Data collection and sequencing

We collected all data in the 2016/17 influenza season as described in [17]. Sequencing was performed as described in [19]. Raw Illumina reads were trimmed with Trimmomatic 0.36 [20]. Alignment of paired-end reads was done by using bowtie 2.2.3 [21], using strain A/New York/18/2014 as a reference. The aligned reads were sorted by using samtools 1.2 [22]. Variants were called and filtered by using lofreq 2.1.2 [23]. Variant calling was done for sites with a coverage of at least 100. Sites with a coverage of less than 100 were assumed to be unknown and were denoted as N, that is any possible nucleotide (Details on the exclusion of sequences are described in S1 Text). Exact input specification can be found at https://github.com/nicfel/FluBaselPhylo/tree/master/Sequences. The consensus sequences from this study were deposited in GenBank (numbers MN299375-MN304713).

Timed phylogenetic tree based on the HA segment

We combined the Basel sequences with all sequences (as of the 17th of Juli 2018) from https://www.gisaid.org sampled between January 1st 2016 and December 31st 2017 for which at least the segments HA, NA and MP were available. We first aligned all consensus sequences using muscle v3.8.3129. We then built an initial phylogeny from the HA segment alone by using RaxML 8.2.1 [24] and obtained branch lengths in calendar time via timetree [25] using the nextstrain pipeline [26]. The inferred time tree is available here https://github.com/nicfel/FluBaselPhylo/tree/master/TimeTree. The tree was plotted using ggtree [27].

Initial clustering based on nucleotide differences

We then calculated the average nucleotide difference between any of the sequences and sequences from Basel. In order to split the dataset into manageable pieces, we first grouped any two sequences from Basel together if they were within an average nucleotide difference of 0.0025 per position. If the full genome for two sequences was available, this would correspond to about 32 different positions on the full genome. For an average clock rate of 2.9*10^-3 per site and year, this would correspond to a pairwise phylogenetic distance of just below 1 year. Sets of sequences from Basel are only split into two groups if the two closest related sequences of each group exceeds this distance. Based on this initial grouping, we added sequences that were not from Basel to each cluster if they were at maximum 0.0025/2 mutations per position away from any of the sequences from Basel. The factor of 2 is only to reduce the number of non-Basel sequences in each of these initial clusters to be computationally tractable.

Phylogenetic trees of initial clusters

We next estimated rates of evolution for each genomic segment using the SRD06 model [28] and a strict clock model from 200 full genome influenza A/H3N2 sequences sampled in California, New York and Europe between 2010 and 2015 in Beast 2.5 [29]. These sequences were downloaded from fludb.org, were not used otherwise and are an independent dataset. We allowed each segment to have its own phylogeny in order to avoid reassortment to bias the estimates of evolutionary rates. Each of the segments, as well as the first two and third codon position was allowed to have its own rate scaler. We ran 10 independent MCMC chains each for 10⁸ iterations and then combined them after a burn-in of 10%. These estimated evolutionary rates are long-term rates of influenza A/H3N2. Since the effects of selection over short time periods are smaller compared to longer time periods. The evolutionary rates can be expected to be faster for shorter time windows [30]. We therefore expect the pairwise distances estimated for our data from the 2016/17 outbreak using these rates to be an overestimate of the actual divergence times. The xml and log files for the analysis can be found here https://github.com/nicfel/FluBaselPhylo/tree/master/EvolutionaryRates.

We next reconstructed the phylogenetic trees of all initial clusters by using the full genomes of all samples in the initial clusters. We fixed the evolutionary rates to be equal to the mean evolutionary rates as estimated previously, with the mean evolutionary rate being 2.9 * 10⁻³ per site and year, as well as fixed the rates of the SRD06 site model to the rates estimates using the influenza A/H3N2 datasets samples over many years. As a population prior, we used a constant coalescent model with an effective population size being shared among all initial clusters. We then estimated a distribution of phylogenies for each initial cluster, assuming that all segments share the same phylogeny. If reassortment happened, it would increase the distance between samples. Due to using fixed evolutionary rates as estimated in the previous analysis, reassortment will not bias evolutionary rates. Hence, reassortment events will increase the pairwise distance between isolates separated by reassortment, but will not bias the distance between isolates that are not.

Local cluster identification

To identify sets of sequences from Basel that were likely transmitted locally, we used the phylogenetic tree distributions for each initial cluster and reconstructed the ancestral states using parsimony. We made some modifications to the standard algorithm for ancestral state reconstruction. To reflect our prior belief that Basel is unlikely to act as a relevant source of influenza on a global scale, we classified internal nodes that are not exclusively classified to be in Basel as not in Basel. Since the flu season is only a few months long, we additionally assumed that lineages are unlikely to persist in Basel for more than 0.1 years without being sampled. To reflect that assumption, we classified internal nodes that are more than 0.1 years from a sample from Basel to be either in a location other than Basel, or to be in an unknown location. We then defined sequences to be in the same local cluster if all their ancestors are inferred to be in Basel. We get these local clusters for each iteration of the MCMC. From the grouping of sequences into local clusters as described above, sequences can be classified into different local clusters over the course of the MCMC. For the estimation of effective reproduction rates we however require each sequence to be in a distinct local clusters. To do so, we randomly picked an iteration of the MCMC and then chose the local clusters present in that iteration. In order to account for uncertainty in the local transmission cluster assignment, we repeated each analysis 10 times with randomly chosen iterations. The exact workflow, including BEAST2 input files can be found at https://github.com/nicfel/FluBaselPhylo/tree/master/LocalClusters. While alternative model based approaches exist to reconstruct locations of internal nodes (e.g. [31]), these approaches themselves make strict assumptions that are violated when studying the spread of diseases on a city scale. Also, it is unknown how well they perform when migration between individual locations is very strong.

Estimation of the effective reproduction number and sampling probability

We then estimated the effective reproduction number through time as well as the sampling proportions and phylogenies from all these local clusters jointly using BDSKY [32]. We assumed the effective reproduction number to be piecewise constant in intervals of 2 days and allowed it to change every 2 days. We then assumed the difference between the log effective reproduction number in interval t (log R_eff(t)) and in interval t-1 (log R_eff(t-1)) to be distributed around N(0,σ), with σ being estimated in the MCMC [33]. Additionally, we assume the log R_eff at the most recent time interval and the one at the very last time interval to be normally distributed in log space around N(-0.6931,0.1). This means that we assume the R_eff to change in a continuous way, which can lead to an underestimation of differences in R_eff, if the R_eff changes abruptly. This adapted version of BDSKY is available on https://github.com/nicfel/bdsky.

We assumed the rate at which an infected individual transitions to being non-infectious to be 0.25 per day. The birth-death model assumes the number of samples over time to be informative about the population dynamics, meaning that the results can be biased if the sampling proportion of individuals would change over time. Since we, however, followed the same procedure for inclusion of patients throughout the epidemic season, our assumption of sampling over time should hold. The BDSKY model additionally conditions on survival [32], meaning that it computes the probability of observing a phylogenetic tree conditional on observing at least one lineage and assumes the host population to be unstructured. We further assume, as is standard, that there is no transmission rate variability between individuals, such as would, for example, be caused by having super spreaders. Additionally, we do not model the process of how lineages are introduced into Basel and how this might change over time.

The weather data used for the correlation analysis was obtained from www.meteoblue.com. This data is based on measurements of weather stations which are then used in simulations to estimate local weather variables (see https://content.meteoblue.com/nl/specifications/data-sources).

Defining connectedness between individuals

We define two individuals to be connected if their pairwise phylogenetic distance is less than 0.1 years. If we assume two individuals to be isolated at the same time, this cutoff would correspond to a common ancestor that was at most 18.25 days ago. Considering that the evolutionary rates we used to perform these inferences are long term rates and therefore lower than the actual short term rates [30], we expect that the cutoff values are effectively lower in reality. This means that if we use a cutoff of 0.1 years, even individuals that are at an inferred pairwise distance of 0.1 years are very likely more closely related than that. To avoid biases originating from these cutoff values, we repeated all analyses that are based on cutoffs with of 0.05, 0.15, 0.2 and 0.3 years as the cutoff value.

Connectedness across age and family status groups

We estimate the average number of connections members from each of the six categorial age/family status groups have, according to the above definition of connectedness. To do so, we model the number of connections an individual from a group has as a negative binomial distribution. This allows us to model the number of connections an individual from a group has, while taking the variance of the relationship between the group label and the number of connections into account. This is in contrast to, for example, the poisson or geometric models. We assess overall model fit with an ANOVA and then perform Tukey contrasts, comparing all pairs of age groupings [34]. We correct for multiple testing by using Schaffer’s method, which is similarly conservative to bonferroni, but takes into account the dependencies enforced in a linear modelling framework [35].

Age mixing patterns

To identify mixing patterns between the six categorial age/family status groups, we again use the definition of connection of two patients.

We use two different approaches to estimate how different groups are connected to each other. First, we use multinomial logistic regression to estimate the probability that a member from one group is connected to a member from another group. As weights, we use the inverse number of samples from each group. This implicitly assumes that individuals from each group have the same probability of being infected. Children however might have higher rates of infection, and we therefore expect this weighting to underestimate the role of children and to overestimate the role of adults.

Second, we use a permutation approach. Between any two groups a and b, we compute the probability of them being associated with one another as follows: For each combination of groups a and b, we count the number of pairs that are associated with one another. We then randomly permute the age labels 10⁶ times. For each permutation, we calculate if the number of pairs between these groups is greater or smaller than what we observed. From these values, we then compute the probability that age groups a and b are positively () or negatively () associated with one another as: With the number of pairs we observe in the data and the number of pairs we observe after we permute the group to patient labels. Because this test does not have an underlying model for how many connections there are between individual groups, we here use a Bonferroni correction instead of the Schaffer’s method to correct for multiple testing. In order to test the sensitivity of these estimates, we repeated this analysis using cut-off values of 0.05, 0.1, 0.15, 0.2 and 0.3 years.

Introduction of new lineages into the city drive the local epidemic

We first assessed how the 663 sampled cases in Basel compare to 11,000 sequences sampled from around the world between January 2016 and December 2017, by inferring a phylogenetic tree using the hemagglutinin (HA) sequences. The Basel sequences span the existing global diversity (see Fig 1a), suggesting strong exchange (most likely importation into Basel) of viruses with other areas around the globe. We, however, did not find isolates in Basel that were part of the same clade as the vaccine strain in that season (i.e. 3c2), which is consistent with very few cases of this clade in that season (see Fig 1A).

Download:

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

Fig 1. The local spread of influenza H3N2 in Basel in the 2016/2017 season.

A Time tree of HA segment reconstructed from sequences from this study and sequences from around the world. The HA sequences sampled in this study (red) are dispersed across almost all clades present between 2016–2018. This indicates that the diversity of samples in Basel is similar to the diversity of samples around the globe. B Number of sequenced weekly cases for the three most abundant HA clades. The number of cases for the corresponding clade is shown in color and the overall number of cases is shown in grey. C Number of introductions depending on how many random sequences from Basel are used. D Estimated proportion of sampled individuals averaged over ten different BDSKY runs using different classifications of local sequences into local clusters. Estimates for individual classifications are shown in S1 Fig. E Estimates of tree heights of local clusters, which can be used as a lower bound to how long lineages persisted in the city. F Estimates of the effective reproduction number through time inferred from all local cluster jointly by using BDSKY. The black line is the mean estimate and the red area denotes the 95% interval. The grey bars denote the number of cases per week that were sequenced and used in the BDSKY analysis. G Comparison of the inferred effective reproduction number with temperature and relative humidity and if a day is a school day or not. The effective reproduction number curves are averaged over the 10 different classifications of sequences into local clusters. Comparisons for the individual random classifications are shown in S2 Fig.

https://doi.org/10.1371/journal.ppat.1008984.g001

The number of sampled sequences in Basel peaked at the end of 2016, with a smaller peak at the beginning of February of 2017 (see Fig 1b). The sequenced cases were dominated by the 3c2 sub-clades A1, A1a and A3, and we observe that a peak in A1a cases mainly contributed to the peak in February of 2017.

Basel sequences cluster into local transmission clusters within the global diversity (see Materials and Methods). We obtained around 240 local clusters (see Fig 1c and https://nextstrain.org/community/jameshadfield/basel-flu/1), suggesting that the sampled sequences were the result of around 240 influenza introductions from areas outside of Basel. In order to investigate if this number is a strict lower bound for the number of introductions, we use random subsets of the 663 sequences to re-estimate the number of introductions. We find that the number of estimated introductions grows approximately linearly with the number of sequences in a subset (see Fig 1c). This suggests that with additional sampling effort, we would have captured more introductions. With its own international airport, the airport of Zürich nearby, and a major rail hub, Basel is well connected to the rest of Europe and the world. As such, people working in Basel often do not live in the city and commute daily from elsewhere in Switzerland, Germany and France. Basel is a tourist destination and often hosts international conferences, attracting people from all over the world. This connectedness likely drives these introductions of influenza into the city.

Quantification of the overall local epidemic following the introduction into the city

After introduction into the city, we next study how influenza is transmitted locally. To get an estimate of how long lineages persist locally, we additionally estimated the tree heights of local transmission clusters that had at least 2 sampled sequences (see Fig 1e). We estimated that the average tree height of a local outbreak clusters with at least 2 sampled sequences was around 30 days. These tree heights provide an estimate of the lower bound to how long lineages spread locally on average and suggest that the average local transmission cluster persists for at least 1 month.

In order to quantify the amount of local transmission, we estimated the effective reproduction number (R_eff) to be between 1 and 1.5 for most of the season, which agrees with previous estimates of the effective reproduction number for seasonal influenza [36]. The R_eff peaks in December and in February (see Fig 1f) with the 95% credible interval excluding and R_eff of 1 in December. In January, we inferred a drop in the effective reproduction number with the estimated median being below 1 (see Fig 1f).

The trend in the overall number of cases is similar to the trends in other places in Switzerland (http://meldesysteme.bagapps.ch/sentinella/publikationen/2017%20Saisonbericht%20Grippe%202016_2017_d.pdf), where doctoral consultations for influenza like illnesses also peaked in early January. Further, these estimates are comparable to the overall trend of influenza cases during the 2016/2017 season in Europe (https://ecdc.europa.eu/en/publications-data/summary-influenza-2016-2017-season-europe).

We next investigated potential factors determining the changes in R_eff. The number of influenza cases over the years show a strong seasonality, with the majority of cases occurring in the winter months in both the northern and southern hemisphere [7]. Relative humidity and temperature have been described to drive influenza transmission [37]. Additionally, the effect of school closures on the spread of pandemic influenza has been discussed [38, 39]. Thus we investigated potential correlations of R_eff with temperature, relative humidity and school days (i.e. days when children go to school). As we only studied one season, these correlations have to be interpreted with caution, and analyses of other seasons are needed to confirm the potential correlations. Neither humidity nor school days showed a significant correlation: relative humidity stayed fairly constant over the season, and both low and high R_eff are found during times when schools were open (Fig 1g and S4, S5 and S6 Figs). To account for autocorrelation, we performed the correlation analysis, averaging the temperature, relative humidity and mean number of school days over 4, 6, 8 and 10 days instead of just 2 days. We find the mean temperature to be significantly correlated with the mean R_eff in both scenarios (see S3 and S4 Figs).

Viral shedding of viruses has been shown previously to be increased at lower temperatures in animal models [37] and higher absolute humidity has been shown to favor transmission on a population level [40]. We here observe lower effective reproduction numbers at lower temperatures. The correlation of the effective reproduction number with the temperature however is not necessarily causal, as it, for example, could be due to social behavior being different at lower temperatures. Also, the computed p-values could be inflated due to unaccounted autocorrelation, artificially inflating the number of independent data points that are actually in the datasets.

Along with the R_eff, we co-estimated the sampling probability, that is the probability of an infected individual being sampled. Since we followed the same procedure for inclusion of patients throughout the epidemic season [17], we assumed that this probability is constant throughout the influenza season. We estimated the sampling probability to be between 3% and 5% (see Fig 1d and S2 Fig). In contrast to the R_eff estimates, this value is more sensitive to the procedure of clustering of sequences into sets of locally transmitted sequences (see S2 Fig). Additionally, the prior probability on the effective reproduction number, as well as the assumed becoming un-infectious rate can influence this estimate [41]. With 663 samples from different patients included in this analysis, this would suggest that between 13260 and 22100 people in Basel or between 8% and 13% of its population of about 171000 were infected with influenza H3N2 during the 2016-17 season. The city limits of Basel are however in reality not fixed and the metropolitan area around the city is substantially larger than the city itself. Furthermore we have sampled patients who went to a doctor or hospital in the city of Basel, but live in the surrounding areas or other parts of the world. We therefore expect the estimate of between 8% and 13% to be an estimate for the upper bound of the number of infected people, rather than the true percentage of infected individuals. These estimates of the overall attack rate of seasonal influenza in Basel are broadly consistent (though on the lower end) with estimates of the attack rate derived from un-vaccinated individuals [42], which range from approximately 20% in children to 10% in adults and assuming a vaccination rate of 12% in the 2016/2017 season in Switzerland [43].

Overall, our analyses suggest that transmission occurred with an effective reproduction number varying between 1-1.5 throughout the season, overall infecting at most 8–13% of the population.

Importance of age groups and family status in the local epidemic spread

After having determined the importance of introductions of influenza into the city and the overall rate at which influenza is spread in the city, we next studied the effect that age and family status has on the overall spread. Fig 2a and 2b and show the patient age distribution within our samples. In order to study the role of age and family status in spreading influenza, we next subdivided our Basel patients into four different age groups, preschoolers (<7 years old), school-aged children (7 to 17 years), adults (18 to 65 years) and the elderly (>65 years old). We further categorized adults into three subgroups corresponding to family status: adults for whom we know that they live in the same household as children, adults for whom we know that they do not and adults for whom we do not have this information. We thus have overall six different categories of patient groups.

Download:

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

Fig 2. Age distribution and differences in the number of connections to other patients that members of the different categorical patient groups have.

A Distribution of patient ages in this study in age group intervals of 10 years. B Proportion of different age groups in this study compared to in the city of Basel. C Model-based confidence intervals for the difference in average number of connections between each pair of age groups from a negative binomial model. Upper and lower bounds represent 95% confidence intervals for the average fold-difference in connection number between two groups corrected for multiple hypothesis testing using the Tukey method. This means that confidence intervals that do not include 1 are statistically significant. We see that the average connection number for elderly patients is twice the average connection number for preschoolers, adults without children, and adults with unknown status. These values are estimated for a cutoff values of 0.1 years. Estimates for different cutoff values are shown in S7 Fig. Also, the estimates of the mean number of connections in each group are shown in S8 Fig.

https://doi.org/10.1371/journal.ppat.1008984.g002

For each individual infected with influenza in each of these categorical groups, we determined the number of patients with influenza viruses isolated below a certain phylogenetic distance. This number, we then define as the number of connections a patient has. A connection exists if the pairwise phylogenetic distance between viruses isolated from two patients is at most 0.1 years. We then evaluate the mean number of connections of a negative binomial distribution for all individuals from each of the six groups. We later repeated the analyses using different cutoff values. This way of defining two viral isolates to be connected is therefore done independently of the above used definition of local transmission clusters. Using pairwise distances allows us to use distance in the transmission chain. Defining two individuals as connected if they are from the same cluster, on the other hand, only says if two sequences originated from the same introduction. Additionally, we confirm with a simulation that the average number of connections we observe empirically are similar to the number of connections we would expect when simulating under a simple SIR model with a 4% sampling proportion (see S9 Fig).

We find that school-aged children are on average connected to more individuals, than preschoolers (see Fig 2c). This difference is statistically significant after multiple hypothesis testing at a cutoff value of 0.1 years, but not other cutoff values. We further, but not significantly, find that adults that reported to live in the same household as children are on average connected to more patients than those that do not live in the same household as children. For the elderly, we find that they have significantly more connections compared to adults with unknown household status, adults without children, and preschoolers. They do not have more connections on average compared to adults living in the same household as children and school aged children. In summary, there is signal for school children and elderly having more connections to other individuals compared to the three groups unknown household status, adults without children, and preschoolers. The adult living in the same household as children group show tendencies to be connected more in average than the latter three groups, though the data is not informative enough, respectively we do not have enough data, to provide strong evidence for that.

That school aged children and elderly are connected to more individuals than the other groups can have different explanations. The most obvious one is that individuals of these groups are more likely to participate in transmission events, either as a donor or recipient; alternatively, strong mixing within a group and a higher probability of visiting a doctor upon infection and therefore a higher sampling probability could act as an explanation (if members from any group are equally likely to transmit or receive influenza to and from members of any other group, higher sampling probability would increase to number of connections of every group and not just one). Indeed, sampling cases from the Hospital and private practitioners will lead to more severe cases being more likely included in the analysis. In paticular, the elderly and pre-school aged children are overrepresented in our dataset compared to their relative proportion in the population of Basel (see Fig 2A).

In order to assess the potential explanations, we investigated how strongly or weakly different age groups are connected with each other. In particular, instead of just looking at how many patients from any age groups an individual is connected to, we now assess how patients from age groups are connected between each other. We did so using two different approaches. First, we estimated the probability that an individual from a group is connected to a member from the same or a different group by using multinomial logistic regression with the inverse number of samples from each group as weights. Second, we use permutation testing to estimate the probability that the number of connections between different groups is significantly higher or lower than expected if all groups would be equally connected. To do so, we again use the definition of a connection between pairs of patients from the last section.

For both approaches, we find that mostly school-aged children are associated with other school-aged children, and the elderly are associated with other elderly people. We additionally find some indication of higher association between children of any age and adults living in the same household as children than children of any age and adults without children (see Fig 3). Adults living in the same household as children, on the other hand, are estimated to have low association to other adults without children and much higher association to children, whereas adults without children are mostly associated to other adults without children (see Fig 3a).

Download:

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

Fig 3. Mixing of different age groups.

Here we show the mixing patterns between the different categorial patient groups. In contrast to Fig 2, we here ask between which groups connections exists and not just if individuals within these groups have more or less connections than individuals within other groups. We define pairs of patients to be connected if their pairwise phylogenetic distance was below 0.1 years. Results for other thresholds are shown in S10 and S11 Figs. A Probability that an individual from the group in each row is connected to a random individual from the group in a column. These probabilities were calculated by using the inverse number of samples from each group as weights. Upper and lower bounds correspond to 95% confidence intervals around the estimated probability. B The color of each tile in the heatmap corresponds to the p-value for either positive (red) or negative (blue) associations. These p-values are bonferroni corrected for the number of comparisons (42). We estimate these p-values by randomly permuting the group to patient labels and then comparing the number of pairs of interactions we observe in the data vs. when randomly permuting.

https://doi.org/10.1371/journal.ppat.1008984.g003

Increased sampling of the elderly relative to the other age groups is likely to occur, since the elderly are more likely to visit a doctor when in case of infections with influenza [44]. The elderly are indeed over-represented in this study (see Fig 2b). Thus, strong mixing within group and high sampling, might explain the increased connectivity of the elderly.

The second group that we found to have many connections to other patients where school-aged children (see Fig 2c). When looking to which groups these were connected, we found them to be associated with other school-aged children. However, they are unlikely to suffer from more symptoms than preschoolers [45] and therefore should not be overrepresented in our dataset compared to pre-schoolers. Also, based on Fig 2b, we do not see evidence for oversampling of this group compared to preschoolers. We therefore interpret our results as school-aged children being involved in more transmission events compared to the other patient groups, including preschoolers. Furthermore, adults living in the same household as children might get mainly infected by the children and not by other adults, which does however not mean that adults do not play a crucial role in introducing novel lineages into the city. Indeed, children have been previously reported to be a strong driver of influenza transmission [14, 15, 39].

These interpretations however are based on the analysis of influenza isolates from one season and city and will therefore need to be repeated in different seasons and cities to get a more complete understanding of the transmission patterns of influenza across age groups.