Sustained ART has been documented to allow broad immune restoration, though in individuals who initiate ART in the chronic stage of infection, the CD4:CD8 ratio often does not restore.^31,32,33 Individuals with poor CD4⁺ T cell reconstitution and/or sustained low CD4:CD8 ratios experience higher rates of HIV co-morbidities despite durable virus suppression.^32,34,35

ART does not prevent CD4⁺ T cells already infected with HIV from entering latency. These latently infected cells reactivate stochastically and drive HIV rebound following ART interruption.^36,37 HIV sequencing has identified identical or near-identical HIV proviruses in a person many years apart.^38,39,40 These observations support a model where decay of the reservoir of latently infected cells is slowed in PWH through homeostatic maintenance of long-lived CD4⁺ T cells.

HIV entry into latency can occur at all stages of infection, though, on average, the majority of the CD4⁺ T cell HIV reservoir is derived from cells infected in and around the time of ART initiation.^41,42,43 These studies suggest that changes wrought by ART initiation facilitate stabilization of infected cells entering latency. Moreover, these studies suggest that targeted therapies near the time of ART initiation could block HIV reservoir stabilization, decrease overall reservoir size, and accelerate HIV cure strategies.^42,44 Other approaches to HIV cure include therapies and vaccines to increase circulating innate and adaptive immunity against HIV.⁴⁵

Altogether, these studies suggest that the post-ART window is critical for the restoration of immune homeostasis, health in PWH, and entry of HIV-infected cells into the stable reservoir. Here we used a combination of mass cytometry, T cell receptor (TCR) sequencing of CD4⁺ T cells and functional proliferation assays to document the kinetics of immune restoration in PWH in the first 500 days following ART initiation. Overall, we found that immune restoration occurs with variable kinetics across lymphocyte lineages, including T cell subsets, and continues well after virus suppression has been achieved. Most surprisingly, over this period, dominant CD4 TCR clonotypes retained their overall frequency and rank order, suggesting that clonal homeostasis of dominant TCRs is independent of homeostatic process that regulate CD4⁺ T cell numbers, phenotype, and function.

Results

Study cohorts

We compared the phenotype and function of blood immune cells over a 17- to 26-month window of two HIV-seropositive cohorts (Figures 1A and 1B). In the first cohort, A5248, participants, all diagnosed in the chronic stage of HIV infection, were sampled up to 12 times in the 504 days from ART initiation. ART, which included an integrase inhibitor, resulted in rapid decreases in viral load (VL) and full suppression within 12 weeks in all participants (day 84) of treatment (Figure 1C).⁴⁶

Figure 1.

Experimental workflow

(A) A5248 participants were sampled 0–504 days following antiretroviral therapy (ART) initiation. In the long-term (LT)-ART group, participants were sampled 3 times over 17–26 months.

(B) Participant peripheral blood mononuclear cells (PBMCs) from A5248 and LT-ART (n = 10/group) were phenotyped by mass cytometry and analyzed using unsupervised and manual gating methods, and changes were quantified. TCRβ clonotype sequencing of isolated CD4⁺ T cells was performed over time in A5248 (n = 4, 5 time points) and LT-ART (n = 3, 3 time points). Proliferative capacity of CD4⁺ and CD8⁺ T cells was also assessed in both cohorts using flow cytometry (n = 8/cohort).

(C and D) HIV VL, absolute CD4⁺ and CD8⁺ T cell count, and CD4:CD8 ratio of A5248 participants (C) and LT-ART participants (D).

All but one participant, 611172, experienced ART-mediated CD4 T cell recovery (CD4 count ≥ 400 cells/mm³ on day 504) (Figure 1C). Overall CD8 T cell counts decreased over the study window; however, the differences in means on day 0 and day 504 were not significant (Figure 1C; p = 0.3750, Wilcoxon signed-rank test).^47,48 The CD4:CD8 ratio increased but, relative to ratios reported in healthy individuals, remained inverted (Figure 1C; Table S1).⁴⁹ While the A5428 follow-up window was limited, 611172 exhibited both a low frequency of naive CD4⁺ T cells (611172 1.9%, cohort average 9.2%, n = 10) and high PD-1⁺CD4⁺ T cells (611172 70.6%, cohort average 39.6%, n = 10), suggesting he was an immunological non-responder (INR).^50,51,52 Notably, 611172 was not an outlier for other measures of immune reconstitution or T cell proliferation, as detailed below.

The second cohort comprised PWH also diagnosed with chronic HIV infection and who have been on ART for an average of 6.7 years (range, 1.8–20.1). All participants were durably suppressed and had stable CD4:CD8 ratios (Figure 1D). Throughout this paper, we refer to the ART initiation cohort as A5248 and to the long-term ART-suppressed cohort as LT-ART. Both cohorts were predominantly male with comparable pre-ART VL and pre-ART CD4⁺ counts (Table S1).

ART initiation elicits time-associated changes across leukocyte populations

We first performed unbiased metaclustering of 31-marker mass cytometry data using the PhenoGraph algorithm.⁵³ PhenoGraph identified phenotypically similar CD45⁺ cell populations shared across participants and time points in each cohort (Figures S1A–S1D). Annotation of the metaclusters was based on expression of 25 proteins (Figures S1B and S1D, “cell types”). For A5248, all but two metaclusters (C10 and C15) were confidently annotated. C10 and C15 were CD45⁺ but otherwise lacked identifying lineage markers (Figure S1B). The successfully annotated cell populations included naive and memory CD4⁺ and CD8⁺ T cell subsets, γδ T cells, NK cells, and B cells and classical, non-classical, and intermediate monocytes.

We trained a penalized linear regression model, elastic net (EN), to ascertain whether changes in metacluster-derived features were predictive of the CD4⁺ T cell count from each participant (Figure 1C), but no features were predictive (median R² = −3.241; Figure S2). Additionally, no correlation was detected between activation marker frequency on CD4⁺ T cells and baseline CD4 count or pre-ART VL (Figures S2B and S2C). EN model training did identify immune features in A5248 participants that were predictive of time on ART.⁵⁴ This analysis focused on expression patterns within annotated clusters of protein markers (CD127, Bcl-2, PD-1, CD57, Ki-67, CCR5, CD38, and HLA-DR) that are dysregulated in T cells following HIV infection.^{2,23,55,56,57} Changes across multiple leukocyte lineages were associated with the time following ART (median R² = 0.466 across cross-validation folds; Figure 2A). Notably, several protein markers showed similar expression changes across different lineages. We add that caution must be taken with interpretation of monocyte subset data given that monocyte frequency may be impacted by the freeze-thaw process.⁵⁸

Figure 2.

ART initiates immune restoration across multiple immune lineages

(A) EN regression analysis showing the descriptive mass cytometry immune features from unsupervised metaclusters. Shown are immune features that have a non-zero coefficient (median with whiskers reaching the minimum and maximum values across cross-validation folds) and are predicted to be associated with time post ART initiation. A positive coefficient indicates an increase and a negative coefficient indicates a decrease over time. Δ = difference between time point > 0 from time point = 0.

(B) Manual gating analysis of total NK, γδ T, CD4⁺ T, CD8⁺ T, and CD19⁺ B cells, showing average %frequency of the indicated markers from day 0 to day 504 post ART initiation; n = 10 mean ± SD. Note: NK cell data are combined CD16⁺CD56⁻, CD16⁺CD56lo, CD16⁻CD56^bright populations (see Table 1).

In the LT-ART cohort, no features were predictive of time (median R² = −0.169), suggesting greater stability of leukocyte populations when individuals had been durably suppressed (Figure S1).

Fixed-effects model identified consistent phenotypic changes in immune lineages following ART initiation

To quantify trends observed in the EN model analysis (Figure 2A), we used quasi-binomial fixed-effects (QBFE) regression for analysis of %frequencies (Figures S3–S5). Note that, while higher-frequency cell populations were identified using unsupervised and supervised (manual gating) methods, we found improved discrimination of lower-frequency T cell subpopulations using manual gating (data not shown). Therefore, manual gating was used for all subsequent analyses. In QBFE regression analysis, changes over time are denoted as the average percentage point increase (+) or decrease (−) per 30-day window. Study numbers were too small to subdivide post-ART data and separately calculate changes in slope within this study period.

Total T cells

In A5248, the frequency of CD4⁺ T cells (day 0 range: 10.5%–41.4%) increased, on average, +0.68 percentage points (%pts) every 30 days post ART initiation (day 504 range: 20.1%–56.4%), consistent with increases in absolute CD4 T cell counts within most participants (Figure 1D). The relative frequency of CD4⁺ T cells continued to increase in the LT-ART group (+0.17%pts). This suggested, consistent with other reports, that, even after years of virus suppression, CD4⁺ T cell absolute numbers continued to restore.^59,60 CD8⁺ T cell frequencies concomitantly decreased (−0.59%pts), and smaller but significant decreases were observed in the LT-ART cohort (−0.22%pts; Table 1).

Table 1.

Quantification of ART-induced changes in leukocyte populations over time in PWH

Lineage	Annotation	Manual gating	A5248a,b(%pts/30 days)	LT-ARTa,c (%pts/30 days)
CD4+ T cells	CD4+ T cells	CD3+ CD4+	+0.68	+0.17
CD8+ T cells	CD8+ T cells	CD3+ CD8+	−0.59	−0.22
B cells	memory B cells (total)d	CD19+ CD27+ CD38−	−0.15	–
	plasmablasts	CD19+ CD27+ CD38+	−0.08	–
	B cells	CD19+ CD27−	+0.20	–
	Ki-67	CD19+ Ki-67+	−0.17	–
	CD38	CD19+ CD38+	−0.29	–
	PD-1	CD19+ PD-1+	−0.25	–
	Bcl-2	CD19+ Bcl-2+	+0.95	–
	CD127	CD19+ CD127+	–	–
NK cells	cytotoxic NK cells	CD3− CD14− CD19− CD56lo CD16+	+0.48	–
	pro-inflammatory NK cells	CD3− CD14− CD19− CD56bright CD16−	–	–
	CD56− NK cellse	CD3− CD14− CD19− CD56− CD16+	−0.52	–
	Ki-67	CD3− CD14− CD19− CD56+ CD16+ Ki-67+	−0.06	–
	CD38	CD3− CD14− CD19− CD56+ CD16+ CD38+	−0.62	–
	PD-1	CD3− CD14− CD19− CD56+ CD16+ PD-1+	−0.23	–
	Bcl-2	CD3− CD14− CD19− CD56+ CD16+ Bcl-2+	+1.03	–
	CD127	CD3− CD14− CD19− CD56+ CD16+ CD127+	+0.36	–
γδ T cells	γδ T cells	CD3+ TCRγδ+	−0.05	–
	Ki-67	CD3+ TCRγδ+ Ki-67+	−0.21	–
	CD38	CD3+ TCRγδ+ CD38+	−1.57	–
	PD-1	CD3+ TCRγδ+ PD-1+	−0.35	−0.12
	Bcl-2	CD3+ TCRγδ+ Bcl-2+	+0.59	–
	CD127	CD3+ TCRγδ+ CD127+	+0.31	–

A QBFE model with indicators for participant ID was used to examine the association between markers of canonical lineages and non-T cell subpopulation dynamics and time in months (treated as 30 days) (see fixed-effects regression). A positive value denotes an increase, and a negative value denotes a decrease. The model was calculated from days 0–504 for A5248 and from T1–T3 for LT-ART. –, no change (p > 0.05).

^aA5248 includes PWH (n = 10) in the 504 days following ART initiation.

^bp values and confidence intervals are listed in Table S2.

^cLT-ART includes PWH (n = 10) who have been virus suppressed for an average of 6.7 years.

^dCD21 was not included in this mass cytometry panel, and therefore we were unable to resolve resting memory and activated memory B cells.⁶¹

^eCD56⁻ NK cells are rarely observed in healthy individuals but have been described in proinflammatory conditions, including HIV infection.^10,62

B cells

In A5248, B cell frequencies significantly increased over time (+0.20%pts). This was consistent with an observed increased frequency of Bcl-2⁺ B cells (+0.95%pts), altogether suggesting greatly improved B cell survival following ART (Figures 2A and 2B; Tables 1 and S2). Cell cycling, measured by Ki-67, which is expressed in all active stages of the cell cycle,^63,64 and activation markers, CD38 and PD-1, also significantly decreased in B cells. We lacked markers to comprehensively assess B cell subpopulations, but, consistent with previous reports, we observed post-ART decreases in frequencies of plasmablasts.^65,66 Increases in resting memory and naive B cell frequencies and decreases in frequencies of immature transitional and mature activated B cells have been described previously ∼12 months post ART.⁶⁷ Here, total memory CD27⁺ CD38⁻ B cells increased post ART initiation, but we were unable to discriminate between mature, activated, and resting memory B cells (Table 1).

NK cells

CD56 is a canonical marker of NK cells. However, in HIV infection, dysfunctional CD56⁻CD16⁺ NK cells have been described.^10,62 Accordingly, changes in CD56 frequency were resolved into CD56^bright and CD56^lo expression as well as CD56⁻ NK cells and examined.^68,69 Consistent with previous descriptions of ART restoration of NK cells, frequencies of cytotoxic CD56^lo NK cells, which mediate strong cytotoxicity in response to receptor engagement, increased significantly following ART (+0.48%pts),⁷⁰ offsetting comparable decreases in dysfunctional CD56⁻ NK cells. Similar to B cells, significant increases in the frequency of Bcl-2 frequency (+1.03%pts) as well as decreases in Ki-67, CD38, and PD-1 frequencies (−0.06%pts, −0.62%pts, and −0.23%pts, respectively) were observed following ART in total NK cells (Figure 2B; Table 1). While no changes in the frequency of CD127⁺ B cells were detected, strong and sustained increases in CD127⁺ NK cells were observed.

γδ T cells

Overall γδ T cell frequencies exhibited a minor but significant decrease post ART (Figure 2B; Table 1). Similar to the changes observed in other lymphocytes, ART initiation resulted in significant decreases in cell cycling (−0.21%pts) as well as CD38⁺ (−1.57%pts) and PD-1⁺ (−0.38%pts) frequencies and increases in Bcl-2⁺ (+0.59%pts) and CD127⁺ (+0.31%pts) populations in γδ T cells.

These dynamic and broad post-ART changes contrasted with longitudinal studies in the LT-ART cohort, in which minimal changes across B cell, NK cell, and γδ T cell populations were observed over 17–26 months (Table 1; Figure S1).

In summary, modern ART regimens induced profound and parallel immune restoration across multiple leukocyte lineages, impacting expression levels of multiple survival and activation markers.

ART initiation induced dynamic changes in CD4⁺ T cell subpopulations

We next focused on the impact of ART on 35 CD4⁺ T cell subpopulations in our study cohorts. Different models were used to examine changes in %frequency of cell subsets (QBFE regression) and in mean signal intensity (MSI) of proteins (γ-fixed effects [GFEs]) over time (Tables 2 and S3–S5). Phenotypic changes were observed in CD4⁺ T cells in the A5248 cohort, coinciding with strong, early decreases and initial VL suppression by day 84 (Figure 1C). For the Ki-67⁺ and CD38⁺ CD4⁺ T cell populations, we observed significant differences in %frequency between day 0 and day 84 (p < 0.05, Wilcoxon signed-rank test; Figure 3A) that continued through the 500-day study window (Table 2). For CD38 and other markers associated with cellular activation (HLA-DR, PD-1, and CCR5), significant decreases in MSI were also observed from day 0 to day 84 (p < 0.05, Wilcoxon signed-rank test; Figure 3B). In contrast, Bcl-2⁺ CD4⁺ T cells increased by day 84 and continued to increase over the full study window (Figure 3A; Table 2), coinciding with an increase in MSI of Bcl-2 (Figure 3B).

Table 2.

Quantification of ART-induced changes in CD4⁺ and CD8⁺ T cell populations in PWH

Annotation	Manual gating	CD4+ T cells		CD8+ T cells
		A5248a,b (%pts/30 days)	LT-ARTb,c (%pts/30 days)	A5248a,b (%pts/30 days)	LT-ARTb,c (%pts/30 days)
Senescent	CD57+	−0.28	–	+0.32	–
Activated	CD25+	–	–	–	–
Activated	CD38+ HLA-DR+	−0.34	–	−0.93	–
Activated	CD38+	−0.65	–	−2.45	–
Activated	HLA-DR+	−0.40	–	−0.71	–
Activated	CCR5+	−0.21	–	−0.60	–
Exhausted	PD-1+	−0.64	–	−1.35	–
Cell cycling	Ki-67+	−0.11	–	−0.11	–
Cytotoxic	CD56+	ND	ND	+0.11	–
Survival	Bcl-2+	+0.37	–	+0.74	–
Longevity	CD127+	+0.57	–	+0.91	–
RTE	CD45RA+ CD31+	–	ND	//d	//
RTE CD127+	CD45RA+ CD31+ CD127+	+0.33	ND	//	//
Naive	CD45RA+ CCR7+ CD28+ CD95−	+0.33	–	+0.32	–
Naive CD127+	CD45RA+ CCR7+ CD28+ CD95− CD127+	–	–	–	–
Stem cell memory	CD45RA+ CCR7+ CD28+ CD95+	–	–	+0.02	+0.05
Stem cell memory CD127+	CD45RA+ CCR7+ CD28+ CD95+ CD127+	+0.27	+0.10	+0.40	−0.09
Central memory	CD45RA− CCR7+ CD28+ CD95+	–	–	+0.10	+0.07
Central memory CD127+	CD45RA− CCR7+ CD28+ CD95+ CD127+	+0.63	–	+1.22	–
Transitional memory	CD45RA− CCR7− CD28+ CD95−	−0.13	–	+0.08	+0.06
Transitional memory CD127+	CD45RA− CCR7− CD28+ CD95− CD127+	+0.65	–	+1.56	–
Effector memory	CD45RA− CCR7− CD28− CD95−	−0.27	–	−0.93	–
Effector memory CD127+	CD45RA− CCR7− CD28− CD95− CD127+	–	–	+0.59	+0.17
Terminal effectors	CD45RA+ CCR7− CD28+ CD95−	−0.03	–	+0.26	−0.15
Terminal effectors CD127+	CD45RA+ CCR7− CD28+ CD95− CD127+	–	+0.25	+0.29	+0.14
CD99hi	CD99hi	–	–	–	–
CD99hi CD127+	CD99hi CD127+	+0.57	–	+0.73	–
Th1	CXCR3+ CCR4− CCR6− CCR10−	–	–	ND	ND
Th2	CXCR3− CCR4+ CCR6− CCR10−	+0.12	–	ND	ND
Th17	CXCR3−CCR4+CCR6+ CCR10−	–	–	ND	ND
Th17.1	CXCR3+ CCR4− CCR6− CCR10−	−0.07	–	ND	ND
Th22	CCR4+CCR6+CCR10+	+0.01	–	ND	ND
Th9	CCR4− CCR6+	−0.10	–	ND	ND
Treg	CD25+ CD127− FoxP3+	−0.08	–	ND	ND
Tfh	CXCR5+ PD-1+CD45RA−	–	–	ND	ND

A QBFE model based on manual gating with indicators for participant ID was used to examine the association between markers of CD4⁺ and CD8⁺ T cell subpopulation dynamics and time in months (treated as 30 days) (see fixed-effects model). A positive value denotes an increase, and a negative value denotes a decrease. The model was calculated from days 0–504 for A5248 and from T1–T3 for LT-ART. ND, not done; –, no change, p > 0.05; RTE, recent thymic emigrant.

^aA5248 includes PWH (n = 10) in the 504 days following ART initiation.

^bp values and confidence intervals can be found in Tables S3 and S4.

^cLT-ART includes PWH (n = 10) who have been virus suppressed for an average of 6.7 years.

^dCD31+CD45RA+ gating was used to identifiy CD4+ RTEs but cannot be applied for identification of CD8+ T cell RTEs.⁷¹

Figure 3.

CD4⁺ T cell immune reconstitution in PWH following ART

(A) Mean %frequency (±SD) of selected CD4⁺ T cell subpopulations in the first 90 days post ART in A5248 (n = 10). %Frequencies at day 0 versus day 84 were compared using Wilcoxon signed-rank test.

(B) Mean log2 fold change in mean signal intensity (MSI) of markers on CD4⁺ T cells over time in A5248 relative to day 504 post ART (day 504 = 0/gray). Average %frequency is denoted by circle size. On the right, significance is indicated for two comparisons: MSI on day 0 versus day 84 (Wilcoxon signed-ranked test) and %change in MSI between day 0 to day 504 (GFE regression; see Table S5).

(C) Mean %frequency (±SD) of CD127⁺ cells in CD4⁺ T cell subsets post ART in A5248 (black) compared with LT-ART (n = 10, gray) over time points T1–T3 spanning 17–26 months.

(D) %Frequency of CD4⁺ T cell subpopulations in A5248 participants on day 504 post ART (light blue) compared with the LT-ART cohort (dark blue, average of 3 longitudinal time points, T1–T3, collected over 17–26 months).

(E) Correlation between years virally suppressed and %CD127⁺ on CD4⁺ T cell subsets in the LT-ART cohort.

All data were gated manually. In (C) and (D), %frequencies on day 504 post ART and LT-ART (average T1–T3) were compared using a Mann-Whitney U test. In (E), associations were tested using Spearman’s rank correlation. For all analyses: (A)–(D), ∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001, ∗∗∗∗p ≤ 0.0001. For (B), -- is p >0.05, elsewhere p > 0.5 left blank.

Over the full 500-day study window, a wider number of phenotypic changes, including many highly significant, were observed across CD4⁺ T cell subsets (Table 2; Figure 3; Tables S3 and S5). We observed that the %frequency of activation and cell cycling CD4⁺ T cell populations (HLA-DR⁺, CD38⁺, CCR5⁺, CD57⁺, PD-1⁺, and Ki-67⁺) decreased, consistent with a decrease in MSI of these proteins over the first 500 days following ART (Figure 3B; Table 2).

MSI changes in CD95 and CD28, which mark the transition from naive to memory T cells,⁷² were consistent with observed increases in naive CD4⁺ T cells and decreases in transitional memory, effector memory, and terminal effector CD4⁺ T cells subsets (Table 2). No change in the frequency of recent thymic emigrants (RTEs), stem cell memory, or central memory CD4⁺ T cells was detected. In total and multiple memory CD4⁺ T cell subsets, the %frequency of CD127 increased (Figure 3C; Table 2). CD127⁺ recent thymic CD4⁺ T cell emigrants also increased significantly (Table 2). Notably, the proportion of CD127⁺ CD99^hi CD4⁺ T memory cells also increased.⁷³ Altogether, increased CD127 frequencies are consistent with improved memory maintenance in CD4⁺ T cells following ART.²⁴

Relative to other changes, CD4⁺ T cell helper subsets, defined by surface expression of chemokine receptors that have been linked to function,^74,75,76,77 exhibited modest changes following ART initiation. Frequencies of Th1, Th17, and Tfh cells did not change, but we observed modest but significant increases in Th2 (+0.12%pts) and Th22 (+0.01%pts), whereas T regulatory (Treg; −0.08%pts), Th17.1 (−0.07%pts), and Th9 (−0.10%pts) populations decreased (Table 2). These changes reflected changes in the MSI of population markers post ART. CCR4, used to gate Th2, significantly increased, whereas the Treg cell lineage marker FoxP3 decreased (Figure 3B).

Overall, 500 days of ART produced sustained recovery of naive CD4⁺ T cells but had modest effects on frequencies of major memory and helper CD4⁺ T cell subsets. Within RTEs and less differentiated memory subsets, significant increases in the proportion of CD127⁺ cells were observed. ART induced early and sustained reduction of activated and cycling markers while increasing Bcl-2⁺ CD4⁺ T cells, altogether consistent with ART-mediated immune reconstitution of CD4⁺ T cells.

CD4⁺ T cell populations were highly stable in LT-ART individuals

In contrast to A5248, in the LT-ART cohort, CD4⁺ T cell memory subsets, helper subsets, and activation markers on CD4⁺ T cells were stable over 17–26 months with no significant changes detected over time in the frequency of measured populations, except CD127⁺ memory T cells (Table 2). Frequencies of CD127⁺ stem cell memory and CD127⁺ terminal effector CD4⁺ T cells increased over time (+0.10%pts and +0.25%pts, respectively) (Figure 3C; Tables 2 and S3; Figure S1).

To examine the level of restoration of CD4⁺ T cell population frequencies post ART, we compared means between A5248 day 504 and the average across time points 1–3 in the LT-ART cohort. In total CD4⁺ T cells, CD38 as well as CD38/HLA-DR frequencies were significantly higher on day 504 compared with the LT-ART cohort. We note that %HLA-DR⁺ and CD38⁺ correlated with time on ART in the LT-ART cohort (Figure S2D). Conversely, the CD127 and Bcl-2 frequencies were significantly lower (Figure 3D). We further examined CD127 expression in memory CD4⁺ T cell populations, again finding that frequencies of CD127 were significantly lower in naive, central, transitional, effector, and terminal effector memory populations (Figure 3C). We further noted, in the LT-ART cohort, that CD127⁺ frequency on effector memory and terminal effector CD4⁺ T cells directly correlated with years suppressed (p < 0.05, Spearman rank; Figure 3E). Altogether, this suggests that normalization of cellular activation and recovery of CD4⁺ T cell memory maintenance post ART is ongoing but slow, occurring over years.

CD8⁺ T cells exhibit more rapid memory restoration post ART than CD4⁺ T cells

We observed significant changes in frequencies of 22/25 CD8⁺ T cell populations in A5248 post ART (Figure 4A; Table 2) that were coupled with highly significant changes in MSI (Figure 4B).

Figure 4.

Memory and functional phenotypes of CD8⁺ T cells restore earlier than CD4⁺ T and NK cells in PWH following ART

(A) Mean (±SD) %frequency of CD8⁺ T cell subpopulations in the first 90 days post ART initiation in A5248 (n = 10). %Frequency on day 0 versus day84 was compared using Wilcoxon signed-rank test.

(B) Mean log2 fold change in MSI (yellow to blue) of markers on CD8⁺ T cells over time in A5248 (n = 10) relative to day 504 post ART. Average %frequency is denoted by circle size. On the right, significance is indicated for two comparisons: MSI on day 0 versus day 84 using Wilcoxon signed-ranked test and %change in MSI between day 0 to day 504, examined using gamma fixed-effects regression with log link (full details in Table S5).

(C) Comparison of the %frequency of CD8⁺ T cell subpopulations in A5248 participants on day 504 post ART (light blue, n = 10 participants) compared with durably suppressed (average 6.7 years) participants in the LT-ART cohort (dark blue, n = 10 average of 3 longitudinal time points, T1–T3, collected over 17–26 months).

(D) Mean frequency (±SD) of CD127⁺ cells in naive and memory CD8⁺ T cell subsets post ART in A5248 compared with LT-ART (n = 10, gray) over time points T1–T3 spanning 17–26 months.

(E) Average frequency (±SD) of NK cell subsets post ART in A5248 compared with LT-ART (n = 10, gray) over visits T1–T3.

%Frequencies on day 504 post ART and LT-ART (average T1–T3) in (C)–(E) were compared using a Mann-Whitney U test. For all analyses (A–E), ∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001, ∗∗∗∗p ≤ 0.0001. For (B), -- is p > 0.05, elsewhere p > 0.5 left blank.

Early decreases were observed in the proportions of CD8⁺ T cells expressing the activation markers Ki-67, CD38, HLA-DR, and PD-1 by day 84 (p < 0.05, Wilcoxon signed-rank test; Figure 4A) that continued across the 504-day study window (QBFE regression; Table 2). Other cellular markers of activation, CCR5 and CD38/HLA-DR, also decreased over the full study window.

Increased frequencies of Bcl-2⁺ CD8⁺ T cells were detected by day 84 and continued to increase across the full study window (Table 2), consistent with increased Bcl-2 MSI (Figure 4B) and previous studies.^23,78 Similar to CD4⁺ T cells, no early changes in the frequency of CD127⁺ CD8⁺ T cell were detected (Figures 4A and 4B), but over 500 days, the percentage and MSI of CD127 increased significantly on CD8⁺ T cells (Table 2; Figure 4B; Table S4). Many other phenotypic markers showed either highly significant increases (CD45RA, CCR7, CD28, and CCR4) or decreases (Ki-67, CD38, HLA-DR, PD-1, CD95, and CCR5) in cellular expression following ART initiation. Notably, the percentage and MSI of CD56, which has been associated with regain of lytic capacity in CD8⁺ T cells,⁷⁹ increased strongly over the 504-day study window (Figure 4B; Table 2).

Naive, stem cell memory, central memory, transitional memory, and terminal effector CD8⁺ T cell frequencies significantly increased, whereas the frequency of effector memory CD8⁺ T cells strongly decreased post ART initiation (Table 2). These changes in memory subsets reflected highly significant changes in expression of memory markers (CD45RA, CCR7, CD95, and CD28) in CD8⁺ T cells over the study window (Figure 4B).

When comparing the changes across time of CD4⁺ and CD8⁺ T cells, we noted the same trends (increases or decreases) between CD4⁺ and CD8⁺ T cells subsets; however, CD8⁺ T cell subpopulations exhibited more rapid changes than CD4⁺ T cells. CD8⁺ T cells showed a steeper average increase in Bcl-2 (+0.74%pts) and CD127 (+0.91%pts) compared with CD4⁺ T cells (+0.37% and +0.57%pts, respectively) as well as steeper average decreases in expression of activation markers (CCR5⁺, PD-1⁺, CD38⁺HLA-DR⁺) (Table 2). The frequency of CD127 in memory CD8⁺ T cell subsets also increased more strongly than in CD4⁺ T memory cells (Table 2). The one exception between CD4⁺ and CD8⁺ T cells was the activation/senescence marker CD57, which significantly increased in CD8⁺ T cells but decreased in CD4⁺ T cells (Table 2). Overall, the patterns of CD127 and CD57 expression in CD8⁺ T cells post ART complement previous in vivo labeling studies finding that HIV infection results in the loss of two long-lived CD8⁺ T cell populations, CD127⁺ central memory and CD45RA⁺CD57⁺ effector memory CD8⁺ T cells, in PWH.⁸⁰

We next compared CD8⁺ T cell subsets 504 days following ART in A2458 with the LT-ART group. We note first that, similar to CD4⁺ T cells in the LT-ART group, only small changes in CD8⁺ T cell subsets, mostly memory subsets, were observed over time. Average frequencies of CD38⁺ and HLA-DR⁺ but not CD38/HLA-DR+ CD8⁺ T cells in A5248 participants were significantly higher than in the LT-ART cohort (Figure 4C). Frequencies of Bcl-2⁺ and CD127⁺ CD8⁺ T cells did not significantly differ (Figure 4C). Within memory CD8⁺ T cell populations, CD127 frequencies also did not significantly differ between A5248 day 504 and LT-ART (Figure 4D), though we note, similar to CD4⁺ T cells, an ongoing increase in CD127⁺ CD8⁺ T cell frequencies with time on ART in the LT-ART group (Figure S2). Altogether, while our data suggest elevated CD8⁺ T cell activation 500 days post ART, markers of CD8⁺ T cell memory and survival were largely re-established.

Cytolytic NK cells restore more slowly than CD8⁺ T cells post ART

CD56^lo NK cell frequencies on day 504 in the A5248 group (average 16.1%, 1.5%–47.6%) were significantly lower than in the LT-ART cohort (average 43.7%, 24.6%–70.4%, p = 0.0003, Mann-Whitney U test), suggesting that, similar to CD4⁺ T cells, cytotoxic NK cells are slower to restore than CD8⁺ T cells post ART (Figure 4E).

VL decrease did not impact dominant CD4⁺ T cell clonotype frequency

Our phenotyping had shown that, in addition to absolute increases in CD4⁺ T cell counts (Figure 1D), CD4⁺ T cell populations post ART were undergoing significant immune reconstitution, albeit at slower rates than other immune lineages. Given these observations, we hypothesized that CD4⁺ T cell immune restoration would result in dynamic changes, likely expansion, of CD4⁺ T cell clonotypes post ART.

We sequenced the TCR of 10⁵ CD4⁺ T cells 0, 7/10, 21, 84, and 504 days post ART initiation from the A5248 (n = 4) cohort and 0, 4–5, and 17–19 months after time point 1 in the LT-ART group (n = 3). Morisita-Horn analysis showed no overlap between participants across both cohorts (Figure S6A). Of note, day 84 of 363043 exhibited lower similarity to the other four time points from this participant, suggesting a skewing of the TCR repertoire, possibly caused by an intercurrent infection.

We next compared the first and last visit in both cohorts (Figures 5A–5G, left column). Unsurprisingly, the majority of clonotypes sequenced (>85% across participants in both cohorts) were detected at only one time point. This reflects sample depth (10⁵) relative to the overall theoretical maximal TCR diversity of 10¹¹⁻¹⁵ and estimates of within-person diversity of 10⁶–10⁸.^81,82,83,84 For clonotypes detected at both time points, two distinct populations were observed. Higher-frequency clonotypes (≥0.1%) showed strong rank-order correlations, and lower-frequency clonotypes (<0.1%) showed a weaker, albeit strongly significant, correlations. We next examined whether dominant clonotypes (≥0.1% at the first and last visit) were stable over time and whether clonotype hierarchies were maintained (Figures 5A–5D, second column; Table S6). A Spearman’s rank correlation matrix was generated for frequencies of dominant clonotypes at all time points. All correlations were positive, producing significant r values between 0.684 and 0.991, except the day 0–504 comparison for participant 291374 (r = 0.714, p = 0.06, n = 8 clonotypes) (Figures 5A–5D, third column; Table S7). It is notable that, while the absolute number of dominant TCRβ clonotypes in each A5248 participant was relatively small (8–30 of tens of thousands of clonotypes), their cumulative total was substantial, contributing between 12.5% and 18.9% of all clonotypes per participant across all time points (Table S7).

Figure 5.

CD4⁺ TCR clonotypes and immunodominance hierarchies are maintained over time on ART

(A–D) Personal Identifier (PID) 611183 (A), 363043 (B), 363044 (C), and 291374 (D). Left: correlation between frequencies of TCRβ clonotypes on days 0 and 504 post ART in A5248 participants. Data were transformed (STAR Methods) prior to graphing. Percentages identify clonotypes detected on day 0 or 504 only (fall on dotted orange lines); clonotypes were detected at frequencies of less than 10⁻¹ or 10⁻¹ or greater. Spearman rank (r) correlation provided for clonotypes of less than 10⁻¹ or 10⁻¹ or greater. ns, non-significant. Center: frequency of dominant TCRβ clonotypes (≥0.1% at first and last visit) over time in A5248 participants. Right: matrix displaying correlations between frequencies of dominant TCRβ clonotypes on days 0, 7/10, 21, 84, and 504 in A5248. The absolute CD4 count (cells/mm³) post ART at each time point is displayed below.

(E–G) PIDs 728 (E), 674 (F), and 861 (G). Left: correlation between frequencies of TCRβ clonotypes at T1 and T3 in the LT-ART group. Center: dominant TCRβ clonotypes (≥0.1% at the first and last time point) over time in LT-ART participants. Right: matrix displaying correlations between frequencies of dominant TCRβ clonotypes at T1–T3 in LT-ART. The absolute CD4 count (cells/mm³) at each time point is displayed below.

(H and I) CD4⁺ and CD8⁺ T cell proliferation following stimulation with PHA, HCMV, pp65/IE1 peptides, or HIV Gag/Nef peptides in A5248 (H) and LT-ART (I).

(J) Day 420 T cell proliferation in A5248 and average proliferation in LT-ART for each participant.

Proliferation data (% CellTrace Violet [CTV]^lo) are shown as fold change between PHA/peptide-stimulated and mock-stimulated cultures (n = 8/group), showing mean ± SD. Spearman’s rank correlations were computed between frequencies of TCRβ clonotypes on different days (A–G), and log2 fold changes from A5248 and LT-ART cohorts in (J) were compared using a Mann-Whitney U test, where ∗p ≤ 0.05, ∗∗p ≤ 0.01, and ∗∗∗p ≤ 0.001.

Data from the three LT-ART participants were similarly plotted and analyzed (Figures 5E–5G). Similar to A5248 participants, dominant clonotypes in LT-ART participants were stable and largely maintained rank order over time (Figures 5E–5G; Tables S6 and S7).

We analyzed the diversity and evenness of all clonotypes, using Shannon entropy and Pielou’s evenness. Measures were similar between A5248 and LT-ART participants and did not exhibit any consistent trends, either increasing or decreasing, over time (Figure S6B). In summary, longitudinal TCR sequencing suggests that ART initiation has minimal impact on shifts of clonal homeostasis of CD4⁺ T cells.

Specificity of CD4⁺ T cell TCR clonotypes

We examined the VDJ database, which houses TCR sequences for unique human epitopes (1,080 TCRs at time of search) to determine whether dominant TCRβ clonotypes (≥0.1%) matched known clonotypes.^85,86 While human cytomegalovirus (HCMV)-, Epstein-Barr virus (EBV)-, and influenza-specific TCR clonotypes were detected in all participants, and HIV-1-specific TCR clonotypes were detected in four of seven participants, all clonotypes were low frequency (0.00044%–0.0072%) (Figure S6C).

HIV-specific CD4⁺ T cell proliferation was undetectable in the first year post ART

We next investigated the proliferation capacity of CD4⁺ T cells (n = 8/group). Our aim was to examine whether proliferation better reflected the dynamic phenotypic changes observed in CD4⁺ T cells post ART or, the relatively stable clonotypic structure of CD4⁺ T cells. Over the first year following ART initiation, strong CD4⁺ T cell proliferation was observed in response to the mitogen phytohemagglutinin (PHA) (Figure 5H). This suggests that total CD4⁺ T cells harbored proliferation capacity. HCMV-specific but not HIV Gag/Nef-specific CD4⁺ T cell proliferation in the first 14 months post ART was detected. By contrast, HIV-specific CD8⁺ T cell proliferation was detected in all participants and exhibited early kinetics similar to PHA- and HCMV-specific CD8⁺ T cells (Figure 5H).

In the LT-ART group, and as reported previously, both CD4⁺ and CD8⁺ T cells proliferated to HIV peptides, HCMV peptides, and PHA and were stable over time (Figure 5I).^87,88 Notably, in this group, the levels of HIV- and HCMV-specific proliferation in CD4⁺ T cells were comparable (Figure 5I). The significant differences in proliferation between the A5248 (420 days post ART) and LT-ART groups (Figure 5J) suggested a time-dependent restoration of HIV-specific CD4⁺ T cell proliferation capacity. We note that HLA-DR expression in the post-ART window ranged between 54% and 98% on monocytes and B cells, suggesting that peptide presentation to HIV-specific CD4⁺ T cells was not rate limiting in A5248.

Discussion

Chronic HIV infection causes extensive dysregulation of immune homeostasis.^89,90 In this study we documented the early kinetics of ART-mediated immune restoration, comparing data with LT virus-suppressed PWH. In A5248, first-line ART decreased HIV viremia rapidly across all study participants, producing a greater than 1-log decrease in virus loads within 10 days.^46,91 In all but one participant, absolute CD4⁺ T cell counts increased over 500 days, and CD4:CD8 ratios improved but remained inverted.

Against this background of changing absolute T cell numbers, ART produced broad changes in the frequencies and phenotypes of circulating leukocytes. Within 7 days of ART initiation, decreases in cell cycling were detected across NK cells, B cells, and γδ and αβ T cells. Cell cycling continued to decrease across all lineages over the first year of ART, after which minimal variation was observed. ART also induced early, albeit more gradual, decreases in cellular activation, most notably decreases in PD-1 and CD38 expression across B cells, NK cells, and γδ and αβ CD4⁺ and CD8⁺ T cells. CD38, which can mediate pro-inflammatory cytokine production,⁹² showed dynamic changes, with the percentage of CD38⁺ cells decreasing between 11.3% (CD19⁺ B cells) and 40.7% (CD8⁺ T cells) in leukocyte populations over the first 500 days following ART. For CD4⁺ and CD8⁺ T cells, the frequency of CD38⁺ cells was significantly higher 500 days post ART compared with PWH who had been suppressed, on average, for 6.7 years.

Strong and sustained increases in Bcl-2 expression, which induces anti-apoptotic signaling,^93,94 were observed across all lymphocyte lineages, specifically B, NK, and αβ and γδ T cells. The frequency of T and NK cells, but not B cells, expressing CD127 increased following ART initiation. In T cells, CD127 expression is necessary for the maintenance of LT T cell memory.⁹⁵ The frequency of CD127⁺CD8⁺ T cells was comparable with the LT-ART group 500 days following ART, whereas CD127⁺CD4⁺ T cell frequencies remained significantly lower for total CD4⁺ T cells as well as multiple memory subpopulations. Additional analysis of the LT-ART cohort showed ongoing increases in the proportion of CD127⁺ effector memory and terminal effector CD4⁺ T cells over time on ART. Altogether, this suggested that full reconstitution of LT CD4⁺ T cell memory takes years.

ART-mediated immune restoration occurred more slowly in CD4⁺ than in CD8⁺ T cells, consistent with ongoing inverted CD4:CD8 ratio observed in many PWH on ART.^31,32 Low CD4 counts and an inverted CD4:CD8 ratio are strongly associated with significantly elevated co-morbidities in PWH.^32,34,35 T cell populations compete for IL-7 and other homeostatic cytokines.^96,97 The relatively higher numbers of CD8⁺ T cells at ART initiation may out compete CD4⁺ T cells, slowing their recovery. Other mechanisms could contribute. CD4⁺ T cells also restore more slowly than CD8⁺ T cells following myeloablation and hematopoietic stem cell transplantation (HSCT), suggesting intrinsic differences between CD4⁺ and CD8⁺ T cells.^98,99,100 HCMV seropositivity has also been associated with maintenance of elevated CD8⁺ T cell frequencies in ART-treated individuals.^31,47,101

We observed that HIV-specific T cell proliferation was profoundly impaired in untreated HIV infection, as reported previously.^102,103 HIV-specific CD8⁺ but not HIV-specific CD4⁺ T cell proliferation was detected in the first 500 days following ART and was stable in LT-ART participants.^104,105 Detection of HIV-specific CD4⁺ T cell proliferation in LT-ART participants suggests that HIV-specific CD4⁺ T cell clonotypes were present but functionally unresponsive in the early post-ART period.¹⁰⁶ Again, these proliferation data are consistent with the slower restoration of functional phenotypes, including CD127⁺ T cell memory observed in CD4⁺ versus CD8⁺ T cells post ART.

TCRβ clonotyping of CD3⁺ T cells in HIV infection has been reported previously, describing modestly increased clonotypic diversity following ART.¹⁰⁷ Here we performed a longitudinal analysis of CD4⁺ T cell TCRβ clonotypes pre and post ART treatment. Our results were surprising. At the pre-ART time point, we observed individual dominant TCRβ clonotypes ranging from 0.1%–11% of all CD4⁺ T cells sequenced that together contributed between 16.2% and 23.6% of all clonotypes at this time point. These dominant TCRβ clonotypes were stably detected with similar relative frequencies over the following 500 days post ART initiation. This pattern mirrored detection of stable CD4⁺ T cell clonotypes in LT-ART participants in whom CD4⁺ T cell phenotypes and function were mostly stable over time. Across all participants studied, pre/post ART and LT-ART, the rank order of dominant TCRβ clonotypes was strongly maintained over time, suggesting that clonal homeostasis of dominant clonotypes is maintained independent of dysregulated CD4⁺ T cell absolute numbers, T cell memory, helper frequencies, and function. We note, that because of sampling depth, we cannot draw similar conclusions regarding lower-frequency clonotypes (<0.1%).

An important question is what maintains the rank order of dominant CD4⁺ T cell TCRβ clonotypes over time. While we were unable to identify the specificity of dominant clonotypes using the VDJ database, several studies have reported dominant clonotypes specific to persistent viral HCMV and EBV antigens.^108,109,110 In PWH cohorts, dominant clonotypes could include HIV-specific clonotypes that are maintained during ART following stochastic reactivation and antigen presentation of HIV-infected cells. Persisting HIV-specific CD4⁺ T cell clonotypes would be consistent with our observations that HIV specific proliferation, while undetectable in the year post ART, was detectable years later. Our current studies are investigating alternative approaches to define the specificity of these dominant clonotypes as well as their transcriptomic profiles at different stages of immune restoration.

Stable expanded TCR clonotypes have also been reported in a range of immune conditions, including individuals in good health,^111,112,113 individuals with autoimmune systemic lupus erythematosus,¹¹³ cancer patients undergoing myeloablation and HSCT,^114,115 and even non-human primates following CMV infection.¹¹⁶ While these studies did not perform deep immune profiling, they represent a broad range of immune states consistent with a model where clonal homeostasis is independent of other facets of T cell homeostasis and suggest that our observations in PWH have broad application.

This study has several implications for therapeutic strategies targeted against HIV. Longitudinal sequencing of the HIV provirus has identified identical or near-identical HIV provirus years apart in ART-treated individuals.^38,39,40 Given the very high mutation rates of HIV,^117,118 these observations are best explained by LT survival or homeostatic maintenance of the latently infected CD4⁺ T cells. Other HIV sequencing studies have shown that HIV reservoir virus can reflect virus archived throughout infection but that, on average, most virus is archived in and around the time of ART initiation.^41,42,43 Here, high-frequency CD4⁺ T cell clonotypes were observed at ART initiation, in agreement with a previous study describing expanded HCMV-specific CD4⁺ clonotypes in HIV-viremic individuals.¹⁰⁸ We speculate that, despite the absolute loss and dysregulation of CD4⁺ T cells in untreated HIV infection, CD4⁺ T cells, particularly those expressing dominant clonotypes, are maintained over time. HIV infection of even a small proportion of these CD4⁺ T cells would be consistent with observations that the HIV reservoir virus can date from early infection.

The overrepresentation in the HIV reservoir of viruses circulating in and around the time of ART initiation is also consistent with restoration of CD4⁺ T cell longevity observed in our phenotyping studies. We have proposed that this early post-ART period may be a therapeutic window during which to destabilize and decrease the HIV reservoir,^42,44 particularly by targeting CD4⁺ T cell immune reconstitution, blocking the transition of CD4⁺ T cells from more short-term to long-lived cells. Our data, which show slow CD4⁺ T cell immune reconstitution occurring over years, suggests that a therapeutic window may be wide. In addition, CD4⁺ T cell-targeted therapies will likely impact multiple leukocyte lineages that also undergo immune reconstitution in the post-ART window.

A direct translational implication of these data relates to the design of antiretroviral treatment interruption (ATI) studies in PWH. These studies examine whether therapeutics, small-molecule inhibitors and immunotherapies, can delay time to virus rebound in PWH following ATI. We show that, even when CD4 counts have restored to greater than 400 within 1 year, CD4⁺ T cell phenotype and function exhibit persistent dysregulation. Therefore, interrupting ART in the first year post ART initiation may slow this restoration even further. Second, approaches that improve cytolytic activity of CD8⁺ T and NK cells and even γδ T cells are being investigated to target and clear HIV-infected cells and to cure hepatitis B in PWH in the post-ART window and during treatment interruption.^{119,120,121,122,123} Here, we show clear differences in the rates of immune reconstitution in terms of frequencies and phenotypes of cytotoxic cells post ART. Most notably, the frequency of CD56^lo NK cells that mediate cytolytic NK cell activity recovered more slowly than CD8⁺ T cells. These differences in kinetics of immune reconstitution will help inform the timing of immune interventions for HIV cure. Altogether, these studies suggest that harnessing endogenous immunity may be more effective when phenotypic and functional homeostasis has been re-established following several years of ART.

Limitations of the study

In this study, we show the global changes wrought by HIV infection and that host immunity likely takes years to restore post ART, akin to reforestation following a fire. To inform ATI design, additional post-ART studies, specifically 2–6 years post ART, that include more detailed phenotyping of leukocytes are needed to fully document recovery of host immunity. Last, participant numbers limited our ability to examine changes in slope of immune populations within the post-ART period. However, the highly consistent changes observed between participants underscores the global immune dysregulation caused by uncontrolled HIV viremia and the robust, albeit slow and steady, effect of modern ART regimens to restore immune homeostasis.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
CD45 HI30 111Cd	Standard BioTools	Cat# 3111001B
CCR6 11A9 141Pr	Standard BioTools	Cat# 3141014A
CD19 HB19 142ND	Standard BioTools	Cat# 3142001B; RRID: AB_2651155
CD45RA HI100 143ND	Standard BioTools	Cat# 3143006B; RRID: AB_2651156
CD4 RPA-T4 145ND	Standard BioTools	Cat# 3145001B; RRID:AB_2661789
CCR4 L291H4 149Sm	Standard BioTools	Cat# 3149029A; RRID: AB_2938871
CD14 M5E2 151Eu	Standard BioTools	Cat# 3151009B; RRID: AB_2810244
CXCR5 RF8B2 153Eu	Standard BioTools	Cat# 3153020B
PD-1 EH12.2H7 155Gd	Standard BioTools	Cat# 3155009B; RRID: AB_2811087
CXCR3 G025H7 156Gd	Standard BioTools	Cat# 3156004B; RRID: AB_2687646
CD27 L128 158Gd	Standard BioTools	Cat# 3158010B; RRID: AB_2858231
FoxP3 259D/C7 159Tb	Standard BioTools	Cat# 3159028A; RRID: AB_2811088
CD28 CD28.2 160Gd	Standard BioTools	Cat# 3160003B; RRID: AB_2868400
Ki-67 B56 162Dy	Standard BioTools	Cat# 3162012B; RRID: AB_2888928
CD95 DX2 164Dy	Standard BioTools	Cat# 3164008B; RRID: AB_2858235
CD127 A019D5 165Ho	Standard BioTools	Cat# 3165008B; RRID: AB_2868401
CCR7 G043H7 167Er	Standard BioTools	Cat# 3167009A; RRID: AB_2858236
CD8 SK1 168Er	Standard BioTools	Cat# 3168002B; RRID: AB_2892771
CD25 2A3 169Tm	Standard BioTools	Cat# 3169003B; RRID: AB_2938861
HLA-DR L243 170Er	Standard BioTools	Cat# 3170013B; RRID: AB_2888929
CCR5 NP-6G4 171Yb	Standard BioTools	Cat# 3171017A
CD38 HIT2 172Yb	Standard BioTools	Cat# 3172007B; RRID: AB_2756288
CD56 N901 176Yb	Standard BioTools	Cat# 3176009B, RRID: AB_2811096
CD16 3G8 209Bi	Standard BioTools	Cat# 3209002B; RRID: AB_2756431
CD45 HI30 89Y	Standard BioTools	Cat# 3089003B; RRID: AB_2938863
CD31 WM59 144ND	Standard BioTools	Cat# 3144023B
CD3 UCHT1	Biolegend	Cat# 300443; RRID: AB_2562808
CD99 3B2/TA8	Biolegend	Cat# 371302; RRID: AB_2572177
Bcl-2 100	Biolegend	Cat# 658702; RRID: AB_2562958
TCRgd B1	Biolegend	Cat# 331202; RRID: AB_1089222
CD57 HCD57	Biolegend	Cat# 359602; RRID: AB_2562403
CCR10 314305	R&D systems	Cat# MAB3478; RRID: AB_2275692
CD3 UCHT1 AF488	Biolegend	Cat# 300415, RRID: AB_389310
CD4 OKT4 BV650	Biolegend	Cat# 317436, RRID: AB_2563050
CD8 SK1 BV510	Biolegend	Cat# 344732, RRID: AB_2564624
CD14 M5E2 PerCP Cy5.5	Biolegend	Cat# 325622, RRID: AB_893250
CD19 HIB19 PerCP Cy5.5	Biolegend	Cat# 302230, RRID: AB_2073119
CD56 HCD56 PerCP Cy5.5	Biolegend	Cat# 318322, RRID: AB_893389
CD16 3G8 PerCP Cy5.5	Biolegend	Cat# 302028, RRID: AB_893262
Biological samples
Human PBMCs	Human	This paper
Chemicals, peptides, and recombinant proteins
Phytohemagglutinin	Thermofisher	Cat# 10576015
Cell-ID Intercalator-Ir	Standard BioTools	Cat# 201192A
Cell-ID Cisplatin-195Pt	Standard Biotools	Cat# 201195
Critical commercial assays
115 Indium (conjugated to CD3)	Millipore Sigma	Cat# 203440
Maxpar X8 Antibody Labeling 147Sm	Standard BioTools	Cat# 201147A
Maxpar X8 Antibody Labeling 152Sm	Standard BioTools	Cat# 201152A
Maxpar X8 Antibody Labeling 173Yb	Standard BioTools	Cat# 201173A
Maxpar X8 Antibody Labeling 175Lu	Standard BioTools	Cat# 201175A
Maxpar X8 Antibody Labeling 166Er	Standard BioTools	Cat# 201166A
Ca2+ Mg2+ free PBS	Rockland	Cat# MB008
Maxpar Cell Staining Buffer	Standard BioTools	Cat# 201068
Human TruStain FcX	Biolegend	Cat# 422301
FoxP3 Fixation/Permeabilization kit	Thermofisher	Cat# 00-5523-00
Zombie NIR Fixable Viability Kit	Biolegend	Cat# 423106
CD3/28 Human Dynabeads	Thermofisher	Cat# 11131D
CD4+ T cell isolation kit	Miltenyi Biotec	Cat# 130-096-533
RNeasy Plus Micro kit	Qiagen	Cat# 74034
Qubit RNA Assay kit	Invitrogen	Cat# Q32855
High Sensitivity RNA ScreenTape	Agilent	Cat# 5067-5579
SMARTer Human TCR a/b Profiling Kit v2	Takara Bio USA	Cat# 634479
Qubit DNA Assay kit	Invitrogen	Cat# Q32854
D5000 ScreenTape	Agilent	Cat# 5067-5592
10% PhiX control v3	Illumina	Cat# FC-110-3001
Deposited data
Preprocessed mass cytometry data	This study	https://zenodo.org/record/7495836
TCR sequencing data	This study	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSM7830997
Software and Algorithms
Cytobank v9.3	Sigma Aldrich	https://premium.cytobank.org/cytobank/
Prism v9	Graphpad	https://www.graphpad.com/
FlowJo v10.8.1	BD Biosciences	https://www.flowjo.com/
Python v3.6		https://www.python.org/
Mixcr v2.1.9.6	Bolotin et al.124,125	https://github.com/milaboratory/mixcr/tree/v2.1.9
R v4.2.2	R Core Team	https://cran.r-project.org/

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Dr. Nilu Goonetilleke (nilu_goonetilleke@med.unc.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

The preprocessed mass cytometry data used in this study are publicly available in the Zenodo repository: https://doi.org/10.5281/zenodo.7986013. All source code for the reproduction of the results, including R code for the fixed-effects regressions and Python code for the unsupervised clustering and elastic net regression are publicly available in the laboratory GitHub repository: https://github.com/glab-hiv/immune-recovery. TCR sequencing data from this study are available in the following repository: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSM7830997. Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Experimental model and study participant details

Study participants

A5248 ART initiation cohort: HIV seropositive adults (≥18 years) with plasma HIV-1 RNA levels >10,000 to ≤300,000 copies/mL were enrolled in AIDS Clinical Trial Group (ACTG) protocol A5248. Participants received raltegravir and tenofovir disoproxil fumarate plus emtricitabine (#NCT00660972).⁹¹ At each visit, viral loads were assessed by Amplicor HIV-1 Monitor, version 1.5, UltraSensitive protocol (≤50 copies/mL; Roche Molecular Systems, Branchburg, NJ), CD4⁺ T cell counts collected and self-reported adherence assessments submitted. Peripheral blood mononuclear cells (PBMC) and plasma were isolated from blood collected at days 0 (pre-ART), 2, 7, 10, 14, 21, 28, 56 and weeks 12, 16, 20, 24, 36, 48, 60 and 72 (Table S1; Figure 1A).

LT-ART, durable viral suppression group: Adults (≥18 years) who initiated ART in chronic HIV infection were enrolled through the UNC Chapel Hill HIV-1 Clinical Trials Unit or the Women’s Interagency HIV-1 Study (WIHS) UNC Chapel Hill Site. HIV infection was defined as a positive licensed HIV ELISA test confirmed by either Western blot (WB), multispot HIV assay, HIV-1 RNA, or two documented plasma HIV-1 RNA >1000 copies/mL. For this study we investigated participants who met the following additional criteria: i) a stable ART regimen for ≥12 months defined as one or more nucleoside/nucleotide reverse transcriptase inhibitors plus a non-nucleoside reverse transcriptase inhibitor, integrase inhibitors, or a protease inhibitor without interruption ii) missing no more than 2 consecutive days or no more than 4 cumulative days of ART doses in the 24 weeks prior to the first study timepoint ii) plasma HIV-1 RNA <50 copies/mL for ≥12 months prior to sample collection as defined by a minimum of two tests (limit of detection determined by assay employed). For each study participant that had met these criteria, PBMC had been collected at three time points over a 17-26-month window (Table S1; Figure 1B). Average time on ART at time point 1 was 6.7 years and at time point 3 was 8.6 years.

Study approval

A5248

The institutional review board at each of the participating clinical research sites approved the study protocols, and all subjects provided signed informed consent.⁹¹

LT-ART

Review and implementation of all protocols utilized for the collection of samples for this analysis were approved by the University of North Carolina at Chapel Hill Biomedical Institutional Review Board (IRB) (ethics numbers: 14–0741, 11–0228, and 13–3613, 15–1626). All participants provided written informed consent.

Method details

Culture media

R10+

RPMI (Corning, MT10040CV) + 10% FBS (VWR, 97068-088, Lot# 079K16) + 10mM HEPES (Corning, 20-060-CL) + 2mM L-glutamine (Corning 25005CL) + 1mM sodium pyruvate (Gibco 11360-070) + penicillin-streptomycin (Sigma, P0781).

RAB10

RPMI (Corning, MT10040CV) + 10% Human AB Serum (Gemini Bio, 100–512, Lot# H121059) + 10mM HEPES (Corning, 20-060-CL) + 2mM L-glutamine (Corning 25005CL) + 1mM sodium pyruvate (Gibco 11360-070) + penicillin-streptomycin (Sigma, P0781).

Mass cytometry staining

Optimization and standardization of the mass cytometry panel was detailed previously.¹²⁶ PBMC were thawed and rested overnight in R10+ at 37°C with average viability of 89% (71–97%). 10⁷ cells/mL PBMC suspension was prepared in Ca²⁺ Mg²⁺ free PBS (Rockland, Cat# MB008) and cell-ID Cisplatin 195 (Fluidigm Cat# 201195) was added. Staining was quenched with Maxpar Cell Staining Buffer (CSB) (Fluidigm Cat# 201068) and PBMC then incubated with Fc blocking antibody for 10 min (Biolegend Cat# 422301). Surface antibodies (see key resources table) were added to PBMC, incubated at 4°C for 30 min and then washed. PBMC were fixed and permeabilized using a FoxP3 Fixation/Permeabilization kit as per manufacturer’s protocol (Thermofisher Cat# 00-5523-00). Intracellular antibodies (see key resources table) were added and PBMC incubated at 4°C for a further 30 min. PBMC were washed then incubated in 1 mL 2% paraformaldehyde in CyPBS (Fluidigm Cat# 201058) for 10 min at room temperature. Cell intercalation solution (Fluidigm Cat# 201192A) was added and each sample gently mixed. Cells were kept overnight at 4°C and the following day washed with CSB and then cell acquisition solution (Fluidigm Cat# 201240). Samples were acquired on the Helios (Fluidigm). 700,000 events were acquired for each sample (on average 380,000 Live+ DNA+ cells) that enabled acquisition of at least 10,000 CD4⁺ T cells/sample including participants with absolute CD4 counts of less than 100 cells/mm³. 10,000 CD4⁺ T cells enabled detection of CD4⁺ T cell subpopulations at a frequency of 1% with a coefficient of variation of 10%.¹²⁷

Manual gating of data

Cytobank (9.3) was used to manually gate cell populations. Plots were ArcSinh transformed with a co-factor of 5 and then gates were set using either unstained or antibody backbone-stained controls.¹²⁶ The unstained tube was used to confirm there was no environmental metal contamination and to appropriately identify CD45⁺ cells. The antibody backbone tube which consisted of all major lineage markers (CD45, CD3, CD4, CD8, TCRγδ, CD19, CD14, CD56 and CD16) was used to set gates downstream of these lineages markers. We report %frequencies and MSI, the latter being the equivalent of mean fluorescence intensity (MFI) in flow cytometry.

Flow cytometry proliferation assays

PBMC were rested overnight then incubated with 5μM of Cell Trace Violet (CTV) (Invitrogen Cat# C34571) for 20 min at 37°C and washed with 40mL of 2% FBS/PBS. PBMC (2 x10⁶/mL) were then stimulated with pools of peptides spanning either human cytomegalovirus (HCMV) proteins pp65/IE1 (15-mer overlapping by 11 amino acids, 1 μg/mL), HIV Gag/Nef HIV (18-mer overlapping by 10 amino acids, 1 μg/mL) or phytohemagglutinin (PHA, 3 μg/mL) in RAB10 and cultured for 5 days at 37°C. Cells were washed then stained with viability dye (Zombie NIR Biolegend Cat #423106) and CD3 AF488 (Biolegend Cat # 300415), CD4 BV650 (Biolegend Cat # 317436), CD8 BV510 (Biolegend Cat # 344732), CD14 PerCP Cy5.5 (Biolegend Cat # 325622), CD19 PerCP Cy5.5 (Biolegend Cat # 302230), CD56 PerCP Cy5.5 (Biolegend Cat # 318322), and CD16 PerCP Cy5.5 (Biolegend Cat # 302028). Cultures acquired at day 5 had at over 90% viability of CD4⁺ and CD8⁺ T cells and at least 10,000 CD4⁺ T cells were acquired on BD LSRFortessa. Data were analyzed with FlowJo (version 10.8.1). The fold-change in CTV^lo cells of PHA/antigen-stimulated cells over mock stimulated cultures was calculated, then log2 transformed. A positive proliferative T cell response was defined as 3-fold over mock (1.58 when log2 transformed).

CD4⁺ T cell isolation

CD4⁺ T cell isolation was performed per manufacturer’s instructions (Miltenyi Biotec Cat #130-096-533). In brief, PBMC were washed in PBS +0.5% Bovine Serum Albumin + 2mM EDTA and incubated with CD4⁺ T cell Biotin Antibody Cocktail at 4°C. Then, CD4⁺ T cell MicroBead Cocktail was added and incubated at 4°C. Cells were then put through a magnetic column (Miltenyi Biotec Cat #130-042-401). Average CD4⁺ T cell purity was 96% (91–99%) and viability was 90% (83.5–96.1%).

TCR sequencing

After isolation, CD4⁺ T cells were stimulated with CD3/28 beads at 1:1 to increase transcription (Thermofisher Cat# 113.31D) for 4 h at 37°C. Note: The 4-h stimulation is not long enough for CD4⁺ T cells to divide.¹²⁸ Stimulated CD4⁺ T cells were then washed with PBS, lysed and total RNA extracted using the RNeasy Plus Micro kit (Qiagen 74034). RNA was quantified using a Qubit fluorometer (Qubit RNA Assay kit, Invitrogen Q32855), and quality was assessed with the Agilent TapeStation using High Sensitivity RNA ScreenTape (Agilent 5067–5579) to measure the integrity of the extracted RNA. All samples had a RIN score >8. Extracted RNA was stored at −80°C.

Libraries were prepared for TCR sequencing using the SMARTer Human TCR a/b Profiling Kit v2 (Takara Bio USA, Inc., Cat# 634479). Typically, 30ng total RNA was used as input for cDNA synthesis using Takara’s oligo dT primers (TCR SMART UMI Oligo). β-chain repertoire of the TCR variable region was targeted for library construction. Library quantity, purity and size selection were assessed using Qubit fluorometer (Qubit DNA Assay kit, Invitrogen Q32854) and an Agilent Bioanalyzer (D5000 ScreenTape, Agilent 5067–5592) with acceptable library size distribution between 600 and 1150bp. Sample TCR libraries were pooled with 10% PhiX control v3 (Illumina, Cat# FC-110-3001). Sequencing was performed on an Illumina NextSeq2000 sequencer with 2 × 150 base paired end read chemistry through the UNC High-Throughput Sequencing Facility.

TCR repertoire analysis

TCR inference from TCR amplicon sequencing FASTQ files was performed with Mixcr (v2.1.9–6),^124,125 where paired-end reads were subject to alignment in default mode, followed by contig assembly and export. Sample diversity metrics were calculated using the Python package scikit-bio (v0.5.6). The process of FastQC (v0.11.7) quality checks, MiXCR inference, and diversity metric calculation was facilitated by the in-house tool FASTMixcr (v0.5) running on Python (v3.6). Following FASTMixcr, further analysis was run using the outputs of Mixcr. Morisita-Horn Overlap Index was used to measure the similarities between samples.

TCR clonotype frequencies were transformed by adding 0.00001 to each clonotype frequency (%) and then plotted on log10 scale.

Quantification and statistical analysis

Quantitative analyses

Statistical analyses were performed using R (v 4.2.2) to fit the fixed-effects regressions (see fixed-effects regression), Python (v3.6) to fit the elastic net regressions (see elastic net regression), and GraphPad Prism Version 9 for all other analyses. Non-parametric approaches were utilized for unpaired and paired comparisons between cohorts (Mann-Whitney U test, Wilcoxon signed rank test) and for correlation analyses (Spearman’s rank). All tests were two-sided, and no adjustments were made for multiple testing.

Unsupervised analysis data preprocessing

All mass cytometry samples were first preprocessed by gating on live, singlet cells, then ArcSinh transformed with a co-factor of 5 prior to analysis. To identify a limited subset of representative cells for more efficient computation, an equal number of cells ($n=2500$) were selected from each profiled sample using kernel herding sketching and then vertically concatenated into a single matrix for clustering.¹²⁹ Kernel herding sketching performs principled downsampling to select cells across all major populations by approximating the kernel mean embedding of the original dataset. The approach selects prototypical cells that are representative of the original distribution of cell type frequencies, while ensuring rare cell types are sufficiently sampled.

PhenoGraph

To partition cells into biologically cohesive cell populations in an automated manner, computation-efficient metaclustering was implemented using the PhenoGraph algorithm⁵³ with Leiden clustering¹³⁰ (PhenoGraph v1.5.7 (k = 30 neighbors in the k-nearest neighbor graph, resolution parameter = 1.5)) in a two-step process. First, cell populations were identified in each profiled sample based on the expression of 25 cell surface proteins (see Table S7, highlighted in blue). To identify cell populations shared across patients over time within a cohort, sample-specific cell populations were subsequently grouped into metaclusters using the PhenoGraph algorithm on the cluster centroids from each sample based on the median expression across all features.

Engineering mass cytometry immune features from unsupervised clusters

Following metaclustering, four categories of descriptive features were computed for each profiled patient sample. First, cell type frequency features were defined for each metacluster as the proportion of the profiled sample’s cells assigned to a metacluster under the PhenoGraph metaclustering approach (analogous to frequency features defined through manual gating). Next, to quantify changes in cell type abundance over time, the difference in the total number of cells within a metacluster was computed between the initial ($t=0$) and subsequent ($t>0$) samples collected from a patient. Per cluster functional features were defined as the median expression of the following protein markers: PD-1, Bcl-2, Ki-67, CCR5, CD38, HLA-DR, CD57, and CD127. Lastly, to quantify changes in signaling responses, the difference in the median expression of a functional protein was computed between the initial ($t=0$) and successive timepoints ($t>0$). The four categories of metacluster-derived immune features were then horizontally concatenated into a $N\times 4C$ feature matrix, where $N$ is the number of profiled samples and $C$ is the number of metaclusters.

Elastic net regression

To ascertain immune cell phenotype and functional changes associated with initiation of antiretroviral therapy or durable suppression, a predictive elastic net regression model was trained on the metacluster-derived features from each patient cohort. The elastic net is a regularized linear regression model that uses $L_1$ and $L_2$ penalties to constrain the coefficients to ensure that the model is sparse, while also allowing for correlated features.⁵⁴ Given a matrix of metacluster-derived immune features for patients within a cohort, $X$, and a response vector corresponding to the time since receiving antiretroviral therapy, $y$, elastic net regression was used to estimate the vector of coefficients, $\beta$, that minimize the sum of squared error between the time following therapy and the predictions.

$$L\left(\beta \right)={\Vert y-X\beta \Vert }^2+{\lambda }_1{\Vert \beta \Vert }_1+{\lambda }_2{\Vert \beta \Vert }_2^2$$ (Equation 1)

To test the generalizability of the elastic net regression model and account for dependencies between samples acquired from the same patient over time, a leave-one-out cross validation procedure was used. For each cross-validation fold, all samples collected from one patient were excluded to train the model. Elastic net regularization parameters, ${\lambda }_1$ and ${\lambda }_2$, were then tuned over a grid search with leave-one-out cross validation prior to estimating the coefficients for immune features and predicting the time following antiretroviral therapy for the excluded patient. Model fit was subsequently assessed by computing the coefficient of determination, $R^2$.

Fixed-Effects regression

Quasi-binomial to quantify changes in percent frequency

To examine longitudinal changes in the proportion of different cell populations (e.g., CD4 cells) expressing certain biomarkers (e.g., CCR5+) or a combination of biomarkers (e.g., CCR7+CD45RA+), quasi-binomial fixed-effects regression models¹³¹ were fit separately for data from each cell population and each cohort (LT-ART and A5248). Specifically, for the $i^{th}$ timepoint in the $j^{th}$ patient ($j=1,\dots ,m$), the counts $Y_{ij}$ of cells in the population expressing the biomarker(s) were modeled as

$$\log \left\{p_{ij}/\left(1-p_{ij}\right)\right\}={\beta }_{1}Month{s}_{ij}+{\alpha }_{1}I\left(PI{D}_j=1\right)+{\alpha }_{2}I\left(PI{D}_j=2\right)+\dots +{\alpha }_{m}I\left(PI{D}_j=m\right)\text{,}$$ (Equation 2)

where $E\left(Y_{ij}|Month{s}_{ij},{PID}_j\right)=n_{ij}{p}_{ij}$ and $Var\left(Y_{ij}|Month{s}_{ij},{PID}_j\right)=\phi n_{ij}{p}_{ij}\left(1-p_{ij}\right)$ are the mean and variance of the counts conditional on time (in months) and participant ID. In these formulas, $p_{ij}$ denotes the proportion of the cell population having the biomarker, $n_{ij}$ denotes the total number of cells, and $\phi$ denotes the dispersion parameter. The parameter $\phi$ reflects additional variance that may be present in the data that could not otherwise be described by a simple binomial regression. In Equation 2, ${\beta }_1$ represents the change in the log odds of the cell expressing certain biomarker(s) for every month increment (30 days), the $\alpha$ parameters represent the patient-specific intercepts for $m$ total patients, and the indicators are such that $I\left(PID=m\right)=1$ if participant ID is $m$ and 0 otherwise. Notice that there are as many indicators as there are patients ($m$ total), so an overall intercept is not included to avoid overparameterization. Based on the fitted model, the average marginal effect of the $Months$ variable was estimated (using the margins package in R¹³² and reported as a percentage point (%pt) change (i.e., the difference between two percentages). The average marginal effect quantifies the average change in the outcome per unit change in the predictor variable (in this case, $Months)\text{.}$ For example, if the estimated %pt change was 0.5, this would imply that the proportion of cells in population expressing the biomarker(s) increases by 0.5% per month on average according to the fitted model.

Gamma to quantify changes in mean signal intensity

Longitudinal changes in proteins expressed on the cell surface and intracellularly as measured by mean signal intensity (MSI) were also examined. Since values of MSI are positive and continuous, a gamma fixed-effects regression with log link and indicators for PID was fit separately to the data from each expressed protein and each cohort (LT-ART and A5248), and deviance residuals were inspected for normality. Specifically, for the $i^{th}$ timepoint in the $j^{th}$ patient ($j=1,\dots ,m$), ${MSI}_{ij}$ was modeled per expressed protein as

$$\log \left\{{\mu }_{ij}\right\}={\beta }_{1}Month{s}_{ij}+{\alpha }_{1}I\left(PI{D}_j=1\right)+{\alpha }_{2}I\left(PI{D}_j=2\right)+\dots +{\alpha }_{m}I\left(PI{D}_j=m\right)\text{,}$$ (Equation 3)

where ${MSI}_{ij}\sim \text{Gamma}\left(k_{ij},{\theta }_{ij}\right)$, $E\left(MS{I}_{ij}|Month{s}_{ij},PI{D}_j\right)={\mu }_{ij}=k_{ij}{\theta }_{ij}$ and $Var\left({MSI}_{ij}|Month{s}_{ij},PI{D}_j\right)=\phi {\mu }_{ij}^2=k_{ij}{\theta }_{ij}^2$ are the mean and variance of the MSI conditional on time (in months) and participant ID. In these formulas, $k_{ij}$ and ${\theta }_{ij}$ denote the gamma shape and scale parameters, respectively, and the variance is a function of the mean, so any shifts in the mean will impact the variance. The remaining covariables are as described after Equation 2. The estimate that we derive from the model in Equation 3 is expressed as a percent change. For example, if the expression of the protein XYZ as measured by MSI is 200 one month and 210 the next month, then there is a 5% increase in expression of protein XYZ. This value is calculated as (210-200)/200 × 100%, and the model in Equation 3 is used to estimate this average percent change (e.g., 5%) for each protein.

These analyses were chosen over mixed-effects models for two reasons: 1) fixed-effects models have been shown to be less susceptible to bias for a small number of clusters (in this case, patients),¹³³ and 2) variation between individuals from the random effects can introduce confounding bias in the coefficient estimates if key time-invariant covariates are omitted from the model.¹³⁴ Fixed-effects methods, which rely only on variation within individuals, control for all unmeasured characteristics as long as those characteristics do not change over time.¹³⁵

Published: November 9, 2023