Inferring pathogen superspreading potential using early spatial spread patterns

Superspreading driven by individual variation in transmissibility shapes novel pathogen emergence and the effectiveness of control measures. Current approaches to estimating superspreading typically rely on cluster size distributions informed by contact tracing or genomic data, limited by data availability and potential sampling bias. Here, we examined the impact of superspreading on the early spatial spread of novel pathogens using a branching process model incorporating inter-county mobility in the United States (US). To represent individual transmission heterogeneity, we modeled the number of secondary infections using a negative binomial distribution with a mean reproduction number $R_0$ and a dispersion parameter $r$ that quantifies superspreading potential. Simulations suggest that stronger superspreading tends to slow early spatial invasion, with a larger variation in epidemic growth between different realizations. Using a graph neural network designed for epidemic inference, we demonstrated that $r$ can be reliably inferred from early spatial spread patterns, robust to the spatiotemporal variation of $R_0$ and case underreporting. Application to early COVID-19 data in the US revealed strong superspreading prior to nationwide lockdown ( $r=0.50$ , 95% CI: [0.23 - 1.20]), followed by weaker superspreading afterward ( $r=1.3$ , [0.64 - 3.18]). Our study offers a new approach to quantifying pathogen superspreading potential using population-level observations.