Pharmacometrics and Machine Learning-Based Modeling to Assess the Impact of Vaginal Microbiome on Tenofovir Exposures in the Female Genital Tract of African Women

Objectives: Antiretrovirals are effective for HIV pre-exposure prophylaxis (PrEP), but treatment outcomes vary by genders [ 1]. Truvada, a U.S. FDA-approved oral combination of tenofovir (TFV) and emtricitabine (FTC), reduces HIV infection risk by over 70% with only two doses per week in men who have sex with men [ 2]. However, despite similar adherence, similar benefits have not been observed in heterosexual women [ 3]. This discrepancy may partly be attributed to differential mucosal tissue penetration and other host factors. Being important in mucosal health, the vaginal microbiome is emerging as a potential modulator of drug responses. However, its high dimensionality poses a challenge to identify specific taxa influencing drug exposure, especially in the context of other clinical variables. This study aimed to develop a combined pharmacometrics (PMx) and machine learning (ML) framework to investigate the impact of the vaginal microbiome on TFV cervical tissue exposure. Methods: Data were collected from a single-visit clinical study (NCT03377608) involving 50 premenopausal Uganda women living with HIV and stable on a TFV disoproxil fumarate (TDF)/ lamivudine regimen [ 4]. Steady-state TFV concentrations in plasma and TFV diphosphate (TFVdp, its active metabolite) concentrations in cervical tissues were collected at a single time point. Vaginal microbiome compositions were acquired using 16S ribosomal RNA sequencing. A published TFV pharmacokinetic (PK) model incorporating plasma and cervical tissue compartments was extracted [ 5]. After validating model performance via visual predictive checks, empirical Bayes estimates (EBEs) were derived using Maximum A Posterior (MAP) Bayesian method in mapbayr package [ 6]. Simulations using EBEs were conducted to derive individual TFV exposures (TFV AUC0-240 hours in plasma and TFVdp AUC 0-240 hours in cervical tissues) under a 2-1-1 dosing on-demand PrEP regimen (600 mg before sex followed by two 300 mg once daily doses). Microbiome data were transformed using centered log-ratio methods. Correlation heatmaps provided an initial assessment of genus-level associations with drug exposures. Ridge and LASSO regressions (with λ optimized via leave-one-out cross-validation) were applied to identify impactful predictors of TFVdp cervical AUC. Results: The PK model adequately characterized the observed plasma and cervical concentrations. EBEs were derived to drive the simulation of Individualized exposures. Correlation heatmaps demonstrated both positive (e.g., Gemela, β=0.41) and negative (e.g. Megasphaera, β=-0.30) associations between genera and TFVdp cervical AUC. Besides TFV plasma AUC (β=0.083) and weight (β=-0.048), LASSO analyses identified Gemella, Falsiporphyromonas, Streptobacillus, and Megasphaera based on penalized regression coefficients (βs: 0.38, -0.0023, -0.097 and -0.23) as influential genera on TFVdp cervical AUC. Conclusions: This study presents a feasible two-step PMx-ML approach to investigate the impact of high-dimensional microbiome on drug exposure. Our analyses identified several vaginal microbiome genera that may significantly affect TFV tissue exposures, providing insights into the observed variability in PrEP efficacy among women.