Assessing the Ecotoxicologic Hazards of a Pandemic Influenza Medical Response

Background: The global public health community has closely monitored the unfolding of the 2009 H1N1 influenza pandemic to best mitigate its impact on society. However, little attention has been given to the impact of this response on the environment. Antivirals and antibiotics prescribed to treat influenza are excreted into wastewater in a biologically active form, which presents a new and potentially significant ecotoxicologic challenge to microorganisms responsible for wastewater nutrient removal in wastewater treatment plants (WWTPs) and receiving rivers. Objectives: We assessed the ecotoxicologic risks of a pandemic influenza medical response. Methods: To evaluate this risk, we coupled a global spatially structured epidemic model that simulates the quantities of antivirals and antibiotics used during an influenza pandemic of varying severity and a water quality model applied to the Thames catchment to determine predicted environmental concentrations. An additional model was then used to assess the effects of antibiotics on microorganisms in WWTPs and rivers. Results: Consistent with expectations, our model projected a mild pandemic to exhibit a negligible ecotoxicologic hazard. In a moderate and severe pandemic, we projected WWTP toxicity to vary between 0–14% and 5–32% potentially affected fraction (PAF), respectively, and river toxicity to vary between 0–14% and 0–30% PAF, respectively, where PAF is the fraction of microbial species predicted to be growth inhibited (lower and upper 95% reference range). Conclusions: The current medical response to pandemic influenza might result in the discharge of insufficiently treated wastewater into receiving rivers, thereby increasing the risk of eutrophication and contamination of drinking water abstraction points. Widespread drugs in the environment could hasten the generation of drug resistance. Our results highlight the need for empirical data on the effects of antibiotics and antiviral medications on WWTPs and freshwater ecotoxicity.


GLobal Epidemic and Mobility modeler
The GLobal Epidemic and Mobility (GLEaM) computational modeler is a georeferenced metapopulation model based on three data/model layers as previously discussed (Balcan et al. 2009a; Balcan et al. 2009c). The first layer is based on the high resolution population census areas defined with a Voronoi tessellation procedure around transportation hubs (Barrat et al. 2004;Colizza et al. 2007;Colizza et al. 2006a;Colizza et al. 2006b). These census areas define the subpopulations of the metapopulation model (Supplemental Figure 1). In particular, we identify 3,362 subpopulations centered around IATA airports in 220 different countries (Balcan et al. 2009c;Colizza et al. 2007). GLEaM integrates short scale mobility between adjacent subpopulations by considering commuting patterns worldwide as obtained from the data collected and analyzed from more than 29 countries in 5 continents across the world (Balcan et al. 2009a;Balcan et al. 2009c). The third layer is the epidemic dynamic model that defines the evolution of the infectious disease inside each subpopulation.

Supplementary Figure 1. Schematic illustration of the stochastic metapopulation model.
GLEaM integrates high resolution demographic and mobility data defining geographical census areas connected by commuting patterns and air travel flows. The same resolution is used worldwide.
The model simulates the mobility of individuals from one subpopulation to another by a stochastic procedure in which the number of passengers of each compartment traveling from a subpopulation j to a subpopulation l is an integer random variable defined by the actual data from the airline transportation database. Short range commuting between subpopulations is modeled (Balcan et al. 2009a;Balcan et al. 2009c) with a time scale separation approach that defines the effective force of infections in connected subpopulations. The baseline infection dynamics takes place within each subpopulation and assumes the classic influenza-like illness compartmentalization in which each individual is classified by a discrete state such as susceptible, latent, infectious symptomatic, infectious asymptomatic or permanently recovered/removed. The model therefore assumes that the latent period is equivalent to the incubation period and that no secondary transmissions occur during the incubation period (Supplemental Figure 2). The following section will provide a detailed description of the compartmentalization. All transitions are modeled through binomial and multinomial processes to preserve the discrete and stochastic nature of the processes. The model generates in silico epidemics for which we can gather information such as prevalence and number of secondary cases for each subpopulation and with a time resolution of 1 day.

Supplemental Table 1. Parameter description and values.
Parameter Value Description β β β β transmission rate AVP L 0, 10 -3 , 10 -2 daily rate of AV distribution for prophylaxis of duration L L 2w, 4w duration of AV prophylaxis intervention; w=week χ χ χ χ (10 d) -1 inverse duration of 1 course of AV prophylaxis; d=day ε ε ε ε ( between those who can travel I j t , probability p t , and those who are travel-restricted I j nt , with probability 1-p t . All the infectious persons permanently recover with probability t ∆ µ , entering the recovered compartment R j in the next time step. Consult Supplemental Figure 1 for the compartmental model adopted and Supplemental Table 1 for the parameter values used. 8 In each subpopulation the variation of the number of individuals in each compartment [m] can be written at any given time step as is drawn upon to model these effects and quantify the antiviral efficacy (AVE) as in the following (Longini et al. 2004;Longini et al. 2005): AVE S represents the AV efficacy for susceptibility to infection and corresponds to the reduction of probability of infection for a susceptible taking AV drugs with respect to a susceptible not under prophylaxis; AVE D represents the AV efficacy for symptomatic disease given the infection occurs and corresponds to a reduction of the probability of developing symptoms with respect to a person not under prophylaxis; AVE I represents the AV efficacy for infectiousness and corresponds to the reduction of the probability of transmission of an infectious individual taking AV drugs. The latter effect is valid for both prophylaxis and treatment and also assumes a reduction by 1 day in the average duration of the infection period.
Prophylaxis is administered to a given fraction c AVP of the susceptible population (Flahault et al. 2006;Kernéis et al. 2008), who enter the S AVP compartment and receive a single course of AV for a total duration of 10 days. At the end of this period, they re-enter the susceptible class S in case infection did not occur, since the risk of infection is expected to resume shortly after cessation unless combined with immunization, as for seasonal prophylaxis (World Health Organization 2004b). The parameter c AVP represents the daily coverage of antiviral prophylaxis and is assumed to be equal in all census basins in the world (Flahault et al. 2006;Kernéis et al. 2008). If susceptibles contract the infection while under prophylaxis, they enter the latent class L AVP , which is the analogue of the compartment L and has the same average . In the case of symptoms, the individual is assumed to be immediately treated, with no further delay, entering the I AVT compartment and starting another course of AV drugs by doubling the dose per day with respect to the prophylaxis.
Both I a AVP and I AVT compartments are characterized by a reduced infectiousness of a factor ) 1 ( I AVE − due to the AV drugs, and by an average duration shortened by 1 day with respect to an infectious individual not taking drugs, i.e. . We assume that all individuals who get infected while under prophylaxis have access to medical care and get treated with antivirals. In absence of antivirals, they follow the same dynamics as the nontreated infectious individuals (see dashed arrows in Supplemental Figure 1).
Therapeutic treatment is modeled by assuming that: (i) all symptomatic infectious individuals under prophylaxis will receive treatment entering the compartment I AVT immediately after the onset of symptoms; (ii) symptomatic infectious individuals not under prophylaxis will enter the compartment I AVTd with probability p AVT , a measure of the prompt case detection and rapid drug delivery. The I AVTd compartment includes both the infectious period before the drug administration and the one during the treatment, for a total average duration of , with a reduction by 1 day with respect to the mean infectious period with no intervention. We assume that the drugs are administered with a delay of 1 day from the symptoms onset, and that the individual has an effective reduced transmissibility r AVTd that takes into account the pre-treatment period with full infectiousness and the treatment period with reduced infectiousness ) 1 ( Influenza-associated pneumonia. Additional compartments and transition rates were introduced to take into account the development of influenza-associated pneumonia (Balcan et al. 2009b). Symptomatic infectious individuals have a certain probability p P of developing influenza-associated community acquired pneumonia (IA-CAP). The incidence frequency of IA-CAP varies from 2% to 40%, and is dependent on  Based on recent studies on the impact of oseltamivir treatment on influenza-related lower respiratory tract complications for seasonal flu (Kaiser et al. 2003;Yu et al. 2010), we assume that under AV treatment the probability of developing pneumonia is reduced by a factor α , i.e. p p-AV = α p p , with α =54%. We also assume that all pneumonia cases are treated with antibiotics.  Table 2). It is assumed that a person will start showing symptoms of IA-CAP after the infectious period is over, entering the compartment P I , as pneumonia symptoms typically appear during the early convalescent period of the influenza infection (Lim 2007). This corresponds to the stage CURB-65=0 and 1, i.e., non-severe 13 pneumonia and outpatient care. After an average period of

Supplemental
days, p II = 11% of the patients in P I are estimated to experience worsening of symptoms (U.S. Department of Human and Health Services (HHS) 2005), thus requiring hospitalization and entering the compartment P II corresponding to CURB-65 = 2, whereas the rest will recover (with probability 1-p II ). Unimproved patients (p III = 1.65%) will develop severe pneumonia and progress to P III corresponding to CURB-65 = 3 to 5. All three stages of pneumonia course (P I , P II , and P III ) have an average duration of Estimates assume no antiviral prophylaxis in any scenario, no antiviral treatment in the mild transmission scenario, and antiviral treatment of 30% of cases in the moderate and severe transmission scenarios (R0=1.9, R0=2.3). A single course of antibiotics is defined as the combination of antimicrobial drugs considered in the treatment regimen for the suggested duration, following the UK guidelines (Lim 2007).

Supplemental Figure 4. Tamiflu (black bar) and Relenza (red bar) stockpiles in national pandemic plans expressed in % of population coverage.
This is not an exhaustive list, but has been adapted from the literature (Tierney and Reddy 2007) and media reports and should be used only as a guide. In this paper the antiviral use was assumed to be composed of 100% Tamiflu owing to uncertainties regarding the conditions for Relenza use.

Pharmaceutical model
Approved antivirals for therapy and prophylaxis of influenza are the neuraminidase inhibitors, i.e., oseltamivir ethylester phosphate (Tamiflu) and zanamivir (Relenza). Relenza stockpiles are increasing due to concerns of Tamiflu resistance (Influenza Project Team 2008; Lackenby et al. 2008;Meijer et al. 2007), however, Tamiflu is currently much more widely stockpiled globally due to the easier administration of the drug (Supplemental Figure 2). Here we consider the use of Tamiflu for therapy and prophylaxis. The consideration of a combination therapy to decrease the risk of developing drug resistance would require a more sophisticated epidemic model that takes into account multiple strains of the virus, and would be considered as a further development of this approach. Individuals under antiviral prophylaxis are given a single course of Tamiflu for 10 days, 1 tablet per day, whereas AVT employs 2 tablets per day for a total duration of 5 days.
During pneumonia stages, patients are given antibiotics, according to the chemotherapy guidelines for treating IA-CAP sanctioned by the British Infection Society, British Thoracic Society and the Health Protection Agency (Lim 2007). It is assumed that empirical antibiotic therapy will be used to treat secondary infections as a result of the overwhelming surge in patients during a pandemic (Lim 2007). In contrast, inter-pandemic IA-CAP treatment would often include laboratory tests to increase the efficiency of the treatment and decrease the likelihood of generating antibiotic resistance. Guidelines recommend the use of 7 days of appropriate antibiotics for patients with non-severe and uncomplicated pneumonia (i.e. compartments P I and P II ), whereas a 10 days treatment is proposed for those with severe pneumonia (P III ). Supplemental Table 3 reports the detailed recommendations for the empirical antibiotic treatment regimes.

Scenarios
All epidemic simulations were initiated with a single symptomatic infectious individual and let evolve for a duration of 1 year. Only the runs with a global outbreak, defined as the generation of new symptomatic cases in more than one country, were considered for the analysis. Initial conditions assume that the pandemic starts in Hanoi on the first of October (Colizza et al. 2007). Simulations with three different values of the reproductive number, R 0 =1.65, 1.9, 2.3 have been carried out.
Simulations consider an unlimited stockpile of antibiotics in the UK. In this study, it was assumed that the U.K. will ultimately achieve an antibiotic stockpile sufficiently large to treat all patients as detailed in the guidelines (Lim 2007), which in some cases would mean a further increase in the stockpile (U.K. Department of Health 2009a).
Supplemental Table 3. Preferred and alternative empirical antibiotic treatment regimens for pneumonic influenza-associated complications (Lim 2007

Excretion and Dilution of Pharmaceuticals in WWTP
Drug excretion was determined from the pharmacological data within the literature (Dollery 1999;Wishart et al. 2006) as the percentage of drug released in the faeces and urine as the parent chemical, or biologically active metabolite (Supplemental Table 4). The dilution of drug released into the WWTP per day was evaluated using: 1) the mean of all WWTP dry weather flows (m 3 /d) within the Thames river basin as calculated from LF2000-WQX (230 L/head/d; see Supplemental Material Section 2.2); and 2) the actual dry weather flow for each WWTP as found within LF2k-WQX (Supplemental Figure 3). As the per capita consumption of water per day does not vary by more than a factor of 2-3 between industrialised countries (Supplemental Figure 4) 1999;Alexy et al. 2004;Benotti and Brownawell 2009;Brain et al. 2004;Gartiser et al. 2007;Junker et al. 2006;Kümmerer et al. 2000), with half-lives of days to weeks (Benotti and Brownawell 2009;Christensen 1998). It is for these reasons that we feel there is justification in assuming a conservative pharmaceutical biodegradation model of zero degradation/sorption. Photolysis is a major contributor to the degradation of otherwise persistent chemicals in the environment (Andreozzi et al. 2003;Lam et al. 2004). However, laboratory estimates on photodegradability can be misleading as such photo-sensitive chemicals might adsorb into the sediment and organic matter where they remain protected from UV and biological degradation (Andreozzi et al. 2003;Burhenne et al. 1997;Kümmerer and Henninger 2003).

Supplemental
As a result of this uncertainty, we have chosen not to include photolysis in calculating environmental concentrations of pharmaceuticals, conservatively assuming that the antibiotics persist in the environment for more than 1 day.
Tamiflu and oseltamivir carboxylate (OC), the active antiviral of the prodrug Tamiflu this study, which assumed no loss to sorption, degradation or photolysis, and that all pharmaceuticals would pass through the WWTP into the receiving river in which it will persist for more than 1 day.

Low Flows 2000 -Water Quality Extension (LF2k-WQX)
The Low Flows 2000 (LF2000) WQX (Water Quality eXtension) model is an extension to the LF2000 system (Young et al. 2003). LF2000 is a decision support tool designed to estimate river flow at gauged and ungauged sites and to assist regional water resources and catchment management.
The LF2000-WQX software (Keller and Young 2004;Williams et al. 2009) is a geographical information-based system that combines hydrological models with a range of water-quality models, including a catchment-scale water-quality model. This model generates spatially explicit statistical distributions of down-the-drain chemicals for both conservative and degradable compounds. It uses a Monte Carlo mixing-model approach to combine statistical estimates of chemical loads at specific emission points (e.g. WWTP) with estimated river flow duration curves for the whole river network of interconnected model reaches (a reach is the river stretch between model features, e.g., major tributaries, sewage treatment works).
Thus working from the low order streams at the head of the river network to the outlet of the river basin, the model accounts for the accumulation of point loads and the accumulation of water in which these loads are diluted. Degradable chemicals may be removed from the river water by a non-specific dissipation process assuming first-order kinetics. Details of the model are given in two recent applications: (1) to assess the risk posed to fish by excretion of steroid oestrogens to rivers via sewage treatment works in England and Wales ), and (2) an initial assessment of the likelihood of cytotoxic drugs reaching drinking water in the River Thames basin in south east England (Rowney et al. 2009).
In this study the model was applied to predict the concentrations of Tamiflu and nine antibiotics in the Thames river basin based on per capita excretion of these drugs by the population. Pharmaceutical load in the WWTP influent per head per day was assumed constant and fixed at the mean peak value for the pandemic. Removal rates in the WWTP were set to zero and the drugs were considered to be stable in the water column (i.e., they were not degraded). The WWTPs were assumed to be the only source of the drugs within the catchment and the background concentrations in the river stretches modelled and lateral inflows to these river stretches was set to zero. The variations of all the model parameters that were defined by a mean and standard deviation were assumed to be described by a normal distribution. Each Monte Carlo simulation comprised 1,350 model runs (shots). For each modelled river reach data were saved in a spatially referenced data file (ArcMap shapefile), which amongst other information, contained the predicted mean concentrations, the standard deviation and the 90th percentile concentration (the concentration which would be predicted not to be exceeded for 90 percent of the time) for each of the modelled drugs.

Background level of pharmaceuticals in the Thames River
A very wide range of pharmaceuticals are in constant use in a population and thus will be present in the wastewater during a pandemic (Kümmerer 2009a, b). If one was to use the average annual antibiotic usage within England during an inter-pandemic period (NHS BSA 2008) as a first approximation, 62 µg antibiotics/L would be present in the UK WWTPs, which was assumed to be diluted by the median volume of wastewater in WWTP within the Thames Catchment, 230 L/head/day (Supplemental Table 5).  values with the corresponding confidence intervals populated the LF2000-WQX model to calculate the concentration of drugs in the Thames catchment. Since LF2000-WQX is not a dynamic model and thus acquires as an input a single value of drug usage in µg/head/day, we considered the median and confidence interval of the peak drug usage rate to assess the highest toxicity in the rivers, given the scenario under study. The day of peak drug usage was used as a worst case scenario for predicted river water concentrations.

Toxicity Assessment
We focus our ecotoxicologic analysis on antibiotics, as Tamiflu itself has not been shown to exhibit accute toxicity ( for a mixture of one given composition, while the ratio of clavulanic acid to amoxicillin in the environment will differ from that composition. This will lead to an underestimation of the overall toxicity, however clavulanic acid was included in calculating the values found in Table 1 in the main text. Thus, the effects of 8 antibiotics were studied (amoxicillin, cefotaxime, cefuroxime, clarithromycin, doxycycline, erythromycin, levofloxacin and moxifloxacin).
All MIC values for one given antibiotic were first aggregated into species sensitivity distributions (SSD) (Newman et al. 2002).

Establishment of species sensitivity distributions
A total of >1,000,000 MIC values were included, including between 21 and 100 bacterial species per antibiotic. SSD curves were constructed to gather the sensitivities of all bacterial species for one given antibiotic in one graph, ordered by their MIC value. The SSD is the cumulative density function giving the "potentially affected fraction" (PAF) of bacteria that are inhibited at a given MIC on a scale between 0 and 1.
For the construction of the SSD curves, three different approaches were applied. Second, the SSD distributions were fitted to two-parameter log-logistic and two-parameter Weibull functions (Ritz and Streibig 2005) with package 'drc' (Ritz and Streibig 2005). The quality of the curve-fits was judged by their residual variance. In order to compare these residual variances to the log-normal distributions, the residual variation of the SSDs based on the mean MIC values on the log-normal distribution was calculated.
For the SSDs based on the normalized MIC distributions, the asymmetric Weibull function (f(x) = exp (-exp(b(log(x)-e)))) ), with e denoting the logarithm of the inflection point of the

Calculation of the combined effects of the antibiotic mixture
For the calculation of the multi-substance PAF (msPAF, ie., all antibiotics present simultaneously), two mixture toxicity models were applied to the SSD dose-effect curves Backhaus et al. 2000;Bliss and Mexico 1939;De Zwart and Posthuma 2005;Loewe 1927;Loewe and Muischnek 1926). The first model, concentration addition, is generally assumed to hold for compounds with an identical mechanism of action, such as binding to the same receptor. The second, independent action, is assumed to hold for compounds with a different mode of action.
Antibiotics belonging to different classes were assumed to act by independent action (also called dissimilar action), whereas the joint toxicity of antibiotics from identical classes were As all antibiotics finally inhibit bacterial reproduction, which could be seen as common mode of action, and as their joint toxicity might thus fall in between predictions from the two models, calculations were also performed for the assumptions that all antibiotics would act according to concentration addition or independent action (calculations were then based on Weibull curve fits only for the 5 th percentile MICs and normalized MIC distributions).
From the possible combinations of the three different SSD curves with the three mixture toxicity models (concentration addition, independent action and the mixed model), we adopted the one with the most refined assumptions: SSD curves based on the complete normalised distribution of MIC values, and a mixed mixture toxicity model with a similar mechanism of action only for antibiotics from the same class. We thus received one toxicity prediction for each location (WWTP/river stretch) and each scenario.
The results were then aggregated by calculating all WWTPs and river stretches that lead to a risk for > 5% of the total bacterial community (potentially-affected fraction: PAF). Fivepercent was chosen as a threshold as the definition of predicted no-effect concentrations (PNEC) for industrial chemicals in the European Union, is based on <5% potentially affected fraction of the total species present (based on NOECs, no observed effect concentrations, and taking into account a 50% confidence interval)(European Commission 2003).

Impact of different toxicity measures
We examined the effect of different microbial toxicity model assumptions on simulated mean toxicity in WWTPs. We compared the influence of changes in both the establishment of the SSD curves and the mixture toxicity assumptions to the chosen, most refined model. Results show that the overall behaviour with respect to the WWTP is left unchanged, with the absolute values of the toxicity varying by about 10-15% (Supplemental Figure 7). We adopted the most refined assumptions: SSD curves based on the complete distribution of MIC values per species, and a mixed mixture toxicity model with a similar mechanism of action only for antibiotics from the same class. Figure 9. Assessment of different toxicity model assumptions as a function of wastewater treatment plants. The scenario considered corresponds to no prophylaxis and AVT=0.3. The influence of mixture toxicity models is shown in PAF IA (all antibiotics independently acting) and PAF CA (all antibiotics acting according to concentration addition), and the influence of the representation of the MIC distributions in the SSD curves in PAF 5P and PAF M (the best fit of the 5 th percentile respectively the loglogistic fit of the median of the sensitivity of each bacterial species to one given antibiotic). These are compared to the chosen model PAF refined (based on the normalized distribution of the sensitivity of each species and a combination of both mixture toxicity models, with concentration additivity only assumed for antibiotics from the same class)

Toxicity of background levels of antibiotics
It was beyond the scope of this study to fully incorporate the toxicity of the pharmaceuticals in constant use (Supplemental Table 5) into the ecotoxicity model. First, pharmaceuticals will not be the only compounds possibly inhibiting bacterial growth, and calculations based on estimates of antibiotics will underestimate toxicity. Moreover, it is uncertain how usage would shift in light of a pandemic. As a first approximation, the toxicity of the 8 antibiotics used during a pandemic was also calculated for combined pandemic and inter-pandemic use.
This will again be an underestimation of the total antibiotic toxicity, as only 8 out of all antibiotics present were taken into account.
The background toxicity of the 8 investigated antibiotics in inter-pandemic times (as in Supplemental Table 5) was calculated to amount to a PAF of 13% for a typical WWTP with a 230 L/head/day dilution (4 to 17% for all WWTP). As the WWTP in the London area are still functioning, an estimated toxicity of 4 to 17% does apparently not cause functional failure.
There are several possible reasons: 1) because of community tolerance acquired upon chronic exposure to antibiotics (Blanck 2002); 2) due to other factors that might reduce toxicity, such as sorption and degradation, 3) the main functionally active bacteria have lower sensitivities than these concentrations.
If the concentrations of the 8 antibiotics from inter-pandemic and pandemic use are added up and their toxicity calculated, the lower and the upper RR of WWTP toxicity for the R 0 =1.65 scenario range from 3.5 to 13 and 4 to 20% PAF, respectively, reflecting a marginal increase in the background toxicity. For R 0 =1.9, the lower RR of all WWTP except 5 would exceed the threshold value of 5% for all possible scenarios. The range of the upper bound of the RR across all different WWTP would increase from 2 to 14% (pandemic use only) to 5 to 25% (inter-pandemic and pandemic antibiotics combined) for the R 0 =1.9 scenario. In the R 0 =2.3 scenario, maximum toxicities for combined antibiotic use amount from 11% to 37%, compared to 8 to 32 % for pandemic use only. When comparing the median toxicity between the background and the combination of pandemic use and background, the percentage of growth-inhibited species is elevated by 0.1 to 16% in the latter (median of all WWTPs), with the highest increase for the R 0 =2.3 scenario. This is in the same range as the increase in toxicity by pandemic use only (0.03 -23%).
As the reason for the apparent tolerance of WWTP consortia to background antibiotic concentrations is unknown, it is not possible to precisely estimate how much toxicity is added to the background toxicity by pandemic use. In case the bioavailable concentrations of antibiotics were reduced by e.g. sorption and degradation, one would need exact bioavailability figures to estimate toxicity. There are no toxicological models for an estimation of 'extra' toxicity for the case that the WWTP consortia have become tolerant to inter-pandemic concentrations. Last, if the sensitivity of WWTP consortia is in the upper range of the established SSDs, the likelihood of functional effects increases with increasing percentages of PAF. For these reasons, we decided to focus on the effects of the pandemic antibiotics only. The conclusions will, however, be similar between the approximations above and the analysis presented in the main paper, as both the R0=1.9 and the R0=2.3 scenarios increase overall toxicity by a considerable amount.

Uncertainties in characterizing toxicity
The models of microbial growth-inhibition resulting from exposure to projected antibiotic concentrations in WWTP during an influenza pandemic have a high degree of uncertainty.
Most importantly, the relative sensitivity of WWTP microbial species/consortia to antibiotics, as compared to clinically-relevant human pathogens, remains largely untested (Kümmerer et al. 2000;Robinson et al. 2005). The major difference between clinically-relevant human pathogen MICs and those relevant to WWTPs and the environment, is that environmentally-relevant MICs would have functional endpoints (i.e., nitrification, mineralization, etc), whereas clinically-relevant endpoints are typically biostasis or death.
The effect of antibiotics on flock-associated WWTP bacteria will undoubtedly differ from effects on planktonic bacteria, as several mechanisms render biofilm consortia more resistant to chemical stress (Donlan and Costerton, 2002). Cells in the centre of the biofilm will be exposed to reduced concentrations of antibiotics due to antibiotic degradation and diffusion limitations. However, the growth rate for the majority of cells was reported to be affected at concentrations of antibiotics comparable to MIC values of planktonic cells, for some antibiotics (Lewis 2000(Lewis , 2010. Further, the most active cells are present at the flock surface, at least for nitrification (Schramm et al. 1998), where nutrient and antibiotic concentrations are highest. Biofilms might, therefore, be less effective for the protection of WWTP functioning than for the preservation of consortia members located in the centre of biofilms.
Still, a compilation of toxicity data for erythromycin in sewage sludge bacteria allows for a comparison between predictions according to species sensitivity distributions of pathogen MIC values, and experimental data. Experiments were performed with sewage sludge bacterial consortia in batch reactors, and therewith studied relevant microbial consortia under representative conditions. Supplemental Table 6 shows that predictions according to MIC values often lay in between measured effects data for sludge derived from two different WWTP, or were not more than 20% different from the predictions. We thus argue that it is scientifically defensible to use clinically-derived MICs as a first approximation for understanding environmental microbial antibiotic sensitivity.  Table 5), and an even more complex mixture of industrial chemicals and personal care products that are routinely found in wastewater (Alder et al. 2006;Ramirez et al. 2009).

Supplemental
Overall, there is a clear need for empirically-determined thresholds of antibiotic and antiviral toxicity on pilot-scale WWTPs. Significant differences between the microbial consortia indigenous to activated sludge and biofilter systems as well as the heterogeneous distribution of these systems within and between countries ) might require several such pilot-plant studies motivated by different regional or national risks.

Sensitivity Analysis: Rate of complications development
In the main text we assumed the rate of complications to be equal to 15% of the clinical  Figure 9 and S10 for the toxicity in WWTPs and in rivers, respectively.
When a low complication rate is assumed in a severe scenario, about 45% of the WWTPs would reach a toxicity above the 5% PAF threshold, if AVT=0.3 is assumed. A moderate scenario would not raise toxicity values in WWTPs above the threshold in the interventions considered. The situation is dramatically different in case of a pandemic with a complication rate equal to 40%. Almost the whole set of WWTPs would reach toxicity levels above the threshold in both scenarios R 0 =1.9 and R 0 =2.3. A mild scenario R0=1.65 would lead 60% to 80% of the WWTPs to reach a toxicity above the 5% PAF threshold. Toxicity higher than the 5% threshold would be observed in a very small fraction of the river stretches, amounting to a negligible fraction of the total river length in the Thames basin, if the complication rate is assumed equal to 2%. When 40% complication rate is considered, about 30 to 40% of the 37 river stretches would reach a toxicity above threshold in the moderate transmission scenario; the percentage would be twice as much (about 80%) in the case of a severe scenario. When a mild scenario is considered, a very small percentage of the river stretches (about 5 to 10%) would experience toxicity higher than the 5% threshold even in the 40% complication rate assumption.
Supplemental Figure 10: Predicted toxicity in WWTPs. Percentage of WWTPs predicted to exceed the threshold of 5% potentially affected species (PAF). Histogram columns and error bars refer, respectively, to the median and 95% confidence intervals obtained from the drugs usage pattern predicted by the 1,000 stochastic runs of the epidemic model. Complication rate equal to 2% (left panel) and 40% (right panel) are shown.
Supplemental Figure 11: Predicted toxicity in river stretches. Percentage of river stretches in the Thames catchment (top row) and percentage of total Thames catchment length (bottom row) predicted to exceed the 5% PAF threshold. Histogram columns and error bars refer, respectively, to the median and 95% confidence intervals obtained from the drugs usage pattern predicted by the 1,000 stochastic runs of the epidemic model. Calculation of the PAF is based on the antibiotic sensitivity distributions of human pathogens and a combination of two mixture toxicity models. Complication rate equal to 2% (left panel) and 40% (right panel) are shown.

Bacterial biofilms
Pathogenic bacteria have been shown to produce neuraminidases to assist in the colonization of heavily sialyated secretions and surfaces of the upper respiratory tract (Soong et al. 2006).
Limited data exists on the effect of neuraminidase inhibitors on bacterial neuraminidases. Soong et al. (2006) demonstrated the efficiency of oseltamivir carboxylate (active metabolite of Tamiflu) and peramivir to inhibit biofilm formation of the pathogen Pseudomonas aeruginosa (Soong et al. 2006). It remains unclear to what extent neuraminidase inhibitors will inhibit biofilm formation in environmentally-relevant strains of microorganisms and how widespread any observed inhibition might be.
Concern has previously been raised regarding the potential for neuraminidase inhibition in bacteria found within WWTP (Singer et al. 2008). The median concentration of Tamiflu in WWTP (assuming 230 L/head wastewater), under a moderate and severe pandemic (AVT=30% and AVP=0) was projected to be approximately 1.00 µg/L and 21.3 µg/L respectively, with maximal values of 5.31 and 102 µg/L OC, respectively. Both scenarios exceed the viral neuraminidase IC 50 (1.3 nM (0.37 µg/L) (Yamashita et al. 2009), i.e., the concentration at which neuraminidase activity is inhibited by 50%), indicating biologically active concentration of the neuraminidase inhibitors will be present in the WWTP. The concern is that OC might inhibit the capacity for microbial floc and biofilm formation within WWTP. Such inhibition could lead to decreasing wastewater treatment, leading to WWTP failure and contamination of receiving rivers and downstream drinking water abstraction points. Moreover, there is a risk that the effects from neuraminidase inhibitor exposure might be further exaggerated when presented in combination with a high load of antibiotics, as projected in this study. Groundwater (Geoghegan et al. 2008) A cell attachment assay was used to determine the influence of Tamiflu exposure on biofilm formation of environmentally relevant bacterial strains (Djordjevic et al. 2002;O'Toole and Kolter 1998). One clinical and nine environmental microorganisms were selected from our culture collection (University of Sheffield), representing a range of possible bacteria potentially found within a WWTP and/or the environment (Supplemental Table 7). The isolates were maintained on a solid R2A medium (Oxoid) or for E. coli strains on LB agar (Oxoid). Each bacterial strain (Supplemental Table 7 (Geoghegan et al. 2008), stained with 0.1% crystal violet for 10 minutes before the crystal violet was released using 33% acetic acid. The crystal violet was then removed to a clean 96-well plate and the absorbance was measured at 630 nm.

Biofilm inhibition assays Supplemental
There was a statistically significant (p<0.05) decline in biofilm formation relative to the control in 20% and 10% of the cases when the bacterial isolates were exposed to 0.1 mg/L and 1 mg/L Tamiflu, respectively (Supplemental Figure 12). Notably, one bacterial isolate Pseudomonas fluorescens WH2 exhibited a significant increase in biofilm formation suggestive of growth on Tamiflu. Although there was a substantial change in biofilm formation in most isolates, these changes were not significant at the p<0.05 level. Additional research would be required: 1) to examine a wider range of Tamiflu and the active antiviral, oseltamivir carboxylate (OC); 2) to examine a more diverse range of microorganisms to determine how widespread the biofilm disruption may be and the mechanism behind such a disruption; 3) to examine mixed cultures, particularly those of a WWTP and river microbial biofilms to address community dynamics; 4) to examine the added effect of chemical mixtures, i.e., antibiotics and Tamiflu; and 5) to examine all available neuraminidase inhibitors, as Tamiflu is only one neuraminidase inhibitor in a growing class of drugs.
Varying IC's for OC and peramivir indicate the likelihood that each neuraminidase inhibitors will potentially have a different effect on microbial biofilms and toxicity (Soong et al. 2006). Figure 12: Impact of Tamiflu exposure on microbial biofilm formation. Fraction increase or decrease in biofilm formation for a range of ten microorganisms (Supplemental Table 7) in the presence of 1.0 µg L -1 (grey bar) or 0.1 µg L -1 (blue bar) Tamiflu. Changes are statistically significant where denoted by a lighter shaded grey or blue bar (Students t-test at p<0.05). Error bars denote standard error of the mean (n=4).