Empirical PK/PD Model for Differentiating Intrinsic Dosing Rules by Therapeutic Indication With Relevance for HIV‐1 Latency Reversal

Abstract

Romidepsin (RMD), an FDA‐approved histone deacetylase inhibitor for treating cancer, demonstrates potent latency‐reversing activity in vitro but has yielded suboptimal outcomes in kick‐and‐kill HIV‐cure trials when administered as a 5 mg/m² infusion. We hypothesized that distinct pharmacokinetic targets exist between the pharmacodynamic goals of cytotoxicity for cancer therapy and latency reversal for HIV cure. To address this question, we designed an empirical modeling framework incorporating patient pharmacokinetics as a New Approach Methodology. Using clonal‐ and primary cell models of HIV‐1 latency with readouts for different stages of latency‐reversing activity, we performed exposure–response and dose‐fractionation studies to define pharmacokinetic drivers of efficacy in latency, then designed tailored dosing strategies to meet these targets in static‐ and dynamic‐culture systems, including a hollow‐fiber model. Across all systems, the only consistent pharmacokinetic parameter driving latency‐reversing activity, was RMD exposure time, with maximal activity achieved in cells exposed to ~4–6 ng/mL RMD for ≥ 20 hours (AUCs of ~80–120 ng*hour/mL). With current clinical‐dosing strategies achieving AUCs of 240–600 ng*hour/mL, the same outcome occurred only when RMD concentrations greatly exceeded a cytotoxicity threshold, suggesting response was a byproduct of cytotoxic processes rather than targeted latency reversal. These findings indicate that conventional oncologic dosing strategies for RMD might not be directly applicable to HIV‐1 latency reversal, offering a mechanistic explanation for suboptimal clinical outcomes. We provide proof‐of‐concept for RMD's limited clinical efficacy and propose a “low and slow” dosing strategy targeting time‐driven, toxicity‐minimized drug exposure, demonstrating how in vitro pharmacokinetic/pharmacodynamic investigations applying clinical pharmacokinetic data may inform dose optimization for clinical trials of latency reversal.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

Romidepsin (RMD), an FDA‐approved histone deacetylase inhibitor for cancer, is a potent HIV latency reversing agent in vitro. However, HIV‐cure trials applying oncology‐derived dosing strategies have produced suboptimal and inconsistent clinical outcomes, and the pharmacokinetic determinants of RMD‐mediated latency reversal are unknown.

• WHAT QUESTION DID THIS STUDY ADDRESS?

This study tested the hypothesis that distinct pharmacokinetic targets exist for the pharmacodynamic goals of RMD‐mediated cytotoxicity in cancer therapy and latency reversal in HIV cure.

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

Using an empirical PK/PD model incorporating clinical pharmacokinetics, exposure–response studies, dose fractionation, and static and dynamic in vitro systems, we identified RMD exposure time as the primary driver of HIV latency reversal. Maximal activity occurred with sustained exposure to ~4–6 ng/mL RMD for ≥ 20 hours (AUC ~80–120 ng*hour/mL). Clinical dosing strategies produced similar latency‐reversing activity only at substantially higher exposures (AUC ~240–600 ng*hour/mL) where concentrations exceeded a cytotoxicity threshold, suggesting that observed responses arise from cytotoxic mechanisms rather than selective latency reversal.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

These proof‐of‐concept findings suggest that conventional oncologic dosing paradigms may not be appropriate for HIV latency reversal and offer a mechanistic explanation for the limited efficacy observed in clinical studies. More broadly, this work demonstrates how clinically informed in vitro PK/PD modeling can be used to identify indication‐specific pharmacokinetic targets and rationally optimize dosing strategies for drug repurposing.

HIV‐1 infection is incurable due to the presence of the latent reservoir, a population of long‐lived resting memory CD4⁺ T cells harboring stably integrated yet transcriptionally silent provirus that is protected from immunological recognition and clearance. Although antiretroviral therapy (ART) effectively suppresses viremia throughout the lifespan of people living with HIV (PLWH), viral recrudescence from the latent reservoir occurs within weeks of treatment interruption. Thus, without a sterilizing cure, lifelong ART is required to control infection. ¹

Efforts to eliminate the latent reservoir have largely focused on the “kick‐and‐kill” strategy, which proposes to reactivate latent virus, enabling targeted clearance by natural or engineered immunological mechanisms. Several classes of drugs have been assessed as latency reversing agents (LRAs), with histone deacetylase (HDAC) inhibitors (HDACis) among the first tested clinically. ² HDACis are a class of small molecules currently being applied in cancer therapeutics that cause the accumulation of acetylated histones, resulting in chromatin relaxation and modulation of gene expression. ³ In the HIV‐cure setting, HDACis were identified as potential LRAs upon discovery that HDAC recruitment and resulting histone deacetylation are a primary mechanism of transcriptional repression in HIV‐1 latency. In 2012, the HDACi, vorinostat, was shown to induce histone acetylation and HIV‐1 RNA expression in resting CD4⁺ T cells in people living with HIV on antiretroviral therapy, which resulted in borrowing FDA‐approved HDACis for oncology for use as LRAs in HIV cure. ² , ⁴ Pharmacokinetic/pharmacodynamic (PK/PD) modeling was integral in understanding how daily oral vorinostat dosing (400 mg) used for its cancer indication results in PD tolerance (i.e. blunting of effect from homeostatic response) for latency reversal, which could be avoided by separating each dose by 72 hours. ⁵ However, the first randomized clinical trial of the kick‐and‐kill strategy (RIVER), which utilized this optimized dosing frequency for vorinostat, ultimately demonstrated no benefit of the intervention beyond ART alone, prompting a shift in the field toward identifying more potent interventions. ⁶

Romidepsin (RMD) is an HDACi with FDA indication for treating cutaneous T‐cell lymphoma using cycles of 4‐hour 14 mg/m² infusions and latency‐reversing potency estimated > 150× greater than vorinostat based on direct comparison in vitro. ³ , ⁷ , ⁸ In two clinical trials (REDUC and BCN02), RMD was applied as an LRA using 3 cycles of a weekly 5 mg/m² infusion based on its favorable safety threshold at ~36% the clinical dose for cancer and its ability to induce latency‐reversing activity in primary CD4⁺ T cells ex vivo. ⁷ , ⁹ , ¹⁰ Despite postdose increases in histone acetylation and cell‐associated HIV‐1 RNA (quantified as unspliced RNA or via conserved regions of the 5′ LTR and gag) in both trials, induction of plasma viremia was limited (low‐level) and variable between participants and doses. ⁹ , ¹⁰ Furthermore, the only statistically significant relationship observed with 5 mg/m² RMD dosing in BCN02 was with increases in expression of exhaustion markers in CD4⁺ and CD8⁺ cells and apoptosis markers in CD4⁺ cells, but not with histone acetylation levels or cell‐associated HIV‐1 RNA quantified via conserved regions of the 5′ LTR and gag. ¹¹ These suboptimal clinical outcomes beg the question of whether the correct dose was selected for this cancer therapeutic to induce the PD outcome of latency reversal.

Apart from safety considerations, justification for the 5 mg/m² RMD dose largely arose from cell culture modeling using peak serum concentration (C _max) as the targeted clinical PK parameter. However, PK targets driving PD effect may differ between inducing latency‐reversing activity for HIV‐1 cure and cytotoxicity for cancer therapy. Defining PK/PD indices, that is linking PK parameters (e.g. C _max, area‐under‐the‐curve [AUC], or exposure time) with intended drug response (PD) through correlative modeling, has been instrumental for antiretroviral dose selection and may similarly inform LRA dosing strategy. ¹² , ¹³ , ¹⁴ Indeed, without knowing the PK/PD index for latency reversal by vorinostat, it is challenging to ascertain whether an effective PK profile was achieved in the RIVER trial.

Here, we conducted in vitro exposure–response and dose‐fractionation studies spanning the clinical RMD exposure range to identify PK determinants of latency‐reversing activity for dose optimization. We then simulated clinical dosing scenarios targeting these PK determinants in both clonal and primary cell models of latency and compared PD outcomes against those achieved with simulation of the 5 mg/m² dose used in the REDUC and BCN02 HIV‐cure trials.

METHODS

Cellular models of latency reversal

All experiments were performed using published cellular model systems where HIV‐1‐encoded reporter expression is driven by the HIV‐1 LTR promoter. Details are available in Supplemental Methods.

Exposure–response studies

RMD exposure–response studies were performed by incubating ~100,000 viable N6 cells/well in 96‐well plates containing 0.78–600 ng/mL RMD for 24 hours, yielding AUC_0‐24hours of 19–14,400 ng*hour/mL spanning below and above clinical AUCs. Two staggered twofold serial dilutions from a single RMD stock were performed to expand concentration coverage while accounting for pipetting variability. Latency‐reversing activity was characterized by relative light units (RLU) and cytotoxicity by relative fluorescence units (RFU), using commercial assays (Supplemental Methods). A total of three independent experiments were performed on different days, using N6 cells from different stock culture passages. In all experiments, each RMD concentration was assessed in triplicate culture wells, with a single RLU or RFU measure obtained from each well, according to the manufacturer's instructions. To account for variability, all independent RLU or RFU measures were included when fitting sigmoid E _max/I _max models.

Dose fractionation

Dose fractionation utilized N6 cells (1–2 × 10⁶ viable cells/mL), with RLU measured at 24 hours (Luciferase Assay System, Promega) and normalized to exposure‐time‐matched untreated controls processed identically to account for cell‐manipulation effects. First, N6 cells were treated with 20, 60, or 120 ng/mL RMD for 24 hours, with 1–2 × 10⁶ cells sampled at intervals to achieve matched AUC_0‐24hours of 120–480 ng*hour/mL within each concentration category (Table S1 ), then washed and incubated in untreated media until endpoint. Next, N6 cells were treated for 2, 8, or 20 hours with 2.4–300 ng/mL RMD for matched AUC_0‐24hours (48–600 ng*hour/mL) within each time category (Table S2 ); ~1–2 × 10⁶ cells per group were collected in duplicate, washed, pelleted (400 g, 10 minutes), and incubated in untreated media until endpoint. A total of five independent dose fractionation experiments were performed on different days, using N6 cells from different stock culture passages. In the redesigned experiments (Table S2 ), two populations of cells were separately sampled from each culture flask treated with a target RMD AUC within each time category, thus median presented data points represent eight independently treated populations of cells.

Clinical‐dose simulations

Clinical dose simulations compared the 4‐hour 5 mg/m² infusion against low‐dose continuous infusions (CIs) administered ≥ 20 hours in N6 and primary cells (Figure S4 ). RMD was removed at 20 hours (N6 cells) or 8 hours (primary cells) to limit centrifugation of primary cells, as RMD exposure from 8 to 20 hours (0.3 ng/mL; AUC_8‐20h 3.6 ng*hour/mL) was not anticipated to meaningfully impact response. All conditions, including controls, underwent identical washing and centrifugation using treated or untreated media where needed. Following RMD removal, cells were incubated in untreated media until 72 hours (N6 cells) or 48 hours (primary cells) to match total duration. N6 cells (~200,000) were sampled at 0, 8, 20, 24, 28, 32, 48, 52, and 72 hours for ONE‐Glo + Tox Assay (Promega), and primary cells (~1 × 10⁶) at 0, 20, 28, and 48 hours for flow cytometry (Supplemental Methods). All sampled cells were pelleted (400× g, 10 minutes) prior to analysis. Two independent experiments were performed in each model system, where one sample was obtained at each timepoint to minimize potential cell loss over time. A single PD measure was obtained from each sample for experimental feasibility.

Hollow fiber studies

To recapitulate clinical dose simulations in a dynamic system, a hollow‐fiber model was employed. The system consisted of a central (external) reservoir connected to a hollow‐fiber bioreactor (FiberCell Systems) containing semipermeable, hydrophilic PVDF fibers that support media flow and nutrient exchange for densely populated cells in the extracapillary space (ECS). A positive displacement pump continuously circulated media between the central reservoir and bioreactor, enabling rapid equilibration. RMD concentrations were modulated by adding RMD‐treated or untreated media to the central reservoir to simulate two infusion scenarios and, when feasible, RMD's triphasic clearance. Intensive PK sampling was obtained from the central reservoir, with limited ECS to confirm equilibration while minimizing cell loss. Roughly, 5–6 × 10⁷ N6 cells were cultured in the bioreactor at densities of 2–2.5 × 10⁶ cells/mL in the ECS representing tissue. The 4‐hour 5 mg/m² clinical infusion was simulated by adding 27,000 ng RMD to the central reservoir, circulating for 4 hours, then iteratively diluting to mimic clearance (Table S3 ). The 24‐hour 5 ng/mL CI was simulated by adding 1,200 ng RMD. System volume was maintained at 240 mL until near‐total RMD clearance at 24 hours postinfusion‐start. PK sampling was obtained from the central reservoir at 0, 1, 2, 3, 4, 4.75, 5.75, 6.75, 7.75, 8, 20, 24, 25.5, and 28 hours, with additional timepoints (30–72 hours) in a second experiment. For PD analysis, 0.2–1 × 10⁶ live N6 cells were sampled from well‐mixed solution in the ECS in a subset of matched timepoints. Cells were pelleted (400× g, 10 minutes) and analyzed by ONE‐Glo + Tox Assay (Promega), while supernatant was used for ECS PK measurements. Two independent experiments were performed, where one sample was obtained at each timepoint to minimize potential cell loss over time. A single RLU or RFU measure was obtained from each sample for experimental feasibility.

Statistics

All plots were generated in SigmaPlot Software (Grafitti LLC). This software was also utilized for statistical analyses with automatic data normality assessment performed within each procedure, including correlative analysis by Pearson correlation and between group comparisons by Mann–Whitney Rank Sum or Kruskal–Wallis One‐Way ANOVA on Ranks. Measures of central tendency and variability are reported as mean with standard deviation (SD) or median with interquartile range (IQR).

Sigmoid E _max/I _max models were fit for RMD concentration ($C$) vs. efficacy or cytotoxicity from the exposure–response studies using the following equations: $\mathrm{RLU}=E_0+\frac{E\max *C^{\gamma }}{{\mathrm{EC}}_{50}^{\gamma }+C^{\gamma }}$ and $\mathrm{RFU}=E_0-\frac{I\max *C^{\gamma }}{{\mathrm{CC}}_{50}^{\gamma }+C^{\gamma }}$ to estimate maximal effect (E _max/I _max), concentration associated with 50% of maximal effect on RLU (EC ₅₀ ), concentration associated with 50% of maximal effect on RFU (CC ₅₀ ), Hill coefficient (γ), and baseline effect (E ₀ ) using the Phoenix® PK/PD Platform software (Certara). This software was also utilized for noncompartmental analysis (NCA) of observed PK/PD data from clinical dose simulations. NCA of RMD concentration vs. time (PK) data from Hollow Fiber studies utilized the Plasma (202) IV Infusion Model with linear trapezoidal‐linear interpolation for calculating AUC_all and partial AUC_24‐72h. NCA of PD data, as raw and untreated‐control‐normalized RLU‐vs.‐time, utilized the Drug Effect (220) Model, which calculates AUC from observed data by linear trapezoidal‐linear interpolation as a net value above baseline and does not extrapolate to infinity. Uniform weighting was utilized in PK/PD NCA.

RESULTS

RMD‐mediated latency‐reversing activity displays a narrow effective concentration range

Exposure–response studies were conducted with N6 cells to minimize cellular heterogeneity and leverage a rapid, scalable readout of latency‐reversing activity (RLU from virally encoded luciferase) arising from a single integrated provirus. Importantly, the integrated provirus of N6 cells is capable of generating fully elongated transcripts, and LRAs induce latency‐reversing activity in N6 cells in a concentration‐dependent manner for transcription, translation, and cell‐surface expression of integrated reporter genes under control of the HIV‐1 LTR promoter. ¹⁵ , ¹⁶ , ¹⁷

The relationship between RMD concentration and latency‐reversing activity in N6 cells was characterized using 0.78–600 ng/mL RMD over 24‐hour incubation (Figure 1 a,b ), yielding RMD exposures (AUC_0‐24hours) of 19–14,400 ng*hour/mL, spanning below and above clinical exposures reported in latency‐reversal and oncology studies of RMD. Latency‐reversing activity (RLU) increased with RMD concentrations between ~2.3 and 6 ng/mL (AUC_0‐24hours 55.2–144 ng*hour/mL) and decreased with ≥ 9 ng/mL (≥ AUC_0‐24hours 216 ng*hour/mL), presumably due to cytotoxicity (Figure 1 a ). Indeed, after 24‐hour treatment with 600 ng/mL RMD, we observed a 24% loss of viability (as estimated by cell count) driven by a sevenfold increase in the number of dead cells while the number of live cells remained largely unchanged (0.95‐fold). In contrast, after 24‐hour treatment with 0.78 ng/mL RMD, we observed a negligible change in viability (< 1% loss), reflecting approximately equivalent increases in both live and dead cells (1.6‐ and 1.8‐fold, respectively), suggesting physiological turnover. An E _max model (Figure 1 b ) was fit for RLU vs. RMD concentrations constrained to 0.78–6.25 ng/mL (AUC_0‐24hours 19–150 ng*hour/mL) to capture latency‐reversing activity without any putative cytotoxicity. The mean (%CV) parameter estimates were EC₅₀ = 3.2 ng/mL (1.4%), γ = 8.3 ng/mL (14.6%), E ₀ = 746 RLU (86.7%), and E _max = 49,035 RLU (2.7%) with an EC₉₀ (90% effective concentration) calculated at 4.2 ng/mL.

Figure 1.

Concentration–activity and –cytotoxicity relationships associated with HIV‐1 latency‐reversing activity by romidepsin (RMD). First, latency‐reversing activity was measured by luminescence from virally encoded luciferase, shown as relative light units (RLU) in N6 cells after 24‐hour exposure to twofold serial dilutions of RMD (n = 2 experiments shown as X and O symbols): (a) Exposure–response relationship for 0.78–600 ng/mL RMD (AUC_0‐24hours = 19–14,400 ng*hour/mL), (b) E _max regression model fit for 0.78–6.25 ng/mL RMD (AUC_0‐24hours = 19‐150 ng*hour/mL) to estimate EC₅₀ (solid lines). Second, two‐step ONE‐Glo + Tox Assay (Promega) was used to measure latency‐reversing activity (as RLU) and, within the same culture well, a fluorescent signal proportional to the number of viable (membrane‐intact) cells shown as relative fluorescence units (RFU) after 24‐hour exposure to twofold serial dilutions of RMD (n = 1 experiment) using the same dilution scheme as above. Exposure–response relationships for 0.78–600 ng/mL RMD (AUC_0‐24hours = 19–14,400 ng*hour/mL) are shown for latency‐reversing activity as RLU (c) and cytotoxicity as reduction in RFU (e) with mean values for untreated controls shown in dashed lines. E _max regression models were fit for 0.78–6.25 ng/mL RMD (AUC_0‐24hours = 19–150 ng*hour/mL) to estimate EC₅₀ (d) and for 0.78–200 ng/mL RMD (AUC_0‐24hours = 19–4,800 ng*hour/mL) to estimate CC₅₀ (f). Maximal observed latency‐reversing activity is highlighted in red for ease of comparison (c–f). In all experiments, each concentration was assessed in triplicate culture wells, with individual data points showing a single RLU measure obtained from each well.

To define the RMD concentration range associated with cytotoxicity, we repeated these experiments (Figure 1 c–f ) using a commercial two‐step assay (Supplemental Methods) to dually measure latency‐reversing activity (RLU) and cell viability (RFU) within the same population of treated N6 cells (i.e. within the same culture well). First, we characterized nonspecific luciferase expression from generalized cytotoxicity of N6 cells in this system and found it negligible (Figure S1 ). Then, we verified RMD's exposure–response profile for latency‐reversing activity (Figure 1 c ) by fitting an E _max model to RLU derived from N6 cells in this two‐step assay as a function of the same RMD concentration range used previously (Figure 1 d ). Like our previous estimates, the mean (%CV) parameter estimates were: EC₅₀ = 3.7 ng/mL (0.9%), γ = 7.6 ng/mL (3.9%), E ₀ = 177 RLU (20.9%), and E _max = 7,626 RLU (1.3%), with an EC₉₀ calculated at 5.0 ng/mL.

In this experiment, N6‐cell viability (RFU) generally decreased over an RMD concentration range of 0.78–6.25 ng/mL (Figure 1 e,f ), while latency‐reversing activity (RLU) increased (Figure 1 c,d ). Both RLU and RFU decreased with RMD concentrations > 6.25 ng/mL, reaching an inhibitory response plateau for both PD readouts with ≤ 200 ng/mL RMD (Figure 1 c,e ). At > 200 ng/mL, additional increasing cytotoxicity (decreasing RFU) in N6 cells was accompanied by further decreases in RLU, approaching untreated control at 600 ng/mL RMD (Figure 1 c,e ). To determine whether reduced latency‐reversing activity at high RMD concentrations reflected loss of viable cells or possible PD tolerance, we assessed whether N6 cells could exhibit fluorescence changes beyond those observed at high RMD concentrations. In a side‐by‐side N6‐cell experiment using nonspecific detergent (Figure S1 ), 24‐hour treatment with 30 μg/mL digitonin reduced RFU by 92% relative to untreated control, compared to a 64% reduction with 600 ng/mL RMD, which, based on earlier experiments applying this concentration, cannot be attributed to a loss of live cells from baseline. An E _max model in the inhibitory form (Figure 1 f ) was fit for RFU vs. RMD concentrations constrained to 0.78–200 ng/mL (AUC_0‐24hours 19–4,800 ng*hour/mL) to capture cytotoxicity before near‐complete abrogation of latency‐reversing activity from putative PD tolerance. The CC₅₀, or the RMD concentration that reduces N6 viability by 50%, was estimated at 5.1 ng/mL (CV%: 2.4%), with mean (%CV) parameter estimates of γ = −4.4 ng/mL (9.9%), E ₀ = 1.4e⁷ RFU (0.8%), and I _max = 6e⁶ RFU (2.0%), and a CC₉₀ (90% viability reduction) calculated at 8.4 ng/mL.

To relate luminescence on a theoretical single‐cell level, we first verified that fluorescence from the CellTiter‐Fluor component of the two‐step assay was proportional to the number of viable (membrane‐intact) cells (Figure S2A ). We then reasoned that because the same number of cells were initially seeded within each culture well, fluorescence could be calculated per viable cell at endpoint as follows. After 24 hours, mean fluorescence for n = 24 untreated control wells from the experiment shown in Figure 1 c–f , was 1.42e7 (SD ± 5.50e5), corresponding to 78 (SD ± 3) RFU per viable cell based on Trypan‐exclusion cell count showing 1.82e5 live cells/well at endpoint. Within the same n = 24 wells, untreated cells exhibited mean luminescence of 85 (SD ± 17), corresponding to 4.66e‐4 (SD ± 9.08e‐5) RLU per viable cell. Thus, 78 RFU was used as a proportionality factor to relate RLU to the number of single viable cells at endpoint with RMD treatment, since, as discussed above, viability loss with the highest RMD concentration assessed (600 ng/mL) was driven by an increase in dead cells rather than a loss of live cells. Applying this principle, latency‐reversing activity within single viable cells (Figure S3 ) exhibited an exposure–response relationship in parallel with that shown in Figure 1 c , where maximal activity per viable cell on average was observed with 6.25 ng/mL (AUC_0‐24hours 150 ng*hour/mL). These findings were derived from cytotoxicity data obtained from the number of N6 cells per well (1e⁵ cells/well) showing the greatest practical level of assay sensitivity for viability discernment following nonspecific detergent treatment of titrated cells (Figure S2B,C ).

RMD exposure time correlates with HIV‐1 latency‐reversing activity

Since maximal latency‐reversing activity in N6 cells was induced by an RMD exposure ~threefold lower than the mean clinical exposure reported for the 5 mg/m² infusion (AUC_0‐∞ 446 ng*hour/mL, 95% CI: 199–693 ng*hour/mL) ¹¹ , we investigated whether the clinical‐dosing strategy applied for latency reversal achieved the PK targets driving activity in our model. To identify these targets, in vitro dose fractionation of clinical and subclinical AUCs was performed in N6 cells by adjusting either RMD exposure time or concentration within a 24‐hour incubation window (Tables S1 and S2 ). Since ~80% of RMD's clinical AUC_0‐∞ occurs within the 4‐hour infusion period, we employed a pragmatic design for this first level of dose fractionation that modeled a 4‐hour constant‐rate infusion followed by instantaneous clearance in N6 cells using an RMD concentration of 120 ng/mL, the median reported C _max observed clinically. To maximize available dose fractionation data points for correlative analysis, we included RMD concentrations at one‐half and one‐sixth the clinical C _max to achieve the same target AUCs but with longer exposure times within 24 hours incubation. Latency‐reversing activity (RLU) was normalized to exposure‐time‐matched untreated controls and evaluated as a function of RMD treatment concentration, exposure time, and AUC_0‐24hours to identify PK parameters driving latency‐reversing activity through Pearson correlation.

First, RMD exposure time was adjusted to achieve target AUCs within categories of RMD treatment concentration (Table S1 ). Significant positive correlations were observed between RLU and RMD exposure time (r = 0.81, P = 0.001) as well as AUC_0‐24hours (r = 0.77, P = 0.003), but not concentration (r = −0.20, P = 0.535; Figure 2 a–c ). However, given the interdependence of exposure time and AUC in this study design, our observed correlation between proviral‐encoded luciferase expression (RLU) and AUC_0‐24hours may be an artifact. To test this, dose fractionation was repeated by adjusting RMD concentration to achieve the same target AUCs within categories of exposure time. A significant positive correlation was observed between RLU and RMD exposure time (r = 0.72, P = 0.009), but not AUC_0‐24hours (r = −0.35, P = 0.259) or concentration (r = −0.54, P = 0.071; Figure 2 d–f ). Based on these results, we hypothesized that latency reversal may be achievable with subclinical exposures if given enough time.

Figure 2.

Dose fractionation of the 4‐hour 5 mg/m² clinical infusion modeled as constant exposure to romidepsin (RMD) followed by instantaneous clearance. Latency‐reversing activity was measured by luminescence from virally encoded luciferase, shown as relative light units (RLU) at 24 hours in N6 cells following RMD treatment at various concentrations and exposure times to achieve matched AUCs of 120, 240, 360, or 480 ng*hour/mL as described in Tables S1 (a–c) and S2 (d–f). First, RMD exposure time was adjusted to achieve target AUCs within categories of RMD treatment concentration (n = 1 experiment; a–c), then RMD concentration was adjusted to achieve the same target AUCs within categories of exposure time (n = 1 experiment; d–f). Luciferase expression was normalized to exposure‐time‐matched untreated controls and evaluated as a function of RMD treatment concentration, exposure time, and AUC_0‐24hours through Pearson correlation as shown. Individual data points show either a single normalized RLU measure (a–c) or mean normalized values from two independent RLU measures obtained from the same cell lysate (d–f).

RMD reverses latency via a time‐driven effect despite subclinical exposures

Dose fractionation was repeated for an expanded target set of clinical and subclinical AUCs (Table S2 ) by adjusting RMD concentrations spanning the EC/CC₅₀ values and exposure–response plateaus from earlier experiments defining the effective concentration range. Median normalized RLU decreased with time for all clinical AUCs (Figure 3 a ) and increased with time for subclinical AUCs when RMD treatment concentration was ≥ EC₉₀ and < CC₉₀ but not when it was < EC ₅₀ (Figure 3 b ). Importantly, the same maximal RLU was achieved with 4 or 6 ng/mL RMD applied over 20 hours, corresponding to subclinical AUC_0‐24hours of 80 or 120 ng*hour/mL, respectively, as with 300 ng/mL RMD applied over 2 hours, corresponding to a clinical AUC_0‐24hours of 600 ng*hour/mL.

Figure 3.

Dose fractionation of an expanded target set of clinical and subclinical AUCs by adjusting romidepsin (RMD) concentrations spanning the EC/CC₅₀ values and exposure–response plateaus from earlier experiments defining the effective concentration range. Latency‐reversing activity was measured by luminescence from virally encoded luciferase, shown as relative light units (RLU) at 24 hours in N6 cells following RMD treatment at various concentrations and exposure times to achieve matched clinical AUCs of 240–600 ng*hour/mL (a) and subclinical AUCs of 48‐120 ng*hour/mL (b) as described in Table S2 . Median (IQR) luciferase expression normalized to exposure‐time‐matched untreated controls for each AUC level (n = 4 experiments) are shown as well as the maximal expression value across any of the median treatment groups for reference (solid line). Each data point represents eight independently treated populations of cells, as two populations of cells were separately sampled from each culture flask treated with a target RMD AUC within each time category. All sampled cells were cultured in blank media to 24 hours, then lysed and assessed for RLU measured in duplicate.

Statistical assessments of normalized RLU were performed separately on data obtained from clinical vs. subclinical AUC targets. Mann–Whitney Rank Sum test identified significant differences in RLU in cells exposed to RMD for 2 vs. 20 hours only when the treatment concentration was ≥ EC₉₀ (P = 0.013 and 0.003, respectively, for clinical vs. subclinical AUCs). RLU was also evaluated as a function of RMD concentration, exposure time, and AUC_0‐24hours through Pearson correlation. For clinical AUCs, a significant positive correlation was observed with RMD concentration (r = 0.369, P = 0.0002) and a significant negative correlation was observed with RMD exposure time (r = −0.289, P = 0.004), whereas no significant correlation was observed with AUC_0‐24hours (r = 0.086, P = 0.408). In contrast, for subclinical AUCs, the only significant correlation observed was with RMD AUC_0‐24hours (r = 0.443, P = 0.000006), but not with concentration (r = 0.011, P = 0.918) or exposure time (r = 0.156, P = 0.130). However, when the treatment concentration was ≥ EC₉₀ a significant positive correlation was revealed between RLU and RMD exposure time (r = 0.387, P = 0.0004), as well as AUC_0‐24hours (r = 0.395, P = 0.0003), but not concentration (r = −0.090, P = 0.429).

Low‐dose continuous infusion of RMD reverses latency with less risk for clinical toxicity

We hypothesized that a ≥ 20‐hour low‐dose continuous infusion (CI) of RMD would reverse latency comparable to the 4‐hour 5 mg/m² infusion. To test this, N6 cells were utilized for clinical dose simulations performed over 72 hours incubation. The median clinical 5 mg/m² infusion PK profile was simulated as 120 ng/mL applied for 4 hours followed by stepwise dilutions with untreated media to mimic RMD's triphasic clearance for an AUC_0‐72h of 514 ng*hour/mL (Figure S4 ). Low‐dose CIs were simulated as 5 ng/mL applied for 20 hours (AUC_0‐72h 100 ng*hour/mL) and of 3 ng/mL for 72 hours (AUC_0‐72h 216 ng*hour/mL). Roughly 200,000 N6 cells were sampled over time to measure latency‐reversing activity (RLU, normalized to exposure‐time‐matched untreated controls) and cytotoxicity (RFU). This sampling target accounted for potential N6‐cell loss from frequent handling and prolonged RMD exposure while remaining within assay limits for background (media only) or saturated fluorescence (Figure S2B,C ).

The 5 mg/m² simulation induced rapid latency‐reversing activity, with a ninefold RLU increase by 8 hours (first timepoint), peaking at 20 hours before declining 1.4‐fold at 24 hours (Figure 4 a ). Both low‐dose CIs exhibited delayed induction, with no appreciable change at 8 hours; peak expression occurred at 20 hours (20‐hour 5 ng/mL CI) and 32 hours (72‐hour 3 ng/mL CI), followed by gradual declines as of the next timepoint sampled (< 1.2‐fold). Despite these kinetic differences, overall RLU (mean AUC_net,0‐72h, units RLU*h) did not differ significantly between the 5 mg/m² infusion, 20‐hour CI, and 72‐hour CI for either normalized (523, 404, and 350; P = 0.07) or raw RLU (304,500, 271,331, and 249,248; Kruskal–Wallis One‐Way ANOVA on Ranks P = 0.67). Viability changes were also similar, with 64–95% RFU reductions (0.5–1.3 log₁₀ decrease) across conditions including untreated (Figure 4 b ), though the 72‐hour CI exhibited the greatest RFU loss (2.6‐fold greater vs. 5 mg/m² and 20‐hour CI). No significant differences in percent viability change from baseline were observed between the 5 mg/m² infusion, 20‐hour CI, 72‐hour CI, and untreated control (17.7%, 18.6%, 6.9%, and 32.7%; Kruskal–Wallis One‐Way ANOVA on Ranks P = 0.07). Since RFU reflects viable (membrane‐intact) cells, automated cell counts were measured alongside fluorescence in one experimental replicate to verify loss of live cells/mL (Figure S5A,B ). Cell counts confirmed 38–65% (0.2–0.4 log₁₀) loss of live cells/mL at endpoint across all groups, including untreated control (Figure S5B ), and percent viability loss by Trypan exclusion (Figure S5C ) driven by increased dead cells/mL, particularly during frequent handling (Figure S5D ). Interestingly, the 5 mg/m² infusion exhibited the greatest loss of live cells/mL (Figure S5B ) despite similar RFU reductions to the low‐dose CIs, suggesting RFU in the 5 mg/m² infusion arose from fewer viable cells. Although sampling numbers were consistent (mean 1.88e5 total cells, SD ± 2.45e4), centrifugation yielded visibly smaller cell pellets for the 5 mg/m² infusion, likely due to RMD‐induced cellular fragility at concentrations well‐exceeding the N6‐cell CC₉₀.

Figure 4.

Clinical dose simulations performed in N6 and primary cell models of HIV‐1 latency. The median clinical plasma romidepsin concentration vs. time profile observed for the 5 mg/m² infusion was modeled and compared against low‐dose continuous infusions (CIs) administered over ≥ 20 hours as described in Figure S4 . First, simulations were performed in N6 cells sampled over a 72‐hour incubation window for two‐step ONE‐Glo + Tox Assay (Promega) measurement of latency‐reversing activity (as relative light unit, RLU) and, within the same sample, a fluorescent signal proportional to the number of viable (membrane‐intact) cells shown as relative fluorescence units (RFU). Pharmacodynamic readouts are presented as mean (SD) for RLU from virally encoded luciferase normalized to exposure‐time‐matched untreated controls (a) and cell viability loss shown as reduction in RFU from baseline (b) over time (n = 2 experiments). Second, simulations were performed in primary cells sampled over a 48‐hour incubation window for flow cytometric analysis of latency‐reversing activity (virally encoded dreGFP median fluorescence intensity [MFI], c) and viability loss (reduction from baseline, d) over time (n = 2 experiments) shown as mean (SD). Cellular population shifts in dreGFP MFI at endpoint and over time are shown in (e). Representative gating strategy for dreGFP MFI measurement within the live, infected Thy1.2⁺ population based on an inverse gate (red) drawn around representative uninfected donor cells (gray) as described in the Supplemental Methods is shown in (f). One sample per timepoint was obtained in each experiment to minimize potential cell loss over time. A single pharmacodynamic measure was obtained from each sample for experimental feasibility.

Low‐dose continuous infusion of RMD reverses latency in a primary cell model

Clinical dose simulations were verified in a more heterogenous primary CD4⁺ T‐cell model of latency described previously. ¹⁸ , ¹⁹ , ²⁰ We selected this primary cell model of HIV‐1 latency for its heterogenous integration landscape, its use of a protein‐based reporter for latency‐reversing activity (dreGFP) under control of the HIV‐1 LTR promoter, and its inclusion of a second virally‐encoded reporter (murine Thy1.2) to delineate actively (Thy1.2 + GFP+) vs. latently (Thy1.2 + GFP‐) infected cells within the same overall population. In this model system, full‐length proviruses generate viral RNAs spanning unspliced to fully spliced forms, ¹⁹ and latently infected cells vary in chromatin accessibility and responsiveness to latency reversal. ²⁰

Since low‐dose CI simulations produced similar latency‐reversing activity in N6 cells, with slightly greater cytotoxicity in the 72‐hour 3 ng/mL CI (AUC_0‐72h 216 ng*hour/mL), we selected a 48‐hour 3 ng/mL CI (AUC_0‐48h 144 ng*hour/mL) for comparison with the 4‐hour 5 mg/m² infusion in primary cells. Roughly 1 × 10⁶ cells were sampled over 48 hours for flow cytometric analysis of dreGFP. At baseline, 97% of live CD4⁺ T cells were infected (Thy1.2⁺), with predominantly latent phenotype, as dreGFP median fluorescence intensity [MFI] approximated background from uninfected cells (200 vs. 126, respectively). Both dose simulations induced population shifts toward the activated phenotype (Figure 4 c,e ; gating strategy in Figure 4 f ) with a 2.5 and 2.2‐fold increase in dreGFP MFI, respectively, for the 5 mg/m² infusion and 48‐hour CI at endpoint after normalizing to untreated control (Mann–Whitney Rank Sum P = 0.33). Viability changes were also comparable at endpoint (78.7%, 83.8%, and 100.1%, for 5 mg/m², 48‐hour CI, and control; Kruskal–Wallis One‐Way ANOVA on Ranks P = 0.20; Figure 4 d ). However, again, 5 mg/m² infusion samples showed marked cell loss during processing, yielding visibly smaller cell pellets, despite similar initial cell numbers sampled. Thus, the 5 mg/m² infusion required near complete sample injection to acquire enough events to reach our target number in the live‐cell gate, with live cells comprising 38%, 48%, and 72% of total events for the 5 mg/m² infusion, 48‐hour CI, and untreated control, respectively.

RMD clinical‐dose simulations are reproducible in a hollow fiber tissue model

The hollow‐fiber model ¹² , ¹³ , ¹⁴ , ²¹ , ²² offers key advantages over traditional cell culture, including media flow and nutrient exchange mimicking blood flow within the capillary bed of a tissue, as well as the ability to recapitulate in vivo PK. Since tissues are a primary source of latently infected cells, ²³ , ²⁴ , ²⁵ we applied the hollow‐fiber model to directly compare clinical dose simulations in a more physiologically relevant model using N6 cells.

Our dosing scenarios targeted: (1) a clinical AUC of 450 ng*hour/mL delivered via a 4‐hour 113 ng/mL “short‐term” infusion followed by triphasic clearance and (2) a subclinical AUC of 120 ng*hour/mL delivered via a 24‐hour 5 ng/mL CI followed by instantaneous clearance. The latter was designed to allow ~66% experimental variability in RMD concentration, while remaining within the optimal subclinical AUC window for latency reversal observed in dose fractionation (AUC_0‐24hours > 80 and < 200 ng*hour/mL). An untreated control was excluded for simplicity, as the positive displacement pump supported side‐by‐side comparison of two dose simulations. In our experience using this system, no appreciable background latency‐reversing activity is observed in the absence of drug (unpublished data).

RMD concentrations in the central reservoir supplying media and the ECS containing N6 cells showed adequate equilibration, and both dosing scenarios achieved target exposures; however, the 5 mg/m² infusion did not adequately recapitulate clinical PK (Figure 5 a ), individual experiments in (Figure S6A ). During continuous infusion without dilution, RMD concentrations declined by ~31% beginning 1 hour after each bolus, suggesting nonspecific binding within the system, which utilized cartridges containing chemically treated polyvinylidene fluoride fibers that could be a source of nonspecific binding. ²⁶ After theoretical total RMD clearance (~24 hours), a second peak was observed, consistent with release of RMD from a sequestered pool. These findings were reproduced in a repeat experiment with a 3‐day observation window, which showed stable RMD concentrations for another ~46 hours when no dilution media was applied, indicating equilibrium between binding and release. The mean observed AUC_0‐72h for the 5 mg/m² infusion was 767 ng*hour/mL, with a partial AUC_24–72hour of 275 ng*hour/mL during the extended observation window when RMD concentrations had been theoretically cleared. This partial AUC was similar to an AUC_24‐72hour of 168 ng*hour/mL predicted for the 5 mg/m² infusion in the first experiment had the observation window included PK sampling out to 3 days. Finally, the 24‐hour 5 ng/mL CI yielded a mean observed AUC_0‐72h of 110 ng*hour/mL, which was within target despite extended exposure, enabling comparison of latency reversal between our defined optimal subclinical AUC window (80–120 ng*hour/mL) and higher exposures approximating clinically observed AUCs from the 5 mg/m² infusion.

Figure 5.

Clinical dose simulations performed in a hollow fiber tissue model. The median clinical plasma romidepsin (RMD) concentration vs. time profile observed for the 5 mg/m² infusion was simulated and compared against a 24‐hour 5 ng/mL continuous infusion (CI) followed by instantaneous clearance as described in Table S3 . Pharmacokinetic/pharmacodynamic assessment was performed over 28‐hour then 72‐hour observation windows (n = 1 experiment each), with data presented as mean (SD) for matched timepoints (closed circles, solid lines) or individual observations within the extended observation window (open circles, short‐dashed lines). RMD concentrations quantified in the central reservoir over time for the simulated 5 mg/m² infusion (red) and 24‐hour 5 ng/mL CI (blue) are shown in (a), with the median clinically observed plasma RMD concentration‐vs.‐time profile (black solid line) and upper/lower bounds (black dashed lines) from the BCN02 trial (Molto et al.) shown as an overlay. Pharmacodynamic readouts were obtained by two‐step ONE‐Glo + Tox Assay (Promega) measurement of latency‐reversing activity (as RLU) and, within the same sample, a fluorescent signal proportional to the number of viable (membrane‐intact) cells shown as relative fluorescence units (RFU). Pharmacodynamic results are presented as virally encoded luciferase expression (b) and cell viability loss shown as reduction in RFU (c) over time in N6 cells densely seeded in the extracapillary space, with results normalized as change from baseline for the simulated 5 mg/m² infusion (blue) and 24‐hour 5 ng/mL CI (red). One sample per timepoint was obtained in each experiment to minimize potential cell loss over time. A single pharmacodynamic measure was obtained from each sample for experimental feasibility.

Our hollow‐fiber model recapitulated PD trends observed in static‐culture systems (Figure 5 b,c , individual experiments in Figure S6B,C ). The 5 mg/m² infusion and 24‐hour CI induced similar latency‐reversing activity (mean normalized AUC_{net,0‐72hour} 2043 vs. 2,321 RLU*hour, respectively), with earlier induction in the 5 mg/m² infusion. However, after 72 hours, viability loss was 2.7‐fold greater for the 5 mg/m² infusion vs. 24‐hour CI (85% vs. 50% RFU reduction; 0.8 vs. 0.3 log₁₀ decrease), contrasting the static‐culture system, where viability loss was comparable across dose simulations.

DISCUSSION

Here we show that RMD‐mediated latency reversal can be achieved with a low‐dose continuous infusion in a time‐driven effect where overall exposure is ~4‐6‐fold lower than mean clinical exposures found in HIV‐cure trials applying 4‐hour 5 mg/m² infusion (geometric mean AUC_0‐∞ 446 ng*hour/mL, 95% CI: 199–693 ng*hour/mL). ³ , ¹¹ Using EC‐/CC‐based constraints, dose‐fractionation studies demonstrated that RMD exposure time‐not concentration consistently drove latency‐reversing activity. For clinically equivalent AUCs (240–600 ng*hour/mL), maximal activity required RMD concentrations greatly exceeding the CC₉₀ (8.4 ng/mL, Figure 1 ), suggesting response is a byproduct of cytotoxic processes. This maximal activity could be matched by subclinical AUCs (80–120 ng*hour/mL) when RMD concentrations remained below the CC₉₀ but were sustained for 20 hours (Figure 3 ). These data provide the first evidence, to our knowledge, that selectively targeting PK determinants of RMD‐mediated latency reversal can decouple efficacy from cytotoxicity, enabling this cancer therapeutic to be repurposed for epigenetic reprogramming in HIV cure. We therefore designed a “low and slow” dosing strategy to match, if not exceed, latency reversal observed with the clinical 4‐hour 5 mg/m² infusion, while reducing exposure. Across all model systems and regardless of latency heterogeneity, low‐dose continuous infusions produced comparable latency reversal to the clinical dose (Figures 4 and 5 ). These findings suggest that maintaining human plasma concentrations equivalent to the ~3.5–8 ng/mL effective concentration range (AUC_0‐24hours 84–192 ng*hour/mL) we identified in vitro may reduce overall risk of clinical toxicity compared to the “high and fast” intravenous infusion dosing paradigm, while preserving efficacy.

Clinical 5 mg/m² dosing maintains plasma concentrations within the narrow estimated (protein‐adjusted) therapeutic window we identified (~7.5–17 ng/mL total RMD) for roughly 1 hour during clearance (Figure S7 ), implying a concentration‐driven, cytotoxic mechanism driving PD, which may not be optimal for kick‐and‐kill strategies requiring HIV‐1 antigen expression for targeted immune clearance of infected cells. Indeed, in the BCN02 trial, the only statistically significant PK/PD relationships identified with 4‐hour 5 mg/m² RMD infusion were between AUC_0‐∞ and increased T‐cell expression of exhaustion and apoptosis markers, and CD4⁺ T‐cell counts decreased proportionally with increasing exposure. ¹¹ Furthermore, a delayed toxic effect on T cells was observed 3 days after RMD dosing, with cumulative toxicity observed after three doses. ²⁷ Similarly, RMD induces apoptosis in vitro in PBMCs from healthy donors and CD4⁺ T cells from PLWH, with effects most pronounced at later timepoints and minimized with concentrations below our identified CC₉₀ (8.4 ng/mL). ²⁸ , ²⁹ Our clinical dose simulations showed peak viability loss beginning ~day 2 posttreatment; however, since all cells were HIV‐1‐infected in our models, some loss may reflect naturally occurring cytotoxicity with viral reactivation rather than RMD. This distinction may be clarified by comparing uninfected and latently infected cells from the same donor.

We also observed increased cellular fragility following 5 mg/m² infusions, leading to cell loss during centrifugation‐based processing and potential overestimation of latency‐reversing breadth, as a smaller proportion of cells remained. This effect was most evident after flow cytometry processing, which requires multiple centrifugation steps. The 5 mg/m² infusion exhibited the greatest cell loss and a slightly greater reduction in viability (Figure 4 d ), yet also modestly increased latency‐reversing activity per remaining viable cell (Figure 4 c ). In contrast, assays without cell loss from centrifugation showed reduced latency‐reversing activity per viable cell at higher (≥ 9 ng/mL) RMD concentrations (Figure S3 ) alongside increased cytotoxicity (Figure 1 e ). These findings demonstrate that cytotoxicity can inflate apparent PD responses, as the effects of treatment‐induced cytotoxicity on latency reversal are not well represented. When cell numbers were held constant, as in our exposure–response studies, maximal latency‐reversing activity per viable cell occurred with 6.25 ng/mL RMD, which, as an average measure, also suggests maximal breadth of response.

Mechanistically, we hypothesize that prolonged low‐dose RMD exposure may provide enough time for latently infected cells, which are rare, widely distributed, and found within tissues, ²³ , ²⁴ , ²⁵ to encounter RMD molecules and leverage its relatively high affinity for and slow off‐rate from HDAC binding. ³⁰ , ³¹ Whereas, a “high and fast” RMD administration (5 mg/m² infusion) may saturate individual cells, promoting cytotoxicity and limiting effective distribution, where latency reversal is largely a byproduct. Interestingly, high RMD concentrations (600 ng/mL) in our exposure–response studies showed negligible latency‐reversing activity despite the presence of viable N6 cells for activation, suggesting cellular saturation may also drive PD tolerance for latency reversal as observed with vorinostat. ⁵

Data supporting this hypothesis includes an in vitro comparison of 4‐ vs. 24‐hour RMD exposure in latently infected 2D10 cells across a 2‐log concentration range, using GFP as a readout of latency reversal. The EC₅₀ value was higher with 4‐hour than 24‐hour exposure (8.4 vs. 1.7 ng/mL), indicating greater apparent potency with prolonged low‐dose exposure. ⁸ Maximal effects occurred at subclinical AUCs (~100 ng*hour/mL), consistent with our findings. However, short‐ vs. long‐term RMD exposures exhibited similar exposure–response relationships, ⁸ whereas our dose‐fractionation results showed decreasing response with increasing exposure time for AUCs ≥ 240 ng*hour/mL (Figure 3 a ). This discrepancy may reflect differences in readout, as Newhard et al. measured latency reversal live cells only, whereas our study included total cells, where cytotoxicity likely contributes to response. Supporting this, our data show declines in both viability and latency‐reversing activity with RMD concentrations ≥ 9 ng/mL (AUC_0‐24hours ≥ 216 ng*hour/mL; Figure 1 c,e ), suggesting exclusion of nonviable cells may mask cytotoxic effects counteracting latency reversal at high RMD exposures. Notably, clinical RMD exposures in HIV cure (reported above) and oncology settings (geometric mean AUC_0‐∞ 1,549 ng*hour/mL, 95% CI not reported) fall within this cytotoxicity‐associated range (Figure 1 c,e ). ³ , ¹¹

Given RMD's high HDAC affinity and slow dissociation, high RMD concentrations during the 4‐hour 5 mg/m² infusion may promote its sequestration in a subset of cells, driving localized overexposure and cell death (or tolerance), while limiting broader target engagement. This may at least partially explain both suboptimal latency reversal and cumulative T‐cell cytotoxicity observed in BCN02. ²⁷ Indeed, a large‐scale assessment of 19 cell lines treated with RMD or vorinostat identified two HDACi effects: a “chromatin effect” (histone acetylation, gene expression, cell cycle arrest) and a “cell death effect.” The chromatin effect is not inherently cytotoxic but can lead to apoptosis with prolonged HDACi exposure; under optimal conditions, it instead promotes greater gene induction, ³⁰ as shown with 48–72 hours exposure to 1 ng/mL RMD in the cancer setting, consistent with a “low and slow” administration that has been considered clinically. ³² , ³³ In contrast, the cell death effect involves rapid apoptosis only in cells that are in a distinct state of susceptibility, irrespective of acetylation magnitude. This apoptosis‐primed state was elucidated through 6‐hour 25 ng/mL RMD treatment designed to model clinical infusion, resulting in an AUC (150 ng*hour/mL) greater than the highest benchmark we identified for optimal latency reversal with subclinical AUCs (120 ng*hour/mL; Figure 3 ). ³⁰ Finally, HIV‐1 latency is maintained by dynamic repressive mechanisms; thus, “low and slow” RMD administration may increase the probability of “catching” latently infected cells as they cycle out of “deep latency” into a more inducible state, promoting broader latency reversal over time. ¹⁸

We therefore propose that sustained low RMD concentrations approaching the EC₉₀ (~4.6 ng/mL) for ≥ 20 hours may optimize latency reversal within a target AUC of ~84–192 ng*hour/mL. Implementation would require formulation development due to RMD's large size (540.71 g/mol), poor bioavailability (1.6% at 5 mg/kg orally), and P‐glycoprotein efflux. ³ , ³⁴ However, RMD is highly potent for HDAC1/2, which are key regulators of HIV‐1 latency, ³⁵ , ³⁶ with IC₅₀ values of 4 and 0.5 ng/mL, respectively, and comparatively weaker activity against HDAC3‐9 (IC₅₀ ranging 17–5,400 ng/mL) ³⁷ suggesting that enhanced specificity for latency reversal may be engineered through low‐dose administration. However, caution is warranted for potential off‐target effects with HDAC10/11 inhibition (IC₅₀ 6 and 0.2 ng/mL, respectively). ³⁷ A long‐acting delivery device capable of loading drugs with poor aqueous solubility, such as a hydrogel‐forming microarray patch or a refillable nanofluidic implant, could enable sustained low‐level exposure, ³⁸ , ³⁹ Refillable functionality would be advantageous, given RMD's cyclic administration and room temperature stability as a diluted solution (≥ 24 hours). ⁴⁰

HIV‐1 cure trials to date have largely failed to achieve clinical outcomes, potentially due to suboptimal latency reversal, as reflected by inconsistent induction of plasma viremia despite increases in cell‐associated HIV‐1 RNA. ⁹ , ¹⁰ , ⁴¹ , ⁴² This finding suggests pro‐transcriptional effects of LRAs may be mitigated by posttranscriptional blocks, an area of ongoing study. ¹⁹ , ⁴³ , ⁴⁴ , ⁴⁵ Importantly, in addition to binding HDAC enzymes to exert its action, RMD also strongly binds human alpha‐1‐acid‐glycoprotein (AAG) and to a lesser extent human albumin. ³ Thus, only a small fraction (~7%) of RMD circulates in human plasma as the unbound form capable of exerting pharmacological activity. One limitation of our study is that, like many cell culture systems, we included 10% FBS supplementation within our culture media, which may not recapitulate protein‐binding in human plasma. While RMD binding to fetal bovine proteins has not been empirically assessed, drugs that primarily bind human AAG, like RMD, exhibit ~72–96% protein binding within a 10% FBS culture system, thus small amounts of protein in culture media can bind large amounts of drug. ⁴⁶ , ⁴⁷ , ⁴⁸ Applying RMD's protein‐binding characteristics reported by the manufacturer ³ to an estimation approach described by Kolli et al., ⁴⁹ we anticipate the unbound fraction of RMD in our culture system could be as high as ~15%. Given this putative correction factor, our proposed EC₉₀ of ~4.6 ng/mL in vitro could translate to a protein‐adjusted EC₉₀ of 10 ng/mL in human plasma, a threshold which all BCN02 participants fell below within 4 hours after ending the 5 mg/m² infusion (Figure S7 ). ¹¹ Our study is also limited by reliance on virally encoded reporter‐based models generated from clonal virus, which do not capture in vivo proviral heterogeneity or contribution from other permissive cell types (e.g. macrophages), although our primary cell model incorporates heterogeneous T‐cell integration sites. While reporter expression reflects HIV‐1 LTR‐driven transcription and translation, it may not fully represent downstream processes like antigen‐presentation. Future work will compare PD outcomes with “high and fast” vs. “low and slow” RMD dosing, particularly antigen production for targeted immune clearance, across diverse LRAs and in primary cells from PLWH on suppressive ART. Despite these limitations, our rigorous PK/PD characterization provides a foundation for iteratively incorporating in vivo complexity to develop predictive NAMs, including physiologically relevant in vitro systems (e.g. organ‐on‐a‐chip) and computational modeling. ⁵⁰ Here, using a proof‐of‐concept PK/PD model novel to HIV cure, we propose a pharmacological basis for the limited success of current dosing strategies and a mechanistically informed alternative to improve latency‐reversing efficacy.

FUNDING

This work was supported by funding from the National Institute of Allergy and Infectious Diseases: R21 AI157853 supporting M.L.C. and L.A.R.T. and the UNC Center for AIDS Research (UNC CFAR; P30 AI050410) supporting M.L.C., L.A.R.T., A.P., C.G., C.S., and A.D.M.K.

CONFLICTS OF INTEREST

The authors declared no competing interests for this work.

AUTHOR CONTRIBUTIONS

All authors wrote the manuscript. L.A.R.T. and M.L.C. designed the research. L.A.R.T., A.P., A.W., C.G., and C.S. performed the research. L.A.R.T., A.P., and A.W. analyzed the data. A.W. provided primary cells and reagents.

Supporting information

Supplemental Figures and Tables

Supplemental Methods

Supporting Data

DATA AVAILABILITY STATEMENT

Data are available in the Supporting Data values file.