(T-075) Mechanistic Modeling and Unsupervised Learning Reveal Drivers of Inter-Individual Variability in COVID-19 Vaccine-Induced Immunity

Carmine Schiavone – University of Naples Federico II; Joseph Cave – Weill Cornell Medicine; Zhihui Wang – Houston Methodist Research Institute; Vittorio Cristini – Houston Methodist Research Institute; Sergio Caserta – University of Naples Federico II; Prashant Dogra – Houston Methodist Research Institute

Objectives: To characterize inter-individual variability in humoral immune response following COVID-19 mRNA vaccination by integrating mechanistic modeling, unsupervised clustering, and machine learning–guided parameter inference to enable individualized immune response predictions.

Methods: An ODE-based mechanistic model was developed to simulate antigen kinetics and downstream dynamics of CD4+ T follicular helper cells, germinal center B cells, plasma cells, memory B cells, and IgG antibodies. The model was first fitted to population-level clinical data to estimate global parameters and rank parameter sensitivity with respect to IgG response using local and global sensitivity analysis. The top four influential parameters were then estimated for each of 3,750 subjects using longitudinal IgG titer data collected over 360 days. Fitted immune trajectories were clustered using growth mixture modeling, revealing response subgroups. We then trained a regression-based machine learning model to predict the four mechanistic parameters based on demographics and early immune response features. The predicted parameters were used in the original mechanistic model to simulate long-term IgG trajectories in new individuals, enabling prospective immune prediction.

Results: The mechanistic model achieved excellent fit to individual IgG data (Pearson R > 0.9). Growth mixture modeling identified three immune trajectory subtypes—good, intermediate, and poor responders—with distinct parameter profiles. Cluster membership correlated with subject-level features such as age, sex, and comorbidities. The ML model accurately predicted key mechanistic parameters from early timepoint data and demographics, allowing for forward simulation of personalized long-term immune responses.

Conclusions: This hybrid framework leverages machine learning to personalize mechanistic models, enabling simulation-driven prediction of vaccine-induced immunity. By integrating early immunological signals with underlying subject characteristics, our approach bridges data-driven inference and systems pharmacology, offering a scalable tool for precision vaccination strategies and immunological risk stratification.

Citations: [1] Klein SL, Flanagan KL. Sex differences in immune responses. Nature Reviews Immunology. 2016;16(10):626-638. doi:10.1038/nri.2016.90
[2] Lynn DJ, Benson SC, Lynn MA, Pulendran B. Modulation of immune responses to vaccination by the microbiota: implications and potential mechanisms. Nature Reviews Immunology. 2022;22(1):33-46. doi:10.1038/s41577-021-00554-7
[3] Dogra P, Schiavone C, Wang Z, et al. A modeling-based approach to optimize COVID-19 vaccine dosing schedules for improved protection. JCI Insight. 2023. doi:10.1172/jci.insight.169860

Keywords: Quantitative systems pharmacology, machine learning, Unsupervised learning, Vaccine-induced immunity