Synthesis of protective oral PrEP adherence levels in cisgender women using convergent clinical- and bottom-up modeling

Globally, most HIV infections occur in heterosexual women in resource-limited settings. In these settings, female self-protection with generic emtricitabine/tenofovir disoproxil fumarate pre-exposure prophylaxis (FTC/TDF-PrEP) may constitute a major pillar of the HIV prevention portfolio. However, clinical trials in women had inconsistent outcomes, sparking uncertainty regarding risk-group specific adherence requirements and causing reluctance in testing and recommending on-demand regimen in women. We analyzed all FTC/TDF-PrEP trials to establish PrEP efficacy ranges in women. In a ‘bottom-up’ approach, we modeled hypotheses corroborating risk-group specific adherence-efficacy profiles. Finally, we used the clinical efficacy ranges to (in-)validate hypotheses. We found that different clinical outcomes could solely be explained by the proportion of enrolled participants not taking the product, allowing, for the first time, to unify clinical observations. This analysis showed that 90% protection was achieved, when women took some of the product. Using ‘bottom-up’ modelling, we found that hypotheses of putative male/female differences were either irrelevant, or statistically inconsistent with clinical data. Furthermore, our multiscale modelling indicated that 90% protection was achieved if oral FTC/TDF was taken at least twice weekly.

PrEP [19] and VOICE [20] reported no incidence reduction for FTC/TDF-based PrEP in 96 women. In Fem-PrEP about 15% of infected individuals had detectable drug levels; in VOICE, 97 adherence was low (~30%) and intermittent, such that it is unclear whether the product was

102
For FTC/TDF-based oral PrEP, some early studies pointed towards lower average risk 103 reduction in heterosexual women, compared to men-who-have-sex-with-men (MSM) [14,19,104 20]. However, it is unclear to date, whether this putatively lower efficacy is a consequence of 105 intrinsic differences in physiology and pharmacokinetics at the site of exposure, or whether it 106 is an artefact of poorly quantified and differing levels of adherence across studies, as many 107 infected participants in these trials were simply not taking PrEP around the time of HIV 108 exposure.

109
For developing PrEP guidelines, the existence of intrinsic differences would be relevant, as it 110 necessitates specific recommendations on, for example, minimal adherence levels, which can 111 also affect the uptake and persistence of PrEP in the considered risk group [21,22]. Current 112 WHO guidelines for PrEP differentiate between heterosexual women and MSM [23]. While 113 PrEP on demand is considered safe in MSM based on the IPERGAY and PREVENIR studies 114 [24], no such study has been attempted in women; therefore, on-demand PrEP is not 115 recommended for heterosexual women. The reluctance to test on-demand PrEP in women is 116 partially originating in prior mathematical modelling studies which suggest higher adherence 117 in women vs. MSM, for the same level of protection [25]. Notably, these modelling studies 118 were partially motivated by results from the Fem-PrEP [19] and VOICE [20] study that failed 119 to show efficacy in women, but also showed the lowest adherence levels of all investigated 120 studies (hence may be uninformative regarding the adherence-efficacy profile).
In this work, we used two independent approaches to quantify the population-specific 122 adherence-protection relationship for PrEP in women (illustrated in Fig. 1): In a data-driven 123 'top-down' approach, we solely utilized clinical data to quantify PrEP efficacy. In analogy to 124 the analysis of early PrEP studies in MSM [26], we used population pharmacokinetic models, 125 to dichotomize the intervention arms into sub-cohorts of product non-taking vs. product taking. 126 Simulation of these sub-cohorts allowed us to estimate PrEP efficacy and corresponding 127 confidence bounds for each clinical study in individuals who took some of the products. 128 In an independent, hypotheses-driven, 'bottom-up' approach, we tested mechanisms that have 129 been proposed to explain risk group-specific adherence-protection relationships, such as the 130 types of exposure (receptive vaginal vs. anal intercourse), potentially distinct drug potency, as 131 well as drug concentrations at these exposure sites. These mechanisms were integrated into an 132 advanced multi-scale modelling framework that has been consecutively developed over the 133 past 10 years [27][28][29][30][31][32][33]. This model allowed us to synthesize adherence-protection profiles for 134 each proposed hypothesis and all their combinations. 135 Finally, mechanistic predictions were evaluated in the light of the clinical data ('top-down' 136 approach). Most importantly, this final model verification approach allowed us to statistically 137 rule in/out proposed mechanisms on differential PrEP efficacy in heterosexual women and 138 inform minimal adherence requirements in this risk group.

140
Pharmacokinetic Modelling. To fully reflect the pharmacokinetics (PK) of FTC/TDF, we 141 utilized the models by Burns et al. and Garrett et al. [34,35], which were trained on rich 142 pharmacokinetic data from the MTN-001, as well as the NCT010330199 and NCT02357602 143 clinical trials (details in the Methods section). Importantly, these models consider relevant Based on these models, we first established the link between dosing frequency and detectable 151 plasma TFV (> 0.001μM), which was the adherence marker used in Partners-PrEP, TDF2, 152 VOICE and HPTN084. Notably, if a sample is taken from a person, who takes TDF or 153 FTC/TDF once a week, this sample will have detectable plasma TFV was above the limit of 154 detection of 0.001μM with 57% probability (IQR: 46-71%), Supplementary Fig. S2. For a 155 person taking 2 pills a week on average, the probability is already 88%, and 96% for 3 pills a 156 week, Supplementary Fig. S2. In other words, individuals without detectable plasma TFV we assumed a PrEP efficacy of 0%. The proportion of random samples with undetectable 167 plasma TFV in the FTC/TDF intervention arm was 19% in Partners-PrEP and TDF2 study, 168 44% in HPTN084, 64% in Fem-PrEP and 71% in the VOICE ( Fig. 2A-E). We then simulated 169 the intervention arms of these studies by assigning a fraction of observation period (person-170 years) to a "drug undetected" sub-cohort, as well as by assigning a fraction of the infected 171 individuals to the "drug undetected" sub-cohort, if TFV was not detectable ( Fig. 2A-E, Table   172 1). The remaining observation time and number of infected individuals was assigned to the 173 sub-cohort of the intervention arm where individuals had detectable drug.

174
Validity of dichotomization into "detectable" vs. "undetectable" drug. To assess our 175 assumption of 0% efficacy in the "drug undetected" intervention sub-cohort, we first computed 176 incidences in this sub-cohort directly from the respective studies ( Fig. 2A-E). These incidences 177 corresponded well with the incidences in the respective placebo arms, albeit slightly (but 178 insignificantly) higher, indicating that we may safely assume that the efficacy of FTC/TDF in 179 non-adherent individuals ("drug undetected" intervention sub-arm) was 0%. We then simulated 180 the "drug undetected" sub-cohort, as explained in the Methods section, by taking both the 181 follow-up time, as well as "drug undetected" incidences from the respective studies as input 182 (Methods and Supplementary Text S1). Our clinical trial simulation took two sources of 183 variability into account: (i) uncertainty in the incidence rate, as well as (ii) intrinsic 184 stochasticity. While the former is standard procedure, the latter is often ignored, but relevant 185 here because infection is an extremely rare event in the investigated clinical trials, contributing 186 to intrinsic randomness. The resultant number of infections (and their uncertainty) are depicted 187 in Fig. 2A-E and are highly consistent with the clinically reported numbers. As a further 188 validity check, we also depicted the simulation-derived incidence rates in Suppl. Fig. S4 189 together with data-derived placebo and "drug undetected" incidences. Notably, because our 190 simulation took both, intrinsic stochasticity as well as uncertainty in the actual incidence rate 191 into account (Supplementary Text S1), and therefore corresponding confidence intervals are 192 wider than those derived from the clinical data alone (which neglect 'intrinsic stochasticity'), 193 Suppl. Fig. S4. 194 PrEP-efficacy in individuals with detectable drug. Next, we calculated the range of PrEP 195 efficacies in individuals with detectable drug (individuals who took some product). For this, 196 we only used reported data, as well as the simulated "drug undetected" sub-cohorts (previous 197 paragraph). For each simulation, we calculated the number of infections (and its uncertainty) 198 in the subset of individuals with "detectable drug" by subtracting the number of infections 199 derived from the "drug undetected" simulation from the total number of infections reported in 200 the respective clinical study. Since we perform stochastic simulations, this yields a probability 201 distribution. I.e. we derive the probability to observe x = 0, …, N infected individuals in the 202 "drug detected" sub-cohort !"#$ ( ); details in Suppl. Text S1. Using the incidences from the 203 "drug undetected" sub-cohort and the entire range of theoretically possible PrEP efficacies 204 (incidence reductions) , we then calculated the probability of , conditioned on observing x 205 infections in a simulated trial !"#$ ( | ). Combining both, we could deduce information about inferring PrEP efficacy in individuals with detectable drug, but rather points towards higher 213 efficacy. (iii) The remaining studies (HPTN084 and Partners-PrEP) point towards high PrEP 214 efficacy in women taking some of the product. I.e. the vast majority of !"#$ ( ) lies above 215 80% and the median efficacy is ~90%.

216
Importantly, this analysis did not yet make any assumption about adherence in individuals with 217 detectable plasma TFV, other than that individuals with undetectable drug have 0% PrEP 218 efficacy; i.e. similar to placebo ( Supplementary Fig. S4).

243
For evaluation, we consider a combinatorial approach to test the proposed mechanisms alone 244 and in combination. We will designate the respective simulation setting in analogy to a light 245 switch, Fig. 4A, where each of the 'four lights' corresponds to a mechanism that we model:      This scenario predicted high average prophylactic efficacy (98%) in fully adherent women 300 after receptive vaginal intercourse (RVI). Simulation results for periodic, but incomplete 301 adherence are depicted in Fig. 5A and show that if FTC/TDF was taken with an adherence of 302 14% (once weekly) median efficacy was 65% (IQR: 35-90%), while with two and three doses 303 per week adherence median efficacy climbs to 90% (IQR: 75-96%) and 96% (IQR: 90-98%).

304
Before challenging all models with reported and simulated clinical trial outcomes, however, 305 we will first assess which of the aforementioned hypotheses -alone or in combination -alter  CD4+ T-cells. 353 We next investigated the relationship between TFV-DP and FTC-TP concentrations in PBMC 354 vs. local tissue-or cell homogenates. For hypothesis testing, we additionally predicted 355 prophylactic efficacy under the assumption that TFV-DP/FTC-TP concentrations in local tissue 356 homogenates coincided with effect-site concentrations. 357 We identified 3 studies with 8 dosing regimens for FTC-TP and 5 studies reporting 10 dosing 358 regimens for TFV-DP that report local TFV-DP or FTC-TP concentrations [36][37][38][39][40][41][42]. By 359 simulating the respective dosing regimens using our pharmacokinetic models (Methods 360 section), we enabled a direct comparison of measurements between studies and vs. our  To evaluate the hypothesis that local tissue concentrations represent a marker of efficacy, we 375 subsequently used the derived local tissue:PBMC concentration ratios to predict local tissue 376 pharmacokinetics. Using this data, model-predicted prophylactic efficacy was markedly 377 reduced. I.e., best-case PrEP efficacy in fully adherent individuals was only 47% (IQR: 42-378 55%), compared to the baseline scenario (efficacy: ~98%).

379
In summary, our simulations point out that the type of exposure, as well as local dNTP 380 concentrations have little impact on PrEP efficacy when considered in isolation. To the 381 contrary, if tissue homogenates were a marker for the relevant effect-site concentrations, PrEP 382 efficacy was markedly reduced, even in fully adherent individuals. Next, we assess whether 383 the hypothesized mechanisms when considered in combination, impact PrEP efficacy 384 differently, and whether they set certain minimum requirements for PrEP adherence to protect 385 women from acquiring HIV infection.

386
Combined impact of hypothesized mechanisms on PrEP efficacy and adherence requirements. 387 We simulated all combinations of aforementioned hypothesis and assessed the impact of  Next, we will evaluate which predicted adherence-efficacy profiles, and thus which of the 416 tested mechanisms are consistent with outcomes of clinical trials in women.

417
For this, we simulate the clinical trials as before, but scale the trial-specific incidences with the  are inconsistent with clinical data. This allowed to rule out mechanistic hypotheses that were 424 postulated to explain efficacy and risk group-specific adherence requirements in women.

425
Individuals with detectable drug. We simulated the sub-cohort of the intervention arm, where 426 individuals had detectable drug, by sampling from the "drug undetectable" incidence (Suppl.

427
Text S1) and multiplying with the weighted average PrEP-efficacies depicted in with-and without detectable drug concentrations is shown in Table 2.

433
The TDF2, Fem-PrEP and VOICE studies do not allow us to distinguish between any of the 434 hypotheses (Table 2)   With regards to the available clinical studies, our population-pharmacokinetic modelling 483 indicated that individuals having undetectable TFV plasma drug levels (< 0.001μM) must have 484 taken FTC/TDF less than once a week, if at all (Suppl. Fig. S2-3). Since TFV plasma levels 485 have been reported in a random sub-set of the PrEP intervention arm, we were thus able to 486 dichotomize the intervention arms into sub-cohorts 'undetectable drug' (≤ 1 dose per week), 487 and 'detectable drug' (≥ 1 dose per week). Interestingly, an identical dichotomization was performed to analyze trials in MSM, concluding that FTC/TDF-based PrEP is ~90% efficient 489 in individuals who take some product [26]. 490 When we assumed negligible PrEP efficiency for 'undetectable drug' sub-cohort and simulated 491 the corresponding trials, we derived incidences closely matching incidences in the respective 492 placebo arms. While serving as an internal control for our analysis, this finding also indicates 493 that the intervention arm of the distinct studies contains variable 'placebo-like' observation 494 periods, being either individuals that were never protected, or not protected for a period of time, site pharmacokinetic models are still incompletely understood and need to be aligned with more 520 sophisticated experimental (e.g. ex vivo) data, including topically applied PrEP.

521
Interestingly, when we simulated the various clinical sub-cohorts dichotomized for detectable 522 plasma TFV, it appeared much more likely that FTC/TDF efficacy was high (~90%), in favor 523 of hypotheses, in which PBMCs represent a more suitable matrix to determine effect-site drug 524 concentrations (Table 2). To fully reflect the pharmacokinetics of the two drugs, we utilized the previously developed 596 models by Burns et al. and Garrett et al. [34,35]. In both models, the amount of (pro-)drug in internally to avoid unit conversions. The following ordinary differential equations (ODEs) 602 were used to describe the mass-flux between aforementioned compartments, in between two 603 dosing events: The terms ) and * (1/h) denote the absorption-and elimination rate constants respectively.  Table 1 and Fig .2.  following reaction stoichiometries and reaction propensities: convenient to compute the extinction probability W , which is its complement:

709
To compute the extinction probability for a certain regimen, we used the method developed in

710
[33]:     in person-years) was dichotomized into "drug detected" vs "undetected" based on the fraction 828 of samples with detectable plasma TFV. Likewise, the number of infections Inf Tot were 829 proportionally assigned based on the fraction of infected individuals with detectable plasma 830 TFV (Inf Obs ). We then performed clinical trial simulations of the 'drug undetected' sub-cohort 831 (Methods and Supplementary Text S1), assuming 0% PrEP efficacy. Based on these 832 simulations, the number of infections with 'detectable drug' was then calculated as Inf Tot -833 Inf clin_sim (drug undetected), which allowed to estimate uncertainty in this quantity without 834 having to make assumptions about the efficacy of PrEP (Methods and Supplementary Text S1).

835
A placebo arm was not present in the HPTN084 study. Uncertainty in the placebo incidence 836 was computed with corresponding clinical trial simulation (Methods and Supplementary Text 837 S1).

Intervention arm
Placebo arm