Utility of Human In Vitro Data in Risk Assessments of Influenza A Virus Using the Ferret Model

Influenza A viruses (IAVs) are capable of overcoming species barriers to cause sporadic cases of human infection. While most zoonotic-to-human transmission events are self-limiting, there remains the potential during this process for the acquisition of features associated with enhanced mammalian adaptation, which could result in the emergence of a pandemic virus. Risk assessment rubrics (notably the Influenza Risk Assessment Tool [IRAT] and the World Health Organization Tool for Influenza Pandemic Risk Assessment [TIPRA]) have been established to consider virus, host, ecological, and environmental properties to facilitate the assessment of the public health risk posed by novel and emerging influenza viruses (1, 2). These models are informed by in vivo pathogenicity and transmissibility data generated in laboratory models, specifically the ferret. In vitro experimentation is frequently employed concurrently with in vivo assessments (and in the case of TIPRA, is formally considered) to provide supporting information to in vivo data. These in vitro assessments are typically conducted in cultured human respiratory epithelial cell lines but have expanded in recent years to include primary cells, tissue explants, tissue constructs, and organoid cultures (3, 4).

To assess mammalian pathogenicity of novel and emerging IAVs associated with human infection or exposure, serologically naive ferrets are typically inoculated intranasally with a high dose of virus (10⁵ to 10⁷ infectious units) and observed daily for clinical signs of infection, with specimens collected during the acute phase of infection from the upper respiratory tract. However, the variability of experimental conditions between laboratories performing risk assessment studies represents a challenge for contextualizing results in the field (5). Even within controlled laboratory environments, the wide heterogeneity of influenza virus subtypes associated with human infection can lead to complications for their comparative study (6). Study-to-study comparisons of viral growth kinetics in cells derived from the human respiratory tract are often similarly hindered due to the use of different cell types between studies and the limited availability of cultures that are not commercially available (4).

While previous studies have examined the association of in vivo ferret data as they pertain to human pathogenicity and transmissibility (7, 8), there remains a need to understand the strengths and limitations of in vitro characterizations performed at a fixed multiplicity of infection (MOI) to mammalian risk assessment activities. Zoonotic IAVs typically replicate to high and sustained titer at both 37°C and 40°C (the temperature of the avian enteric tract), whereas a shorter time to reach comparable titer metrics at 33°C (the temperature of the human nasal passages) could be indicative of mammalian adaptation (9, 10), although this has not been systematically examined. In vitro results are typically presented as a capacity to replicate to high titer, to indicate relative levels of host adaptation, and as a potential to replicate efficiently in human respiratory tract tissues, often contextualized with additional parameters such as elicitation of host responses (11, 12). As in vitro models become more advanced, so must our understanding of the capacity of these models to recapitulate in vivo environments. In this vein, mathematical models (MMs) of replication dynamics and their use to identify differential growth properties between human and zoonotic influenza viruses in relevant human cell types represent a sophisticated approach to improving our knowledge of host adaptation processes (13). An in vitro MM of IAVs with pandemic potential distills viral titer data into key parameters characterizing their replicative fitness, providing valuable information for both laboratory and epidemiological applications (14).

Here, we perform an exploratory analysis comparing IAV titer measurements and some MM-derived quantities in vitro and in vivo, employing data obtained in the human bronchial epithelial cell line Calu-3 with paired data obtained from virus-inoculated ferrets. Analyses were conducted using titers generated by viral titration during standard replication kinetics experiments in vitro, and MM parameters were estimated from the in vitro infections for each virus strain. The inclusion of contemporary human and zoonotic IAVs that exhibit differential mammalian pathogenicity permits the investigation and identification of features generated from in vitro assessments that are most indicative of in vivo mammalian replication fitness. Collectively, this study provides a necessary first step toward improved quantitative interpretation of in vitro data generated for risk assessment purposes.

Pandemic preparedness efforts encompass a diverse array of approaches to understand not only the properties of novel influenza viruses but also the host and environmental contexts in which these viruses emerge (1, 2). As such, in vitro and in vivo assessments presented in this study represent just one of a multitude of concurrent activities performed when assessing the pandemic potential of influenza viruses. As laboratory assessments of novel and emerging influenza viruses are conducted worldwide, it is critical to understand how different experimental protocols can inform and improve our ability to contextualize results, to achieve the highest confidence possible in conclusions drawn from concurrent in vitro and in vivo evaluations. Here, we use viral titer measurements from IAV infections in vitro to assess the extent to which they correlate with in vivo viral titer measurements in the respiratory tract of ferrets following infection and employ mathematical modeling to examine in finer detail the replication parameters not apparent by traditional virology analyses alone.

The data presented in this study represents an analysis of in vitro and in vivo data generated under consistent laboratory experimental protocols representative of those employed in risk assessment activities, an approach that helps reduce the effect of confounders when comparing data between different laboratory groups (7, 22). Due to extensive heterogeneity in both in vitro (3) and in vivo (5) protocols associated with influenza virus risk assessment activities, analyses investigating different cell types (such as A549 or NHBE cells), titration methods (such as TCID₅₀), and ferret specimens (such as nose/throat swabs) were not included in this study. Calu-3 cells are highly permissive to IAV infection and support high-titer replication of a range of human and zoonotic IAVs without the addition of exogeneous trypsin, making them well suited to evaluate the heterogenous panel of viruses in this study (18). Both titer and infection progression MM-predicted parameters reported here are specific to Calu-3 cells grown under Transwell conditions, and would likely vary should the culture method or cell type be changed. Despite these limitations on the in vitro culture condition and the MM analysis, the utility of MM to validate time point selection during experimentation and to identify features of the replication curve (with regard to both viral titer and infection progression kinetics) that offer the highest predictive value to in vivo-derived measurements highlights the benefit that this analysis approach can contribute to the interpretation of data generated for risk assessment purposes.

Our data set included 52 contemporary IAV evaluated in both in vivo and in vitro models via consistent experimental protocols. An important limitation within this data set was the use of two different titration matrices (eggs and cells). There were numerous instances where high-ranking correlations were identified between in vitro and in vivo measures for viruses titrated in one titration matrix that were low ranking in the other matrix (notably with regard to measures of infection progression). We suspect this is because zoonotic viruses of avian origin tend to be titrated in eggs while mammalian-origin viruses tend to be titrated by plaque assay, as eggs are more sensitive in detecting the presence of avian-origin viruses, which may not grow well in mammalian cells. In contrast, many human and swine influenza viruses are preferentially titrated in mammalian cells (typically MDCK cells), due to robust growth in this matrix and to avoid the potential for egg-adapted mutations (23). In support of this, stratifying the data set by virus origin rather than titration matrix when possible resulted in generally comparable correlation coefficients and maintenance of statistically significant findings, although this was not possible in all analyses. As such, we cannot fully separate the contributing role of titration matrix versus host origin when reporting observed correlations. Nonetheless, considering the dual use of both egg and cell titration of IAV in the field, approaches such as this to better control for and interpret data generated from multiple titration matrices represent a critical and overdue effort.

The virologic and phenotypic heterogeneity among viruses in these data sets represented an additional challenge for comparative study. Striking differences in exponential growth or decay of replicating virus in ferret NW specimens depending on virus host origin (as supported by slope_1,3 measurements) further support the idea that mixed data sets inclusive of both mammalian- and zoonotic-origin IAV may exhibit host-specific features during replication in vivo. Of note, even the contribution of specific molecular determinants of virulence can be challenging to assess in data sets with viruses containing substantial genetic diversity. For example, when cultured at 33°C, the PB2 627K mutation was associated with a higher growth rate in Calu-3 among viruses titrated in eggs (0.13 [0.11, 0.17] h⁻¹ 627K versus 0.095 [0.045, 0.11] h⁻¹ 627E) but not among viruses titrated in cells (0.16 [0.12, 0.2] h⁻¹ 627K versus 0.15 [0.08, 0.19] h⁻¹ 627E); other MM-predicted parameters (e.g., t_1%maxV) exhibited similar, though less striking, trends. In the case of PB2 E627K, we examined one mutation in isolation (Table S2D); it is known that many amino acid substitutions can functionally compensate for one another (24) and that polygenic traits such as pathogenicity and transmissibility are often a reflection of specific mutations present at multiple amino acid residues (25). Limiting analyses to genetically related viruses (e.g., H7N9) often yielded higher correlation coefficients, although this could be at least in part due to the smaller number of viruses in these analyses; nonetheless, this study provides a framework for further investigation of predictive roles for specific molecular features in vitro in predicting in vivo behavior should viruses that possess reduced genetic diversity be employed.

As similar analyses have not previously been conducted, we designed this work to be exploratory in nature, intended to identify avenues and analysis strategies for further investigation rather than to provide definitive findings. Considering the substantial number of metrics evaluated throughout this study (from both raw Calu-3 data and MM derived, many of which were ultimately not included in this report), we do not report the absolute statistical significance of any correlation to avoid overinflation of reported significance. Rather, we sought to focus on biologically meaningful relationships present between parameters of the viral replication curve in vitro and multiple sample types in vivo. Linear associations (identified via Pearson correlations) between viral titers obtained in vitro and in vivo were frequently present, though correlations rarely exceeded r = 0.5 (Table S2), even among our comparisons with the highest statistical significance. Of note, analyses repeated using rank correlations were found to have relative significance (RS-p) and strength (r) largely similar to those obtained with the linear correlations (see Supplemental Methods). Calculation of predictive power scores (PPS) was employed as a complementary analysis approach to Pearson correlations, as this asymmetric nonlinear index can explore predictive nonlinear and asymmetric relationships between in vitro and in vivo data not possible by linear correlation analyses alone (26, 27). PPS analysis of viral titer data in vitro generally agreed with the strength of correlations obtained with Pearson correlations (Table S4). Collectively, these analyses demonstrated that viral titer outcomes in vitro are associated with viral titers from both nasal wash specimens and discrete tissues collected in vivo during the acute phase of infection, even when working with heterogeneous viruses, but can be dependent on the IAV under investigation, culture temperature in vitro, and sampling time and specimen type in vivo. Furthermore, these analyses identified novel areas of investigation to pursue metrics associated with infection progression in both in vitro and in vivo settings.

A MM, by its structure and the biological assumptions it encodes, constrains what range of viral titer measurements are possible over the course of an infection. Experimental measurements then provide further constraints on the shape of the MM-predicted curves, which the MM translates into constraints on its parameters (e.g., rate of virus production by each infected cell, life span of virus-producing cells, etc.). This allows a MM that is well informed by experimental measurements to successfully interpolate and predict unobserved quantities, smooth over experimental variability, and isolate different facets of the replicative fitness of a virus (13, 28). The in vitro data in the analyses considered herein were sparser (titers collected only at 1 to 2 h, 24 h, 48 h, and 72 h) than what is ideally required for MM analysis and often did not include data points during the viral decay phase following titer peak. As such, the MM-predicted viral titer curves were generally well constrained, but less so for the MM parameters.

Mathematical modeling provides an analytical and quantitative approach to examine the magnitude and impact of IAV replication efficiency relative to infection temperature in a more analytical and quantitative way than is possible with standard experimental measures. While experiments only allow us to compare the difference in titers at specific times, MM-derived quantities such as the infection growth rate or the time to reach 1% of peak titer suggested that higher temperatures generally increased the progression rate of the infection (how quickly peak titer is reached) more so than it increased the peak titer (Supplemental Methods). Interestingly, with regard to temperature sensitivity, viruses for which MM-derived peak and mean log₁₀ Calu-3 titers decreased the most when the infection temperature was decreased from 37°C to 33°C were generally found at higher titers at day 3 p.i. In ferret NT, the tissue specimen exposed to the coolest temperatures (Fig. 4I and K) (16, 29). This same drop in peak and mean log₁₀ Calu-3 titers at lower infection temperature was also associated with lower day 1 p.i. NW titers (Table S3F) and a larger (more positive) slope_1,3 (Fig. 4J and L). Taken together, these findings suggest that viruses whose yield is most compromised at lower temperatures in vitro tend to grow more slowly in vivo, peaking at day 3 p.i., the day tissue titers were collected, rather than day 1 p.i. It is also of interest that this was more pronounced among viruses titrated in eggs (largely comprised of avian-origin strains) and not viruses titrated in cells (predominantly mammalian-origin strains), suggesting a potential difference in host adaptation captured by these analyses. To our knowledge, this is the first time a MM has been used to analyze infections conducted at different temperatures and identifies several potential areas of future inquiry. Unfortunately, because individual MM parameters were not well constrained by the analysis, we were not able to ascribe the effect of lower temperature to one or more of the virus replication steps captured by the MM parameters. Despite this, identification of heightened predictive power of 33°C Calu-3 data for upper respiratory tract specimens collected in vivo highlights the utility of employing culture temperatures emulative of the human upper respiratory tract when conducting risk assessment evaluations in vitro.

While it would have been highly desirable, we were not able to expand the same MM analysis to the in vivo infection data. This is in part because most of the ferret infection time course data considered herein either were already at peak titer at the first NW sample collected and decreased thereafter or had a single NW sample measurement collected before reaching peak titer (see Supplemental Methods). While the timing of this sampling is a reflection of appropriate anesthesia schedules for laboratory animals (30), these data lack key kinetic information, specifically the rate at which infection expanded in the host, i.e., the viral titer up-slope, which is required by the MM to robustly extract the replication efficacy of each strain. Furthermore, administration of a high-titer inoculum is necessary when evaluating pathogenicity of IAV with unknown 50% ferret infectious doses, but collection of the first NW specimen 24 h p.i. means that the presence of residual inoculum in this sample cannot be fully ruled out. Despite these challenges, we calculated the growth or decay rate of the ferret NW titer between pairs of measurement time points (see Materials and Methods) and identified one (slope_1,3), which showed strong correlations with mathematical quantities characterizing infection kinetics in vitro. Our findings indicate that viruses exhibiting the most rapid growth in vitro were associated with the greatest decay between NW specimens collected day 1 and 3 p.i., likely driven by mammalian-origin viruses exhibiting robust replication in a mammalian cell line and productively infecting the ferret upper respiratory tract early after infection due to strong binding to α2,6 linked sialic acid receptors. Poorer correlation with measures of titer growth or decay in ferret NW beyond day 3 p.i. may be attributed to elicitation of diverse innate immune responses and other confounders (31); it should be noted that NW specimens are inclusive of multiple independent sites of virus replication throughout the ferret upper respiratory tract, where infection likely proceeds at different rates, peaking at different times (32). MMs incorporating experimental data from in vitro experimentation have been used to inform refinement of in vivo models and kinetics of host response elicited following viral infection, most frequently in mouse models (33, 34), highlighting the need to perform similar analyses in the ferret. As public health efforts increasingly rely on data generated from the ferret model to prevent and mitigate influenza virus infection in humans, it is prudent to similarly explore the use of MMs for data generated in this species, especially as in vitro assessments are often employed before in vivo experimentation in adherence to the 3Rs (replacement, reduction, and refinement) of animal research when employing small mammalian models (35).

This study emphasizes the need for in vivo assessments so long as in vitro experimentation cannot fully recapitulate or predict in vivo virulence phenotypes. That said, it is possible that, under yet-to-be-identified protocols for both in vitro and in vivo experimentation, additional correlates between these experiments could be identified and the predictive power of in vitro data could be significantly increased. For example, the adoption of a more accurate measure of virus sample infectivity (36) or identifying and then better managing in vitro factors that undermine reproducibility across in vitro infection experiments (37) could improve correlations between within-host and in vitro infections (38). Ultimately, this study identifies previously unrecognized correlates between data generated in vitro and in vivo for the purpose of influenza A virus risk assessment and highlights potential future areas of investigation to more quantifiably apply in vitro data toward pandemic preparedness efforts.

We thank the international influenza community for facilitating access to many of the viruses employed in this study.

This project was supported in part by an appointment to the Research Participation Program at the Centers for Disease Control and Prevention (CDC) administered by the Oak Ridge Institute for Science and Education through an interagency agreement between the US Department of Energy and the CDC (to H.M.C. and T.J.K.). This work was supported in part by Discovery Grants 355837-2013 and 2022-03744 (to C.A.A.B.) from the Natural Sciences and Engineering Research Council of Canada (www.nserc-crsng.gc.ca) and by the Interdisciplinary Theoretical and Mathematical Sciences program (iTHEMS, https://ithems.riken.jp/en) at RIKEN (to C.A.A.B.).

The findings and conclusions are those of the authors and do not necessarily reflect the views of the Agency for Toxic Substances and Disease Registry/CDC.