Alexandre Marie (1,2,3,4), Prague Mélanie (1,2,3,4), Yves Lévy (2), Thiébaut Rodolphe (1,2,3,4)
(1) INRIA, SISTM team, Bordeaux, France, (2) Vaccine Research Institute (VRI) Créteil, France, (3) INSERM U1219, Bordeaux, France, (4) University of Bordeaux, ISPED, Bordeaux, France
Introduction: HIV therapeutic vaccines are currently developed with the objective of long-term viral control in the absence of anti-retroviral treatment (ART). To evaluate their efficacy, supervised treatment interruptions (STIs) are often used. Modeling biomarkers before and after STI in HIV-vaccinated individual is needed to gain a deeper understanding of mechanisms of action of vaccines.
Objectives: We aim at modeling viral load rebound in HIV-infected patients after HIV prime-boost vaccination during ART cessation. We investigate which mechanistic model, based on ordinary differential equation (ODE), is best to model the variability of the viral rebound induced by reservoir reactivation.
Methods: Five mechanistic models were studied to describe the viral rebound and CD4+ T cells count dynamics: (1)-(2) 2 and 3 compartments ODE systems featuring uninfected CD4+ T cells, productively infected CD4+ T cells and viruses concentration, respectively with and without proportional relation between infected cells and viruses, (3) 4 ODE system also considering quiescent CD4+ T cells, (4) 5 ODE system also considering immune response by means of effector and precursor immune cells and (5) 5 ODE system introducing eclipse phase and non-producing virions. Practical and theoretical identifiability as well as sensibility analysis were performed to evaluate these models. Parameter estimations were done using the SAEM (Stochastic Approximation of Expectation-Maximization) algorithm as implemented in Monolix. Finally AIC criteria, quality of prediction using cross-validation leave-one-out and external data validation were used as comparison criteria.
Results: Models were evaluated on HIV therapeutic vaccine trial VRI02 ANRS 149 LIGHT, a double-blinded phase II trial including ninety-eight chronic HIV patients with CD4+ T cells counts ≥ 600 cells/μL and HIV RNA < 50 copies/mL for at least 6 months previous to the study and whose nadir CD4 ≥ 300 cells/μL while on ART. Patients were randomized (2:1) to receive either GTU-MultiHIV B at week 0, 4 and 12 followed by LIPO-5 vaccine at week 20 and 24, or vaccine placebos. Prime-boost vaccination strategy was followed by STI at week 36 until 48 and resumption of ART after week 48. Viral load and CD4+ T cells dynamics of the sixty-seven patients of the control group of the ILIADE clinical trial were then considered as external validation sample to evaluate the models (1-5) quality of prediction. In this trial 24 weeks on ART is followed by therapy cessation at week 24 until week 120 and a follow up until week 168, for patients with CD4 cells count above 350 cells/μL. Practical identifiability was achieved by fixing, when needed, some of the parameters of the models. Overall, in most patients, adjustment and prediction where satisfactory, but simple model (2-3) failed to describe specific viral rebound trajectories, such as delayed viral rebound. The best model, which is a model with 5 ODE, has an AIC decreased by about 10% compared to other models. In term of prediction on the ILIADE dataset, we have a good agreement between observations and predictions. Assuming the measurement error similar between LIGHT and ILIADE datasets, we find that about 95% of the observation for CD4 T cell count and 90% of the observation for viral load count are accurately predicted. For the latter, this is lower than the expected nominal value of 95% but no particular trend indicating a model misspecification was detected. We explain most of this inconsistency by the fact that the model assumes perfect adherence.
Conclusions: Although increasing the number of parameters to estimate makes the estimation less identifiable and lead to the necessity of fixing parameters, we demonstrate that ODE models with higher number of compartments are more flexible to adjust viral rebound and CD4 T cell count data and show good prediction abilities on external dataset.
Reference: PAGE 28 (2019) Abstr 8960 [www.page-meeting.org/?abstract=8960]
Poster: Drug/Disease Modelling - Infection