PrEP trials for TDF/FTC-based oral PrEP in MSM

We analyzed publicly available data from iPrEX, HPTN 083, IPERGAY, DISCOVER and PURPOSE 2^{6,15,21,27,28}. All primary endpoints for these trials have been published and ethical approvals and written informed consent were obtained. In brief, iPrEX tested the efficacy of daily oral TDF/FTC-based PrEP vs. a placebo of which 1217 MSM were in the TDF/FTC PrEP intervention arm (1628 person-years of observation time). A total of 64 vs. 34 new infections were recorded in the placebo and intervention arm. TFV plasma concentrations was measured in 43 random samples to assess overall adherence in the intervention arm, and was measured in 34 individuals who became infected in the intervention arm. TFV was detectable in 51% of the the random samples, but only in 9% of individuals who became infected.

HPNT 083 assessed long-acting PrEP with injectable carbotegravir vs. oral TDF/FTC-based daily PrEP (control arm) in 2247 MSM and transgender women (TGW)²⁷ amounting to 3123 person-years of observation time. Adherence in the TDF/FTC control group was assessed both in terms of plasma TFV (acute adherence) and TFV-DP in dried blood spots (long-term adherence marker) based on a total of 390 samples. Plasma TFV was detectable in 86% of the random samples and 6/39 (15%) infected individuals had evidence of recent drug intake in the TDF/FTC arm of the study (see also Supplementary Table 2.).

IPERGAY tested on-demand PrEP in 199 MSM vs. a placebo in 201 MSM, contributing to 221 vs. 223 person-years of observation time. 14 vs. 2 infections were observed in the placebo vs. intervention arm. Adherence was measured using plasma TFV and participants reported using a median of 15 tablets per month¹⁵. A total of 86% of individuals had detectable TFV in the intervention arm and 0% of those infected in the intervention arm had detectable plasma TFV.

DISCOVER assessed TAF/FTC-based daily oral PrEP with vs. oral TDF/FTC-based daily oral PrEP (control arm) in 5387 MSM amounting to 8756 person years observation time²⁸. Adherence in the TDF/FTC control group was assessed both in terms of plasma TFV (acute adherence) and TFV-DP in dried blood spots (long-term adherence marker) in a randomly pre-selected subset of 536 participants. In the TDF/FTC arm of the study, TFV-DP was detectable in all sampled individuals, including those who became infected. Notably, 96% of the participants had TFV-DP levels consistent with ≥2 doses per week. Among the infected individuals, only one (9%) showed levels consistent with recent drug intake (≥4 doses per week), while the remaining had levels indicating 6, of which 920 individuals were included in the TDF/FTC-based arm, amounting to 828 person-years of observation time in the FTC/TDF control arm. A total of 8 new infections were observed in the FTC/TDF control arm. Adherence was assessed through TFV-DP concentrations in red blood cells (RBC). According to the TFV-DP levels detected in RBCs, a total of 20% individuals took one dose per week or less, which corresponds to the adherence strata used for plasma TFV²⁴ (Supplementary Fig. S1 therein). Of the 8 infected individuals, 2 individuals had quantifiable TFV-DP levels.

Incidence rate estimation. HIV incidence is typically calculated as \({r}_{{{\rm{Inf}}}}=\frac{{{\rm{number}}}\,{{\rm{of}}}\,{{\rm{infected}}}\,{{\rm{individuals}}}}{{{\rm{observation}}}\,{{\rm{time}}}}\). However, because the observed number of infections tends to be low in a given study, incidence estimates calculated directly from clinical data are statistically uncertain. We utilized a previously developed method²⁴ to capture this uncertainty. In brief, we assumed the number of observed infections (\({N}_{{{\rm{Inf}}}}\)) during a clinical trial to be binomially distributed and then inverse sampled incidence parameters using the simulation_utils.py module from https://github.com/KleistLab/PrEP_Truvada/. To sample incidence rates from the respective clinical trials, we utilized the number of individuals in the sub-group of individuals with no plasma TFV detectable, the observation time and the number of observed infections from each trial, compare Fig. 1e.

Estimation of prophylactic efficacy from clinical data

We adopted a Bayesian approach to estimate prophylactic efficacy φ ∈ [0, 1] (HIV risk reduction) in the subgroup of individuals who took (some) drug. The posterior probability of PrEP efficacy φ after observing \({{{\rm{Inf}}}}_{{{\rm{obs}}}}\) infections is defined as:

In the following, we assume a uniform prior regarding PrEP efficacy, \(P(\varphi )={{\mathcal{U}}}(0,1)\). We model the likelihood of observing a particular number of infections assuming a binomial distribution:

where N_tot denotes the number of individuals in the drug detected subgroup, \({r}_{\inf }(\varphi )={r}_{\inf }(\varnothing )\cdot (1-\varphi )\) is the infection rate for PrEP efficacy φ and \({r}_{\inf }(\varnothing )\) is the baseline infection rate in the placebo group (or the subgroup of the intervention arm where the drug was undetectable). The marginal likelihood in the denominator ensures proper normalization:

We compared this approach to a previously developed approach²⁴ (online Methods section therein) that utilizes the number of observed infections in the entire intervention arm (with- and without detectable drug). This method confirmed all findings presented in this manuscript.

Information gain. To quantify, whether a clinical trial was informative for quantifying average PrEP efficacy, we computed information gain as

where \({H}_{\max }={\log }_{2}(100)\) denotes the Shannon Entropy of a discrete uniform distribution on the interval [0, 100] (=our prior regarding the prophylactic efficacy φ in %) and \(H(\varphi )={\int }_{0}^{100}P(\varphi ){\log }_{2}(P(\varphi ))\,d\varphi\) denotes the Shannon Entropy of P(φ) derived from the clinical data. I.e., in case of \({\varphi }_{i} \sim {{\mathcal{U}}}(0,100),\,H(\varphi )={H}_{\max }\) and there was no information gain. In contrast, if P(φ_i) was a point distribution, H(φ) = 0 and therefore \(I(\varphi )={H}_{\max }\).

Pharmacokinetics of Oral TDF/FTC

The pharmacokinetic (PK) model was adopted from the models by Burns et al. (2015)³³ and Garrett et al. (2018)³⁴. In brief, FTC and TDF are dosed into respective dosing compartments D and absorbed into the blood plasma with rate parameter k_a. During this process TDF is converted into its circulating form tenofovir (TFV) by first-pass metabolism. The circulation of TFV and FTC in blood plasma- (A₁) and peripheral (A₂) compartments are described by first-order reactions with rate parameters k₁₂ and k₂₁ depending on the direction of flux. TFV and FTC can be taken up from the plasma, and be intracellularly phosphorylated into their active moieties in PBMC (compartment A₃) with rate functions f₁₃ and f₃₁ depending on the direction of flux. For FTC, nonlinear uptake and conversion kinetics with \({f}_{13}=\frac{{V}_{max}\cdot {A}_{1}}{{K}_{m}+A1}\) and linear efflux kinetics were considered with f₃₁ = k₃₁ ⋅ A₃. For TFV, linear uptake kinetics were considered f₁₃ = k₁₃ ⋅ A₁, as well as intracellular elimination f₃₀ = k₃₁ ⋅ A₃.

Calculations were performed with molecular units to avoid conversion factors that account for the distinct masses of the drug metabolites.

To reflect pharmacokinetic variability, we utilized parameters from 1000 virtual individuals as previously described²⁴ (Supplementary Data sets therein) and simulated pharmacokinetic trajectories for the four compartments and any given adherence pattern over a time span of nine month.

Gender differences with regards to PK. Previous studies indicated that no gender differences existed with regards to PK parameters^62,63,64. For verification, we developed two population pharmacokinetic (Pop-PK) models to evaluate sex effects on TDF and FTC PK using data from the DOT-DBS study⁴³ and the HPTN 066 study⁶³. The Pop-PK analysis was performed using NONMEM version 7.5 (ICON Development Solutions, MD, USA) with Perl speaks NONMEM (PsN version 4.9.0) and Pirana as the interfaces. Data visualization, NONMEM dataset preparation, and NONMEM post-run diagnostics, and figure preparation were conducted using R version 4.3.2 (R Development Core Team; http://www.r-project.org/) and R studio (2024.09.1+394). All TFV and FTC plasma concentrations were converted to molar concentrations with its molecular weight (TFV: 287.2 g/mol, FTC: 247.25 g/mol). TFV-dp and FTC-tp concentrations in PBMC were converted from fmol/million cells to umol/L based on the volume of a single PBMC (282 fL/cell). Handling of below-the-limit-of-quantification (BLQ) data varied by analyte. For plasma TFV and FTC concentration, BLQ data were handled using the Beal M3 method. For TFV-dp and FTC in PBMC samples, where each sample had a unique lower limit of quantification (LLOQ), the Beal M1 method was applied to handle BLQ data. We first modeled TFV and FTC in plasma, followed by incorporation of a PBMC compartment into the base model. Model performance was assessed using diagnostic plots, including observed versus population and individual predicted concentrations and conditional weighted residuals versus predicted concentrations or time. The objective function value (OFV) was used for model comparison. Internal validation was performed using visual predictive checks (VPCs). Overall, this analysis concluded with no sex differences detectable as shown in Supplementary Table 1.

Pharmacokinetics in exposed tissue. In a previous study²⁴, we used all available data (16 individual studies) that measured TFV-DP and FTC-TP concentrations in local tissue, as well as in PBMC. By modeling the dosing schedules from each of these studies, we could deduce a ratio of drug levels between local tissue and PBMC that appropriately captured pharmacokinetics in local tissue. For TFV-DP we deduced that concentrations in colorectal tissue homogenates were 2.92 times greater than the concentration in PBMC, whereas FTC-TP concentrations in colorectal tissue were 0.04 times as high as the corresponding concentrations in PBMC. For comparisons between colorectal and vaginal tissues in Fig. 2c we used previously derived vaginal tissue-to PBMC ration of 0.07 and 0.06 for TFV-DP and FTC-TP respectively.

Drug potency & pharmacodynamics

We utilized a previously developed-³⁶ and validated⁶⁶ molecular mechanisms of action (MMOA) model to estimate drug potency (IC₅₀), as well as to model the synergistic inhibition by TFV-DP and FTC-TP³², expressed in terms of their concomitant direct drug effect η(I₁, I₂). Notably, this model takes deoxynucleoside triphosphate (dNTP) levels at the site of action into account. Hence we could use the model with dNTP levels measured in CD4+ cells vs. in rectal tissue cells²³ to predict drug potency if either PBMC were a marker of efficacy (consisting mainly of CD4+ cells), or if rectal tissue was the effect site, as previously described.²⁴

Virus dynamics model and prophylactic efficacy

We utilized the HIV dynamics model from²⁹, which takes into account free viruses (V), early infected T cells (T₁) and late infected T cells (T₂). In brief, the HIV replication cycle is modeled by 6 reactions R₁ − R₆ with propensities a₁−a₆ and stoichiometry as outlined below.

where the first three reactions describe the clearance of free virus by the immune system with rate parameter CL or by unsuccessful infection with parameter \(\left(\frac{1}{{\rho }_{rev,\varnothing }}-(1-\eta ({I}_{1},{I}_{2}))\beta \cdot {T}_{u}\right)\), the clearance of the pre-integration complex δ_PIC or early infected cells with parameter δ₁, and the clearance of late infected cells (T₂) with parameter δ₂. The last three reactions describe the infection of a previously uninfected T-cell T_u by free virus with rate parameter (1−η(I₁, I₂))β ⋅ T_u⋅V, the propagation of an early-infected cell to a late infected cell by proviral integration and cell reprogramming with rate parameter k, and the generation of novel viruses from late infected cells with rate parameter N_T. In the equations above, η(I₁, I₂) denotes the direct effect of TFV-DP (I₁) and FTC-TP (I₂), depending on their target site pharmacokinetics (either local tissue or PBMC). All other parameters were taken from Zhang et al. (2021) (Table 1 therein).³⁵

We calculated the prophylactic efficacy per exposure from the viral dynamics model in terms of the relative reduction in infection probability per exposure for a prophylactic regimen S compared to the absence of drugs \(\varnothing\) after virus challenge Y at time t (Y_t).³⁵

where \(P({Y}_{t},\varnothing )\) was calculated analytically.³¹The variable P_I(Y_t, S) was calculated from its complement extinction probability P_I(Y_t, S) = 1 − P_E(Y_t, S). The latter was computed using the method developed in ref. ³⁵ (Probability Generating System) which, for the considered virus dynamics model, results in the following system of coupled ODEs, which is solved backwards in time by considering how an individual pharmacokinetic trajectory affects reactions a₁ and a₄ through time. The backwards integration is performed in matrix form, such that solutions for N (here: N = 1000) virtual patients for a given adherence profile are computed simultaneously using standard ODE solvers.

In this system of ODEs, \({P}_{E}({Y}_{t}=\widehat{V}),\,{P}_{E}({Y}_{t}=\widehat{{T}_{1}})\) and \({P}_{E}({Y}_{t}=\widehat{{T}_{2}})\) denote the probability of viral extinction if only one virus, one early infected cells, or one late infected cells were present. The probability of extinction for any possible state of the viral dynamics model is calculated by assuming statistical independence, i.e.

where #V,  #T₁ and #T₂ denotes the initial number of viruses, early- and late infected T-cells.

Receptive Anal Virus Exposure. Modeling of receptive anal intercourse (RAI) was implemented in this system by considering the exposure-specific inoculum size⁶⁷, see also ref. ²⁴ (Supplementary Fig. S10 therein) and by either utilizing FTC-TP & TFV-DP pharmacokinetics and pharmacodynamics in rectal tissue or PBMC, according to the employed hypothesis testing scheme.

PrEP adherence. In the HPTN 083, PURPOSE 2, and DISCOVER trial, adherence was directly quantified using TFV-DP concentrations in red blood cells (RBCs). For these trials, we assumed a uniform distribution within the individual adherence categories stated in the original studies (e.g., 3-5 doses/week), as shown in Supplementary Fig. 1. We then computed adherence-averaged PrEP efficacy directly utilizing the adherence-efficacy relationships depicted in Fig. 2c, d. For IPERGAY and iPrEx, only plasma TFV levels were available and we utilized our pharmacokinetic models to compute the probability of TFV plasma levels above the lower limit of quantification (LLOQ; compare Supplementary Fig. S1, 2 in ref. ²⁴) for different adherence strata (number of pills per week). This allowed us to estimate adherence frequencies for these studies as well, as shown in Supplementary Fig. 1.

Clinical Trial Simulation

Outcomes from actual clinical trials (i.e., number of infected individuals) may be subject to intrinsic stochasticity. I.e., hypothetically repeating the same trial may yield distinct outcomes since there is randomness in whether an infection may occur before vs. after the end of the trial’s observation period or patient drop-out. To account for this intrinsic randomness, we simulated clinical trials using Gillespie’s algorithm.⁶⁸ The trial is modeled by two possible events: either an individual is infected during the observation period or drops out (= observation period ended). These events occur with reaction rates; \({r}_{{{\rm{Inf}}}}\) and r_dr-out and stoichiometries:

where I and S denote the infected individuals and the remaining number of susceptible individuals and respectively. To simulate the drug-detected group in the data-driven approach, we assigned an individual-level efficacy value φ to each simulated individual from the PrEP efficacy distributions derived from Bayesian inference as discussed above. Accordingly, we utilized an adapted infection incidence \({r}_{{{\rm{Inf}}}}(\varphi )={r}_{{{\rm{Inf}}}}\cdot (1-\varphi )\), where \({r}_{{{\rm{Inf}}}}\) is the infection incidence in the absence of drug. We simulated each trial 100,000 times by inverse sampling the incidence rate and computing \({r}_{{{\rm{dr-out}}}}=1/\widehat{T}-{r}_{{{\rm{Inf}}}}(\varphi )\), where \(\widehat{T}\) denotes the average follow-up time per individual.

When testing the mechanistic hypotheses, we sampled the simulated time-averaged individual-level efficacy φ from our virtual population (compare Fig. 2). This efficacy was sampled according to the probability of detecting TFV or, when available, reported adherence data in the trial (see Supplementary Fig. 1).

Hypothesis testing. Lastly, the PrEP efficacy estimates from clinical data were used to evaluate the mechanistic hypotheses, as previously described in Zhang et al., 2023.²⁴ In brief, we computed a one-sided P-value as the frequency by which the empirically computed distributions of the number of infected individuals at the end of the respective trials overlapped (100,000 simulations respectively). Note that this denotes a conservative test that compared distributions instead of central values (means, medians).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.