Modeling of In-Utero and Intra-Partum Transmissions to Evaluate the Efficacy of Interventions for the Prevention of Perinatal HIV

Background Antiretroviral treatments decrease HIV mother-to-child transmission through pre/post exposure prophylaxis and reduction of maternal viral load. We modeled in-utero and intra-partum HIV transmissions to investigate the preventive role of various antiretroviral treatments interventions. Methods We analysed data from 3,759 women-infant pairs enrolled in 3 randomized clinical trials evaluating (1) zidovudine monotherapy, (2) zidovudine plus perinatal single-dose nevirapine or (3) zidovudine plus lopinavir/ritonavir for the prevention of mother-to-child transmission of HIV in Thailand. All infants were formula-fed. Non-linear mixed effect modeling was used to express the viral load evolution under antiretroviral treatments and the probability of transmission. Results Median viral load was 4 log10 copies/mL (Interquartile range: 3.36–4.56) before antiretroviral treatments initiation. An Emax model described the viral load time-course during pregnancy. Half of the maximum effect of zidovudine (28% decrease) and lopinavir/ritonavir (72% decrease) were achieved after 98 and 12 days, respectively. Adjusted on viral load at baseline (Odds ratio = 1.50 [95% confidence interval: 1.34, 1.68] per log10 copies/mL increment), antiretroviral treatments duration (OR = 0.80 [0.75, 0.84] per week increment) but not the nature of antiretroviral treatments were associated with in-utero transmission. Adjusted on gestational age at delivery (<37 weeks, OR = 2.37 [1.37, 4.10]), baseline CD4 (Odds ratio = 0.79 [0.72, 0.88] per 100 cells/mm3 increment) and predicted viral load at delivery (OR = 1.47 [1.25, 1.64] per log10 copies/mL increment), single-dose nevirapine considerably reduced intra-partum transmission (OR = 0.32 [0.2, 0.51]). Conclusion These models determined the respective contributions of various antiretroviral strategies on prevention of mother-to-child transmission. This can help predict the efficacy of new antiretroviral treatments and/or prevention of mother-to-child transmission strategies particularly for women with no or late antenatal care who are at high risk of transmitting HIV to their offspring. Trial Registration This analysis is based on secondary data obtained from three clinical trials. ClinicalTrials.gov. NCT00386230, NCT00398684, NCT00409591.


Introduction
Mother-to-child transmission (MTCT) of HIV can occur during pregnancy (in-utero), labor/ delivery (intra-partum) or breastfeeding (post-partum). In the absence of antiretroviral treatment (ART), the transmission rate is 10% during pregnancy, 15% during labor and delivery and 10% during breastfeeding [1]. Viral load (VL) is the main predictor of MTCT [2]. Over the last two decades, studies have demonstrated that antiretroviral treatment during gestation, intra-partum, and breastfeeding dramatically reduces MTCT [3][4][5][6]. ART reduces MTCT through two complementary mechanisms: ART can reduce viral load and decrease the infant exposure to maternal viruses. Antiretrovirals can also cross the placenta and provide pre/postexposure prophylaxis to the fetus and infant [7,8]. The 2013 World Health Organization (WHO) consolidated guidelines on the use of antiretroviral drugs for treating and preventing HIV infection [9] recommend that all HIV infected pregnant and breastfeeding women should initiate antiretroviral therapy as early as possible in pregnancy and maintain it at least for as long as the child is exposed to HIV.
Due to the high efficacy of current strategies leading to transmission rates as low as 2% or less, the evaluation of new drugs or drug combinations or new strategies for prevention of mother to child transmission (PMTCT) of HIV requires very large sample sizes to demonstrate efficacy improvements or non-inferiority. Modelling the efficacy of ARTs on MTCT, taking into account known risk factors, becomes increasingly important to gain prior information and optimize clinical trial design.
The objective of this work was to model in-utero and intra-partum HIV transmissions and, after adjusting for known risk factors, to investigate the role of various antiretroviral drug interventions for the PMTCT.
Scholarship of Thailand for her PhD. Co-author Billy Amzal is employed by LASER Analytica. LASER Analytica provided support in the form of salary for author BA, but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The specific role of this author is articulated in the 'author contributions' section.
Competing Interests: The author BA is employed by Laser ANALYTICA. However, this company does not work on HIV, and did not sponsor this research. This does not alter the authors' adherence to all PLOS ONE policies on sharing data and materials.

Material and Methods Patients
Data were collected from pregnant women and infants who participated in three perinatal HIV prevention trials in Thailand.
PHPT-1. (NCT00386230) [4], carried out between 1996 and 2000, was a randomized, double-blind equivalence trial which compared the efficacy of zidovudine (ZDV) starting at 28 weeks' gestation plus 6 weeks of treatment in infants (the reference, "long-long" regimen) versus zidovudine starting at 35 weeks' gestation, with 3 days of zidovudine in infants ("shortshort" regimen), and long-short and short-long regimens.
PHPT-2. (NCT00398684) [10], carried out between 2000 and 2004, was a randomized, double-blind trial of three treatment regimens, which evaluated the efficacy of single dose nevirapine (sdNVP) in mother during labor and in neonates or in mother only in addition to zidovudine during the third trimester of pregnancy and at least one week in children. Women enrolled in two PHPT-2 sub-studies (i) an open-label study for women who presented after 28 weeks gestation and (ii) a nevirapine pharmacokinetic study [11] were also included.
PHPT-5 first phase. (NCT00409591) [12], carried out between 2008 and 2010, was a randomized, 3-arm, double-blind trial. The three ARV strategies initiated during the third trimester were (i) maternal ZDV plus sdNVP at onset of labour and two infant NVP doses (at birth and 48 hours of life), (ii) maternal ZDV and two infant NVP doses, (iii) Maternal ZDV plus lopinavir/ritonavir (LPV/r), with no maternal or infant NVP.
CD4 cell count and viral load were performed before starting antiretrovirals and during pregnancy. Viral loads were repeated at variable times during pregnancy and at delivery. All infants were formula fed. Of the 3,948 confirmed HIV positive pregnant women, 71 were lost to follow up, withdrew consent or died before delivery. Therefore, 3,877 women delivered at the PHPT hospital sites. After exclusion of 28 mothers of a stillborn child, 73 women who had no VL evaluation and 17 women who were receiving HAART for their own health, the aggregated dataset included 3,759 women with at least one VL sample during pregnancy.

Maternal plasma HIV-1 RNA levels
In all studies the maternal HIV-RNA measurement was planned prior to antiretroviral prophylaxis/treatment initiation to assess risk factors of transmission. In PHPT-1 and PHPT-2, VL samples were primarily collected for measurements at entry and delivery, while for PHPT-5, VL was measured monthly to assess HIV-RNA kinetics on antiretroviral drugs.
Plasma VL was assessed at the central PHPT laboratory in Chiang Mai University. Samples from PHPT-1 and PHPT-2 studies were tested using Cobas Amplicor HIV-1 Monitor kit version 1.5 (Roche Molecular Systems, USA) with a limit of quantification of 400 copies/mL; and samples from PHPT-5 first phase using the Abbott m2000 RealTime© HIV-1 assay (limit of quantification 40 copies/mL).

HIV status in infants and timing of transmission
To determine HIV infection status in infants, peripheral blood was drawn and spotted onto filter papers, dried and stored at -20°C before shipment to a central laboratory. Each of the PHPT samples were collected at birth, 6 weeks, 4 and 6 months. PHPT-5 samples were collected at birth, 7-10 days, 1, 2, 4 and 6 months of age.
In the original trials, infants were considered confirmed HIV-infected if samples obtained on two separate occasions were found positive by HIV-1 DNA PCR and confirmed HIV-uninfected if samples obtained on two separate occasions after one month of age were negative [13].
When only one DNA PCR was available and positive, infants were considered unconfirmed HIV-infected. When only one DNA PCR was available after the 1 st week of life and negative, infants were considered unconfirmed HIV-uninfected. When only one DNA PCR was available within the 1 st week of life and negative, infants were considered as indeterminate [13]. In the present analysis, unconfirmed HIV infected infants were considered HIV-infected and unconfirmed HIV-uninfected infants were considered HIV-uninfected, while indeterminate infants were excluded.
Infants with a positive DNA PCR result during the first week of life were considered to be infected during pregnancy ("in-utero" transmission); infants with negative HIV-DNA PCR results during the first week of life but with a subsequent positive result were considered to be infected during labor or delivery ("intra-partum" transmission) [13]. Twins were considered a single entity and discordant twins were counted as one infected infant.

Ethics
Each of the PHPT perinatal study protocols and their amendments, as well as the use of data for this analysis received ethical clearance from the Thai Ministry of Public Health, the Harvard School of Public Health and Chiang Mai University Faculty of Medical Associated Sciences Ethics Committees. The consent procedures were reviewed and approved by the ethics committees. Before enrollment, all women provided written informed consent for their participation and that of their infants.

Modeling of viral load time course during pregnancy
An Emax model was chosen since it is based upon pharmacological principles, i.e. the theory of drug action mediated by ligand-receptor interaction which translates in an hyperbolic equation. Because ZDV and LPV/r have 2 distinct sites and mechanisms of action, the effects were considered to be additive [14]. A proportional effect Emax model [14] was applied to describe the viral load at time T during pregnancy (VL, expressed in log values). The model took into account VL before treatment initiation (VL 0 ) and the duration and nature of the 2 ARTs including ZDV and LPV/r (S1 Dataset) was composed as follows.
Model parameters were VL 0 : VL before treatment E MAX : Treatment maximum effect γ: Hill coefficient for treatment effect T 50 : Treatment duration to reach half of E MAX Since ZDV and LPV/r have distinct sites and mechanisms of action, the effects of ZDV and LPV/r administered in combination were assumed to be additive [14]. Because the inhibition cannot exceed 100%, E MAX,LPV was deduced from E MAX,ZDV by Where E MAX,ZDV : ZDV maximum effect E MAX,LPV : LPV/r maximum effect Interindividual variability was modeled using an exponential error model, with η i being the interindividual random effect with mean 0 and variance ω 2 .

Modeling of in-utero and intra-partum transmissions
In-utero and intra-partum transmissions were treated as independent outcomes. Logistic regression models with random effect (η) were developed to predict in-utero and intra-partum transmission according to relevant risk factors. For each odds ratio, point estimate and 95% confidence intervals are provided. We investigated known risk factors for mother-to-child transmission of HIV [2,15,16]. In-utero transmission was assumed to depend on VL 0 (log 10 copies/mL), drug(s) treatment duration(s) in weeks, CD4 count before treatment (CD4 BASELINE ) and gestational age (GA) at treatment initiation and at delivery (S2 Dataset). The risk of intra-partum transmission was dependent on VL at delivery (VL DELIVERY ), itself predicted by the VL time-course model, perinatal NVP (sdNVP administered at onset of labor or during the first hours of life or both), maternal ZDV loading dose during labor, delivery mode, premature labor (Gestational age (GA) <37 weeks) and ART(s) administered to infants during the first weeks of life (S3 dataset).

Data analysis
VL time-courses and MTCT events were analysed using a non-linear mixed effect modeling approach. Parameters were estimated by computing the maximum likelihood estimators without any approximation of the model (no linearization) using the stochastic approximation expectation maximization (SAEM) algorithm combined to a Markov Chain Monte Carlo (MCMC) procedure. Data were analysed using MONOLIX (version 4.1.2, http://www.lixoft. com/) [17,18]. VL counts were log transformed and residual variability was described by an additive error, whereas an exponential model was used for between-subject variabilities (η). Data below the limit of quantification were left-censored [19]. The effect of a covariate on a structural parameter was retained if it produced a decrease in the Bayesian Information Criterion (BIC) compared to the baseline model i.e. the covariate-free model. A smaller BIC value signifies a model that better fits the data [20].
The logistic model for MTCT events analysis was written in a MLXTRAN script file (S1 MLXTRAN scripts); the random effect η was assumed to be normally distributed. In the univariate analyses, variables that both decreased BIC and had acceptable relative standard error (RSE<50%) were considered as significant factors to be included in the multivariable analyses. In the multivariable analysis, these variables were added one by one considering the largest drops in the BIC value to define the final model.

Visual predictive check (VPC) evaluation
Simulated VL time-courses were compared with the observed data to evaluate the performance of the model. The vector of model parameters from 400 replicates of the database was simulated. The 5 th , 50 th and 95 th percentiles of the simulated dependent variables at each time were then overlaid on the observed data. The proportion of observed MTCT with their confidence intervals were plotted as a function of significant predictors. The 5 th , 50 th and 95 th percentiles of the model predictions were simultaneously plotted. Visual inspection was used to confirm that the observed proportions were included in the limits defined by the percentiles curves. The residual sum of square (RSS) was provided in addition to graphical check.

Characteristics of the study population
A total of 3,759 HIV-infected pregnant women enrolled in the PMTCT studies from 1996 to 2010 were included (Fig 1). Table 1 presents the baseline and delivery characteristics of the women included in the analysis and the treatments they received.

VL time-course modeling
A total of 5,576 VL measurements in 3,759 subjects were available for modeling (median 1 measurement per patient, range 1 to 6). Sixty five percent of the women had only 1 measurement (all but 5% of these at ART initiation), and 35% had at !2 measurements. Median VL was 4 log 10 copies/mL (IQR: 3.36-4.56) before ART initiation and 3.51 log 10 copies/ mL (IQR: 2.89-3.34) at delivery. VL at any time point during pregnancy was dependent on baseline VL and ZDV and LPV/r treatment durations and was well described by a combined Emax model. The Hill coefficient for ZDV effect (γ ZDV ) γ ZDV was close to 1 and thus was fixed to 1. The η parameters for γ ZDV , LPV/r duration to reach half of E MAX,LPV (T 50,LPV ) and the Hill coefficient for LPV/r (γ LPV ) γ ZDV were not statistically significant. Removing them from the model did not alter the quality of the fit or further decreased the BIC value (final model, BIC = 2113.33). None of the other covariates, including CD4 cell count and GA at baseline, had a significant effect on model parameters. All parameters were well estimated with RSE below 30% ( Table 2). The model estimated that half of the maximum effect of ZDV (28% VL decrease from baseline) and LPV/r (72% decrease) were observed after 98 and 12 days respectively. Using the population parameter, the RSS was 747.39 while it was 369.20 when using individual parameter. The observed vs. model-predicted plots are shown in Fig 2A and 2B (top). The visual predictive checks are shown in Fig 2C and 2D (bottom).

MTCT modeling
The MTCT models were built step by step from the basic model (without explanatory variable). Viral load at delivery (VL DELIVERY ) was estimated through the VL time course final model using individual parameters.
In-utero transmission. Upon univariate analysis, CD4 count, gestational age, VL before treatment initiation and ZDV duration caused a drop in the BIC, indicating significant effects of these variables. In the multivariable analysis, only ZDV duration and VL before treatment initiation remained independently associated with in-utero transmission (Table 3). LPV/r duration and baseline CD4 count were not significantly associated with in-utero transmission.
The final model was: where ZDV DURATION is the ZDV duration during pregnancy (weeks). The probability of in-utero HIV transmission as a function of zidovudine duration is shown in Fig 3A,   Intra-partum transmission. Perinatal sdNVP, duration of ZDV, premature labor (GA at delivery<37 weeks)viral load at delivery and CD4 cell count were significantly associated with intra-partum transmission in the univariate analyses, while mode of delivery, ZDV loading dose and infants ZDV prophylaxis were not.
In the multivariable analysis, final model included VL at delivery, perinatal sdNVP administration, CD4 cell count and premature labor ( Table 4). The duration of ZDV was no longer significant when other factors were included.  Fig 3C and 3D show the probability of intra-partum HIV transmission as a function of viral load at delivery without and with single dose nevirapine, respectively."

Discussion
The VL time-course model during pregnancy developed as a function of the type of treatment administered, its duration and the VL level at baseline, provided a good prediction of the VL level at delivery. This predicted VL could be used in the MTCT models. VL at treatment initiation and treatment duration were the main determinants of in-utero transmission, regardless of the ARV regimens used. High VL at delivery, absence of perinatal sdNVP and premature labor were associated with intra-partum transmission.
As previously shown in PACTG 076 [21], although ZDV monotherapy had only a slight effect on maternal VL (only 0.43 Log decrease in this analysis), it was very effective in reducing in-utero transmission. This is consistent with the accepted concept that ZDV, which cross the placenta freely, exerts its prophylactic effect largely through pre-post exposure prophylaxis [21]. In the late 1990, it was hypothesized that in-utero transmission would occur late in pregnancy [22]. This justified for the launch of several short ZDV course trials in developing countries [23][24][25]. However, this was not confirmed by the PHPT-1 trial where the rates of in-utero transmission were 1.6 versus 5.1% with long and short ZDV treatments, respectively [4,15]. This supports the WHO PMTCT 2013 guidelines recommending ART initiation as early as possible during pregnancy [9]. Although ZDV plus LPV/r was much more effective than ZDV alone in reducing VL, adding LPV/r to ZDV did not decrease further in-utero transmission [26].  Although it had no effect on in-utero transmission, the addition of LPV/r, which has limited perfusion through the placenta, had as expected a major effect on the VL at delivery (mean reduction, 2.18 log 10 copies/mL), and thus a major effect on intra-partum transmission. More importantly, after adjusting for all factors associated with transmission, perinatal sdNVP, in the mother only, the mother and her infant, or the infant only, markedly reduced the risk of intrapartum transmission at all viral load levels (Fig 3C and 3D).
Modeling of VL during LPV/r plus ZDV treatment ( Fig 2D) showed that with a treatment duration less than 8-10 weeks before delivery, VL at delivery remained detectable in a large proportion of women. Accordingly, when mothers start HAART late in pregnancy, it is advisable to intensify the ART regimen by providing sdNVP to both mother and infant and a brief course of combined ART to the newborn in order to reduce the probability of intra-partum transmission [27][28]. Interestingly, in the presence of perinatal sdNVP, intra-partum transmission was only weakly associated with VL at delivery, indicating the potent pre/post exposure prophylactic effect of this drug.
Prematurity has been found to be associated with perinatal transmission in several studies [15,16,29]. It has been debated whether prematurity was a consequence of in-utero transmission or if premature infants were more vulnerable to HIV infection [30]. The fact that in this study prematurity was associated with intra-partum but not with in-utero transmission supports a higher vulnerability of premature children.
Several studies reported that planned caesarean section (CS) [31,32] reduced the risk of intra-partum transmission in particular when VL at delivery is high but, in this study, the percentage of women with planned C-section was too low (8%) to assess this factor. This study has several limitations. In our definition of HIV status of the infants, we considered as infected or uninfected, infants with unconfirmed HIV status and excluded those who were indeterminate. However, a sensitivity analysis restricted to infants with confirmed HIV status provided very close results (data not shown). Datasets from other trials with different antiretroviral prophylaxis regimens could have been incorporated into the model but this would have perhaps offset the advantages of using data from trials performed in the same setting by the same team. Also, all subjects were from Thailand although there is no indication that risks of transmission and intervention effectiveness differ across ethnic groups.
In conclusion, our models provide insights on the respective roles of pre-exposure prophylaxis and maternal viral load reduction in preventing mother-to-child transmission according to the preventive strategy. With the regimens considered in this analysis, while the preventive effect of ART during pregnancy was essentially driven by transplacental pre/post exposure prophylaxis, both viral load reduction by the time of delivery and infant prophylaxis were important in preventing intra-partum transmission. Given the high efficacy of current interventions, clinical trials to test the efficacy of new antiretrovirals or PMTCT strategies have become more and more difficult to implement. A Bayesian approach with data from previous clinical trials could reduce sample sizes and help optimize trial design.