Evolutionary trajectories are unique for different subtypes

Wild birds were live-caught and sampled across North America between 2008 and 2018 and the 888 influenza sequences generated (including H13: n = 23 sequences and H16: n = 58 sequences) were compared with global reference sequences for the three subtypes. The HA gene phylogeny showed markedly different evolutionary dynamics for each of the three subtypes: H13, H16 and HPAI H5. Molecular dating of the gull-associated subtypes revealed persistence of H16 since approximately 1969 (95% HPD: May 1966 –April 1972) and of H13 since approximately 1973 (95% highest posterior density, HPD: December 1971–June 1974). This suggested a similar time of origin for the two gull-associated subtypes. In comparison, HPAI H5 diversified rapidly over a much shorter time frame since 1996 (95% HPD: January 1996 –December 1996).

For H13, viruses belonged to three major clades that continue to be detected across a wide geographic range including all hemispheres (S1A Text). Concurrent circulation of multiple clades of H13 contrasts with the evolution of H16. For H16, most circulating viruses belong to a single dominant clade that has persisted via localised transmission over a smaller geographic area for successive years (S1B Text). The pattern of endemic circulation is demonstrated in Alaska where the same lineage has persisted for approximately a decade since 2009 (95% HPD: November 2009–June 2010). The mean evolutionary rate (or substitution rate) of H13 was estimated to be 5.81 x 10⁻³ substitutions per site per year (95% HPD: 5.36–6.29 x 10⁻³), faster than H16 for which we estimated a mean evolutionary rate of 4.61 x 10⁻³ (95% HPD: 4.12–5.15 x 10⁻³). These results indicate that the evolutionary dynamics of IAV subtypes can be unique and distinct even when hosted by the same avian reservoir, such as H13 and H16 in gulls.

The evolution of HPAI H5 indicates extensive diversification over time (S1C Text) consistent with the known clade designations determined by the WHO/OIE [28] resulting in the most recent clade 2.3.4.4 (95% HPD: February 2009 –April 2011) that expanded into continental Europe, Asia and Africa and a rare introduction event into North America via the Pacific-Central flyways in approximately 2014 (95% HPD: November 2013 –May 2014). Most lineages predominately circulate in East and Central Asia with the exception of clade 2.2 that since 2005 (95% HPD: November 2004 –July 2005) has become endemic in the Black Sea-Mediterranean and Africa (S1C Text). Diversification of clade 2.3 into multiple co-circulating sub-clades has been associated with a geographic range spanning Eurasia but with no evidence of occurring along the Atlantic seaboard. We estimated that HPAI H5 evolved at a rate of 5.65 x 10⁻³ (95% HPD: 5.21–6.15 x 10⁻³), comparable to the evolutionary rates of H13 and H16.

Hotspots rarely overlap for subtypes adapted to the same host

We reconstructed the global phylogeography of the three subtypes: H13, H16 and HPAI H5 using a discrete trait model based on 11 geoclusters of the globe. Two geographic areas emerged as major sources of H13: the Mississippi-Atlantic flyway in the US (26.0% Markov jumps) and the Black Sea-Mediterranean region in Eurasia (22.0% Markov jumps: Fig 1A). The Mississippi-Atlantic acted as a source of spread for regions north (North Atlantic) and south (South America). The Black Sea-Mediterranean acted as a source of H13 for continental Europe and East Asia. Using Markov rewards to estimate the proportion of time H13 spent in each region revealed that these viruses spent comparable amounts of time across regions. H13 spent most time in the Mississippi-Atlantic (24.2%) followed by the Black Sea-Mediterranean (15.5%), continental Europe (12.3%), South America (11.3%) and East Asia (10.3%). These viruses spent least amount of time in Alaska and the Pacific-Central flyway of the US. H13 transmission between the eastern and western hemispheres was rare with the exception of East Asia acting as a repeated source of viruses for Alaska, highlighting this region of connectivity across the Pacific Ocean.

Download:

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

Fig 1. The contrasting pattern of geographic jumps, hotspots and global distribution of influenza A viral subtypes.

The heatmaps: A) H13 (green), B) H16 (purple) and C) highly pathogenic avian influenza H5 (red), indicate the frequency of transitions between locations estimated using a discrete trait phylogenetic model and the number of location transitions was determined using Markov jumps. The proportion of time the virus spends in each location (center panel) is indicated by the bar charts with each subtype showing a uniquely different geographic distribution. The global distribution of each subtype is depicted by shading on the maps (right panel). Lines indicate the route of inter-hemispheric virus flow and represent the branching pattern of internal and external nodes of the underlying phylogenetic tree. Basemap made with Natural Earth (https://www.naturalearthdata.com).

https://doi.org/10.1371/journal.ppat.1010062.g001

For H16, high latitude regions including Northern Europe (41.0% Markov jumps) and to a lesser degree, the North Atlantic coast of the Americas (23% Markov jumps) were important sources of virus (Fig 1B). Northern and continental Europe were strongly connected (24.1% Markov jumps), as were the North Atlantic and Mississippi-Atlantic flyway of the US (16.8% Markov jumps), indicating localized patterns of H16 spread. Alaska was also identified as a source of H16 for locations further south along the Pacific and Central flyways of the US. Estimation of the time H16 spent in each region using Markov rewards indicated a pattern highly skewed towards a few locations: Northern and continental Europe (35.9% and 25.2%) and Alaska (16.0%). H16 spent the least amount of time in the Black Sea-Mediterranean and South America. H16 transmission between hemispheres occurred infrequently via Northern Europe into the North Atlantic coast of the Americas, and inter-hemispheric transmission was otherwise limited.

Geographic distribution of HPAI associated with proximity to endemic hotspots

HPAI H5 revealed a geographic distribution centred in the eastern hemisphere with transmission primarily seeded by the Black Sea-Mediterranean (29.45% Markov jumps), East Asia (17.77% Markov jumps) and Central Asia (17.37% Markov jumps) (Fig 1C). The Black Sea-Mediterranean, Africa and Central Asia were important sources of HPAI H5 for continental Europe, while further transmission to Northern Europe was rarely detected. Viral flow from East to Central Asia was strongly supported by the large number of Markov jumps estimated (17.5%). Consistent with these results, analysis of time spent by HPAI H5 in each geographic location indicated East Asia was the primary site of circulation (33.1%). Furthermore, the amount of time spent by these viruses in each location was based on proximity to East Asia; i.e. Central Asia (24.5%), Black Sea-Mediterranean (15.3%) and Africa (9.5%). Transmission between the eastern and western hemispheres was infrequent, but also appeared to be dictated by proximity to East Asia due to evidence of the HPAI H5 circulating primarily along the Pacific-Central flyways of the US, compared to the Atlantic coast.

Dispersal between hemispheres occurs via the Pacific and Atlantic Rim

Transmission between North America and Eurasia was consistently rare for all three subtypes, relative to transmission within each region. For H13, viral migration between East Asia and Alaska indicated spread via the Pacific Rim (Fig 1A). A similar route of introduction into North America was estimated for HPAI H5, whereby East Asia was identified as a source for the incursion of virus into the Pacific-Central flyways (Fig 1C). For H16, a different route of introduction into North America was identified involving the Atlantic Rim. Northern Europe acted as a viral source for the Northern Atlantic seaboard of the Americas—a pair of locations in proximity (Fig 1B). In summary, both the Pacific and Atlantic Rims acted as gateways for the introduction of IAVs into North America with a correlation between source and sink locations that are in geographic proximity along the continental margins.

Competition and co-existence underlies subtype distribution

We focused on Alaska to address how multiple subtypes interact and behave when co-circulating in the same population. The H16 subtype dominated in gulls that migrate and breed in Cordova, Alaska as evidenced by detection of H16 during each of the 10 years of the study (Fig 2). H16 consisted of two geographically distinct clades (Northern Europe and Mississippi-Atlantic US). The temporal distribution of these H16 clades based on multiple components analysis (MCA) indicated that they rarely co-circulated in Cordova, Alaska (S2 Text), suggestive of competitive exclusion occurring between lineages. Multiple introductions of H13 from East Asia and to a lesser degree South America were also observed but persisted for only 1–2 years before they were no longer detected in the gull population. Temporal overlap between H13 from East Asia and H16 from Mississippi-Atlantic US based on MCA, suggested co-circulation and a lack of competition between these two geographically distinct lineages.

Download:

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

Fig 2. The occurrence of influenza A viral subtypes over a 10-year period in Alaska reflect neutral and non-neutral evolutionary processes.

A) The temporal pattern consists of two H16 clades (blue: Northern Europe, and purple: Atlantic-Mississippi) that alternate, with introductions of H13 from East Asia (green) and South America (yellow). Geographic origins were determined by most recent common ancestor (MRCA) phylodynamic analysis. B) Genetic diversity of H13 through time C) relative to H16. D) Cross-reactivity of anti-influenza antibodies from Cordova gulls and Minto Flats ducks determined by hemagglutination inhibition (HI) assay. Many antibodies against H13 and H16 were cross-reactive, while few sera recognized H5 low pathogenic virus. E) Antigenic properties of H13, H16 and H5 clades mapped onto two-dimensional space. Virus clades positioned centrally (i.e. H13 from South America) were more cross-reactive than viruses located at the edges (i.e. H5 from Pacific-Central, H13 from East Asia). Circle markers (solid) are size-scaled according to the prevalence of each clade in the Cordova gull population, or their absence (outline). F) Conceptual diagram of the antigenic distance between subtypes relative to prevalence. Gulls are proposed as the reservoir host for H13 and H16, which are antigenically related and overlap resulting in cross-reactivity and competition between H13 and H16. G) Conversely, ducks (and geese and swans) are proposed as the reservoir for highly pathogenic avian influenza H5. No antigenic overlap occurs between HPAI H5, H13 and H16, such that gulls are largely immuno-naïve to HPAI H5. Animal silhouettes from http://phylopic.org.

https://doi.org/10.1371/journal.ppat.1010062.g002

The genetic diversity of H13 circulating in Cordova, Alaska was low compared to H16, but rapidly accrued with the introduction of the South American clade, as inferred from Bayesian skyride plots that showed changes in viral effective population size over time (Fig 2B and 2C). After introduction of the South American clade, the genetic diversity of H13 markedly declined within one year indicating local extirpation or exclusion of this clade from the viral population. In contrast the diversity of H16 was two-fold higher and relatively stable over time (Fig 2C). We found that despite turnover of the two H16 clades (Northern Europe and Atlantic-Mississippi) over time, the genetic diversity did not increase or diminish suggesting they were genetically equivalent and occupied similar spaces in the viral fitness landscape defined by gulls in Cordova, Alaska.

Frequent reassortment between prevailing subtypes

To evaluate the degree of overlap and interaction between H13 and H16, we estimated rates of reassortment by quantifying transitions of the PB2 internal segment between subtypes (S3 Text). H16 was a frequent source (or “donor”) of PB2 for the H13 subtype infecting the Cordova gull population, indicating reassortment and competition between subtypes at the individual host level. Conversely, H13 also donated the PB2 segment to H16, but at a rate with lower statistical support (S3 Text). Other HA subtypes that reassorted with H16 included the H1, H2, H5 and H6 group, while H13 reassorted with the H8, H9, H12 group. However, to a large degree the H13 and H16 subtypes showed limited reassortment with other HA subtypes, which primarily swapped PB2 internal segments among each other as a distinct and separate viral pool.

Subtypes evolve by antigenic selection and geographic structuring

Results from serological testing of Alaskan gulls and ducks also supported interaction between H13 and H16 with evidence of antigenic cross-reactivity between the two subtypes depending on geographic lineage. For example, sera against an H16 clade from Northern Europe was strongly cross-reactive with an H13 clade originating from South America (Fig 2D). Additionally, sera against the H13 clade from South America recognized two other virus clades H13 (from East Asia) and H16 (from the Atlantic-Mississippi) and was therefore the most broadly cross-reactive strain according to its central position in the MCA map (Fig 2E). This result implied antigenic overlap may have caused the low levels of H13 from South America detected in the study population (one out of 10 years). In contrast, the H16 clade from Northern Europe predominated for 7 out of 10 years (Fig 2A) and was not cross-reactive against H13 from East Asia (Fig 2E), the second most common clade in the gull population. The two most common viral clades never temporally overlapped. The sera results indicate that geographic structuring rather than antigenic overlap was responsible for the lack of co-circulation observed.

Unlike the gull-associated subtypes of H13 and H16, the H5 subtype (LPAI) was antigenically unrelated and did not cross-react with any of the other viruses based on the serological analysis. Sera from only one duck recognized the H5 virus and at a relatively low titer of 1:32 hemagglutinating units (Fig 2D). The H5 virus occupied a distant position at the edges of the MCA map that confirmed lack of antigenic overlap with other viruses tested (Fig 2E). Ducks as a host group generated sera that rarely recognized H13 and H16, relative to gulls. Two duck sera recognized the H16 clade from North America, and one recognized the H13 clade from South America (Fig 2D). These findings concur with the view that ducks are largely immuno-naïve against gull-associated viruses, and likewise, gulls are largely immuno-naïve against duck-associated viruses. Overall, this supports that host barriers to virus transmission exist between gulls and ducks, which can subsequently cause immunologically naïve populations with heightened susceptibility to antigenically distinct viruses (Fig 2F and 2G).

Spillover of gull-associated virus is uncommon

To compare which species were involved in host jumping, we reconstructed the phylodynamics of H13, H16 and HPAI H5 using a discrete traits model focused on taxonomic order. For H13 and H16, gulls acted as a source of virus to a limited number of other species, primarily wild ducks (H13: 42.0%; H16: 32.9% host jumps) and to a lesser degree shorebirds (H13: 20.7%; H16: 22.2%, Fig 3). Rarely did these two subtypes cross into domestic poultry, with the exception of a low number of H13 host jumps between gulls and gallinaceous poultry. Further, shorebirds (H13: 21.0%; H16: 23.1% host jumps) followed by ducks (H13: 10.3%; H16: 16.6% host jumps) were the only two hosts identified as sources responsible for transmission back to gulls (Fig 3). This was reflected in the high transition rates from wild ducks to gulls (H13: mean: 1.45, range: 0.49–2.41; H16: mean: 1.60, range: 0.48–2.72) and shorebirds to gulls (H13: mean: 1.38, range: 0.51–2.25; H16: mean: 2.36, range: 0.96–3.76). This pattern was remarkably consistent for both the H13 and H16 subtypes and indicated a limited host range involving aquatic wild bird reservoirs.

Download:

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

Fig 3. Host jumps and spillover of influenza A viral subtypes at the wild-domestic interface reflects co-evolution with the reservoir host.

The heatmaps: A) H13 (green), B) H16 (purple) and C) HPAI H5 (red), indicate the frequency of transitions between host pairs estimated using a discrete trait phylogenetic model and the number of transitions between species/taxa was determined using Markov jumps. Host interfaces between wild and domestic birds (middle panel) are inferred based on host transitions that are well supported by Bayes Factors (BF). Dashed lines indicate known but infrequently sampled transmissions that could not be included in the analysis due to low sample size. Transition rates between host pairs are shown in the box and whisker plots (only the most strongly supported rates are shown). Large variation in transition rates is observed for HPAI H5N1 with highest rates between wild birds, followed by domestic poultry. In contrast, transition rates for H13 and H16 are lower with consistent patterns observed between host pairs for each subtype. Animal silhouettes from http://phylopic.org.

https://doi.org/10.1371/journal.ppat.1010062.g003

HPAI H5 spillover is multi-faceted involving more than ducks

For HPAI H5, a much wider range of avian hosts contributed to transmission dynamics, both domestic and wild. Our phylodynamic reconstruction largely focused on the evolutionary pathways of the most recent clades, including clade 2.2.1 and clade 2.3.4 and showed that spillover is multi-faceted involving transmission chains other than just wild to domestic ducks. Within domestic species, geese played a key role as source of virus for domestic ducks (8.2% host jumps), that in turn served as a source of virus to chickens and turkeys (11.1% host jumps). Host jumps at the domestic-wild interface primarily occurred from domestic geese to land birds (6.0% host jumps, ie. corvids, passerines and raptors) and domestic ducks to gulls (2.4% host jumps). However, this type of domestic to wild spillover occurred at much lower frequency than transmission from wild to domestic hosts.

Host jumps from wild to domestic birds primarily involved dispersal from wild geese and swans to their domestic counterparts (7.8% host jumps). Additionally, land birds were an important source to HPAI H5 to domestic ducks (8.6% host jumps) and to a lesser degree gallinaceous poultry (3.1% host jumps). Within wild birds, geese and swans played an outsized role in the dispersal to wild ducks (12.4% host jumps), followed by land birds (9.9% host jumps) and gulls (8.5% host jumps). Of interest, gulls infrequently acted as a source of virus back to wild geese and swans (3.3% host jumps). Transition rates for HPAI H5 indicated that dispersal was most efficient between wild-to-wild hosts and domestic-to-domestic hosts, reflecting a barrier between the two host types. For example, the highest transition rates occurred between wild geese/swans to wild ducks (mean: 2.50, range: 1.66–3.34) and domestic ducks to chickens/turkeys (mean: 2.26, range: 1.51–3.01).

Gulls contribute to rapid spread of HPAI and slow spread of LPAI

The rate of spread by different hosts was investigated for each subtype using a joint continuous geographic and discrete host diffusion model. On average, the HPAI H5 subtype spread at a rate of 4250 km/year among wild birds (95% HPD: 3009–5620) (Fig 4A), 4 times faster than H13 that spread at a rate of 1029 km/year (95% HPD: 360–1970) (Fig 4B) and over 8 times as fast as H16 that spread at a rate of 521 km/year (95% HPD: 281–773) (Fig 4C).

Download:

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

Fig 4. Ecologically divergent hosts contribute to the diffusion and geographic expansion of influenza A virus.

A) The diffusion rates of highly pathogenic avian influenza (HPAI) H5 (red), B) H13 (green) and C) H16 (purple). D) Comparison of two dispersal metrics of HPAI H5: viral diffusion (km/year) and viral wavefront distance (km) plotted for each host type (round markers). The red line indicates the negative relationship between viral diffusion and wavefront distance of HPAI H5 based on regression analysis, with grey shading indicating the 95% confidence interval. E) Viral wavefront distances for HPAI H5 fluctuate over time with a large peak during 2005–2006 and a gradual increase from 2010 onwards. Dispersal rates (km/year) and wavefront distance (km) were estimated using a joint continuous geographic and discrete host diffusion model. Wild bird outbreak data are courtesy of the World Organisation for Animal Health. Available from: https://wahis.oie.int/#/home.

https://doi.org/10.1371/journal.ppat.1010062.g004

When diffusion rates were evaluated by each host taxonomic group, wild birds contributed to the fastest rates of geographic spread of HPAI H5 (Fig 4A), with gulls playing a key role in moving the virus more rapidly than any other host (5919 km/year, 95% HPD: 4057–7897). Wild ducks (4426 km/year, 95% HPD: 2848–6286) and geese and swans (4269 km/year, 95% HPD: 3478–5145) were also important for rapid spread, while land birds (2384 km/year, 95% HPD: 1653–3152) moved the virus much slower compared to aquatic species. The spread of HPAI H5 by domestic birds was relatively slow compared to wild aquatic species. Domestic geese contributed to fastest spread (2970 km/year, 95% HPD: 2346–3656), while ducks (1830 km/year, 95% HPD: 1326–2376) and chickens (1423 km/year, 95% HPD: 898–1977) were associated with the slowest rates of HPAI H5 dispersal.

For H13 and H16, which showed a more restricted host range, taxonomic groups were considered at lower resolution to determine species-level contributions to spread. Our results showed substantial variation in the rates of spread between and within wild species. For H13, Glaucous-winged Gulls (Larus glaucescens) spread the virus fastest (1524 km/year, 95% HPD: 567–2603), followed by shorebirds (1224 km/year, 95% HPD: 407–2351), Black-headed Gulls (Chroicocephalus ridibundus) (1179 km/year, 95% HPD: 358–2161), Ring-billed Gulls (Larus delawarensis) (951 km/year, 95% HPD 287–1752) and finally Herring Gulls (Larus smithsonianus) (605 km/year, 95% HPD: 262–1032) (Fig 4B). For H16, spread was much slower. Herring Gulls spread H16 the fastest (635 km/year, 95% HPD: 373–904), followed by Glaucous-winged Gulls (536 km/year, 95% HPD: 262–833) and Black-headed Gulls (385 km/year, 95% HPD: 202–581) (Fig 4C). We lacked samples from shorebirds to estimate dispersal rates for H16 and lacked samples from ducks to estimate dispersal rates for either H13 or H16.

Domestic birds drive slow but steady range expansion of HPAI

To determine host-specific differences in the geographic expansion of the virus, the wavefront distance was estimated from the continuous geographic diffusion model. The wavefront distance tracks the distance from the origin of the epidemic to the location of currently circulating phylogenetic lineages, measured as a great circle distance and averaged across all lineages. For HPAI H5, an inverse relationship was suggested between wavefront distance and diffusion rates (Fig 4D). Gulls (197.25 km) and wild geese and swans (164.68 km) contributed least to geographic expansion of the virus from detected origin. In contrast, domestic geese (676.93 km) and domestic ducks (526.04 km) were responsible for the greatest range expansion of HPAI H5. These trends were evaluated over time and indicated a 2005–2006 peak in expansion coinciding with an increasing frequency in wild bird outbreaks based on global case count data [29]. A second sustained expansion started in 2010 that coincided with the emergence of clade 2.3.4.4 based on molecular dating (95% HPD: 2009.21–2011.41). This period also coincided with an increasing frequency and magnitude of reported wild bird outbreaks across the globe (Fig 4E).

Ethics statement

Procedures for capture and handling of wild birds were approved by the Tufts University Institutional Animal Care and Use Committee (Protocol # G2017-118) and the U.S. Geological Survey Bird Banding Laboratory (Federal Bird Banding Permit: 21963 and 23432).

Sample collection

Wild birds representing the orders Anseriformes and Charadriiformes were sampled in Alaska and Massachusetts in North America between 2008–2018. In Alaska, wild waterfowl primarily dabbling ducks including Mallard (Anas platyrhynchos), Northern Pintail (Anas acuta) and American Green-winged Teal (Anas carolinensis) were sampled from 2008–2012 at Minto Flats State Game Refuge, interior Alaska (lat. 64.90, long. -148.85), a wetland complex that serves as a major breeding, molting and staging area. Birds were captured and swabbed for oral-pharyngeal and cloacal samples, and blood was collected from the jugular or cutaneous ulnar vein in the wing. Detailed methods for capture and influenza sampling are published in Hill, Ma [53].

Gulls and shorebirds were sampled at Cordova, southcentral Alaska (lat. 60.59, long. -145.77) between 2009 and 2018, a breeding and stopover site for wild birds migrating along the Pacific Rim. Glaucous-winged Gulls were captured with a net launcher during the breeding and post-breeding season (April-September), while Least Sandpipers (Calidris minutilla) and Western Sandpipers (Calidris mauri) were captured with mist nets during the spring migration (April-May). Oral-pharyngeal and cloacal swabbing was performed, and blood was collected from the brachial vein. Fecal sampling was performed weekly at congregation sites for gulls and shorebirds from March through October.

On the Atlantic Rim, gulls were captured throughout Massachusetts (lat. 42.36, long. -71.06) at urban habitats that attract migratory and over-wintering gulls. Herring Gulls (L. smithsonianus), Ring-billed Gulls (L. delawarensis) and Greater Black-backed Gulls (L. marinus) were captured with a net launcher and birds were sampled as per Ineson, Hill [51]. Sampling was performed weekly from October to April coinciding with the non-breeding season between 2012 and 2014. All swab samples were stored in VTM (Remel) at 4°C in the field for up to 12 hours and transferred to -20°C or -80°C storage prior to laboratory analysis.

Influenza screening

Viral RNA was extracted from swab samples using the Mag-Bind Viral DNA/RNA 96 Kit (Omega Bio-Tek Inc., Norcross, Georgia, USA) on a semi-automated KingFisher Purification System robot (ThermoFisher Scientific, Waltham, Massachusetts, USA). One-step real-time RT-PCR was used to detect influenza RNA using primers targeting the matrix gene [54] and qScript XLT 1-Step RT-qPCR ToughMix (Quanta BioSciences, Gaithersburg, Maryland, USA). Samples with a cycle threshold (Ct) value less than 45 were considered positive.

To amplify virus, VTM (100 μl) from positive samples was inoculated into the allantoic cavity of 9- to 11-day old embryonating specific pathogen-free chicken eggs (Charles River Laboratories, Wilmington, Massachusetts, USA) and incubated at 37°C for 72 hours or until embryo death, as detected by daily candling. RNA was extracted from the allantoic fluid and matrix rRT-PCR was performed to screen for presence of IAV [55]. The eight genomic RNA segments of IAV were simultaneously amplified using a multi-segment RT-PCR strategy [56]. Illumina libraries were prepared using the Nextera DNA sample preparation kit (Illumina, Inc., San Diego, California, USA) with half-reaction volumes as described previously [57]. Whole-genome sequencing was performed on the DNA amplicons at the J. Craig Venter Institute in Rockville, Maryland, USA as described by Nelson et al. [58] and all sequences deposited into GenBank (S1 and S2 Data). Subtyping was confirmed by BLASTn of the sequence against isolates in GenBank and identifying the subtype match that showed highest percentage identity.

Antigenic characterization of viruses

To test for the presence of influenza antibodies in serum, an Enzyme-Linked Immunosorbent Assay (ELISA) kit (IDEXX AI MultiS-Screen, Westbrook, Maine, USA) was used. Absorbance values were measured with an Epoch Spectrophotometer (BioTek Instruments, Winooski, Vermont, USA) and the mean absorbance value for each sample was used to calculate the sample to negative ratio. Samples with a ratio ≤0.5 were considered positive for influenza antibodies. Seropositive samples from Alaskan gulls and ducks were selected to assess cross-reactivity between strains and subtypes using hemagglutination inhibition (HI) assays as per the methods of Puryear, Keogh [59]. Only sera from flighted juveniles (between 2 and 6 months) were assayed in order to maximize the likelihood that antibodies were de novo (not materially derived) and were generated against first IAV exposure. Serum was defined as positive for HI titers equal to 1:32 or above.

The antigen panel for the HI assay included representatives of H5, H13 and H16 collected from Alaskan gulls and ducks belonging to different geographic origins inferred by phylogenetic analysis of the HA segment. We estimated the geographic source of the most recent common ancestor (MRCA) of selected antigens using ancestral state reconstruction applied to each HA phylogenetic tree. All nodes that were used to infer the geographic origin of the MRCA were statistically supported by a posterior probability >0.95. The MRCA was identified from one of the prominent geographic clusters (‘geoclusters’) from across the globe, and implemented using BEAST v1.10.4 [60,61] following the methods described in detail below.

The HI assay included two viral antigens for each of the three HA subtypes, a total of six antigens. For H5: A/sandpiper/Southcentral Alaska/16MB01156/2016(H5N2) and A/sandpiper/Southcentral Alaska/16MB01220/2016(H5N2) both originating from Pacific-Central United States (US) inferred from the MRCA. For H13: A/Glaucous-winged Gull/Southcentral Alaska/17MB03606/2017(H13N2) originating from East Asia and A/Glaucous-winged Gull/Southcentral Alaska/ 15MB01632/2015(H13N6) originating from South America. For H16: A/Glaucous-winged Gull/Southcentral Alaska/16MB02941/2016(H16N3) originating from Northern Europe, and A/Glaucous-Winged Gull/Southcentral Alaska/20MB02834/2020(H16N3) originating from Mississippi-Atlantic U.S.

To assess the degree of antigenic overlap between viruses, the titers from the HI assay were visualised using a multiple correspondence analysis (MCA) performed using JMP Pro v.14.3 [62]. The MCA uses categorical data as input, which for this study included i) the geographic origin associated with each virus based on MRCA analysis, and ii) the reactivity of the sera against each virus (HI present/absent). Through representation of these variables in two-dimensional Euclidean space, clustering of viruses was assessed and interpreted as antigenic overlap. In contrast, viruses that occupied different spaces in the MCA plot were interpreted as antigenically distant.

Linking viral sequences, geography and host

A total of 888 complete influenza sequences representing 12 HA subtypes were generated from our North American wild bird surveillance between 2008 and 2018. This included H13 (n = 23 sequences) and H16 (n = 58 sequences) used in this study. For comparative analysis, all HA sequences belonging to HPAI H5, H13 and H16 subtypes were obtained from the Global Initiative on Sharing All Influenza Data (GISAID) on 20^th September 2019. Search criteria included complete sequences from avian hosts for which the collection year and month was known, as well as geographic location.

For each geographic location we extracted centroid coordinates for the lowest administrative unit using Geonames.org. All sampling coordinates were input into a k-means cluster analysis to determine groupings of sampling sites based on geographic proximity rather than administrative or political boundaries (implemented using JMP Pro v.14.3 [62]). A total of 11 geoclusters were identified that corresponded to Alaska, Pacific-Central US, Mississippi-Atlantic US, Northern Atlantic, South America, Northern Europe, Central Europe, Africa, Black Sea-Mediterranean, Central Asia, and East Asia (S4 Text). These geographic ‘traits’ were used for subsequent phylodynamic analysis.

Hosts were grouped by two traits: species and family/sub-family according to eBird.org (a recognized global authority on ornithological taxonomy). Any environmental samples were removed from the dataset. Host taxa were further parsed as ‘wild’ or ‘domestic’ and any ambiguous data (ie. “duck”, “goose” or “avian”) were removed from the dataset. Ducks or geese were only included if the metadata specified whether the sample was from a wild or domestic bird, or the full species names was provided, allowing inference based on history of domestication in a region (ie. swan-goose in China was considered ‘domestic’, pink-footed goose in the Netherlands was considered ‘wild’).

Downsampling global datasets

To mitigate sampling bias, datasets for each subtype were stratified by geocluster and downsampled to ensure an equivalent number of samples across the 11 geographic traits. Samples were limited to 2005–2018 but were not stratified and downsampled by year because the spatio-temporal distribution was considered an important feature linked to the epidemiology of each subtype. Alignments for each subtype were performed using default settings in MUSCLE v.3.8.31 [63] and trimmed to the open reading frame. For each geocluster (ie. Africa or East Asia), maximum likelihood phylogenetic trees were constructed using RAxML v8.2.12 [64] using the GTRGAMMA substitution model. A general time reversible model was chosen to account for the complexity of the evolutionary process that governs how influenza viruses mutate. Each tree was downsampled to approximately 40 taxa while preserving the maximum amount of genetic diversity using the Phylogenetic Diversity Analyzer tool (www.cibiv.at/software/pda) following the approach of Trovao, Suchard [30]. This downsampling approach was repeated for each subtype stratified by host traits (S5 Text). Each of the subtype datasets was analyzed separately. Analyses were performed to assess spatial spread according to geoclusters: H13: n = 398, H16: n = 191, and HPAI H5: n = 295. Separate analyses were performed to assess spread according to host taxa: H13: n = 298, H16: n = 202, and HPAI H5: n = 367.

Bayesian phylogeny reconstruction

All phylogenetic tree reconstructions using Markov chain Monte Carlo (MCMC) sampling analyses were performed using BEAST v1.10.4 package [60,61] in combination with the BEAGLE library [65]. The dataset included viral sequences collected between 1972 and 2019 to ensure accurate rooting for the tree topology, but sequences prior to 2005 were masked in the phylogeographic analysis as per Hicks, Dimitrov [66]. This allowed us to analyze IAV evolution over a more recent time period that also represented the bulk of available data. Phylogenies were reconstructed using a General Time Reversible nucleotide substitution model with gamma distribution of substitution rates, a Gaussian Markov Random Field Skyride coalescent model and an uncorrelated lognormal clock. Four independent MCMC runs of 200 million chains were performed for each segment. Runs were combined to ensure an effective sample size (ESS) over 200 was achieved in Tracer v1.6.0 and the maximum clade credibility (MCC) tree was determined. A subset of the last 500 trees from the posterior distribution was used to generate an empirical tree distribution used in the subsequent phylogeographic analysis, an approach that reduced computational time and burden [30,67].

Phylodynamics incorporating geography and host

To estimate the spread of virus between geoclusters, a joint discrete trait and continuous geographic diffusion model was performed [30]. For discrete traits, the asymmetric substitution model was selected to estimate pairwise bi-directional viral flow between geoclusters. A Bayesian stochastic search variable selection (BSSVS) was used to summarize the diffusion rates and Bayes factors (BF) [68] were estimated using SPREAD v1.0 [69]. Rates were considered statistically supported when BF >3.0, strongly supported when BF > 10 and decisively supported when BF >100 [70].

To estimate the Markov jump counts between geographic states along the nodes of the trees we applied the continuous-time Markov chains (CTMCs) model [71]. The number of jumps between states was expressed as a proportion of the total number of transitions occurring across the tree and depicted by heatmaps. Heatmaps were plotted using JMP Pro v.14.3 (SAS Institute Inc, Cary, North Carolina, USA). To quantify the proportion of time spent by the virus in each of these geographic states, Markov rewards were also estimated [71].

To complement the discrete traits analysis, we estimated the spread of virus through space using the continuous traits (latitude and longitude) of sampled taxa. To account for variation in the diffusion rates along branches, a Cauchy relaxed random walk model of evolution was selected [72] with 0.001 random jitter added to spatial coordinates to reduce overlapping. The reconstructed dispersal history in continuous space was visualised using SPREAD v1.0 and qualitatively assessed to understand the degree of inter-hemispheric transmission among the H5, H13 and H16 subtypes, and specifically trans-Atlantic and trans-Pacific Ocean movements of the virus.

To elucidate the role of different avian hosts in the movement of viruses, we quantified host-specific diffusion rates and geographic expansion [30]. The host-specific diffusion rate (km/year) is the mean speed the virus travels within a host group, and represents a collective estimate made up of multiple individuals and species. The host-specific geographic expansion or ‘wavefront distance’ (km) measures the distance from the origin of the epidemic to the location of currently circulating lineages per host group, measured as a great circle distance and averaged across all lineages [30]. Both measures were estimated from the joint discrete (host) trait and continuous (geographic) diffusion model, performed independently for each of the three subtypes. Although both discrete and continuous process are modelled separately, the joint inference allows host-specific contributions to the spatial dispersal dynamics to be estimated [30]. To account for transmissions among multiple hosts along the length of the branch, the latitude and longitude at each time point is used to estimate statistics from the compound host-specific trajectories.

To evaluate viral flow between different avian host taxa or ‘host jumping’ a discrete trait phylodynamics analysis was performed. For HPAI H5, host taxa were considered at the higher taxonomic resolution of family or sub-family and parsed into ‘wild’ or ‘domestic’ categories (ie. domestic geese). For H13 and H16, subtypes that have a narrower host range, host taxa were considered at both the sub-family and species level. To assess BF support between pairwise host taxa, BSSVS analysis was performed as per geographic discrete traits. Markov jumps counts between host taxa along the nodes of the tree were estimated using the CTMC model. To estimate the contribution of host taxa to the dispersal and geographic range expansion of the three subtypes, we performed a joint analysis of discrete and continuous spatial diffusion. A joint analysis allows the delineation of host-specific rates of spread from the continuous spatial dynamics model as per Trovao, Suchard [30].

Estimates of viral interactions at the local scale

A separate local-scale analysis of H13 and H16 was performed, that included 134 viral sequences exclusively from our study site in Cordova, Alaska collected between 2009 and 2018. Phylogenetic trees of H13, H16 and PB2 were estimated using identical methods as described above. Changes in relative genetic diversity for the two HA subtypes were estimated using the uncorrelated lognormal relaxed-clock with Gaussian Markov random field smoothing of the Bayesian skyride plot [73]. These plots showed changes in the viral effective population size (Ne) over time and were visualised in Tracer v.1.6.0. and plotted using JMP Pro v.14.3 [62].

To evaluate the degree of overlap and interaction between H13 and H16 within Cordova, we estimated rates of reassortment by quantifying transitions of the PB2 internal segment between subtypes. This was modelled using a a discrete trait phylodynamic analysis, in which the traits were represented as different HA subtypes. To evaluate BF support between pairwise HA subtypes, BSSVS analysis was performed (as per discrete traits analyses described above) and results were visualized as heatmaps.