(a) Co-infection and reassortment processes

Significant progress has been made in understanding within-host reassortment processes in spillover host models (i.e. poultry and mammals) using experimental inoculation techniques [16,17]. Rates of de novo reassortment within hosts can be determined by co-infecting single hosts with known parental genotypes and measuring the proportion of shed virions that are reassortants. Such experiments have helped to develop assays for studying reassortment rates within hosts, and demonstrated substantial and highly variable rates of reassortment within-host individuals ([18–22]; e.g. 8.7% of viruses recovered from ferrets co-infected with human and avian strains were reassortants [18], to approximately 50% of viruses recovered from co-infections of chickens with other chicken-derived subtypes [19], to 86% in swine and guinea pigs co-infected with swine or human strains [20,21]). The differences are thought to be owing to functional incompatibilities between viral genome segments or differing receptor binding specificities or other host restrictions, leading to low levels of co-infection in the same cells.

Within-host studies have determined that transmission route, timing of exposures, dose, strains and strain competition all can influence co-infection and reassortment rates [16,17,21]. Higher doses lead to higher rates of co-infected cells [21]. When exposure with two different strains is offset by a short period of time (12 h), co-infection and reassortment rates are higher than when the host is exposed to the two strains simultaneously [21]. Delayed introduction of the second strain likely increases co-infection rates because it allows the first strain time to infect many cells, thus maximizing co-infection potential by the second strain [21]. By contrast, exposures with longer time lags can reduce co-infection because of super-infection exclusion mechanisms that occur later post-infection (i.e. host innate immune function, and viral processes such as receptor interference) [17] and because death of infected cells occurs. As expected, effects of co-infection time lags on reassortment rates can differ depending on host-strain combinations and experimental design [19]. Importantly, these advances in our understanding of within-host reassortment provide a foundation for predictive models to account for co-infection and reassortment processes mechanistically in epidemiological models. An important gap for experimentation is that little work has focused on understanding reassortment in reservoir hosts under natural transmission conditions (figure 1A); how the lessons learned from animal models apply to wild birds infected by environmental transmission remains unknown.

(b) Environmental transmission

Models suggest that environmental reservoirs play an important role in the ecology of IAVs because they allow infection between hosts that infrequently come into direct contact, such as migratory wild waterfowl on breeding grounds [13,14,23,24]. Experimental systems have shown the occurrence of environmental transmission of IAVs among reservoir hosts and from reservoir hosts to spillover hosts [25–28]. However, the role of indirect transmission though the environment relative to other transmission mechanisms is not well understood empirically, which is important for informing models aimed at predicting epidemiological risk. One study identified a correlation between host population density and environmental levels of IAV [29]. Environmental transmission experiments combined with molecular characterization and disease-dynamic models would provide powerful tools for improving our quantitative knowledge of environmental transmission, co-infection and reassortment processes.

One key driver of environmental transmission is persistence of viral viability after shedding into the environment. There is substantial variation in environmental durability of IAVs (between 1 and 600 days) depending on abiotic factors, especially temperature, salinity, pH and mineral content of the medium [30–35] and viral genotype [33,35], including under natural conditions [36]. While some information is available for water and bird faeces, studies with soil [37], sediments [29,33,38] and plant material are few. Variation in viral environmental persistence could be an important component of fitness [39] and should be included in models predicting the ecological and evolutionary dynamics of IAVs. A final important knowledge gap is understanding the relationship between environmental virus concentrations and infectivity.

(a) Environmental sampling

A major challenge with environmental surveillance is capturing the strain diversity of IAVs in the environment [38]. Questions include: do you sample water or soils or both? How many samples per site? How deep do you sample? How many technical replicates? A gold standard for capturing, isolating and identifying true IAV diversity from the environment is still not established [45]. Similarly, the effects of common environmental variables such as water turbidity, rates of dispersion and diffusion and spatial heterogeneity across water or other environments have not been examined, although they affect environmental DNA capture from aquatic organisms [46]. Because knowledge of how to sample the environment for target RNA is still in its infancy, pilot studies structured to optimize sampling protocols are critical (e.g. [47]).

Another major sampling challenge is RNA preservation [48]. Filtration to capture virions/RNA requires hours at a single site (e.g. 800 L) to achieve adequate probability of virus detection [49]. Decontamination of pumping equipment at each site is necessary, further increasing sampling time. While these methods provide a useful means of collecting environmental genetic sequences [46], methodological advancements in RNA preservation/collection techniques are needed to improve the efficiency of collecting high-quality RNA viromes from environmental samples.

(b) Generating sequence data from environmental samples

An initial field-based screen can rapidly prioritize localities for further sampling efforts. Technological advances will soon allow diagnostic PCR assays in the field, with results within hours to minutes. Companies such as Biomeme (Philadelphia, Pennsylvania, USA) have developed field-based nucleic acid extraction kits (M1 Sample Prep Kits) and a hand held real-time PCR thermocycler that attaches to a smart phone (two3™ Real-Time PCR Thermocycler). Both of these have been deployed in the field to identify mosquito pools positive for RNA viruses in less than 2 h to target samples for more thorough sequencing [50].

New sequencing platforms (e.g. MinION from Oxford Nanopore Technologies) hold promise for field-based surveillance of IAV owing to their rapid library prep and run times, and elimination of PCR amplification. The MinION generates long reads that allow sequencing of entire gene segments in a single read, single-molecule sequencing and direct RNA sequencing without conversion to cDNA and PCR amplification [51,52]. For field-based pathogen surveillance, it is possible to run a sequencer from a smartphone (e.g. SmidgION from Oxford Nanopore Technology). All of these technologies currently lack high-throughput capacities, have cost-prohibitive library prep, low sensitivity and high error rates for application to broad-scale surveillance. Despite these hurdles, there is a continual push to improve mobile diagnostic capabilities [53,54], which will inevitably improve efficiency, accuracy and precision of obtaining IAV genomic data from environmental samples.

All contemporary sequencing platforms require a minimum concentration of input nucleic acids, and water samples tend to fall below this threshold, requiring enrichment of target IAV RNA [55]. Centrifugation following filtration has been effective at concentrating viral particles [56]. The gold standard for IAV enrichment is virus culture in either embryonated chicken eggs or Madin–Darby canine kidney cells. These methods work well for clinical samples, but there can be high costs and challenges with isolating the virus from environmental samples [57–59]. Additionally, using virus isolates to study reassortment can be counter-productive because of subtype and genomic selection bias in media used to propagate isolates. PCR is another method for enriching viral genomes, however, PCR can preferentially amplify specific strains and introduce nucleotide errors that can be difficult to distinguish from true, low-frequency virus variants. The ideal approach is direct sequencing of viral genomes without PCR (see below).

Another major challenge with viromic data from environmental samples is delineation of individual genomes—i.e. how do we distinguish potential reassortants from multiple parental genotypes or viral mixtures? A common method is to isolate individual genotypes using limiting dilution plaque assays and subsequent whole-genome sequencing. However, plaque assays can cause biased replication levels between multiple strains because of tissue tropism [60], in vitro strain mutation [61] and overlapping plaques [62], all which can obscure the delineation of genotypes. Serial dilution can reduce plaque overlap but is very time consuming.

The advent of single-virus sequencing techniques [63,64], similar to single-cell genome sequencing techniques of eukaryotic and bacterial genomes [65], holds promise for circumventing several potential biases from viral isolation techniques while providing sequence data from individual virons. Using flow cytometry to isolate virions has improved sequencing of single-virus genomes [63,66], but RNA input requirements for library prep are still too high to allow the sequencing of a single, non-enriched virus genome from environmental samples. However, methods for conducting single-virus RNA sequencing on a Nanopore or similar single-molecule sequencing platform will undoubtedly be realized in the near future, providing a valuable method of delineating IAV genotypes that have not been subject to selective media.

(c) Assembling genes and genomes for reassortment analysis

The most widely used sequencing platforms generate short reads, which makes genome assembly challenging because there may not be enough overlap between reads to determine alignments with high enough certainty. A common first step is to remove non-target sequence reads. Environmental samples include nucleic acids from non-target organisms, for which sequenced genomes are not available. The effectiveness of filtering is thus reduced, which increases computing time for genome assembly. The two main approaches for viral genome assembly are reference-based mapping and de novo assembly. Mapping reads to a reference genome can simplify assembly, but reference genome choice is not trivial and IAV genomic diversity in wild birds is very high, which can bias results when trying to recover viral population diversity [67]. De novo assembly of segmented viral genomes is complicated because of incomplete and uneven coverage of different segments, which can lead to skewed results or failed assemblies [68]. A useful approach could be de novo assembly to build reference genomes directly from the sample reads then map all the reads back to this genome to recover the viral population diversity, infer haplotypes and estimate their frequencies [67,69,70]. The ultimate goal is to accurately assemble all genome variants in systems with high levels of diversity and determine dominant and low-frequency strains, which can greatly influence epidemiologic outcomes.

(b) Bottom-up approaches: disease-dynamic modelling

Within-host viral dynamics of IAV have now been reasonably well studied using experimental data [82,83], providing a strong foundation for developing predictive models of reassortment dynamics. Important factors determining within-host viral kinetics for IAV include target-cell availability [84], immune system factors [85], spatial distribution of target cells [86] and initial dose [87]. While these viral kinetic models have been developed using data from human infections [82,83], their insights and methods have begun to be applied to avian hosts [88]. The most important factor for reassortment is co-infection [16,17]. Thus, initial dose, target-cell availability and the distribution of target cells will play a primary role in predicting de novo reassortment rates. Once reassortment has occurred within a host, selection within the host and the size of the transmission bottleneck will determine whether a reassortant is shed into the environment [16] (figure 2a). The recent advances in understanding of intracellular and within-host processes that govern reassortment [16,17] provide conceptual and quantitative knowledge for models of within-host viral and reassortment dynamics. In the larger context, the reassortment processes and within-host viral dynamic models can be nested in between-host models that account for selection, transmission bottlenecks and environmental transmission processes (e.g. [39]) to develop a predictive framework that can be validated with experimental data (figure 2a).

Controlled experiments that allow transmission to occur naturally coupled with data-driven modelling frameworks (as above, figure 2a) together form a valuable stepping-stone toward inferring epidemiological risk from genetic samples. Controlled experiments are important for restricting the range of potential genotypes and environmental factors so that models can be designed to capture epidemiological behaviour based on a limited and known set of factors. In high-dimensional systems such as IAV, model-guided fieldwork [89]—i.e. iterative model development, prediction and experimentation (figure 2a)—can be applied for efficient experimental design and knowledge building. By this approach, models could be used to predict which processes are the most important to assess. Then experiments would test our knowledge of how ill-understood processes such as environmental transmission contribute to the observed outcome. For example, a poor-fitting model of diversity-pressure-phenotyping would suggest that there are ill-defined processes occurring in the system that are not well captured or are absent in the model. Then the model and experimental design could be adjusted to elucidate these ill-understood mechanisms. This concept is not new [89] but it is surprisingly underused for designing empirical studies of complex ecological and epidemiological systems with ample process uncertainty. The iterative approach of hypothesis generation → model building and prediction → data collection → model evaluation aims to reduce process uncertainty by targeted learning. Thus, it is a powerful approach for identifying important gaps in our ability to predict reassortment in environmental reservoir-host systems (figure 1B), and to design effective experiments that will best inform the identified gaps. Once this type of mechanistic epidemiological model is developed and validated using experimental data from controlled studies with natural transmission, it could be used to predict epidemiological risk from surveillance of diversity-pressure-phenotypes data in environmental reservoir-host systems.

As an illustration of how these cross-scale processes could be explored empirically, we describe a possible experimental study to measure diversity-pressure-phenotypes and resulting infection risks in an environmental reservoir-host system. A potential study would simulate a scenario where multiple hosts infected with different virus strains would be allowed to contaminate a quasi-natural environment. Infected hosts would then be removed and susceptible individuals would be exposed to the contaminated environment to allow for infection. Hosts and the environment would be sampled longitudinally at regular intervals throughout the experiment. All samples would be sequenced to determine de novo reassortment rates and diversity-pressure-phenotypes (figure 1; which measures the frequency dynamics of particular genotypes and their associated phenotypes, i.e. reassortant emergence). Testing both reservoir and spillover host species (e.g. mallards and chickens) would allow for determination of potential host differences in susceptibility and the impacts of diversity-pressure-phenotypes on epidemiological dynamics. A similar design could be applied to conditions that only allow direct contact between hosts. Comparing results from both experiments would help quantify transmission mechanisms and determine which mechanisms lead to higher reassortment risk.

The findings and conclusions in this preliminary publication have not been formally disseminated by the U.S. Department of Agriculture and should not be construed to represent any agency determination or policy. The authors thank three anonymous reviewers for helpful comments and the organizers of a workshop on pathogen spillover in MT 20–23 February, which helped with idea development, and DARPA for funding the workshop.