Background

Successive Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) variants of concern (VOCs) have exhibited stepwise increases in relative epidemic growth rates [1–3]. The selective advantages of successive VOCs are the result of complex epistatic relationships arising from combinations of mutations occurring in the viral genome. Selective advantage is frequently associated with mutations in the region encoding the spike protein. To increase the growth rate of a VOC within a community, these mutations must confer improved evasion of antibody-mediated immunity, increased intrinsic transmissibility, or a combination of both. Understanding how interrelated factors including pathogen biology and host characteristics—such as preexisting immunity—impact on replication and transmission of SARS-CoV-2 is essential to inform ongoing Coronavirus Disease 2019 (COVID-19) planning and response.

Over the course of a SARS-CoV-2 infection, the viral load in the nasopharynx increases after infection until reaching a peak and then declining. The dynamics of the viral load over time are known as viral kinetics. Viral kinetics are quantified by performing reverse transcription PCR (RT-PCR) on nasal swabs, this returns a cycle threshold (Ct) value that measures the concentration of viral RNA in the nasopharynx [4]. Thus, the cycle thresholds inversely correlate with higher RNA viral concentrations. Such values measure viral RNA in the nasopharynx, which correlate well with live virus as measured by plaque-forming unit (PFU)/ml in the early stages of infection [4]. Viral kinetics will vary between individuals due to host factors such as age, sex, or prior immunity or due to changes in the pathogen related to the emergence of new VOCs [4–14].

Viral kinetics may change between variants and epidemic waves [11,14], necessitating updating of recommendations as new variants become dominant. However, as population immunity accumulates via infection and vaccination, it will become harder to generalise descriptive insights from one population to another. Studies that can identify factors influencing viral kinetics and likely infectiousness are therefore crucial for developing up-to-date public guidance and understanding risk of transmission for patients.

Changes in viral kinetics have implications for transmission since higher viral loads are frequently associated with increased host transmissibility. There are several reasons higher transmissibility could be observed after the emergence of a new VOC, including higher immune escape; faster initial replication; higher peak viral load; longer period of viral shedding; or a different symptom profile, and hence different host behaviour. As of 2023, individuals within populations have a range of life course exposure histories due to their past infections and vaccinations. These complex life course exposure histories make it challenging to determine whether changes in observed viral kinetics are due to host factors such as prior immunity or changes in the pathogen itself. Studies that can separate such factors influencing viral kinetics, and hence likely infectiousness, are therefore crucial for developing up-to-date public guidance and understanding risk of transmission for patients.

Previous studies have sought to relate within-host viral kinetics with other measurable properties such as disease severity, mortality, VOC, and the probability of onward transmission [4–6,8–10,15]. However, they have often had one of 2 key limitations. Either they have used serial sampling following an initial diagnostic swab, conducted only after symptom onset, thus missing the early rates of viral replication [13]. Or, they used single sampling with Ct values that correspond to a single time point per individual, meaning that variation in the viral load over the course of an infection is not distinguishable from the variation in viral load between individuals [16,17]. As a result, any conclusions about differences in Ct value between individuals may be an artefact of factors such as diagnostic protocols or underlying epidemic dynamics.

Repeatedly testing many individuals throughout their infection, including in the early presymptomatic period, makes it possible to separate individual Ct dynamics over time from the between-individual variation in viral load [11,18]. In cohort studies where individuals have been repeatedly tested to infer the viral dynamics of their infection, this between-individual variation has previously been compared to single characteristics, including the infecting VOC [5,11,12], age [5,7], and vaccination status [5,12]. However, there can be considerable confounding between different characteristics. For example, each sequential VOC has spread against a different background of overall population vaccine coverage, including numbers of doses received, confounding any comparisons based only on VOC. Failure to adjust for such confounding factors risks attributing differences in viral load to different VOCs that are in fact associated with factors that influence viral replication, such as the presence of vaccine-induced immune effectors such as neutralising antibody levels. Accounting for both interindividual variation in Ct values and complex host characteristics—such as vaccination status, exposure history, and age—is therefore crucial for obtaining reliable estimates of the relationship between VOCs and viral kinetics [19].

To address this complexity, we developed a Bayesian hierarchical modelling framework that estimates the unobserved Ct trajectories for each individual, using data from a subset of a prospective cohort undergoing weekly occupational health PCR screening for SARS-CoV-2. As well as accounting for different covariates jointly, this modelling approach pools data across individuals to identify the main drivers of variation in Ct dynamics and estimate the accompanying uncertainty. Using this model, we examined the relationship between Ct dynamics and different VOCs, including timing and peak viral loads, and quantified the impact of key individual-level characteristics on these dynamics.

Methods

Clinical cohort composition

The Legacy study is a prospective observational cohort study of healthy adult volunteers undergoing weekly occupational health PCR screening for SARS-CoV-2 at University College London Hospitals or at the Francis Crick Institute (NCT04750356) [20,21]. The Legacy study was approved by London Camden and Kings Cross Health Research Authority Research and Ethics committee (IRAS number 286469) and is sponsored by University College London Hospitals. Written consent was given by all participants. Participants provided baseline demographics including age, sex, and medical comorbidities, with data collection on subsequent vaccine doses and COVID-19 infections. Participants reporting either a positive test result for SARS-CoV-2 or onset of symptoms consistent with COVID-19 were sent same-day swabs via courier up to day 10 post positive test or symptom onset. They received further swabs every 2 to 3 days until day 10 (Fig 1A). All participants had a face-to-face postinfection study visit before day 30 postinfection, at which point an additional swab was collected, and a symptom diary was completed [22]. We linked diagnostic swab data with earlier swab data from the occupational screening pipeline at the Francis Crick Institute, enabling estimation of the presymptomatic period between the last reported negative screening swab and onset of SARS-CoV-2 infection.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic of the study design and modelling procedure.

(A) Schematic of the cohort and data collection procedure used for the subset of the Legacy study. (B) Typical Ct value data and model fits for 2 representative individuals with different covariates. (C) Schematic of the individual-level and covariate-level model fits. The model fits each participant’s viral kinetics and pools the estimates in a statistically robust manner. All three of these images were drawn by the authors.

https://doi.org/10.1371/journal.pbio.3002463.g001

We included all infection episodes reported by Legacy study participants between 3 June 2021 and 25 April 2022, which had at least 2 positive RT-PCR results. The total size of the cohort up to the maximum date considered was 680 participants, of which 152 had 2 positive results and, as such, were included in this analysis. Five individuals had 2 recorded reinfections, resulting in a total of 157 infection episodes with at least 2 positive PCR test results per episode. For each infection episode, whole-genome sequencing or a combination of date of infection and viral genotype was used to assign the VOC that caused the infection.

Estimating the incubation period of SARS-CoV-2 variants

In total, 136/157 (86.6%) of infection episodes had symptoms reported, reported by their date of symptom onset (Table 1). Participants were a subset of the wider post-vaccine infection cohort within the Crick, where serial sampling data were available [22]. We used the difference between the estimated time of exposure and the time of symptom onset for each individual to fit 2 parameters of an assumed log-normal form of the incubation period for SARS-CoV-2 (Fig 1B and 1C). We specified semi-informative priors for both parameters using existing estimates of the incubation period [24].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Number of participants in each covariate category used in our study.

Covariate categories used in the baseline model include VOC status, symptom status, total number of exposures at time of infection, and age group. The other covariates are investigated in one of the various alternative models considered.

https://doi.org/10.1371/journal.pbio.3002463.t001

Estimating viral kinetics by variant and host characteristics

We constructed a Bayesian model that jointly infers individual-level Ct trajectories, population (covariate)-level trajectory parameters, parameters of a log-normally distributed incubation period, and multiplicative effect size parameters following a linear regression. We regressed against several covariates for both the population-level Ct trajectory and incubation period parameters. Covariates used are described in detail in the next section. The model allows for the addition (or removal) of covariates through a formula interface.

For the individual-level Ct trajectories, we extended and modified an existing model of individual-level Ct dynamics over the course of entire infections of any respiratory disease measurable by RT-PCR test and applied this to our cohort of participants infected with SARS-CoV-2 [11,12]. The individual-level dynamics are modelled with a piecewise linear curve, on a logarithmic scale, with a single breakpoint, where the breakpoint represents the peak of the trajectory (Fig 1B). The logarithmic scale corresponds to assuming an exponential increase in viral load from the point of exposure until the peak, then an exponential decrease at a different rate to the point of clearance.

Each individual trajectory was parameterised using 4 parameters with the following epidemiological interpretations, each corresponding to events that were not directly observed: the timing of initial exposure event, timing of the peak (lowest) Ct value, Ct value at the peak, and the timing the trajectory hits the limit of detection (LOD, highest Ct value) (Fig 1B and 1C), which for this assay was Ct = 40. The reported LOD is a censored value, meaning in theory, a more sensitive machine could detect SARS-CoV-2 at higher Ct values. As such, we also estimate a true (censored) LOD for each individual, both at the start and end of each infection episode. We report differences in the peak Ct value, the timing of the peak, and the timing trajectories hit the censored LOD.

The viral kinetic model generated the expected Ct value at a given time since exposure for each individual, which was compared to the measured Ct values using a normally distributed observation process. The variance parameter of this term was shared across all individuals and represents the measurement error from the RT-PCR sampling and testing process. Given that we fit to data at 3 gene targets (ORF1ab, S, and N-gene targets), we include a term in the Bayesian linear regression adjusting for differences in gene target (Fig A in S1 Text).

Given the complex nature of the exposure history of each individual (Fig 2A), we adjusted for the time since the last exposure for each individual. We did so by calculating the time since the last infection or vaccination (whichever was sooner) for each individual as well as estimating effect sizes for the total numbers of exposures for each individual. This adjustment was performed in the Bayesian multiple linear regression component of the model, and, as such, it did not include explicit antibody dynamics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Summary of longitudinal cohort dataset analysed.

(A) Ct value distribution for all participants, stratified by the interaction between the number of vaccines and the infecting VOC, the interaction between the number of infections and the infecting VOC, and, finally, by the 4 covariate categories used throughout the study: VOC, symptom status, number of exposures, and age. (B) The individual-level Ct data, including the timing of each test, the Ct value of the result, given by the colour and the number of vaccinations, infections, and the total exposures for each participant on the left of their individual timeline. Data underlying the graphs in this figure can be found in https://zenodo.org/doi/10.5281/zenodo.10223385.

https://doi.org/10.1371/journal.pbio.3002463.g002

Results

Differences in peak Ct value by variant and host characteristics

There was considerable variation in both individual characteristics and Ct dynamics across the cohorts (Fig 2B). Applying our Bayesian model to these data, we estimated that the Ct trajectories for participants in the reference regression group (Omicron BA.1, symptomatic, 4 exposures, and aged 35 to 49 years) peaked at 15.9 (95% credible interval [CrI]: 14.8 to 16.9) (Fig 3A and S3 Table). In comparison, we estimated that the trajectories of the Delta- and BA.2-infected subsets have peak Ct values of 14.8 (95% CrI: 13.5 to 16.2) and 14.9 (95% CrI: 13.9 to 16.0), respectively (Fig 3A and S3 Table), resulting in differences of 0.93 (95% CrI: −0.41 to 2.40) and 0.89 (95% CrI: −0.17 to 2.0) Ct values from baseline individuals (respectively). We estimated that asymptomatic individuals have a peak Ct value of 16.6 (95% CrI: 15.1 to 18.2)—a peak Ct value 0.75 (95% CrI: −0.91 to 2.3) higher than baseline individuals (Fig 3B and S4 Table)—corresponding to a lower peak viral load. We find that individuals with 3 exposures in total have a peak Ct value of 17.0 (95% CrI: 15.3 to 18.8). These individuals and the baseline individuals have peak Ct values 1.2 (95% CrI: −0.81 to 3.2) and 2.3 (95% CrI: −1 to 3.7) lower (respectively) than individuals with 5+ exposures (who have a peak Ct value of 18.2 (95% CrI: 16.9 to 19.6)), corresponding to substantially higher viral load for both 3- and 4-exposure individuals (Fig 3C and S5 Table). Estimates for the 20 to 34 age group are almost identical to the estimates for baseline individuals. However, 50+ year olds have peak Ct value estimates of 17.7 (95% CrI: 16.4 to 18.9), 1.8 (95% CrI: 0.5 to 3.2) Ct values higher than baseline individuals (Fig 3D and S6 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Fitted covariate-level Ct trajectories and the differences in the parameter values stratified for each covariate.

All left panels: Model fits of the viral kinetics over the course of whole infections, beginning from the inferred exposure time. After fitting the model, we simulated 10,000 trajectories, using parameters sampled from the inferred posterior distributions for each covariate considered. We then calculate the median and 95% CrI for each covariate and plot the resulting trajectories as coloured regions, with a sample of 100 trajectories plotted transparently beneath. All right panels: The median and 95% estimates of the inferred parameter values for the peak Ct value, the timing of the peak, and the time at which the trajectory hits the LOD. Dashed line shows the value of the baseline participants inferred parameter values. (A) Trajectories and parameter values for the VOCs considered. (B) Trajectories and parameter values for the symptom statuses considered. (C) Trajectories and parameter values for the numbers of exposures considered. (D) Trajectories and parameter values for the age groups considered. Data underlying the graphs in this figure can be found in https://zenodo.org/doi/10.5281/zenodo.10223385. CrI, credible interval; Ct, cycle threshold; LOD, limit of detection; VOC, variant of concern.

https://doi.org/10.1371/journal.pbio.3002463.g003

Estimating the relationship between timing of peak shedding and symptom onset

Combining the fitted Ct trajectories with incubation period estimates, we estimated the temporal dynamics of viral shedding and their relation to the timing of symptom onset for each covariate. To compare dynamics, we selected a Ct threshold relating to a time point corresponding to comparatively high viral load (Ct = 20). We found that individuals infected with BA.2 reach the assumed Ct threshold on their upwards trajectory earlier than those infected with BA.1 and that BA.1 individuals reached the Ct threshold earlier than Delta-infected individuals. Similarly, using our incubation period estimates, we estimated that symptom onset was earlier for BA.2-infected individuals than for BA.1-infected individuals, and earlier for BA.1-infected individuals than for Delta-infected individuals (Fig 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Bivariate density plots of median incubation periods against the time at which trajectories surpassed an assumed Ct value corresponding to high viral loads, for each covariate.

All panels: We plot the posterior distributions for the estimated incubation periods, stratified by all covariates considered, against the time at which the simulated Ct trajectories, as shown in Fig 3, cross an assumed Ct value threshold for the first time (Ct value = 20). This Ct value represents a time point in the covariate-level trajectories of high viral load. (A) Parameter values for the 3 VOCs considered. (B) Parameter values corresponding to symptomatic infections. (C) Parameter values corresponding to the numbers of exposures considered. (D) Parameter values for the age groups considered. An alternative Ct value threshold (Ct = 25) was considered in an analogous analysis in the Supporting information (Fig N in S1 Text). Data underlying the graphs in this figure can be found in https://zenodo.org/doi/10.5281/zenodo.10223385.

https://doi.org/10.1371/journal.pbio.3002463.g004

Estimates stratified by VOC revealed that these 2 timescales have become shorter as variants have emerged (Fig 4A). For example, the marginal estimates for the time at which the Ct trajectories reached the assumed threshold move from 5.6 (95%: 3.8 to 7.3) days to 5.2 (95%: 4.1 to 6.2) days to 4.5 (95%: 3.3 to 5.5) for the VOCs in the order in which they evolved: Delta ➔ BA.1 ➔ BA.2. Similarly, the incubation periods are 6.0 (95%: 4.2 to 7.5) days to 5.0 (95%: 4.3 to 5.6) days to 4.5 (95%: 3.3 to 5.5) days for the same ordering (Fig 4A). As asymptomatic individuals do not have an incubation period, such estimates by symptom status were only possible for symptomatic individuals, which were used as the baseline set of individuals and therefore match those of the BA.1 estimates (Fig 4B). We did not find substantial deviations from the reference group when investigating this relationship by total numbers of antigenic exposures (Fig 4C). Lastly, when stratified by age group, we estimated the time trajectories pass the assumed threshold moves from 5.7 (95%: 4.3 to 7.0) days to 5.2 (95%: 4.1 to 6.2) days to 4.0 (95%: 2.9 to 5.0) days, and the onset of symptoms occurs at 5.8 (95%: 4.0 to 7.2) days to 5.0 (95%: 4.3 to 5.6) to 4.1 (95%: 3.0 to 5.0) days when age groups are reported by decreasing order: 50+ ➔ 35 to 49 ➔ 20 to 34 (Fig 4D). Our estimated incubation periods, sampled using both the inferred mean and standard deviation parameters, are included in the Supporting information (Fig O in S1 Text).

Discussion

We estimated differences in viral shedding dynamics over the course of infections with different SARS-CoV-2 VOCs by combining data from a large cohort of healthy adults with a statistical model that accounted for both individual- and group-level variation in Ct value. Adjusting for a changing background of host characteristics over time, we identified differences in Ct kinetics that were associated specifically with VOC status and found that Delta and BA.2 infections tended to peak at higher viral loads compared to BA.1 infections. We also found that age and number of prior exposures had a strong influence on Ct peak, with lower shedding among older individuals and those who had at least 5 exposures to vaccination and/or infection. Moreover, we found evidence of a correlation between the speed of early shedding and duration of incubation period when comparing different VOCs and age groups.

The unique dataset and model developed here allow us to expand upon previous findings from recent studies that investigated specific aspects of viral kinetics of SARS-CoV-2 infections, both in a human challenge model [4] and in the community [5,7,11–14,19,25]. First, we collected serial swabs from participants in a large, well-defined cohort, which is broadly representative of the working-age UK population and who were undergoing regular surveillance RT-PCR testing. Second, we captured both asymptomatic and symptomatic infections with high confidence, linked to the last negative test, in comparison with other studies [25] that relied upon self-initiated symptom-based testing. This knowledge of prior negative test timings allowed us to estimate the duration of presymptomatic infection with greater confidence and directly compare the peak and trajectory of symptomatic and asymptomatic infections. One of the major limitations of most studies of SARS-CoV-2 infection has been the unknown time between infection and symptom onset. While the human challenge model approach has provided detailed information about viral kinetics following inoculation, participants were unvaccinated, healthy younger adults infected with a small dose of ancestral “pre-VOC” virus [4], limiting inferences about the kinetics of emergent VOCs within the majority-vaccinated general public. Using the last known negative test in our participants, through to their first positive test, and then through serial tests up to 30 days afterwards, we were able to fully characterise viral kinetics over the course of infection rather than focusing exclusively on the initial post-onset period [11,12]. Furthermore, by linking to dates of symptom onset and conditioning on each participant’s estimated exposure time, we jointly—alongside each individual’s viral kinetics—estimated an incubation period for each covariate. Finally, we included participants with Delta, Omicron BA.1, and Omicron BA.2 infections, enabling direct comparison between VOCs that have circulated since May 2021 in the United Kingdom [15].

Our analysis made it possible to evaluate several putative relationships between host characteristics, disease profile, infecting variant, and viral kinetics. First, it has previously been suggested that both vaccination and milder or asymptomatic infection may be associated with lower viral loads and faster viral clearance [12,26]. Our finding of higher Ct values after 5+ viral exposures (whether vaccine or infection) is consistent with enhanced host immune responses due to previous exposure to spike protein. Among asymptomatic infections, there was some—albeit weak—evidence pointing to lower and earlier peak viral load. We did not find evidence that people with asymptomatic infections cleared the virus more quickly, with predicted Ct remaining below 30 for 14 days after exposure. These findings are consistent with those in the human challenge model, where viral loads remained high regardless of symptom severity [4]. Within our cohort, effects of vaccine doses or symptom status on the value or timing of the peak Ct value were small, suggesting that the impact of isolation periods or likelihood of transmission would not differ based on presence of symptoms or number of previous exposures. However, we did not have data from participants with no prior exposures, limiting our ability to fully explore the role of heterogeneity in immunity.

Second, while a report on a large community-based study reported that Omicron BA.1 caused shorter symptomatic illness compared to the Delta variant [27], our estimated viral load kinetics did not indicate faster clearing of Omicron BA.1. This is consistent with our previously reported data on symptom severity and duration, which found these to be the same across all VOC infections considered here [22]. It has been suggested that different VOCs have altered tropism for higher or lower portions of the respiratory tract [4,28], with BA.1 preferring upper portions and Delta those deeper. These differences in anatomical preference may alter Ct dynamics in the nasopharynx, and, hence, the symptom constellation may also be different with coryzal symptoms favoured in BA.1 and BA.2 infections compared to Delta [22,29]. We found a clear difference in time to LOD for Delta infections compared to Omicron BA.1 infections (in 35 to 49 yeas old with 4 previous exposures), but the majority of people had recently received their third vaccine dose at the time of infection and were more likely to be asymptomatic, suggesting a role of faster viral clearance for the nasopharynx in the immediate post-vaccine period [4,30]. We also found evidence of an earlier peak Ct and higher peak viral load in BA.2 infections, consistent with a recent study of serial viral loads by [27,31] and the observed rapid spread of BA.2 across the globe during the latter stages of the BA.1 epidemic [32].

Third, the arrival of each new VOC in the UK typically resulted in a wave of infections concentrated in specific age groups, while at the same time vaccine doses have been prioritised towards older age groups [15,33]. We found the peak Ct was lower in younger adults across all VOC, with older adults peaking later, and having a shorter trajectory to the LOD, despite similar vaccination doses at the time of infection. Indeed, the older adults, who received vaccine doses significantly earlier than the younger cohort, had less viral shedding. While not formally included in the model, older adults typically have an attenuated neutralising antibody response after COVID-19 vaccination [20,21,34]. One might expect in this scenario that viral replication would therefore be more active and prolonged. Further study to understand why we observed the opposite is needed.

There are some additional limitations in our analysis. Error in Ct values is known to decrease at higher viral loads and increase at lower viral loads [11] (Fig B in S1 Text), but to ensure model identifiability, we assumed a constant measurement error across the range of reported Ct values. Such an approximation would be particularly important to revisit if we were analysing timescales beyond the relatively short postinfection period considered here and, hence, a larger range of Ct values. We also used prior exposures as a proxy to capture the potential effects of existing immunity. With more detailed linked serological and PCR data, it would be possible to examine this relationship in more detail, although this would require a substantially more complex model structure. Our finding of higher Ct values after 5+ viral exposures (whether vaccine or infection) is consistent with enhanced host immune responses due to previous exposure to spike protein. However, the nonincreasing relationship in Ct estimated between 3 and 4 exposures indicates that immunological dynamics, particularly following initial vaccinations and infections at different points in time, may be more complex than can be fully captured in a single exposure total.

To test the relative predictive performance of several model candidates, we performed a leave-one-out (LOO) model selection procedure for 8 candidate models (see S7 Table for the candidate models considered). The candidate models have either different design matrices, specifying which set of covariates are included in each model, or different priors to the model used in the main analysis. The predictive power of all models included in the LOO analysis with different design matrices are within 4 standard errors of one another, which is broadly the rule of thumb for a model with significantly lower predictive power than another. Therefore, we can conclude that including covariates known to influence viral kinetics is justified, given that the models have similar predictive power with or without their inclusion. The model with noninformative priors, which has the lowest predictive power compared to all other candidates—relative to the model with semi-informative priors used in the main analysis (S7 Table)—is still within the range where it is difficult to conclude an overall difference in predictive power. As such, we also performed a complementary sensitivity analysis in which we plot the effect sizes of the model with noninformative priors (Fig T in S1 Text). The central effect size estimates are broadly similar in both the semi- and noninformative models (Fig 3 and Fig T in S1 Text). However, the resulting uncertainty is higher in the noninformative prior case. Combining the interpretation from the LOO and sensitivity analysis around the influence of semi-informative priors, we conclude that their inclusion in the results reported in the main analysis is justified, given that the predictive power is higher with them, the central estimates are broadly similar, and prior information is only used where reliably reported elsewhere in published literature.

Previous studies inferring viral kinetics have concentrated mostly on differences between VOCs circulating at the time in question. As such, they did not consider the host of other possible confounding variables, making their results difficult to attribute solely to differences in infecting VOC [5,11,30,35]. In interpreting effect sizes between viral kinetic estimates, we controlled for the complex emerging picture around individual life-course exposures to SARS-CoV-2, both as infections and vaccines. Differences in numbers of infections, the VOC causing the infection, the numbers of vaccinations, and many other individual-level variables introduce many sources of possible confounding. In particular, the timing of the sampling within our study differs compared to other studies, and, hence, Ct dynamics by variant could vary substantially across populations if these factors are not adjusted for.

Moreover, we find less difference in viral kinetics by age in comparison to other studies [5,7]. Again, this is likely to be influenced by a number of factors. Given the complex relationship between age, symptom severity, immunity, and exposure risk, studies that do not account for these characteristics could attribute differences in shedding to age, rather than another factor that correlates with age. However, combining the underlying viral kinetics model with a Bayesian hierarchical framework allowed such information to be jointly incorporated into our analysis. Additional factors that could influence observed differences in effect size estimates include calibration differences between PCR pipelines resulting in differences in the absolute Ct values; differences in sampling procedure with previous studies primarily relying on symptom-triggered sampling; complex exposure histories following both the vaccine rollout and the emergence of new VOCs; and the ability of the underlying model structure to account for the numerous covariates suspected to affect viral kinetics or the inherent individual-level heterogeneity in response to exposure to SARS-CoV-2. We combine data from a study designed to capture viral load across entire infections, including sampling before and after exposure, with a model structure complex enough to adjust for all necessary covariates and individual-level variation in viral kinetics and able to inform model parameters by jointly fitting to data where possible. Therefore, the differences between our effect sizes and previously published studies may be explained by the combined adjustment for all of these complexities.

The flexibility provided by our framework provides a methodological foundation for future work using similar datasets. Given the ongoing antigenic turnover of SARS-CoV-2 variants of concern since late 2020, our results show that it will be increasingly important to adjust for a range of changing host factors when quantifying viral kinetics and considering implications for potential infectiousness in future.

Supporting information

Acknowledgments

The authors would like to thank all the study participants, the staff of the NIHR Clinical Research Facility at UCLH including Miguel Alvarez and Marivic Ricamara. We would like to thank Bobbi Clayton, Gita Mistry, Simon Caiden, and the staff of the Scientific Technology Platforms (STPs) and COVID-19 testing pipeline at the Francis Crick Institute. We thank Prof. Wendy Barclay of Imperial College and the wider Genotype to Phenotype consortium for the Alpha and Delta strains used in this study, and Max Whiteley and Thushan I de Silva at The University of Sheffield and Sheffield Teaching Hospitals NHS Foundation Trust for providing source material. We thank Prof. Gavin Screaton of the University of Oxford for the Omicron BA.1 strain used in this study.