Early Prediction of SARS-CoV-2 Variant Dynamics: Modelling Virus Interactions with Antibodies and ACE2 Receptor

The continuous emergence of new SARS-CoV-2 variants presents significant challenges to public health efforts. In response to epidemic outbreaks, one critical challenge is to predict early which variants have a high transmission potential before they spread widely. Early identification allows for timely public health measures, including vaccine development. However, the capacity for early prediction in the current methods is limited. The most widely used predictive approaches, such as multinomial logistic regression (MLR) or mechanistic models (compartmental models for infectious disease transmission), often ignore population immunity or actual viral traits. Current predictive models cannot capture the trade-off between ACE2 binding and immune escape that has been observed.Recent advances in deep mutational scanning have provided valuable data on two important virus traits: immune escape and the binding affinity to the ACE2 receptor. Biophysical modelling can map these two phenotypes to viral fitness in growth. Hence, we aim to develop an early prediction framework for SARS-CoV-2 variant dynamics by embedding a biophysical model based on mutational effects in the spike protein, derived from deep mutational scanning data, alongside population immunity.In this proposal, we first demonstrate how to formulate viral fitness depending on two infection stages: the initial stage after susceptible individuals are exposed to the virus, and the infectious period when infected individuals can transmit the disease. This shows a trade-off pattern between ACE2 binding and immune escape. We then construct a cross-immunity function using these traits and population immunity under biophysical principles in a transmission model. The ACE2 binding of each variant will be estimated using a computational tool accounting for measurement errors and epistasis. We train the model using Bayesian Markov Chain Monte Carlo and perform model comparison. The complexity of the candidate models will be systematically increased along two dimensions: biophysical constraints and human heterogeneity. Model validation will be conducted on unseen (out-of-time) data to assess predictive performance.Finally, we refine the proposed model by incorporating unknown mutational effects outside the receptor-binding domain of the spike protein. Sensitivity analysis will be performed to assess the influence of uncertainties at different levels, from viral trait measurements to heterogeneity in the host population, and to quantify model robustness using failure-rate metrics. Both parameter and structural uncertainties are addressed. The robust early predictive tool, mechanistically guided by biophysical principles, will help ensure timely vaccine production and establish a quantitative framework for understanding viral evolutionary dynamics and the role of host heterogeneity.