Clarisse Schumer 1, Matthias Pierre 1,2, Julie Bertrand 1, Jérémie Guedj 1
1 IAME, Inserm (Paris, France), 2 Simulation Plus PTO (Lancaster, USA)
Introduction: Within-host models are widely used to characterize viral kinetics, quantify host-pathogen interactions and to support the evaluation of antiviral strategies [1–3]. Increasingly, such models are developed across pathogens to predict and compare biological mechanisms. However, model selection remains largely empirical and the increasing availability and heterogeneity of data make reproducible model selection challenging. As a result, differences reported between viruses may reflect modeling choices rather than true biological variability. A standardized framework is therefore needed to ensure reproducible model selection and enable meaningful comparisons across viruses.
Automated model selection offers a systematic and reproducible framework to explore model assumptions and quantify model uncertainty. While automated procedures are well established in pharmacokinetics [4,5], their application to within-host viral dynamics remains limited. Here, we adapted a decision tree-based automated framework [6] to explore structural assumptions, parameter identifiability and inter-individual variability, and applied a common model selection procedure to two respiratory viruses using human challenge data.
Objectives:
1. Develop a reproducible automated framework for within-host model selection
2. Apply this framework to SARS-CoV-2 and influenza A virus (IAV) human challenge data [7,8]
3. Discuss selected mechanisms, identifiability and resulting viral dynamics across viruses
Methods: We analyzed human challenge datasets including 12 SARS-CoV-2 infected participants and 44 IAV infected participants. Viral RNA and infectious virus measurements were available for SARS-CoV-2, whereas only infectious virus measurements were available for IAV. Model parameters were estimated using nonlinear mixed effects modeling in Monolix.
Model selection relied on a decision tree (DT) algorithm applied sequentially:
1. Structural features: A search space defined by 9 features, including core infection processes (inoculum loss, eclipse phase, refractory cells and target-cell regeneration) and innate or adaptive effects implemented as saturating modulation of viral parameters.
2. Parameter identifiability: Parameters with relative standard error (RSE) greater than 50% were considered as unidentifiable, and alternative fixed values within a predefined range were explored.
3. Parameter inter-individual variability. Random effects were removed sequentially from smallest to largest variance.
At each step, the models with a corrected Bayesian Information Criterion (BICc) within 2 units of the best model were retained [9,10]. This procedure resulted in a set of candidate models rather than a single model. Predictions were therefore obtained using model averaging across the selected models [11].
Results: The structural search space defined 512 candidate models that could be tested using virological data only. The structural selection step required approximately 12 hours for both viruses on a computing cluster. Subsequent identifiability and variability steps were faster, ranging from minutes to a few hours as they were applied to a reduced set of candidate models.
For IAV, where we had limited data, the same structural model was selected regardless of the order of the features, corresponding to a simple target-cell limited model. However, 4 of the 7 parameters were poorly identifiable, leading to multiple combinations of fixed parameter values and 15 candidate models with similar BICc (ΔBICc < 2). The variability selection step retained only one random effect. These models were combined using BICc-based weights to generate population predictions and derived virological metrics.
In contrast, SARS-CoV-2 required more complex structures including innate and adaptive immune responses, and model selection was more sensitive to the order of tested features. Three closely related models were retained, with only a few parameters requiring to be fixed. Unlike IAV, inter-individual variability was largely supported by the data.
These models were then used to compare virological metrics across viruses. IAV exhibited faster kinetics, with a time to infectious peak of 2 days post-exposure (dpe) (95% Prediction Interval PI: 0-3 dpe) compared 5 dpe (95% PI: 2-8 dpe) for SARS-CoV-2.
Conclusion: We applied a reproducible and flexible framework for automated selection of within-host viral dynamic models to IAV two respiratory viruses. By formalizing structural assumptions, identifiability decisions and variability selection, the approach reduces subjective choices and accelerates model selection. While IAV dynamics were compatible with a simple model, SARS-CoV-2 required more complex immune mechanisms.
References:
1. Asher J, Lemenuel-Diot A, Clay M, et al. Novel modelling approaches to predict the role of antivirals in reducing influenza transmission. PLoS Comput Biol 2023; 19:e1010797.
2. Iyaniwura SA, Ribeiro RM, Zitzmann C, Phan T, Ke R, Perelson AS. The kinetics of SARS-CoV-2 infection based on a human challenge study. Proceedings of the National Academy of Sciences 2024; 121:e2406303121.
3. Phan T, Zitzmann C, Chew KW, et al. Modeling the emergence of viral resistance for SARS-CoV-2 during treatment with an anti-spike monoclonal antibody. PLOS Pathogens 2024; 20:e1011680.
4. Sibieude E, Khandelwal A, Girard P, Hesthaven JS, Terranova N. Population pharmacokinetic model selection assisted by machine learning. J Pharmacokinet Pharmacodyn 2022; 49:257–270.
5. Ying Y, Duan J, Wang C, Wang Y, Huang C, Xu B. Automated Model Selection for Time-Series Anomaly Detection. 2020; Available at: http://arxiv.org/abs/2009.04395. Accessed 23 February 2026.
6. Implementation and comparison of six algorithms for automated model building in Monolix: two simulation studies and ten applications. PAGE Meeting (Population Approach Group Europe). Available at: https://www.page-meeting.org/Abstracts/implementation-and-comparison-of-six-algorithms-for-automated-model-building-in-monolix-two-simulation-studies-and-ten-applications/. Accessed 16 February 2026.
7. Killingley B, Mann AJ, Kalinova M, et al. Safety, tolerability and viral kinetics during SARS-CoV-2 human challenge in young adults. Nat Med 2022; 28:1031–1041.
8. Canini L, Carrat F. Population Modeling of Influenza A/H1N1 Virus Kinetics and Symptom Dynamics. Journal of Virology 2011; 85:2764–2770.
9. Delattre M, Lavielle M, Poursat M-A. A note on BIC in mixed-effects models. Electronic Journal of Statistics 2014; 8:456–475.
10. Kass RE, Raftery AE. Bayes Factors. Journal of the American Statistical Association 1995; 90:773–795.
11. Gonçalves A, Mentré F, Lemenuel-Diot A, Guedj J. Model Averaging in Viral Dynamic Models. AAPS J 2020; 22:48.
Reference: PAGE 34 (2026) Abstr 11995 [www.page-meeting.org/?abstract=11995]
Poster: Drug/Disease Modelling - Other Topics