Inference of person-to-person transmission of COVID-19 reveals hidden super-spreading events during the early outbreak phase

The study addresses whether super-spreading events (SSEs) occurred during the earliest phase of the COVID-19 outbreak in China. SSEs, known from prior outbreaks (SARS, MERS, Ebola), involve individuals infecting disproportionately many secondary cases relative to the mean reproduction number (R0). Early identification of SSEs is critical for outbreak control, yet traditional detection methods relying on epidemiological tracing can miss events. The authors aim to reconstruct early person-to-person transmission using viral genomes and Bayesian epidemiological models to infer the offspring distribution and assess the presence and impact of SSEs during the first two months of the outbreak.

Prior work has documented SSEs across multiple pathogens, with implications for outbreak speed and control strategies. Traditional identification approaches depend on contact tracing and can yield false negatives. Lloyd-Smith et al. proposed modeling the number of secondary infections per case (y) as a random variable from a distribution with mean R0, where SSEs correspond to the right tail and skewness/overdispersion can be quantified (e.g., via the negative binomial dispersion parameter). Early reports suggested the possibility of SSEs for COVID-19, but their role in the earliest phase remained uncertain. Phylogenetic data have been shown to enable accurate estimation of dispersion parameters compared to purely epidemiologic data.

- Study scope: Focused on the first two months of the outbreak in China (earliest sampling date plus 60 days) to balance limited genetic diversity with sufficient signal for inference. - Data: Downloaded SARS-CoV-2 genomes with high coverage from GISAID. Inclusion criteria were complete genomes with exact collection dates (day-level). Sequences flagged as error-prone were excluded. A total of 208 genomes from China during the early outbreak were analyzed. - Sequence processing and phylogenetics: Terminal regions were trimmed and 36 sites were masked (per Supplementary Data). Recombination was assessed (RPBAY) with no evidence detected. The substitution model was selected using jModelTest (HKY+I). Time-calibrated phylogenies were inferred under a molecular clock using Bayesian MCMC; maximum clade credibility (MCC) trees were summarized after discarding burn-in. Posterior tree distributions were retained to assess phylogenetic uncertainty. - Transmission inference: Combined the dated phylogeny with a Bayesian epidemiological model to reconstruct who-infected-whom transmission trees via MCMC. The generation interval (time from infection to onward transmission) was modeled with a mean around 3.5 days and SD about 3.4 days. For each posterior sample, the offspring distribution (number of secondary infections per case) was assumed negative binomial; its mean (R0), variance, and dispersion parameter (k) were computed. - Validation and sensitivity analyses: 1) Computed bidirectional probabilities of direct transmission between all case pairs; identified high-probability direct transmission pairs (>0.5). 2) Conducted cross-validation by removing one case from top-probability pairs and re-running analyses to test robustness of inferred links. 3) Assessed sensitivity to phylogenetic uncertainty by repeating parameter estimation on ten randomly selected posterior trees versus the MCC tree. MCMC runs used ~50,000 iterations with 20–30% burn-in removed. - Lineage assessment: Pangolin v1.2.8 was used to contextualize genomic diversity of sampled viruses. - Data availability: All sequences and accession numbers are in GISAID (Supplementary Data).

- Transmission tree and direct links: Reconstructed early transmission revealed considerable uncertainty but identified 18 patient pairs with bidirectional probability for direct transmission >0.5. Cross-validation by removing one case from high-probability pairs showed probabilities of alternative direct transmissions dropped toward zero, supporting robustness of these links. - Offspring distribution and reproduction number: The mean offspring number (R0) was 1.23 (95% CI: 1.09–1.39). The variance of the offspring distribution was 8.31 (95% CI: 5.06–13.39), indicating strong overdispersion. - Dispersion parameter (k): Estimated k was 0.23 (95% CI: 0.13–0.39), substantially below 1, evidencing a highly skewed distribution consistent with SSEs. Analyses using randomly selected posterior trees yielded slightly higher k than the MCC tree, suggesting phylogenetic uncertainty may overestimate k and thus underestimate SSE importance. - Additional quantitative notes: Across patient pairs, metrics related to direct transmission ranged from ~0.002 to 15.60 (mean ~8.60), consistent with many unsampled infections and missing links in the reconstructed chain. - Overall, the transmission pattern during the early outbreak was highly heterogeneous, with SSEs playing an important role.

The genomic-epidemiologic reconstruction demonstrates that COVID-19 transmission in the earliest phase in China was dominated by substantial heterogeneity, with a small fraction of infections causing many secondary cases. This aligns with patterns seen in SARS and MERS and underscores the potential for SSEs to accelerate spread and strain health systems. The findings were robust to phylogenetic uncertainty and validated direct transmission links, despite low sampling density. The estimated dispersion parameter (k ~0.23) implies that targeted measures to prevent conditions conducive to SSEs (e.g., crowding, poorly ventilated indoor spaces, high-contact settings such as hospitals and communities) could have a disproportionate impact on controlling spread. Asymptomatic and presymptomatic transmission, combined with variable incubation periods (mean ~5.2 days; 95% up to ~12.5 days), complicates traditional contact tracing and may mask SSEs in epidemiologic data alone; genome-based approaches help reveal these dynamics. The authors infer SSEs likely occurred within the first month of the outbreak, potentially before large-scale interventions limiting movement and gatherings took effect.

By integrating early SARS-CoV-2 genomic data with Bayesian epidemiological modeling, the study reconstructs person-to-person transmission and quantifies offspring distribution heterogeneity, providing strong evidence for super-spreading events during the initial COVID-19 outbreak in China. The low dispersion parameter and overdispersion indicate SSEs were important drivers of early spread. These insights support prioritizing interventions that reduce opportunities for SSEs and highlight the value of genomic surveillance for early outbreak assessment. Future research should identify environmental, behavioral, and host factors underpinning transmission heterogeneity to better target prevention strategies and refine model assumptions (e.g., incubation and generation intervals).

- Sampling constraints: Low sampling frequency and incomplete case ascertainment during the early outbreak mean many infections and transmission links were unsampled, contributing to uncertainty. - Phylogenetic resolution: Limited genetic diversity early on hindered precise phylogeny; although sensitivity analyses were performed, residual uncertainty may bias parameter estimates (e.g., k may be overestimated by some phylogenies). - Model assumptions: Estimates depend on assumptions about the generation interval and incubation period; longer assumed incubation can inflate k and underestimate SSEs. - Temporal and geographic scope: Analysis restricted to the first two months in China; findings may not generalize to later phases or other regions with different interventions and contact patterns. - Reliance on genomic data: Epidemiologic metadata were limited; without detailed contact data, some direct transmission inferences remain probabilistic.