Background/Aim: Protective immunity after vaccination is still commonly inferred from single antibody titers, despite protection emerging from coupled innate activation, T follicular helper (Tfh) support, germinal center dynamics, immune memory, and waning. We aimed to develop a personalized immunodynamics digital twin that quantifies protection as a mechanistic latent state and predicts individual-specific durability of vaccine-induced immunity and optimal booster timing.
Methods: We harmonized longitudinal human influenza vaccination cohorts from ImmPort, using SLVP015-associated studies (SDY312, SDY314, SDY311, SDY112, and SDY315) as the primary modeling backbone, with SDY478 incorporated as an auxiliary aging-focused cohort where overlapping immune variables and sampling windows permitted harmonized calibration. Repeated measurements included neutralizing and HAI titers, serum cytokines, and immune cell subsets across adult age groups. A mechanistic delay ordinary differential equation (ODE) system was constructed to model innate priming, Tfh expansion, germinal center B-cell dynamics, plasma cell output, immune memory formation, and antibody waning. Individual-level parameters were inferred using hierarchical Bayesian calibration with partial pooling, incorporating age, sex, baseline inflammatory tone, and antigenic distance as covariates. Protective immunity was defined as a posterior probability that a composite latent protection state, integrating effective neutralization capacity, memory reserve, and T-cell support, exceeded an empirically learned threshold. Booster timing was derived using stochastic optimal control to maximize cumulative protection reserve while constraining excessive innate activation.
Results: Across held-out influenza seasons, the digital twin predicted high serological response at day 28 with an AUROC of 0.83 (95% CI 0.80–0.86), outperforming titer-based benchmarks (AUROC 0.71; 95% CI 0.67–0.74; p<0.001). Posterior predictive coverage for antibody and cellular trajectories reached 92%, indicating robust calibration. Individuals in the highest quartile of baseline inflammation showed lower inferred germinal center efficiency (−18%; 95% CrI −24% to −11%) and shorter protection half-life (median 94 vs 141 days), with an increased hazard of falling below the protective threshold (HR 1.62; 95% CI 1.34–1.95; p<0.001). Personalized booster scheduling improved season-long protection-reserve area under the curve by 21% (95% CI 16–27%) while reducing booster frequency by 0.37 doses per person-year compared with fixed-interval strategies (p<0.001). Sensitivity analyses showed stable parameter identifiability across cohorts, with posterior CVs <20%, and protection-half-life estimates remaining within ±9 days under variable missingness and alternative waning prior specifications.
Conclusions: This personalized immunodynamics digital twin defines protective immunity as an integrated mechanistic state. By estimating individual protection half-life and uncertainty-aware booster timing, the model enables precise, risk-stratified vaccination strategies.