Skip to main content
PMC home page
eLife logoLink to eLife
. 2015 Sep 2;4:e07969. doi: 10.7554/eLife.07969

Mapping influenza transmission in the ferret model to transmission in humans

Michael G Buhnerkempe 1,2,*,†, Katelyn Gostic 1,†, Miran Park 1, Prianna Ahsan 1, Jessica A Belser 3, James O Lloyd-Smith 1,2
Editor: Mark Jit4
PMCID: PMC4586390  PMID: 26329460

Abstract

The controversy surrounding 'gain-of-function' experiments on high-consequence avian influenza viruses has highlighted the role of ferret transmission experiments in studying the transmission potential of novel influenza strains. However, the mapping between influenza transmission in ferrets and in humans is unsubstantiated. We address this gap by compiling and analyzing 240 estimates of influenza transmission in ferrets and humans. We demonstrate that estimates of ferret secondary attack rate (SAR) explain 66% of the variation in human SAR estimates at the subtype level. Further analysis shows that ferret transmission experiments have potential to identify influenza viruses of concern for epidemic spread in humans, though small sample sizes and biological uncertainties prevent definitive classification of human transmissibility. Thus, ferret transmission experiments provide valid predictions of pandemic potential of novel influenza strains, though results should continue to be corroborated by targeted virological and epidemiological research.

DOI: http://dx.doi.org/10.7554/eLife.07969.001

Research organism: human, other

eLife digest

Every year, thousands of people develop influenza (flu). After being infected by the influenza virus, the immune systems of most people adapt to fight off the virus if it is encountered again. However, there are many different strains of influenza, and new strains constantly evolve. Therefore, although someone may have developed resistance to one previously encountered strain, they can still become ill if another strain infects them.

Different strains of the influenza virus have different abilities to spread between people and make them ill. One way that scientists assess whether a particular strain of influenza is a threat to people is by studying ferrets, which develop many of the same flu symptoms as humans. However, questions have been raised over how accurately ferret studies reflect whether a particular virus strain will spread between humans.

Controversy has also arisen over experiments in which ferrets are infected with genetically engineered strains of influenza that mimic how a strain that has evolved in birds could adapt to cause a pandemic in humans. In 2014, the United States government suggested that such research should be temporarily stopped until more is known about the risks and usefulness of these studies. Now, Buhnerkempe, Gostic et al. have compared the results of 240 ferret and human studies that aimed to assess how easily strains of influenza spread. Specifically, the studies looked at how often a healthy ferret or human became ill when exposed to an animal or human infected with a particular strain of influenza.

The results of the ferret transmission studies matched well with transmission patterns observed in human studies. Ferret studies that assessed how the influenza virus is transmitted through the air via sneezes and coughs were particularly good at predicting how the virus spreads in humans. But Buhnerkempe, Gostic et al. caution that ferret studies are not always accurate, partly because they involve small numbers of animals, which can skew the results. There also needs to be more effort to standardize the procedures and measurements used in ferret studies.

Still, the analysis suggests that overall, ferret studies are a useful tool for making an initial prediction of which influenza strains may cause a pandemic in humans, which can then be verified using other methods.

DOI: http://dx.doi.org/10.7554/eLife.07969.002

Introduction

The emergence of deadly animal-origin influenza viruses in human populations, such as influenza A(H5N1) (Chan, 2002; Li et al., 2004), and influenza A(H7N9) (Gao et al., 2013; Gong et al., 2014; Li et al., 2014), has underscored the need to rapidly determine the pandemic potential of novel strains found in humans or in zoonotic reservoirs. Although characterizing human transmissibility of emerging influenza viruses is a perpetual challenge, animal models are often used to characterize transmission among mammals, which can be viewed implicitly as a preliminary screen for pandemic potential in humans. Ferrets are the preferred animal model for influenza transmission studies because clinical signs, pathogenesis and sialic acid distribution are similar in ferrets and humans (Maher and DeStefano, 2004; Shinya et al., 2006; Bouvier and Lowen, 2010). Consequently, the ferret model has been used to assess numerous aspects of influenza transmission potential including: phenotypic traits associated with transmission (Belser et al., 2008, 2013; Song et al., 2009; van Doremalen et al., 2011; Blumenkrantz et al., 2013), transmission under antiviral prophylaxis (Oh et al., 2014), and the relative transmissibility of drug resistant (Herlocher et al., 2004; Hurt et al., 2010; Kiso et al., 2010; Seibert et al., 2010; Duan et al., 2011; Hamelin et al., 2011), emerging (Maines et al., 2006; Itoh et al., 2009; Belser et al., 2013; The SJCEIRS Working Group, 2013; Watanabe et al., 2013; Zhu et al., 2013; Xu et al., 2014), or lab-created isolates (Herfst et al., 2012; Imai et al., 2012; Sutton et al., 2014).

Despite the widespread use of ferrets to assess transmission of influenza, the suitability of ferrets to assess pandemic potential in humans remains unknown, because the relationship between transmission in ferrets and in humans has never been assessed quantitatively (Palese and Wang, 2012; Casadevall and Imperiale, 2014; Lipsitch, 2014). In fact, conspicuous differences in ferret and human transmissibility for influenza A(H7N9) have cast doubt on the validity of the ferret model for assessing transmission in humans (Lipsitch, 2013). As a consequence, ferret studies can only be interpreted, strictly, in terms of general mammalian transmissibility (Herfst et al., 2012; Imai et al., 2012; Casadevall and Imperiale, 2014; Casadevall et al., 2014).

Furthermore, the recent controversy surrounding ‘gain-of-function’ (GOF) experiments on highly pathogenic avian influenza A(H5N1) in ferrets (Herfst et al., 2012; Imai et al., 2012) and proposed GOF experiments on A(H7N9) viruses (Fouchier et al., 2013) has led to ethical questions about influenza GOF experiments and scientific questions about the use of ferrets to assess transmission (Morens et al., 2012; Casadevall and Imperiale, 2014; Casadevall et al., 2014; Lipsitch, 2014; Lipsitch and Galvani, 2014; Russell et al., 2014). With the U.S. government halting funding and calling for a voluntary moratorium and period of review on such experiments as of October 2014 (White House Office of Science and Technology Policy, 2014), groups on all sides of the debate have issued renewed calls for studies on the link between influenza transmissibility in ferrets and in humans (Morens et al., 2012; Lipsitch, 2013, 2014; Casadevall and Imperiale, 2014). Here, we address this gap by compiling ferret transmission studies and comparing their results to estimates of influenza transmission in humans.

Results

Comparing ferret and human secondary attack rates

To assess the quantitative relationship between influenza transmission in ferrets and in humans, we assembled data from all published ferret transmission studies that met our inclusion criteria, including ferret experiments designed to test transmission in the presence of direct contact (co-housing) or by respiratory droplets (adjacent housing allowing air exchange). For each experiment, we calculated the secondary attack rate (SAR), which is defined as the probability of infection for a susceptible individual following known contact with an infectious individual (Halloran, 2005). To match the close contact found in ferret studies, we reviewed estimates of SAR in humans obtained from household contact data (Figure 1).

Figure 1. Boxplots of influenza SAR estimates by subtype.

(A) Human SAR, (B) ferret respiratory droplet SAR, and (C) ferret direct contact SAR. Solid, black lines represent the subtype medians. Boxes give the inter-quartile range with whiskers extending out up to 1.5 times this range. Points represent extreme values. The number of estimated SARs for each subtype is given above each box-and-whisker plot (n). Subtypes were ordered according to the mean human SAR value in all panels. Shading depicts the known human transmission pattern of the subtypes (red—supercritical; blue—subcritical).

DOI: http://dx.doi.org/10.7554/eLife.07969.003

Figure 1—source data 1. Estimates of human household SAR.
DOI: http://dx.doi.org/10.7554/eLife.07969.004
elife07969s001.docx (106.1KB, docx)
DOI: 10.7554/eLife.07969.004
Figure 1—source data 2. Ferret influenza transmission studies via respiratory droplets using human isolates.
DOI: http://dx.doi.org/10.7554/eLife.07969.005
elife07969s002.docx (120.4KB, docx)
DOI: 10.7554/eLife.07969.005
Figure 1—source data 3. Ferret influenza transmission studies via direct contact using human isolates.
DOI: http://dx.doi.org/10.7554/eLife.07969.006
elife07969s003.docx (101.8KB, docx)
DOI: 10.7554/eLife.07969.006
Figure 1—source data 4. Ferret influenza transmission studies via respiratory droplets and direct contact using avian isolates.
DOI: http://dx.doi.org/10.7554/eLife.07969.007
elife07969s004.docx (29.3KB, docx)
DOI: 10.7554/eLife.07969.007

Figure 1.

Figure 1—figure supplement 1. Analysis supporting inclusion of SAR estimates from isolates generating using reverse genetics.

Figure 1—figure supplement 1.

Comparison of ferret SAR for wild-type influenza isolates and their counterparts engineered using reverse genetics under (A) respiratory droplet and (B) direct contact transmission. Because experiments were not paired, SAR estimates for a wild-type isolate were plotted against the mean SAR for the reverse genetic derived isolate and vice-versa. Estimates from the same isolate are joined by a line with the isolate name given. Both supercritical (red) and subcritical (blue) isolates are shown. The dashed gray line denotes a one-to-one relationship between the two. Note that some points are jittered for clarity (see Figure 1—source data 2, 3 for full data).
DOI: http://dx.doi.org/10.7554/eLife.07969.008
Open in a new tab

When comparing estimates of human and ferret SAR across subtypes, we found that, as expected (Lakdawala and Subbarao, 2012), ferret SAR estimates from current experimental designs do not quantitatively align with human SAR estimates—ferret SAR estimates are typically higher than the corresponding human estimate. However, ferret and human SAR estimates are correlated. For respiratory droplet experiments, the ordering of subtypes by ferret SAR was similar to that in human SAR (Figure 1A,B), and mean ferret respiratory droplet SAR explained 66% of the variation in mean human SAR estimates across subtypes (p = 0.003, Figure 2A). Direct contact transmission in ferrets was not significantly related to human SAR at the subtype level (p = 0.14, Figure 2A), suggesting that for estimates of human-to-human transmissibility, direct contact experiments may have less value than respiratory droplet experiments.

Figure 2. Analysis of subtype-specific SAR.

(A) Comparison of human SAR and ferret SAR for ferret respiratory droplet (black squares) and direct contact (red circles). Data points are the mean human SAR by subtype vs the weighted mean ferret SAR by subtype, where weights are determined by the number of ferrets used in each experiment. Lines give the best fit weighted linear regression models with weights given by the number of human SAR estimates. The solid line indicates a significant relationship between ferret respiratory droplet SAR and human SAR described by the given equation (significant terms are bolded; p = 0.003), while the dashed line indicates a non-significant relationship (p = 0.14) for ferret direct contact transmission. (B) The degree of overlap in the distributions of ferret respiratory droplet SAR estimates for each subtype. Dark purple indicates subtypes with complete overlap, while white indicates no overlap.

DOI: http://dx.doi.org/10.7554/eLife.07969.009

Figure 2.

Figure 2—figure supplement 1. Analysis of subtype-specific SAR including avian isolates for H5N1 and H7N9.

Figure 2—figure supplement 1.

(A) Comparison of human SAR and ferret SAR for ferret respiratory droplet (black squares) and direct contact (red circles). Lines give the best fit weighted linear regression models with weights given by the number of human SAR estimates. The solid line indicates a significant relationship between ferret respiratory droplet SAR and human SAR described by the given equation (significant terms are bolded; p = 0.004), while the dashed line indicates a non-significant relationship (p = 0.18) for ferret direct contact transmission. (B) The degree of overlap in the distributions of ferret respiratory droplet SAR estimates for each subtype. Dark purple indicates subtypes with complete overlap, while white indicates no overlap.
DOI: http://dx.doi.org/10.7554/eLife.07969.010
Open in a new tab

Despite the strong relationship observed between mean ferret and human SAR estimates (Figure 2A), distributions of ferret SAR estimates for each subtype overlapped substantially (Figure 2B). These overlaps prevent the result from any given ferret experiment (e.g., on a novel, uncharacterized strain) from being unambiguously aligned with the transmission potential exhibited by any particular, previously-characterized subtype.

Using ferret SAR to characterize human pandemic potential

To improve the power to assess pandemic potential, we specified two clusters of subtypes with distinct transmission patterns in humans: subtypes with sustained human-to-human transmission (i.e. supercritical; H1N1, H3N2, H2N2 and pH1N1) and subtypes without sustained human-to-human transmission (i.e. subcritical; H7N9, H5N1, H7N7, H7N2, H7N3 and H9N2). Using logistic regression, we identified ranges of ferret SAR that characterize supercritical and subcritical influenza viruses (Figure 3). Ferret respiratory droplet SAR was a significant predictor of the probability that a virus is supercritical or subcritical in humans (p < 0.0001; Figure 3A, Table 1). By accounting for the uncertainty in this relationship, we identified ranges of ferret SAR that indicate a high probability of strains being identified as supercritical or subcritical (Figure 3A). However, a range of intermediate ferret SAR values yielded equivocal results (i.e. the 95% confidence interval for classification included a classification probability of 0.5). Direct contact transmission was also a significant predictor of supercritical or subcritical transmission in humans (p = 0.01; Figure 3B, Table 1). Information theoretic model comparisons showed marginal support for a bivariate model using both respiratory droplet and direct contact transmission data (Table 1). Considering the bivariate distribution of SAR estimates, however, it is clear that respiratory droplet SAR has the potential for greater specificity in predicting supercritical transmission (Figure 3—figure supplement 1).

Figure 3. Weighted logistic regression predicting the probability of a supercritical classification based on ferret SAR.

(A) Respiratory droplet SAR and (B) direct contact SAR. Solid black line gives the fit of the weighted logistic regression, where model weights are given by the number of ferrets in each experiment. Dashed black lines give the 95% confidence interval on the model predictions. Shading in the prediction interval represents values of SAR for which the 95% confidence intervals for predicted model fit do not overlap a probability of 0.5 (the dashed red line) indicating a high probability of being supercritical (red shading) or subcritical (blue shading). The gray shading represents SAR values where the 95% CI on the prediction overlaps 0.5, providing equivocal classification. Circles show the individual ferret SAR estimates (See Figure 1—source data 2, 3) for supercritical (top in red) and subcritical viruses (bottom in blue).

DOI: http://dx.doi.org/10.7554/eLife.07969.011

Figure 3.

Figure 3—figure supplement 1. Comparison of ferret SAR via respiratory droplet and direct contact transmission for single influenza isolates.

Figure 3—figure supplement 1.

Each point represents a single set of experiments that tested an isolate for transmission in ferrets under both respiratory droplet and direct contact transmission with other experimental protocols held fixed. Isolates belonging to subcritical subtypes are depicted by blue squares, and supercritical subtypes are depicted by red circles. Note that some points are jittered for clarity (see Figure 1—source data 2, 3 for full data).
DOI: http://dx.doi.org/10.7554/eLife.07969.012
Figure 3—figure supplement 2. Effect of uncertainty in ferret SAR on its relationship with the probability of being classified as supercritical.

Figure 3—figure supplement 2.

(A) Respiratory droplet SAR and (B) direct contact SAR. To assess the impact of binomial uncertainty in ferret SAR estimates, we simulated 1000 datasets by taking binomial samples from each data point. Here, the binomial probability for each was given by the observed SAR and the number of trials was the number of ferrets used. To introduce binomial uncertainty into those experiments with an SAR of 0 or 1, we set the binomial probability to 0.1 or 0.9, respectively. The solid line is the average model fit to all of the simulated datasets and is nearly identical to that in Figure 3. Dashed lines give the 97.5 and 0.025 percentiles of the upper and lower bounds, respectively, of the 95% confidence intervals on model predictions from each of the simulated datasets. These indicate much more uncertainty in model predictions across datasets that generates a larger equivocal region of ferret SARs than observed in Figure 3. However, values of ferret SAR indicative of subcritical and supercritical strains still exist, indicating that our qualitative results are robust to binomial uncertainty.
DOI: http://dx.doi.org/10.7554/eLife.07969.013
Figure 3—figure supplement 3. ROC curves for classifying pandemic potential using different definitions of transmission and transmission routes.

Figure 3—figure supplement 3.

Receiver operating characteristic (ROC) curves and area under the curve (AUC) using (A) seroconversion and/or viral isolation or (B) viral isolation alone as evidence for transmission in ferrets when classifying influenza isolates as either supercritical or subcritical in humans. Lines indicate ferret respiratory droplet SAR (red) or ferret direct contact SAR (black). Curves were calculated from raw data shown in Figure 3, using a range of SAR classification thresholds from 0 to 1. Numbers indicate the threshold values for which the true positive rate (i.e. the sensitivity) and false positive rate (i.e. the complement of the specificity) changed. Threshold values intermediate to any of those depicted have true positive and false positive rates identical to that of the next lowest value shown. The dashed gray line is the one-to-one line corresponding to random classification. AUC values are shown in the figure legend with higher values corresponding to higher predictive power.
DOI: http://dx.doi.org/10.7554/eLife.07969.014
Open in a new tab

Table 1.

Parameter estimates for the weighted logistic regression relating human transmission class to ferret SAR

DOI: http://dx.doi.org/10.7554/eLife.07969.015

Data Model β0 βRD βDC ΔAIC
Full data Direct contact −4.39 - 6.30 -
Respiratory droplet −3.52 6.10 - -
Restricted data Respiratory droplet + direct contact −1.76 8.72 −3.76 0
Respiratory droplet −3.77 6.42 - 3.623
Direct contact −3.07 - 3.74 57.348
Open in a new tab

Bolded estimates are significant at the α = 0.05 level. Due to differing data between ferret respiratory droplet and direct contact transmission experiments, no model selection was done on the full data. Instead, model selection was done only for studies where authors performed respiratory droplet and direct contact transmission experiments on the same isolate.

The classification thresholds we identified for likely supercritical or subcritical subtypes account for uncertainties arising from the structure of our model, but not for uncertainties arising from the experimental data used to inform the model. Binomial uncertainties in ferret SAR data can be substantial, as ethical and logistic considerations limit sample sizes in these experiments (Nishiura et al., 2013). By re-fitting our logistic regression model to 1000 simulated datasets generated by binomial re-sampling of each data point, we found that the relationship between ferret SAR and a supercritical classification is quite robust to this uncertainty (Figure 3—figure supplement 2). However, while our analysis was fairly insensitive to binomial uncertainty within the aggregate data, attempts to classify SAR estimates from any individual experiment will be more sensitive to binomial uncertainty. For example, we applied our model to the most transmissible strains from two recent GOF studies on H5N1 avian influenza (Imai et al., 2012; Herfst et al., 2012; Figure 4—source data 1). All three strains had a ferret SAR that fell into the supercritical range, but the confidence intervals for the SAR estimates overlapped with the subcritical and/or equivocal ranges, preventing definitive classification (Figure 4A). Similarly, we found that studies on 1918 pandemic H1N1, a known pandemic strain, had ferret SAR estimates indicative of supercritical transmission, but again wide confidence intervals overlapped the subcritical and equivocal ranges (Figure 4B). SAR estimates for H7N9, known to be subcritical in humans, spanned the supercritical, subcritical, and equivocal ranges (Figure 4C). Even if results across all ferret respiratory droplet trials for H7N9 were aggregated into a single SAR estimate (representing 42 ferrets in all), we found an equivocal classification of human transmission pattern (Figure 4C). Consequently, care must be taken to avoid over-interpreting the results of ferret transmission studies.

Figure 4. Predictions of the transmission pattern for current and historical isolates of concern.

(A) Gain-of-function experiments with H5N1 avian influenza (Herfst et al., 2012; Imai et al., 2012), (B) the reconstructed 1918 pandemic H1N1 strain (Tumpey et al., 2007; Imai et al., 2012), and (C) H7N9 avian influenza. Solid black curves and shading represent the logistic regression fit and likely transmission pattern, respectively, as depicted in Figure 2. Horizontal lines give the 95% Wilson-score interval for each binomial estimate. In all panels, transmission is defined using seroconversion and viral isolation in nasal washes. In (C), green triangles represent individual experiments, while the green square is the aggregated data across all twelve H7N9 transmission experiments in ferrets. Notice that 6 data points are represented at a SAR of 0.33 and 3 at a SAR of 1. See Figure 1—source data 2 and Figure 4—source data 1 for full data.

DOI: http://dx.doi.org/10.7554/eLife.07969.016

Figure 4—source data 1. Ferret influenza transmission studies via respiratory droplets using strains from gain-of-function experiments with H5N1 avian influenza and the reconstructed 1918 pandemic H1N1 strain.
DOI: http://dx.doi.org/10.7554/eLife.07969.017
elife07969s005.docx (21KB, docx)
DOI: 10.7554/eLife.07969.017

Figure 4.

Figure 4—figure supplement 1. Sample size calculations.

Figure 4—figure supplement 1.

Sample size estimates to achieve 80% power at a significance level of 0.05 when testing whether a ferret SAR estimate is greater than the lower limit of the identified supercritical window (0.643, Figure 3A). Sample sizes were calculated using a one-sided binomial exact test.
DOI: http://dx.doi.org/10.7554/eLife.07969.018
Open in a new tab

Discussion

For the first time, we have demonstrated a quantitative link between estimates of transmission efficiency of influenza among ferrets and among humans, at the subtype level. However, there is little power to resolve human SAR using ferret SAR estimates from single experiments. Instead, we observed ranges of ferret SAR distinguishing supercritical from subcritical subtypes that may be useful in identifying influenza viruses that pose greater or lesser risk of pandemic spread—especially for viruses with very high or low ferret SAR. In all analyses, including comparisons of sensitivity and false positive rate of various classification thresholds (Figure 3—figure supplement 3), we found that respiratory droplet transmission in ferrets was a better indicator of transmission in humans than direct contact transmission. However, direct contact experiments used in conjunction with respiratory droplet experiments can provide additional information on transmission in humans.

Sample size is a serious challenge to operational use of the results shown here. The largest sample size we found in our review of transmission studies was twelve ferrets (Herlocher et al., 2002). Even for a supercritical strain with an assumed ferret SAR of 1, 8 ferrets must be tested to classify that strain as supercritical with 80% power at a significance level of 0.05 (Figure 4—figure supplement 1). For an assumed ferret SAR of 0.8—more in line with zoonotic strains of interest (Figure 4), but closer to the lower end of the supercritical range—achieving the same power would require more than 30 ferrets (Figure 4—figure supplement 1). Such a sample size is obviously prohibitive. It is important to note, though, that data from future experiments should refine the relationship in Figure 3, expanding the ranges corresponding to subcritical and supercritical transmission, and hence lowering the sample size requirements somewhat.

Other design changes could also enhance the value of ferret transmission experiments for informing risk assessments. In particular, it is vital to standardize experimental design in order to reduce noise and strengthen inference, beginning with establishing standard definitions of transmission for ferret experiments (i.e. viral titers in nasal washes vs serologic evidence). Discord between viral isolation and antibody data within a single experiment highlights this need and shows that serologic data is often a more sensitive metric of pathogen exposure (Figure 1—source data 2, 3). It has been questioned whether seroconversion always reflects a productive viral infection, but recent imaging studies indicate that seroconversion can detect infections that manifest deep in the respiratory tract, which would be missed by nasal wash measurements (Karlsson et al., 2015). Although all of our results were robust to the choice of transmission definition (‘Results’ not shown), defining transmission by viral isolation alone slightly increased predictive power for direct contact experiments, and slightly decreased predictive power for respiratory droplet experiments (Figure 3—figure supplement 3). Ultimately, this suggests that transmission should be assessed using both serological and viral data to aid in comparisons across experiment types, while allowing for exploration of exposure vs active infection. Dosing protocols can also vary widely across and within studies, in terms of viral titer and volume and even incompatible units. Standardized dosing protocols could reduce variability in ferret SAR estimates substantially. Additional data on time to infection, clinical signs, and mechanistic insights such as receptor binding affinities, none of which are systematically collected under standard protocols, could add value to ferret studies by giving additional power to differentiate among influenza viruses and subtypes with similar transmission outcomes.

Despite these challenges, ferret transmission experiments can contribute distinctive insights into the pandemic potential of novel influenza isolates. Our results show that ferret experiments provide a tool with relatively high sensitivity and specificity for identifying strains that may be supercritical in humans (Figure 3—figure supplement 3). Based on current scientific knowledge, risk screening might also incorporate high-throughput virologic and genetic screens used to identify isolates of concern by looking for genetic changes associated with altered binding affinities and other markers of transmission in mammals (Russell et al., 2012). Ultimately, however, human transmission is a complex and partially understood phenotype that is difficult to predict using these initial screens (Russell et al., 2014). Ferrets can provide a potential link between underlying virologic and genetic changes and potential transmissibility in humans. Future analyses should attempt to simultaneously incorporate data on the presence of specific mutations (e.g., PB2-K627E, Van Hoeven et al., 2009) and virologic factors (e.g., binding to α2-6 sialic acid glycans, Belser et al., 2008) into the present analysis of ferret transmissibility to determine if these genetic and virologic screens provide additional information on human transmission not captured by ferrets alone.

The resolution of our analysis was limited to the subtype level, because human transmissibility data are not available for more specific strains. Some caution is needed when assessing transmission of novel isolates within a subtype, as out-of-sample predictions under this scenario are likely to be particularly hazardous. As a first assessment of the effect of within-subtype variation, we re-ran our analyses in Figure 2 for a broader dataset including H5N1 and H7N9 strains isolated from avian hosts (Figure 1—source data 4; Figure 2—figure supplement 1). The results were consistent with our main findings, giving some confidence that our results are robust to such within-subtype variation (Figure 1—source data 4; Figure 2—figure supplement 1). Additional data on consensus viral sequences within human outbreaks are needed to relate human SAR estimates more specifically to isolates tested in ferrets and clarify the effect of within-subtype variation on predicted human transmission behavior. In the absence of these data, our analysis represents a new null model against which deviations within subtypes can be measured to identify strains that can provide additional information on the molecular features associated with transmissible phenotypes in ferrets and/or humans.

Recently, the obvious disparity between highly efficient H7N9 transmission in some ferret experiments and inefficient H7N9 transmission in humans (see Figure 2B, Figure 4C) led to questions about the general validity of the ferret transmission model (Lipsitch, 2013). Our results at least partially assuage these concerns. In spite of the substantial variation we observed within H5N1 and H7N9 subtypes, our results show that, statistically, isolates more transmissible in ferrets are more likely to be capable of sustained transmission in humans. Yet our data also demonstrate that the ferret transmission model is fallible: for H7N9, an emerging virus of great concern, ferret transmission experiments sometimes yield results that obviously contradict observed patterns in humans. These results are anomalous within the general mapping of ferret transmissibility to human transmissibility and thus, as mentioned previously, may present an opportunity to gain new insight into the molecular drivers of this complex phenotype. However, when screening emerging influenza viruses for pandemic potential, both false negatives and false positives have important consequences for health policy decisions. The deviations of H7N9 from the general correlation between human and ferret transmissibility underscore the importance of corroborating transmission estimates from the ferret model with other lines of evidence. The ultimate evidence to corroborate human transmission comes from epidemiological patterns of infection in humans. For a true pandemic influenza virus, however, such data are likely to come too late, highlighting the need for reliable methods to provide early warning on strains with pandemic potential.

Here we have put forward the first guidelines for translating the results of ferret experiments into a measure of pandemic potential in humans. Given the continued use of ferrets in other areas of influenza research (e.g., vaccine development), this finding enhances the broad value of ferret experiments. However, given pragmatic limitations on sample sizes in ferret studies, uncertainties in ferret SAR estimates are likely to limit the operational utility of these guidelines. This coupled with the biological complexities underlying transmissibility suggests that, at this time, ferret transmission data provide a valuable but imperfect correlate of human transmissibility, and further evidence is needed to assess whether other lines of evidence can improve this predictive capacity.

Materials and methods

Secondary attack rates

Most ferret transmission studies report the number of secondary infections amongst a specified number of naïve ferrets that are exposed to single inoculated individuals. This enables calculation of the SAR, which is the probability of infection for a susceptible individual following a known contact with an infectious individual (Halloran, 2005) and establishes a metric of transmissibility in ferrets that is directly comparable to household SAR in humans.

We obtained estimates of SAR in humans from household contact data using two methods. Ad hoc SAR estimates are obtained by taking the ratio of infected household contacts over total household contacts. This method is widely used, but may overestimate SAR, as it assumes each household experiences only one disease introduction (the index case) and ignores the possibility of multiple household exposures to an exogenous reservoir (Longini et al., 1982). Meanwhile, maximum likelihood procedures for SAR estimation use statistical models to simultaneously estimate the probability of secondary transmission within a household (SAR) and the probability of infection from the community (or other source). Thus, these estimates attempt to correct for the possibility of multiple introductions from an exogenous source (Longini and Koopman, 1982; Longini et al., 1982). However, even these estimates can be strongly skewed by the inclusion or exclusion of specific clusters, especially early in an outbreak when data is limited (Aditama et al., 2012). Furthermore, variation in existing, population-level immunity to specific strains, and the use of different case ascertainment methods in specific studies also inevitably skew estimates made using either procedure. Because each method has unique biases and limitations, we used published estimates of SAR based on either method, or calculated an ad hoc SAR estimate ourselves from data on the total and infected number of household contacts in an outbreak. Human SAR estimates are only considered in our initial regression analysis (Figure 2A), so they do not influence our classification model (Figure 3).

Literature review

To assess the relationship between human and ferret transmissibility of influenza, we reviewed existing estimates of subtype-specific SAR in humans and ferrets. We searched PubMed and Web Of Science [v5.15] databases using the following queries: (influenza AND household AND transmission AND H#N#) and (influenza AND ‘secondary attack rate’ OR SAR AND human AND H#N#) for human studies and (influenza AND transmission AND ferret* AND H#N#) for ferret studies. We repeated searches for subtypes H1N1, H7N9, H3N2, H7N7, H7N9, H7N2, H9N2, H5N1, H7N3, and H2N2. To ensure comprehensive coverage, additional studies were identified using reference lists from search results and additional spot searches were also conducted. We excluded isolates that represented outliers from identified subtypes (i.e. 1918 pandemic H1N1 (Tumpey et al., 2007) and novel swine-origin H3N2 in 2009–10 (Pearce et al., 2012)). Searches were completed on 20 July 2015.

Although the transmission potential of unique isolates within a subtype may vary, SAR in humans was reported only at the subtype level, preventing us from analyzing isolate-specific transmission potential. Overall, we found data for all three measures (ferret direct contact, ferret respiratory droplet, and human SARs) for ten influenza A subtypes: H7N3, H9N2, H7N7, H7N2, H5N1, H7N9, H2N2, pH1N1 (i.e. influenza A(H1N1)pdm09 virus), H3N2, and seasonal H1N1 (Figure 1).

Inclusion criteria for ferret studies

We excluded ferret transmission studies that included serial passage of human isolates in ferrets prior to transmission experiments. To maintain consistency in transmission mechanisms, we excluded studies that inoculated ferrets by routes other than intranasal with a liquid inoculum (e.g., ocular inoculation or aerosol inhalation) and that inoculated ferrets with a lower viral dose than was typical for ferret transmission studies (<103 50% egg infectious dose [EID50]). We excluded studies where naive ferrets were not exposed to inoculated ferrets at 1 day post-inoculation, as was standard, and studies where the duration of contact was restricted. We also excluded trials in which ferrets were vaccinated or administered antiviral drugs for treatment or prophylaxis. If transmission of more than one subtype and/or isolate was tested in a single study (using different sets of immunologically naive ferrets for each isolate), we treated each subtype/isolate-specific data point separately. However, for some analyses, we grouped data from isolates belonging to the same subtype—the one exception being separation of 2009 pandemic H1N1 isolates (pH1N1) and pre-2009 H1N1 isolates (H1N1).

We distinguished between direct contact transmission experiments (in which sentinel ferrets were co-housed with the donor ferret) and respiratory droplet transmission experiments (in which ferrets were housed in adjacent cages designed to allow for airborne exchange, but in which direct or indirect contact between sentinels and donors is not possible). Transmission amongst ferrets was determined in each study using either a viral titer in nasal washes or a positive serologic test (i.e. hemagglutination inhibition assay) or by a combination of both tests. We noted any discrepancies between the two transmission mechanisms (Figure 1—source data 2, Figure 1—source data 3) and conducted analyses that showed our results were relatively robust to the transmission definition used (Figure 3—figure supplement 3).

To promote quality of comparison between ferret and human studies, we only included data from ferret studies that tested one or more wild-type human isolates. While avian and other animal isolates maintain close sequence homology with human isolates (Claas et al., 1998), the transmission of animal isolates into humans is associated with genetic bottlenecks (Zaraket et al., 2015) and considerable within-host adaptation (Linster et al., 2014). These evolutionary barriers lead to avian precursors that have lower mortality in mice, less morbidity in ferrets, and lower viral titers in human epithelial cells (Belser et al., 2013; Watanabe et al., 2014; Zaraket et al., 2015). Thus, these cross-species and within-host barriers have the potential to obscure the relationship between transmission in ferrets and transmission in humans, and we excluded avian and other animal strains from the main analysis as a result. We did, however, compile a database of ferret transmission experiments using avian isolates from subtypes H5N1 and H7N9 (Figure 1—source data 4) to test the validity of this exclusion. Avian isolates in these subtypes have the benefit of contemporary sampling in both space and time with their human counterparts. Supplementary analyses including these avian isolates showed that our results were robust to the exclusion of non-human isolates (Figure 2—figure supplement 1).

We also included wild-type isolates from humans generated using reverse genetics techniques. Although viral isolates rescued through reverse genetic techniques are often assumed to have lower transmissibility, analyses with and without these rescued isolates yielded indistinguishable results. Indeed, for the small number of isolates for which we could make direct comparisons, isolates generated using reverse genetics exhibited similar transmissibility to their wild-type counterparts (Figure 1—figure supplement 1). Thus, our data set contained a total of 81 respiratory droplet (Figure 1B; Figure 1—source data 2) and 76 direct contact transmission trials (Figure 1C; Figure 1—source data 3).

Inclusion criteria for human studies

Because we considered only household SAR, we excluded studies with non-standard household definitions (e.g., dormitories, health care centers, summer camps), and studies where household contacts could not be distinguished from broader community contacts. We also excluded data from studies of zoonotic strains where prior contact with potential livestock or wildlife reservoirs was noted for multiple contacts, thus hindering the distinction between primary and secondary cases. In order to represent a broad range of human SAR estimates, we included both prospective and retrospective household studies that either provided an explicit SAR estimate or reported data sufficient to calculate a SAR. This yielded a total of 83 estimates of human SAR (Figure 1A; Figure 1—source data 1).

Analysis

Quantitative comparison of SAR in ferrets and SAR in humans was performed using linear regression (Figure 2A). Because human SAR estimates are not typically made for individual isolates, the comparison was done at the subtype level using the mean value of all estimates belonging to a subtype. For ferret experiments, we used a weighted mean by subtype, where the weights were given by the number of ferrets used in each experiment; for human estimates, we used the simple mean by subtype. The potential uncertainty in subtype mean SAR was large, especially for human SAR, where several emerging subtypes (i.e. H7N3, H9N2, H7N7, and H7N2) only had one or two estimates (Figure 1A). To allow for this uncertainty, we used a weighted linear regression with model weights given by the number of human SAR estimates.

To create Figure 2B, we developed an empirical measure for the overlap between distributions of ferret SAR estimates for pairs of subtypes that was a simple variant of other overlap indices used in ecology (Ricklefs and Lau, 1980). This was calculated by comparing the more transmissible and less transmissible of each of the subtypes, taking the minimum SAR estimate for the more transmissible subtype and the maximum SAR estimate for the less transmissible subtype, and counting the number of estimates for both subtypes that fell within this range of overlap (normalized by the total number of estimates for both subtypes). This yielded a measure between 0 and 1, where zero indicated that the ranges of observed SAR estimates for two subtypes were completely distinct and one indicated that the ranges completely overlapped, rendering the subtypes indistinguishable on the basis of SAR.

Examination of Figure 2B revealed two distinct clusters of subtypes whose distributions of SAR estimates overlapped almost completely. H1N1, H3N2, pH1N1, and H2N2 are supercritical subtypes with sustained transmission among humans; H7N2, H7N7, H9N2, and H7N3 are subcritical subtypes with weak transmission among humans. This grouping suggests there may be potential to use ferret SAR estimates for broader functional classification of viruses with or without pandemic potential. However, two subtypes of concern, H7N9 and H5N1 (both known to be subcritical in humans), were anomalies within the natural clusters we observed in Figure 2B: H7N9 clustered with supercritical subtypes, while H5N1 was weakly associated with both groups. We interpreted this as important biological variation within the group of subcritical subtypes, but considering the overarching interest in predicting whether particular subtypes might have pandemic potential, for our further analyses we chose to group subtypes according to their observed transmission pattern in humans (i.e. supercritical—H1N1, H3N2, pH1N1, H2N2 and subcritical—H7N3, H9N2, H7N7, H7N2, H5N1, H7N9).

To determine how ferret SAR was related to supercritical and subcritical classifications, we used a weighted logistic regression (Figure 3). Here, model weights are based on the number of ferrets used in each experiment, thus allowing for more confidence in estimates with larger numbers of ferrets. Ferret SAR estimates that corresponded to a high probability of an isolate being classified as supercritical or subcritical (i.e. low probability of being supercritical) were determined by calculating 95% confidence intervals for the predicted model fit and identifying the ranges where these confidence intervals were either wholly above (supercritical) or wholly below (subcritical) a value of 0.5 (representing a random guess of supercritical or not). The sensitivity and specificity of various thresholds in ferret SAR were also assessed (Figure 3—figure supplement 3).

All analyses were done using R Statistical Software version 3.1.2 (R Development Core Team, 2014).

Acknowledgements

MGB and JOL-S are supported by the Research and Policy for Infectious Disease Dynamics (RAPIDD) program of the Science and Technology Directorate, Department of Homeland Security, and Fogarty International Center, National Institutes of Health. KG is supported by National Institutes of Health under the Ruth L Kirschstein National Research Service Award (T32-GM008185). JOL-S is supported by the National Science Foundation (EF-0928690) and the De Logi Chair in Biological Sciences. MP is supported by the National Science Foundation Graduate Research Fellowship under Grant No. (DGE-1144087). Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation, the National Institutes of Health or the Centers for Disease Control and Prevention.

Funding Statement

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Funding Information

This paper was supported by the following grants:

  • National Institutes of Health (NIH)/Fogarty International Center (FIC) Research and Policy for Infectious Disease Dynamics (RAPIDD) to Michael G Buhnerkempe, James O Lloyd-Smith.

  • National Institutes of Health (NIH) Ruth L. Kirschstein National Research Service Award (T32-GM008185) to Katelyn Gostic.

  • National Science Foundation (NSF) DGE-1144087 to Miran Park.

  • National Science Foundation (NSF) EF-0928690 to James O Lloyd-Smith.

  • University of California, Los Angeles (UCLA) De Logi Chair in Biological Sciences to James O Lloyd-Smith.

Additional information

Competing interests

The authors declare that no competing interests exist.

Author contributions

MGB, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article.

KG, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article.

MP, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article.

PA, Conception and design, Acquisition of data, Drafting or revising the article.

JAB, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article.

JOL-S, Conception and design, Analysis and interpretation of data, Drafting or revising the article.

References

  1. Aditama TY, Samaan G, Kusriastuti R, Sampurno OD, Purba Misriyah W, Santoso H, Bratasena A, Maruf A, Sariwati E, Setiawaty V, Glass K, Lokuge K, Kelly PM, Kandun IN. Avian influenza H5N1 transmission in households, Indonesia. PLOS ONE. 2012;7:e29971. doi: 10.1371/journal.pone.0029971. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Belser JA, Blixt O, Chen L-M, Pappas C, Maines TR, Van Hoeven N, Donis R, Busch J, McBride R, Paulson JC, Katz JM, Tumpey TM. Contemporary North American influenza H7 viruses possess human receptor specificity: Implications for virus transmissibility. Proceedings of the National Academy of Sciences of USA. 2008;105:7558–7563. doi: 10.1073/pnas.0801259105. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Belser JA, Gustin KM, Pearce MB, Maines TR, Zeng H, Pappas C, Sun X, Carney PJ, Villanueva JM, Stevens J, Katz JM, Tumpey TM. Pathogenesis and transmission of avian influenza A (H7N9) virus in ferrets and mice. Nature. 2013;501:556–559. doi: 10.1038/nature12391. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Blumenkrantz D, Roberts KL, Shelton H, Lycett S, Barclay WS. The short stalk length of highly pathogenic avian influenza H5N1 virus neuraminidase limits transmission of pandemic H1N1 virus in ferrets. Journal of Virology. 2013;87:10539–10551. doi: 10.1128/JVI.00967-13. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Bouvier NM, Lowen AC. Animal models for influenza virus pathogenesis and transmission. Viruses. 2010;2:1530–1563. doi: 10.3390/v20801530. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Casadevall A, Howard D, Imperiale MJ. Reply to “can limited scientific value of potential pandemic pathogen experiments justify the risks?”. mBio. 2014;5:e02053–14. doi: 10.1128/mBio.02053-14. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Casadevall A, Imperiale MJ. Risks and benefits of gain-of-function experiments with pathogens of pandemic potential, such as influenza virus: a call for a science-based discussion. mBio. 2014;5:e01730–14. doi: 10.1128/mBio.01730-14. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Chan P. Outbreak of avian influenza A (H5N1) virus infection in Hong Kong in 1997. Clinical Infectious Diseases. 2002;34:S58–S64. doi: 10.1086/338820. [DOI] [PubMed] [Google Scholar]
  9. Claas ECJ, Osterhaus AD, van Beek R, De Jong JC, Rimmelzwaan GF, Senne DA, Krauss S, Shortridge KF, Webster RG. Human influenza A H5N1 virus related to a highly pathogenic avian influenza virus. Lancet. 1998;351:472–477. doi: 10.1016/S0140-6736(97)11212-0. [DOI] [PubMed] [Google Scholar]
  10. Duan S, Boltz DA, Seiler P, Li J, Bragstad K, Nielsen LP, Richard J, Webster RG, Govorkova A. Competitive transmissibility and fitness of oseltamivir- sensitive and resistant pandemic influenza H1N1 viruses in ferrets. Influenza and Other Respiratory Viruses. 2011;5:79–89. [PubMed] [Google Scholar]
  11. Fouchier RA, Kawaoka Y, Cardona C, Compans RW, García-Sastre A, Govorkova EA, Guan Y, Herfst S, Orenstein WA, Peiris JSM, Perez DR, Richt JA, Russell C, Schultz-Cherry SL, Smith DJ, Steel J, Tompkins SM, Topham DJ, Treanor JJ, Tripp RA, Webby RJ, Webster RG. Avian flu: gain-of-function experiments on H7N9. Nature. 2013;500:150–151. doi: 10.1038/500150a. [DOI] [PubMed] [Google Scholar]
  12. Gao R, Cao B, Hu Y, Feng Z, Wang D, Hu W, Chen J, Jie Z, Qiu H, Xu K, Xu X, Lu H, Zhu W, Gao Z, Xiang N, Shen Y, He Z, Gu Y, Zhang Z, Yang Y, Zhao X, Zhou L, Li X, Zou S, Zhang Y, Li X, Yang L, Guo J, Dong J, Li Q, Dong L, Zhu Y, Bai T, Wang S, Hao P, Yang W, Zhang Y, Han J, Yu H, Li D, Gao GF, Wu G, Wang Y, Yuan Z, Shu Y. Human infection with a novel avian-origin influenza A (H7N9) virus. The New England Journal of Medicine. 2013;368:1888–1897. doi: 10.1056/NEJMoa1304459. [DOI] [PubMed] [Google Scholar]
  13. Gong Z, Lv H, Ding H, Han J, Sun J, Chai C, Cai J, Yu Z, Chen E. Epidemiology of the avian influenza A (H7N9) outbreak in Zhejiang Province, China. BMC Infectious Diseases. 2014;14:244. doi: 10.1186/1471-2334-14-244. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Halloran ME. Secondary attack rate. In: Armitage P, Colton T, editors. Encyclopedia of biostatistics. Chichester, UK: John Wiley & Sons, Ltd; 2005. [Google Scholar]
  15. Hamelin ME, Baz M, Bouhy X, Beaulieu E, Dubé K, Mallett C, Boivin G. Reduced airborne transmission of oseltamivir-resistant pandemic A/H1N1 virus in ferrets. Antiviral Therapy. 2011;16:775–779. doi: 10.3851/IMP1794. [DOI] [PubMed] [Google Scholar]
  16. Herfst S, Schrauwen EJA, Linster M, Chutinimitkul S, de Wit E, Munster VJ, Sorrell EM, Bestebroer TM, Burke DF, Smith DJ, Rimmelzwaan GF, Osterhaus ADME, Fouchier RAM. Airborne transmission of influenza A/H5N1 virus between ferrets. Science. 2012;336:1534–1541. doi: 10.1126/science.1213362. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Herlocher ML, Carr J, Ives J, Elias S, Truscon R, Roberts N, Monto AS. Influenza virus carrying an R292K mutation in the neuraminidase gene is not transmitted in ferrets. Antiviral Research. 2002;54:99–111. doi: 10.1016/S0166-3542(01)00214-5. [DOI] [PubMed] [Google Scholar]
  18. Herlocher ML, Truscon R, Elias S, Yen HL, Roberts NA, Ohmit SE, Monto AS. Influenza viruses resistant to the antiviral drug oseltamivir: transmission studies in ferrets. The Journal of Infectious Diseases. 2004;190:1627–1630. doi: 10.1086/424572. [DOI] [PubMed] [Google Scholar]
  19. Hurt AC, Nor’e SS, McCaw JM, Fryer HR, Mosse J, McLean AR, Barr IG. Assessing the viral fitness of oseltamivir-resistant influenza viruses in ferrets, using a competitive-mixtures model. Journal of Virology. 2010;84:9427–9438. doi: 10.1128/JVI.00373-10. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Imai M, Watanabe T, Hatta M, Das SC, Ozawa M, Shinya K, Zhong G, Hanson A, Katsura H, Watanabe S, Li C, Kawakami E, Yamada S, Kiso M, Suzuki Y, Maher EA, Neumann G, Kawaoka Y. Experimental adaptation of an influenza H5 HA confers respiratory droplet transmission to a reassortant H5 HA/H1N1 virus in ferrets. Nature. 2012;486:420–428. doi: 10.1038/nature10831. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Itoh Y, Shinya K, Kiso M, Watanabe T, Sakoda Y, Hatta M, Muramoto Y, Tamura D, Sakai-Tagawa Y, Noda T, Sakabe S, Imai M, Hatta Y, Watanabe S, Li C, Yamada S, Fujii K, Murakami S, Imai H, Kakugawa S, Ito M, Takano R, Iwatsuki-Horimoto K, Shimojima M, Horimoto T, Goto H, Takahashi K, Makino A, Ishigaki H, Nakayama M, Okamatsu M, Takahashi K, Warshauer D, Shult PA, Saito R, Suzuki H, Furuta Y, Yamashita M, Mitamura K, Nakano K, Nakamura M, Brockman-Schneider R, Mitamura H, Yamazaki M, Sugaya N, Suresh M, Ozawa M, Neumann G, Gern J, Kida H, Ogasawara K, Kawaoka Y. In vitro and in vivo characterization of new swine-origin H1N1 influenza viruses. Nature. 2009;460:1021–1025. doi: 10.1038/nature08260. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Karlsson EA, Meliopoulos VA, Savage C, Livingston B, Mehle A, Schultz-Cherry S. Visualizing real-time influenza virus infection, transmission and protection in ferrets. Nature Communications. 2015;6:6378. doi: 10.1038/ncomms7378. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Kiso M, Shinya K, Shimojima M, Takano R, Takahashi K, Katsura H, Kakugawa S, Le MTQ, Yamashita M, Furuta Y, Ozawa M, Kawaoka Y. Characterization of oseltamivir-resistant 2009 H1N1 pandemic influenza A viruses. PLOS pathogens. 2010;6:e1001079. doi: 10.1371/journal.ppat.1001079. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Lakdawala SS, Subbarao K. The ongoing battle against influenza: the challenge of flu transmission. Nature Medicine. 2012;18:1468–1470. doi: 10.1038/nm.2953. [DOI] [PubMed] [Google Scholar]
  25. Li KS, Guan Y, Wang J, Smith GJ, Xu KM, Duan L, Rahardjo AP, Puthavathana P, Buranathai C, Nguyen TD, Estoepangestie AT, Chaisingh A, Auewarakul P, Long HT, Hanh NT, Webby RJ, Poon LL, Chen H, Shortridge KF, Yuen KY, Webster RG, Peiris JS. Genesis of a highly pathogenic and potentially pandemic H5N1 influenza virus in eastern Asia. Nature. 2004;430:209–213. doi: 10.1038/nature02746. [DOI] [PubMed] [Google Scholar]
  26. Li Q, Zhou L, Zhou M, Chen Z, Li F, Wu H, Xiang N, Chen E, Tang F, Wang D, Meng L, Hong Z, Tu W, Cao Y, Li L, Ding F, Liu B, Wang M, Xie R, Gao R, Li X, Bai T, Zou S, He J, Hu J, Xu Y, Chai C, Wang S, Gao Y, Jin L, Zhang Y, Luo H, Yu H, He J, Li Q, Wang X, Gao L, Pang X, Liu G, Yan Y, Yuan H, Shu Y, Yang W, Wang Y, Wu F, Uyeki TM, Feng Z. Epidemiology of human infections with avian influenza A(H7N9) virus in China. The New England Journal of Medicine. 2014;370:520–532. doi: 10.1056/NEJMoa1304617. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Linster M, van Boheemen S, de Graaf M, Schrauwen EJA, Lexmond P, Mänz B, Bestebroer TM, Baumann J, van Riel D, Rimmelzwaan GF, Osterhaus ADME, Matrosovich M, Fouchier RA, Herfst S. Identification, characterization, and natural selection of mutations driving airborne transmission of A/H5N1 virus. Cell. 2014;157:329–339. doi: 10.1016/j.cell.2014.02.040. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Lipsitch M. Ferret H7N9 flu model questioned. Nature. 2013;501:33. doi: 10.1038/501033e. [DOI] [PubMed] [Google Scholar]
  29. Lipsitch M. Can limited scientific value of potential pandemic pathogen experiments justify the risks? mBio. 2014;5:e02008–14. doi: 10.1128/mBio.02008-14. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Lipsitch M, Galvani AP. Ethical alternatives to experiments with novel potential pandemic pathogens. PLOS Medicine. 2014;11:e1001646. doi: 10.1371/journal.pmed.1001646. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Longini IM, jr, Koopman JS. Household and community transmission parameters from final distributions of infections in households. Biometrics. 1982;38:115–126. http://www.jstor.org/stable/2530294. [PubMed] [Google Scholar]
  32. Longini IM, Koopman JS, Monto AS, Fox JP. Estimating household and community transmission parameters for influenza. American Journal of Epidemiology. 1982;115:736–751. doi: 10.1093/oxfordjournals.aje.a113356. [DOI] [PubMed] [Google Scholar]
  33. Maher J, DeStefano J. The ferret: an animal model to study influenza virus. Lab Animal. 2004;33:50–53. doi: 10.1038/laban1004-50. [DOI] [PubMed] [Google Scholar]
  34. Maines TR, Chen LM, Matsuoka Y, Chen H, Rowe T, Ortin J, Falcón A, Nguyen TH, Mai LQ, Sedyaningsih ER, Harun S, Tumpey TM, Donis RO, Cox NJ, Subbarao K, Katz JM. Lack of transmission of H5N1 avian-human reassortant influenza viruses in a ferret model. Proceedings of the National Academy of Sciences of USA. 2006;103:12121–12126. doi: 10.1073/pnas.0605134103. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Morens DM, Subbarao K, Taubenberger JK. Engineering H5N1 avian influenza viruses to study human adaptation. Nature. 2012;486:335–340. doi: 10.1038/nature11170. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Nishiura H, Yen H-L, Cowling BJ. Sample size considerations for one-to-one animal transmission studies of the influenza A viruses. PLOS ONE. 2013;8:e55358. doi: 10.1371/journal.pone.0055358. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Oh DY, Lowther S, McCaw JM, Sullivan SG, Leang S-K, Haining J, Arkinstall R, Kelso A, Mcvernon J, Barr IG, Middleton D, Hurt AC. Evaluation of oseltamivir prophylaxis regimens for reducing influenza virus infection, transmission and disease severity in a ferret model of household contact. The Journal of Antimicrobial Chemotherapy. 2014;69:2458–2469. doi: 10.1093/jac/dku146. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Palese P, Wang TT. H5N1 influenza viruses: facts, not fear. Proceedings of the National Academy of Sciences of USA. 2012;109:2211–2213. doi: 10.1073/pnas.1121297109. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Pearce MB, Jayaraman A, Pappas C, Belser JA, Zeng H, Gustin KM, Maines TR, Sun X, Raman R, Cox NJ, Sasisekharan R, Katz JM, Tumpey TM. Pathogenesis and transmission of swine origin A(H3N2)v influenza viruses in ferrets. Proceedings of the National Academy of Sciences of USA. 2012;109:3944–3949. doi: 10.1073/pnas.1119945109. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. R Development Core Team . R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing; Vienna, Austria: 2014. [Google Scholar]
  41. Ricklefs RE, Lau M. Bias and dispersion of overlap indices: results of some Monte Carlo simulations. Ecology. 1980;61:1019–1024. doi: 10.2307/1936817. [DOI] [Google Scholar]
  42. Russell CA, Fonville JM, Brown AEX, Burke DF, Smith DL, James SL, Herfst S, van Boheemen S, Linster M, Schrauwen EJ, Katzelnick L, Mosterin A, Kuiken T, Maher E, Neumann G, Osterhaus ADME, Kawaoka Y, Fouchier RAM, Smith DJ. The potential for respiratory droplet-transmissible A/H5N1 influenza virus to evolve in a mammalian host. Science. 2012;336:1541–1547. doi: 10.1126/science.1222526. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Russell CA, Kasson PM, Donis RO, Riley S, Dunbar J, Rambaut A, Asher J, Burke S, Davis CT, Garten RJ, Gnanakaran S, Hay SI, Herfst S, Lewis NS, Lloyd-Smith JO, Macken CA, Maurer-Stroh S, Neuhaus E, Parrish CR, Pepin KM, Shepard SS, Smith DL, Suarez DL, Trock SC, Widdowson M-A, George DB, Lipsitch M, Bloom JD. Improving pandemic influenza risk assessment. eLife. 2014;3:e03883. doi: 10.7554/eLife.03883. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Seibert CW, Kaminski M, Philipp J, Rubbenstroth D, Albrecht RA, Schwalm F, Stertz S, Medina RA, Kochs G, García-Sastre A, Staeheli P, Palese P. Oseltamivir-resistant variants of the 2009 pandemic H1N1 influenza A virus are not attenuated in the guinea pig and ferret transmission models. Journal of Virology. 2010;84:11219–11226. doi: 10.1128/JVI.01424-10. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Shinya K, Ebina M, Yamada S, Ono M, Kasai N, Kawaoka Y. Avian flu: influenza virus receptors in the human airway. Nature. 2006;440:435–436. doi: 10.1038/440435a. [DOI] [PubMed] [Google Scholar]
  46. Song H, Wan H, Araya Y, Perez DR. Partial direct contact transmission in ferrets of a mallard H7N3 influenza virus with typical avian-like receptor specificity. Virology Journal. 2009;6:126. doi: 10.1186/1743-422X-6-126. [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Sutton TC, Finch C, Shao H, Angel M, Chen H, Capua I, Cattoli G, Monne I, Perez DR. Airborne transmission of highly pathogenic H7N1 influenza virus in ferrets. Journal of Virology. 2014;88:6623–6635. doi: 10.1128/JVI.02765-13. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. The SJCEIRS Working Group Assessing the fitness of distinct clades of influenza A (H9N2) viruses. Emerging Microbes & Infections. 2013;2:e75. doi: 10.1038/emi.2013.75. [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Tumpey TM, Maines TR, Van Hoeven N, Glaser L, Solórzano A, Pappas C, Cox NJ, Swayne DE, Palese P, Katz JM, García-Sastre A. A two-amino acid change in the hemagglutinin of the 1918 influenza virus abolishes transmission. Science. 2007;315:655–659. doi: 10.1126/science.1136212. [DOI] [PubMed] [Google Scholar]
  50. van Doremalen N, Shelton H, Roberts KL, Jones IM, Pickles RJ, Thompson CI, Barclay WS. A single amino acid in the HA of pH1N1 2009 influenza virus affects cell tropism in human airway epithelium, but not transmission in ferrets. PLOS ONE. 2011;6:e25755. doi: 10.1371/journal.pone.0025755. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Van Hoeven N, Pappas C, Belser JA, Maines TR, Zeng H, García-Sastre A, Sasisekharan R, Katz JM, Tumpey TM. Human HA and polymerase subunit PB2 proteins confer transmission of an avian influenza virus through the air. Proceedings of the National Academy of Sciences of USA. 2009;106:3366–3371. doi: 10.1073/pnas.0813172106. [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Watanabe T, Kiso M, Fukuyama S, Nakajima N, Imai M, Yamada S, Murakami S, Yamayoshi S, Iwatsuki-Horimoto K, Sakoda Y, Takashita E, McBride R, Noda T, Hatta M, Imai H, Zhao D, Kishida N, Shirakura M, de Vries RP, Shichinohe S, Okamatsu M, Tamura T, Tomita Y, Fujimoto N, Goto K, Katsura H, Kawakami E, Ishikawa I, Watanabe S, Ito M, Sakai-Tagawa Y, Sugita Y, Uraki R, Yamaji R, Eisfeld AJ, Zhong G, Fan S, Ping J, Maher EA, Hanson A, Uchida Y, Saito T, Ozawa M, Neumann G, Kida H, Odagiri T, Paulson JC, Hasegawa H, Tashiro M, Kawaoka Y. Characterization of H7N9 influenza A viruses isolated from humans. Nature. 2013;501:551–555. doi: 10.1038/nature12392. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Watanabe T, Zhong G, Russell CA, Nakajima N, Hatta M, Hanson A, McBride R, Burke DF, Takahashi K, Fukuyama S, Tomita Y, Maher EA, Watanabe S, Imai M, Neumann G, Hasegawa H, Paulson JC, Smith DJ, Kawaoka Y. Circulating avian influenza viruses closely related to the 1918 virus have pandemic potential. Cell Host & Microbe. 2014;15:692–705. doi: 10.1016/j.chom.2014.05.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
  54. White House Office of Science and Technology Policy . Doing diligence to assess the risks and benefits of life sciences gain-of-function research. 2014. http://www.whitehouse.gov/blog/2014/10/17/doing-diligence-assess-risks-and-benefits-life-sciences-gain-function-research. [Google Scholar]
  55. Xu L, Bao L, Deng W, Dong L, Zhu H, Chen T, Lv Q, Li F, Yuan J, Xiang Z, Gao K, Xu Y, Huang L, Li Y, Liu J, Yao Y, Yu P, Li X, Huang W, Zhao X, Lan Y, Guo J, Yong W, Wei Q, Chen H, Zhang L, Qin C. Novel avian-origin human influenza A(H7N9) can be transmitted between ferrets via respiratory droplets. The Journal of Infectious Diseases. 2014;209:551–556. doi: 10.1093/infdis/jit474. [DOI] [PubMed] [Google Scholar]
  56. Zaraket H, Baranovich T, Kaplan BS, Carter R, Song M-S, Paulson JC, Rehg JE, Bahl J, Crumpton JC, Seiler J, Edmonson M, Wu G, Karlsson E, Fabrizio T, Zhu H, Guan Y, Husain M, Schultz-Cherry S, Krauss S, McBride R, Webster RG, Govorkova EA, Zhang J, Russell CJ, Webby RJ. Mammalian adaptation of influenza A(H7N9) virus is limited by a narrow genetic bottleneck. Nature Communications. 2015;6:6553. doi: 10.1038/ncomms7553. [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Zhu H, Wang D, Kelvin DJ, Li L, Zheng Z, Yoon S-W, Wong S-S, Farooqui A, Wang J, Banner D, Chen R, Zheng R, Zhou J, Zhang Y, Hong W, Dong W, Cai Q, Roehrl MHA, Huang SSH, Kelvin AA, Yao T, Zhou B, Chen X, Leung GM, Poon LLM, Webster RG, Webby RJ, Peiris JSM, Guan Y, Shu Y. Infectivity, transmission, and pathology of human-isolated H7N9 influenza virus in ferrets and pigs. Science. 2013;341:183–186. doi: 10.1126/science.1239844. [DOI] [PubMed] [Google Scholar]
eLife. 2015 Sep 2;4:e07969. doi: 10.7554/eLife.07969.019

Decision letter

Editor: Mark Jit1

eLife posts the editorial decision letter and author response on a selection of the published articles (subject to the approval of the authors). An edited version of the letter sent to the authors after peer review is shown, indicating the substantive concerns or comments; minor concerns are not usually shown. Reviewers have the opportunity to discuss the decision before the letter is sent (see review process). Similarly, the author response typically shows only responses to the major concerns raised by the reviewers.

Thank you for submitting your work entitled “Mapping influenza transmission in the ferret model to transmission in humans” for peer review at eLife. Your submission has been favorably evaluated by Prabhat Jha (Senior Editor), a Reviewing Editor, and two reviewers.

The following individuals responsible for the peer review of your submission have agreed to reveal their identity: Mark Jit (Reviewing Editor) and Marc Lipsitch (peer reviewer).

The Reviewing Editor and the reviewers discussed their comments before we concluded that your manuscript is a topical and valuable contribution that brings clarity to an area that has been in the realm of opinion and aggregated anecdotes. It is an important step forwarding in bridging mammalian transmission data with human epidemiological findings, and we commend you for attempting to tackle this imposing but important topic.

However, we have important reservations about it, particularly around the literature search, source data, its analysis and the description of the analysis in the text. In particular, there are major omissions in the source data tables which underpin the entire analysis. We strongly recommend (but do not insist) that you include as a contributing author a subject matter expert familiar with the breadth and scope of ferret transmission data, which may not be captured by a simple algorithmic PubMed search.

The following comments need to be addressed in a revised submission before it can be considered for publication:

1) More detail needs to be provided about the selection criteria for the analysis conducted in this assessment. While the manuscript states in the Methods that you “excluded subtypes for which there were no estimates of human or ferret SAR” (in the subsection “Literature Review”) and that you “excluded isolates that represented outliers from identified subtypes”, it is unclear to us how published H2N2 ferret studies do not meet the definition of providing estimations of “ferret SAR” or the selection criteria employed to determine which viruses constitute “outliers from identified subtypes”.

Furthermore, while you state (in the subsection “Inclusion criteria for ferret studies”) that “to promote quality of comparison between ferret and human studies, we included data only from ferret studies that tested one or more wild-type human isolates”, this exclusion removes numerous studies with avian viruses which nonetheless share close sequence homology with viruses isolated from humans; please provide additional text justifying this decision.

2) In Figure 1, please better define “n”. Does it reflect the number of individual ferret inoculated-contact pairs included in each ferret study? You state that “the number of estimates for each subtype is given above each box-and-whiskers plot” (Figure 1 legend) yet this does not clarify the true number of ferret inoculated-contact pairs represented in this graph. Do you give more weight to transmission studies which employ more animals or are all transmission studies weighed equally, even if one study uses an n=2 and another uses an n=4-6? This level of simplicity does not accurately reflect the complexity of data in the published literature and must be augmented and/or better described for the presented results if you intend for this data to be meaningful to researchers who work with ferret transmission models.

3) Upon first glance at Figure 1–source data 1, 2 and 3, it appears that there are missing studies and inaccuracies in the studies listed which you used to conduct their analyses. For example, in Figure 1–source data 2, you list respiratory droplet H5N1 studies from 4 published studies, yet exclude ferret transmission data which appears to meet the criteria outlined which appears in studies by Jackson et al. (wild-type Thai/16 virus, PMID 19493997) and Herfst et al. (wild-type Indo/05 virus, PMID 22723413). Also in Figure 1–source data 2, you appear to exclude ferret transmission data for respiratory droplet pH1N1 which meets the criteria outlined in the study by Lakdawala et al. (California/4/09, PMID 22241979). Furthermore, in Figure 1–source data 2 and Figure 1, you list duplicate entries for respiratory droplet transmission of the H7N7 virus A/NL/230/03, yet the publication cited states that only direct contact transmission, and not respiratory droplet transmission, was performed in duplicate for this virus, meaning that there should be an n=2 for this virus subtype for RD transmission per your definitions, not an n=3. Did you consult an outside source (ideally a subject matter expert in ferret transmissibility studies) to verify/confirm their literature search? These are only some of the errors the reviewers identified in studies that do not appear to have been captured in this literature search.

4) The grouping of viruses by surface glycoprotein HA/NA subtype only seems highly simplistic, as this does not account for differences in species origins (i.e. swine vs. human vs. canine lineage H3N2 viruses) or within virus subtypes (i.e. North American vs. Eurasian lineage H7 viruses). This grouping also does not take into account potential differences in receptor binding specificity or the presence of particular mutations/amino acids throughout the virus which might contribute to increased/decreased virus transmissibility by either direct contact or respiratory droplets. You must include in the text a greater emphasis on how the limitations this simplistic grouping of viruses by surface glycoproteins only may adversely affect or otherwise influence your findings – the statement in the Discussion that “none of [these features] are systematically collected under standard protocols” is not sufficient. While it is understandable that some generalizations will need to be made when grouping viruses by subtype only, you should disclose in the text to more detail the potential caveats of such assumptions.

5) You rightfully mention in the third paragraph of the Discussion the need for a standard definition of transmission for ferret experiments (i.e. viral isolation vs seroconversion only vs. both), yet appear to use all definitions interchangeably in your analysis. From the perspective of researchers who perform ferret transmission studies, there is a large difference in transmission robustness between these two criteria, with the presence of both virus isolation and seroconversion considered the best evidence of successful virus transmission. This has notable bearing on the direct contact transmissibility analysis of this study, as viruses which transmit poorly by respiratory droplets are often able to lead to low-level seroconversion of naïve ferrets when placed in direct contact. If you are counting these low-level seroconversions as true virus transmission events, it is not surprising that the data regarding SAR from direct contact transmission exhibits greater variability and less statistical weight. You should disclose more fully this potential caveat in your analysis and interpretation of direct contact ferret transmission SAR.

6) The ferret data are treated in the statistical analyses as being without error, although as you acknowledge they are extremely noisy because of small sample size. Horizontal error bars are shown in Figure 4 but not in Figure 3. In the linear regression context there is something called Type 2 regression to handle this. I suspect there is something comparable in logistic regression, and it would be good to try it or at least point to it.

7) The conclusion at the end of the Discussion and throughout is not quite justified. Why should ferrets be a first-like screening tool, when other virologic studies are cheaper and higher-throughput? In practice ferret studies are done on a pretty restricted subset of isolates thought to be high-risk for other reasons, including epidemiological data and virologic findings. So the questions that arise are: conditional on being considered high-enough risk to merit ferret studies, what additional value do the ferret studies add? Should we really be doing ferret studies first? If so, what should be the means of follow-up?

In essence, the paper shows that there is a good correlation between ferret outcomes and human data. This is very useful. It also quantifies the noise and notes the sources of that noise. That is also very useful. We are not sure that it provides evidence in the current form of the argument for how ferrets can add value to pandemic prediction.

This leads to a broader point that is not much addressed. Pandemic risk prediction is used for the targeting of countermeasures to contain certain strains considered high risk. Thus having a good correlation may help to guide such efforts to on average more threatening strains, but may nonetheless produce many false negatives (failing to spot highly risky strains) and false positives (identifying strains as high risk that are not actually high risk). The rather bland conclusion that ferrets should keep being used seems to skirt this issue. We would like you either to modify your conclusions to present this point, or else to give stronger justification for the use of the ferret model to guide pandemic prediction.

8) In the subsection “Inclusion criteria for ferret studies”, the inclusion of only human isolates studied in ferrets should be directly highlighted. The data on H7N9 (and maybe others too) show that taking a strain from a bird into a ferret may give very different results (no transmission in ferrets) from the same subtype isolated from a human (transmission in ferrets). This has several implications. First, not everyone will note the subtlety but it probably affects the results of the study (it might have looked different if bird isolates tested in ferrets had been included). Second, it highlights the point that subtype is not an ideal level of aggregation, though it is needed for the reasons you discuss. Third, it emphasizes the ambiguity of the assessment of pandemic potential. If a strain that is nontransmissible in ferrets can change in possibly a single human passage into one that is, which readout in the ferrets gives the better estimate of pandemic potential?

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled “Mapping influenza transmission in the ferret model to transmission in humans” for further consideration at eLife. Your revised article has been favorably evaluated by Prabhat Jha (Senior Editor), Mark Jit (Reviewing Editor) and Marc Lipsitch (peer reviewer).

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

1) In the fourth paragraph of the Discussion, you advocate screening strains for virologic and genetic markers of pandemic potential before selecting a subset (presumably positive on all these screens) for ferret testing. There is no evidence that this is a good idea. To demonstrate that, a larger analysis would need to be done on at least the included datasets here (but really one would want to know about strains on which ferret studies were not done, too) to see if that screen would have been a good idea. The fact that one of the most “popular” mutations was not present in H1N1p (PB2 E627K) casts doubt on this strategy. We would suggest you modify this paragraph to speculate that further studies (perhaps replicating their approaches) would be needed to clarify whether virologic and genetic data add anything to ferret findings.

2) Again in the Discussion, you state: “The ultimate evidence to corroborate human transmission comes from epidemiological patterns of infection in humans, but for a true pandemic influenza, this data may come too late highlighting the importance of ferrets in early identification of these strains”. This conclusion seems to beg the question and not to follow from the rest of the paragraph. You have clearly documented that while very often informative, the results of ferret experiments can sometimes be misleading. Human transmission data highlights the need for a reliable early warning system that can reliably identify strains of pandemic potential, but not the “importance of ferrets in the early identification of these strains.” Ferrets can contribute, but a reliable early warning system is still needed.

3) Also in the Discussion, you state: “This coupled with the biological complexities underlying transmissibility suggests that, at this time, ferret transmission data are a valid tool for assessing human transmissibility, but should be corroborated by further virologic and epidemiologic research”. We do not have any valid tool for assessing transmissibility, though we do have useful correlates that improve our estimate. And saying we need further virologic and epidemiologic evidence suggests that this would be dispositive, but we don't have virologic predictors that are reliable, and epidemiologic evidence comes too late as you point out. It may be better to write something like “[…] at this time, ferret transmission data provide a valuable but imperfect correlate of human transmissibility, and further evidence is needed to assess whether other lines of evidence can improve this predictive capacity.”

4) In the second paragraph of the Discussion you state: “Sample size is a serious challenge to operational use of the results shown here. The largest sample size we found in our review of transmission studies was twelve ferrets (Herlocher et al. 2002). Even if 8/8 contact ferrets were infected in an experiment, there is only 80% power at a significance level of 0.05 to classify that strain as supercritical (Figure 4–figure supplement 1).” This is difficult to understand because a power calculation can only be done before specifying the outcome of the binomial experiment, and not after. Something seems to be wrong with this.

eLife. 2015 Sep 2;4:e07969. doi: 10.7554/eLife.07969.020

Author response

1) More detail needs to be provided about the selection criteria for the analysis conducted in this assessment. While the manuscript states in the Methods that you “excluded subtypes for which there were no estimates of human or ferret SAR” (in the subsection “Literature Review”) and that you “excluded isolates that represented outliers from identified subtypes”, it is unclear to us how published H2N2 ferret studies do not meet the definition of providing estimations of “ferret SAR” or the selection criteria employed to determine which viruses constitute “outliers from identified subtypes”.

Furthermore, while you state (in the subsection “Inclusion criteria for ferret studies”) that “to promote quality of comparison between ferret and human studies, we included data only from ferret studies that tested one or more wild-type human isolates”, this exclusion removes numerous studies with avian viruses which nonetheless share close sequence homology with viruses isolated from humans; please provide additional text justifying this decision.

We added additional text justifying our exclusion of avian and other animal isolates (in the subsection “Inclusion criteria for ferret studies”). Briefly, despite sequence homology between avian and human isolates, the cross-species jump from birds to humans involves evolutionary barriers, including transmission bottlenecks and within-host adaptation. Considering only human isolates removes uncertainties generated by some of these barriers, potentially providing a clearer signal on the relationship between transmission in ferrets and transmission in humans. To test the validity of these omissions, we also added data on ferret experiments performed with H5N1 and H7N9 avian isolates. We chose isolates from these subtypes given their position at the boundary between sub- and supercritical subtypes (Figure 2B) and the concurrence in both space and time between isolation in birds and people. However, inclusion of avian strains has little impact on the correspondence between ferret and human transmission (Figure 2 and Figure 2–figure supplement 1), suggesting that an extra review of animal isolates would not contribute added insight to this analysis.

Additionally, we added experiments using H2N2 and H7N3 isolates to the analysis after finding appropriate estimates of ferret and human SARs, respectively, in the literature.

2) In Figure 1, please better define “n”. Does it reflect the number of individual ferret inoculated-contact pairs included in each ferret study? You state that “the number of estimates for each subtype is given above each box-and-whiskers plot” (Figure 1 legend) yet this does not clarify the true number of ferret inoculated-contact pairs represented in this graph. Do you give more weight to transmission studies which employ more animals or are all transmission studies weighed equally, even if one study uses an n=2 and another uses an n=4-6? This level of simplicity does not accurately reflect the complexity of data in the published literature and must be augmented and/or better described for the presented results if you intend for this data to be meaningful to researchers who work with ferret transmission models.

We have clarified that “n” is the number of estimated SARs for each subtype in the legend for Figure 1. Given that there is likely more variability between laboratories (e.g. dosing or cage setups) than between any two ferrets, the unit of this analysis is necessarily an experiment. We do agree with the reviewers that higher weight should be placed on experiments that use more ferrets. We have redone the analyses using weighted means of the ferret SARs in Figure 2A and including the number of ferrets tested as case weights in the logistic regression (Figure 3). A description of this is included in the Methods (in the subsection “Analysis”) and in the legends for Figures 2 and 3.

3) Upon first glance at Figure 1–source data 1, 2 and 3, it appears that there are missing studies and inaccuracies in the studies listed which you used to conduct their analyses. For example, in Figure 1–source data 2, you list respiratory droplet H5N1 studies from 4 published studies, yet exclude ferret transmission data which appears to meet the criteria outlined which appears in studies by Jackson et al. (wild-type Thai/16 virus, PMID 19493997) and Herfst et al. (wild-type Indo/05 virus, PMID 22723413). Also in Figure 1–source data 2, you appear to exclude ferret transmission data for respiratory droplet pH1N1 which meets the criteria outlined in the study by Lakdawala et al. (California/4/09, PMID 22241979). Furthermore, in Figure 1–source data 2 and Figure 1, you list duplicate entries for respiratory droplet transmission of the H7N7 virus A/NL/230/03, yet the publication cited states that only direct contact transmission, and not respiratory droplet transmission, was performed in duplicate for this virus, meaning that there should be an n=2 for this virus subtype for RD transmission per your definitions, not an n=3. Did you consult an outside source (ideally a subject matter expert in ferret transmissibility studies) to verify/confirm their literature search? These are only some of the errors the reviewers identified in studies that do not appear to have been captured in this literature search.

With the help of Dr. Jessica Belser, we have considerably expanded the scope of our literature review. Specifically, we added 23 more estimates of ferret respiratory droplet SAR, 27 more estimates of ferret direct contact SAR, and 7 more estimates of human SAR. These additions also broadened our analysis to include two more subtypes, H2N2 and H7N3. These additional data, along with our weighted regression scheme, have led to narrower confidence intervals in our main analysis.

4) The grouping of viruses by surface glycoprotein HA/NA subtype only seems highly simplistic, as this does not account for differences in species origins (i.e. swine vs. human vs. canine lineage H3N2 viruses) or within virus subtypes (i.e. North American vs. Eurasian lineage H7 viruses). This grouping also does not take into account potential differences in receptor binding specificity or the presence of particular mutations/amino acids throughout the virus which might contribute to increased/decreased virus transmissibility by either direct contact or respiratory droplets. You must include in the text a greater emphasis on how the limitations this simplistic grouping of viruses by surface glycoproteins only may adversely affect or otherwise influence your findings – the statement in the Discussion that “none of [these features] are systematically collected under standard protocols” is not sufficient. While it is understandable that some generalizations will need to be made when grouping viruses by subtype only, you should disclose in the text to more detail the potential caveats of such assumptions.

We agree that grouping at the subtype level may potentially obscure important complexities within select subtypes, including molecular determinants or other phylogenetic features which may contribute to virus transmissibility in mammalian models. However, this level of grouping was necessary to perform the correlations to human SAR and general humans transmission pattern (i.e. subcritical vs. supercritical) which represent the primary focus of this study. There simply are not sufficient data to estimate human transmissibility with resolution below the HA/NA subtype level. Additional text has been included in the Discussion to better clarify and justify this decision. Importantly, the grouping at subtype level employed here has provided a valid null model against which deviations can be identified, thus laying the foundation for potential further analyses investigating virus- or host-specific attributes which may contribute to these deviations, which we acknowledge in the fifth paragraph of the Discussion.

5) You rightfully mention in the third paragraph of the Discussion the need for a standard definition of transmission for ferret experiments (i.e. viral isolation vs seroconversion only vs. both), yet appear to use all definitions interchangeably in your analysis. From the perspective of researchers who perform ferret transmission studies, there is a large difference in transmission robustness between these two criteria, with the presence of both virus isolation and seroconversion considered the best evidence of successful virus transmission. This has notable bearing on the direct contact transmissibility analysis of this study, as viruses which transmit poorly by respiratory droplets are often able to lead to low-level seroconversion of naïve ferrets when placed in direct contact. If you are counting these low-level seroconversions as true virus transmission events, it is not surprising that the data regarding SAR from direct contact transmission exhibits greater variability and less statistical weight. You should disclose more fully this potential caveat in your analysis and interpretation of direct contact ferret transmission SAR.

We added additional data to Figure 1–source data 2 and 3 to address this, specifically assessing transmission by either virus isolation or seroconversion per the raw data provided in individual published studies, in lieu of relying on conclusions of the authors of the individual studies, which could vary between laboratories. Here, we specified experiments where results from viral isolation did not match seroconversion results. Our results were robust to the definition of transmission, with only the analysis of AUC showing any discernable impact of considering viral isolation alone (slightly positive impact for direct contact, negative for respiratory droplet). We added a recommendation that in practice both measures are needed to allow for comparisons across experiment types in the third paragraph of the Discussion.

6) The ferret data are treated in the statistical analyses as being without error, although as you acknowledge they are extremely noisy because of small sample size. Horizontal error bars are shown in Figure 4 but not in Figure 3. In the linear regression context there is something called Type 2 regression to handle this. I suspect there is something comparable in logistic regression, and it would be good to try it or at least point to it.

Because we are interested in predicting pandemic potential from ferret SAR and not in the regression coefficients themselves, model I regression (i.e. ordinary logistic regression) is the preferred statistical model as it minimizes squared residuals in the y direction, whereas model II regression approaches do not (Sokal and Rohlf, 1995, p. 545, Table 14.3; Legendre and Legendre, 1998). However, we agree with the reviewers that error in ferret SAR estimates will have an impact on setting super- and subcritical thresholds using 95% confidence bands on model predictions (Sokal and Rohlf, 1995, p. 545, Table 14.3; Legendre and Legendre 1998). To explore this uncertainty, we used a simulation approach. Here, we simulated 1000 datasets by taking binomial samples from each data point using a probability given by the observed ferret SAR and N equal to the number of ferrets used. To introduce uncertainty into those experiments where the ferret SAR was 0 or 1, we set the binomial probability to be 0.1 or 0.9, respectively. The results of this analysis are given in the subsection “Using ferret SAR to characterize human pandemic potential” and in Figure 3–figure supplement 2 and indicate that while uncertainty in the ferret SAR estimates impacts our results quantitatively, the underlying relationship between ferret SAR and an isolate’s general transmission behavior is robust to such uncertainty.

7) The conclusion at the end of the Discussion and throughout is not quite justified. Why should ferrets be a first-like screening tool, when other virologic studies are cheaper and higher-throughput? In practice ferret studies are done on a pretty restricted subset of isolates thought to be high-risk for other reasons, including epidemiological data and virologic findings. So the questions that arise are: conditional on being considered high-enough risk to merit ferret studies, what additional value do the ferret studies add? Should we really be doing ferret studies first? If so, what should be the means of follow-up?

It was not our intention to suggest that ferret studies should be the very first test done to identify strains of concern. Rather, we aimed to emphasize that they are not the final word. We have changed the language in the fourth paragraph of the Discussion to stress that, while virologic and genetic screening tools can indeed identify strains of concern, transmission is a complex phenotype that cannot be completely characterized using these high-throughput initial screens. Ultimately, transmission is a process occurring at the level of whole animals (and pairs of animals), and given our current state of knowledge, animal experiments continue to contribute distinct insights into transmission biology. Thus, ferrets provide a valuable missing link from virologic and genetic changes to a more complete understanding of their impact on human transmission behavior.

In essence, the paper shows that there is a good correlation between ferret outcomes and human data. This is very useful. It also quantifies the noise and notes the sources of that noise. That is also very useful. We are not sure that it provides evidence in the current form of the argument for how ferrets can add value to pandemic prediction.

This leads to a broader point that is not much addressed. Pandemic risk prediction is used for the targeting of countermeasures to contain certain strains considered high risk. Thus having a good correlation may help to guide such efforts to on average more threatening strains, but may nonetheless produce many false negatives (failing to spot highly risky strains) and false positives (identifying strains as high risk that are not actually high risk). The rather bland conclusion that ferrets should keep being used seems to skirt this issue. We would like you either to modify your conclusions to present this point, or else to give stronger justification for the use of the ferret model to guide pandemic prediction.

We agree that predictions from ferret studies have the potential to identify many false negatives and false positives, although our results do indicate that sensitivity and specificity are relatively high when using ferret SAR to identify supercritical strains. Given the complexities underlying transmission, it is likely that ferret transmission studies, as a holistic measure of transmission, provide higher sensitivity and specificity than initial screens. While the current analysis does not allow us to verify this fact, it does provide the opportunity for future analyses to include all screening tools in predictions of human pandemic potential, thus increasing sensitivity and specificity further. We have included a discussion of this point, and explicit reference to the hazards of false negatives and false positives, in the fourth and sixth paragraphs of the Discussion.

8) In the subsection “Inclusion criteria for ferret studies”, the inclusion of only human isolates studied in ferrets should be directly highlighted. The data on H7N9 (and maybe others too) show that taking a strain from a bird into a ferret may give very different results (no transmission in ferrets) from the same subtype isolated from a human (transmission in ferrets). This has several implications. First, not everyone will note the subtlety but it probably affects the results of the study (it might have looked different if bird isolates tested in ferrets had been included). Second, it highlights the point that subtype is not an ideal level of aggregation, though it is needed for the reasons you discuss. Third, it emphasizes the ambiguity of the assessment of pandemic potential. If a strain that is nontransmissible in ferrets can change in possibly a single human passage into one that is, which readout in the ferrets gives the better estimate of pandemic potential?

We have added further clarification to the Methods about the exclusion of avian and other animal isolates (in the subsection “Inclusion criteria for ferret studies”). We also tested the validity of this exclusion by compiling data on avian H5N1 and H7N9 isolates. The inclusion of these isolates did not impact our results, suggesting that further review of these isolates is not currently necessary. Additionally, this also suggests that within-subtype variation is not obscuring broad-scale patterns. However, we added a discussion of within-subtype variation to address the second and third points raised by the reviewers, and a clarification that despite the subtype aggregation, we view high ferret SARs as indicative of a high likelihood of human-human transmission, regardless of subtype (see the sixth paragraph of the Discussion). While we agree with the reviewers’ final point that evolutionary dynamics within and between hosts are a necessary part of predicting pandemic potential, and indeed this is an important and fast-moving scientific frontier, this is beyond the scope of this study. Gain-of-function studies and studies of natural isolates are all trying to get at this point, and ethical concerns aside, we have shown that the use of ferrets to draw comparisons to humans in these studies is scientifically justified.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

1) In the fourth paragraph of the Discussion, you advocate screening strains for virologic and genetic markers of pandemic potential before selecting a subset (presumably positive on all these screens) for ferret testing. There is no evidence that this is a good idea. To demonstrate that, a larger analysis would need to be done on at least the included datasets here (but really one would want to know about strains on which ferret studies were not done, too) to see if that screen would have been a good idea. The fact that one of the most “popular” mutations was not present in H1N1p (PB2 E627K) casts doubt on this strategy. We would suggest you modify this paragraph to speculate that further studies (perhaps replicating their approaches) would be needed to clarify whether virologic and genetic data add anything to ferret findings.

We agree with the reviewers that this is an important avenue for future research. As suggested, we have modified this paragraph to stress that further studies are needed to integrate genetic data on specific mutations, such as PB2-K627E, and virologic changes, such as sialic acid binding affinity, into the current results to truly assess how high-throughput screens may add value to predictions of human transmissibility from ferret experiments alone.

2) Again in the Discussion, you state: “The ultimate evidence to corroborate human transmission comes from epidemiological patterns of infection in humans, but for a true pandemic influenza, this data may come too late highlighting the importance of ferrets in early identification of these strains”. This conclusion seems to beg the question and not to follow from the rest of the paragraph. You have clearly documented that while very often informative, the results of ferret experiments can sometimes be misleading. Human transmission data highlights the need for a reliable early warning system that can reliably identify strains of pandemic potential, but not the “importance of ferrets in the early identification of these strains.” Ferrets can contribute, but a reliable early warning system is still needed.

We agree the lack of human data highlights the need for a reliable early warning system and not the need for ferret studies specifically, especially given some of the issues we discuss. We have changed this sentence to stress that methods to provide reliable early warning on strains with pandemic potential are needed.

3) Also in the Discussion, you state: “This coupled with the biological complexities underlying transmissibility suggests that, at this time, ferret transmission data are a valid tool for assessing human transmissibility, but should be corroborated by further virologic and epidemiologic research”. We do not have any valid tool for assessing transmissibility, though we do have useful correlates that improve our estimate. And saying we need further virologic and epidemiologic evidence suggests that this would be dispositive, but we don't have virologic predictors that are reliable, and epidemiologic evidence comes too late as you point out. It may be better to write something like “[…] at this time, ferret transmission data provide a valuable but imperfect correlate of human transmissibility, and further evidence is needed to assess whether other lines of evidence can improve this predictive capacity.”

We agree that the suggested change is a valuable clarification and have changed the text accordingly.

4) In the second paragraph of the Discussion you state “Sample size is a serious challenge to operational use of the results shown here. The largest sample size we found in our review of transmission studies was twelve ferrets (Herlocher et al. 2002). Even if 8/8 contact ferrets were infected in an experiment, there is only 80% power at a significance level of 0.05 to classify that strain as supercritical (Figure 4–figure supplement 1).” This is difficult to understand because a power calculation can only be done before specifying the outcome of the binomial experiment, and not after. Something seems to be wrong with this.

We appreciate the reviewers pointing out this confusing wording. To remedy this, we have changed the wording to reflect that the power calculations are in fact a priori instead of post-hoc. We intend for this analysis to demonstrate the sample size necessary to detect a difference between an assumed ferret SAR of a future experiment and the lower threshold for a supercritical classification.

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Supplementary Materials

    Figure 1—source data 1. Estimates of human household SAR.

    DOI: http://dx.doi.org/10.7554/eLife.07969.004

    elife07969s001.docx (106.1KB, docx)
    DOI: 10.7554/eLife.07969.004
    Figure 1—source data 2. Ferret influenza transmission studies via respiratory droplets using human isolates.

    DOI: http://dx.doi.org/10.7554/eLife.07969.005

    elife07969s002.docx (120.4KB, docx)
    DOI: 10.7554/eLife.07969.005
    Figure 1—source data 3. Ferret influenza transmission studies via direct contact using human isolates.

    DOI: http://dx.doi.org/10.7554/eLife.07969.006

    elife07969s003.docx (101.8KB, docx)
    DOI: 10.7554/eLife.07969.006
    Figure 1—source data 4. Ferret influenza transmission studies via respiratory droplets and direct contact using avian isolates.

    DOI: http://dx.doi.org/10.7554/eLife.07969.007

    elife07969s004.docx (29.3KB, docx)
    DOI: 10.7554/eLife.07969.007
    Figure 4—source data 1. Ferret influenza transmission studies via respiratory droplets using strains from gain-of-function experiments with H5N1 avian influenza and the reconstructed 1918 pandemic H1N1 strain.

    DOI: http://dx.doi.org/10.7554/eLife.07969.017

    elife07969s005.docx (21KB, docx)
    DOI: 10.7554/eLife.07969.017

    Articles from eLife are provided here courtesy of eLife Sciences Publications, Ltd

    RESOURCES