A major concern when any pathogen leads to a large increase in case counts is that it has evolved to have a higher virulence or transmissibility. In the case of EEEV, increased transmissibility would refer to a higher infection success in either mosquitoes or birds. Increased virulence would lead to increased case ascertainment, as diagnosis occurs on hospitalized individuals and most human infections are probably asymptomatic¹⁵. We would expect to see a phylogenetic signal for any intrinsic variability in transmissibility or virulence.

We searched the alignment for any nucleotide substitutions, relative to a reference sequence from 2005, which are common to some or all of the 2019 EEEV sequences. There were 17 substitutions across the genome shared by all of the 2019 sequences, a further two that more than 90% of them shared, and an additional two that 75% of them shared. These 21 substitutions, however, were found across the whole phylogeny, in at least 93% of non-2019 EEEV sequences. We found no shared substitutions unique to more than 75% the sequences sampled in 2019. This was expected given that these sequences are spread across the phylogeny (Figure 2D), as there would have to be extremely strong positive selection leading to multiple instances of convergent evolution for unique shared substitutions to be possible. While some of these shared substitutions may have mild phenotypic effects, their spread across the phylogeny implies they are not a main driver of the increase in cases in 2019.

Past population size of viruses can be useful in inferring underlying transmission dynamics. We used a skygrid model¹⁶ to estimate temporal changes in virus effective population size across the whole of the US. We found that there was a steady increase in effective population size in the 1990s, which then plateaued and decreased more recently in the late 2010s (Figure S1B). We then formally incorporated human and horse case data as covariates in the skygrid¹⁷ hypothesizing that increased transmission in the mosquito-bird cycle could lead to more reported cases in humans and horses.

We tested 9 different case-based covariates independently using data from 2003 onwards, but we did not find any significant associations between the viral population size and any of these covariates (Figure S1C). There is, therefore, no simple association between national circulation and reported cases, meaning that regional dynamics are likely more important for regional cases.

Human behavior could be one explanation for the increase in human EEEV cases in 2019, which could be revealed by analyzing the case demographics. We did not however find any meaningful differences in case demographics or timing in 2019 compared to previous years since 2003 (Figure S2)

To further explore whether human behavior may have changed in 2019, leading to a higher EEEV exposure rate, we compared the length of time individuals spent outside in relevant counties in Massachusetts, Connecticut, and New York (Figure S2E). We performed a paired t-test on the indoor activity seasonality metric for each of these counties in June, July, August, and September between 2018 and 2019 (data only available for these years) using data from ¹⁸. We found no evidence of a difference in indoor-ness between 2018 and 2019 in counties that had ever had a human case of EEEV (p=0.56) or only those which had a case in 2019 (p=0.72). Therefore, we found no evidence of time spent outdoors, the timing of cases, or case demographics being major drivers of EEEV outbreaks. Understanding the behaviors and activities linked to human exposure, however, is still an important area of research.

Bird populations likely have a complicated role in EEEV outbreaks. While the presence of competent bird species is a necessity for EEEV transmission and large populations may lead to more infections, conversely, abnormally large bird populations may dilute the virus, decreasing the likelihood of mosquito infection during blood feeding¹⁹. Moreover, when sufficient opportunities exist for mosquitoes to feed on birds, it may limit the pursuit of mosquitoes to feed on alternate sources, such as humans and horses. Finally, bird population immunity is likely a key component of the potential for EEEV perpetuation, but studies that routinely collect bird serology data are difficult and not often conducted.

While data on birds are limited, we examined the effect of the (1) absolute abundance and (2) proportion relative to all species of 8 bird species for which Cs. melanura has a high blood feeding preference^9,20 on (A) human and horse cases and (B) mosquito EEEV infection rate using a Poisson regression. We also compared these metrics for 8 bird species for which a low feeding preference was estimated²⁰ as a control (Figure S3). There were no significant relationships when considering mosquito infection rates regardless of species or state. In New York, there were very few significant relationships in high-preference birds, and several in the low-preference bird species and so this comparison is likely not valid for New York. In Massachusetts and Connecticut, we found several positive relationships between case counts and high-preference bird species case counts (Figure S3) but also with a smaller number of low-preference birds. The relationship between specific bird presence and EEEV transmission is therefore complicated and not easily solved here. There are, however, some low-preference birds with significant relationships (although fewer), meaning that the relationship between specific bird presence and EEEV transmission/cases is complicated and not easily solved here.

Further information and requests for resources should be directed to and will be fulfilled by the lead contact Nathan Grubaugh (nathan.grubaugh@yale.com).

This study did not generate new unique reagents

No human samples or direct clinical data were used in this study. Sequencing of remnant veterinarian samples by the Wadsworth Center were done following institutional protocols. All data are available through aggregated surveillance databases as described below.

In New York State (NYS), since 2012 (including 2019), 13 counties representing all regions of NYS have been surveilled, as previously described ³⁰. In total, >100,000 pooled individuals representing 53 species are tested for EEEV each year, this includes ~15,000 pooled Cs. melaura annually, which are predominantly trapped in the main EEEV/Cs. melanura focus including Oswego, Oneida, Onondaga and Madison Counties. Trapping was completed using a combination of Centers for Disease Control and Prevention (CDC) light traps, gravid traps and diurnal resting boxes. Resting boxes were primarily used in EEEV endemic areas to collect Cs. melanura. Mosquito specimens were sorted by species and pooled for testing. Pools of 10–60 mosquitoes were shipped on dry ice to the NYS Arbovirus Laboratory for EEEV testing by molecular and cellular methods. Specifically, pools containing a zinc-plated steel bead and 1ml mosquito diluent (20% heat-inactivated fetal bovine serum (FBS) in Dulbecco’s phosphate-buffered saline plus 50 μg/mL penicillin/streptomycin, 50 μg/mL gentamicin, and 2.5 μg/mL Fungizone) were homogenized using a mixer mill for 30 sec at 24Hz and centrifuged for 5 min at 6000 rcf. Quantitative reverse transcriptase polymerase chain reactions (qRT-PCR) were performed using two distinct primer and probe sets, following RNA extraction and purification, as previously described³¹. In addition, 100 μL of supernatant from Cs. melanura pools were inoculated on Vero cell culture and monitored for cytopathic effect. EEEV isolates obtained from a single round of amplification were used for sequencing.

In Massachusetts, mosquito surveillance was conducted from mid-May through to mid-October, as previously described^32,33. Trapping was performed by the Massachusetts Department of Public Health (MDPH) in collaboration with 10 local mosquito control projects (MCP) at semi-variable frequencies visiting non-fixed trap sites spread across all 14 counties. Site visitation frequency increased with high volume collections of vector species and narrowed to weekly-biweekly over time in correlation with increased site-specific target mosquito abundance. Targeted sites visitation frequency increased to weekly at minimum when EEEV activity was detected and persisted through the duration of the seasonal surveillance period. Weekly collections were performed at 10 fixed collection sites in Bristol and Plymouth counties known to be historically active EEEV sentinel sites. These site collections increased to twice a week after initial EEEV detection.

Mosquito collection methods varied depending on MCP, however nearly all successful Cs. melanura collections were performed using primarily CDC-Miniature Light traps with a CO₂ source (either dry ice or regulated tank flow ranging from 250–500cc), gravid traps baited with an infusion of lactalbumin-yeast-hay with oak leaves, or resting boxes placed primarily in locales with both deciduous and evergreen freshwater forested swamps. Light traps and gravid traps were placed in the early morning-late afternoon and retrieved 24 hours later, and resting boxes were visited once weekly. Mosquito trap canisters collected from the field were transported to the laboratory in an igloo cooler lined with dry ice, freeze-killed in an ultra-low −80°C freezer, identified by species using a dichotomous key to characterize morphological differences with a stereoscope, and pooled in sample vials of 5–50 female mosquitoes. Sample pools were grouped by species/trap site/date of collection before being submitted to the MDPH Molecular Diagnostics lab for arbovirus testing.

In Connecticut, mosquito trapping and arbovirus surveillance was conducted from the beginning of June through the end of October at 91 fixed collection sites, distributed among all 8 counties, as previously described in ³⁴. Trapping locations, where Cs. melanura were likely to be collected or where there was historical record of EEEV activity, were established in sparsely populated rural settings that included permanent fresh-water swamps (red maple/white cedar) and bogs, coastal salt marshes, horse stables, and swamp-forest border environs. Additional trap sites are located in more densely populated urban or suburban locales, including parks, greenways, golf courses, undeveloped wood lots, sewage treatment plants, dumping stations, and temporary wetlands associated with waterways.

Mosquito trapping was conducted with CO₂ (dry ice)-baited CDC miniature light traps equipped with aluminum domes or gravid mosquito traps baited with a lactalbumin-yeast-hay infusion. Traps were placed in the field in the afternoon, operated overnight, and retrieved the following morning. Trapping frequency was minimally made once every ten days at each trap site over the course of the entire season. Mosquito trapping frequency was increased at EEEV-positive sites to twice per week after the virus was isolated from that site. Adult mosquitoes were transported alive to the laboratory each morning in an ice chest lined with cool packs. Mosquitoes were immobilized with dry ice and transferred to chill tables where they were identified to species with the aid of a stereo microscope (90X) based on morphological characters. Female mosquitoes were pooled in groups of 50 or fewer by species, collection date, trap type, and collection site and stored at −80°C until processed for virus isolation. Processed mosquito pools were inoculated into Vero cell cultures and screened for cytopathic effect (CPE) as previously described³⁵. CPE positive virus cultures and the original mosquito pool were then tested for EEEV by TaqMan RT-PCR assay³⁶.

RNA was extracted on the MagMax-96 Express robot (Applied Biosystems, Foster City, CA) with the Magmax Viral isolation kit (ThermoFisher Scientific, Waltham, MA), according to manufacturer’s specifications. 50 μL of sample was added to 130 μL of lysis buffer containing 20 μL of RNA binding beads diluted 1:1 with wash buffer. RNA was eluted in 90 μL of elution buffer. Primer pairs (Table S5) were used to generate 6 overlapping fragments of approximately 2.5 kb each using one-step superscript III RT–PCR with platinum Taq (Life Technologies, Carlsbad, CA). Each reaction utilized 5 μL of RNA, 1 μL of enzyme, and a 0.4 μM final concentration of primer pairs in a total reaction volume of 25 μL. Thermocycler amplification was completed using the following conditions: 55°C for 30 min; 94°C for 2 min; 40 cycles of 94°C for 15 sec, 55°C for 30 sec, 68°C for 3.5 min; and a final extension of 68 °C for 10 min (Simpliamp by Applied Biosystems, Waltham, MA). Two uL of amplicons were visualized on a 1% agarose gel to confirm size and quality, and subsequently purified using Zymo DNA Clean and Concentrate (Zymo Research, Irvine, CA). Amplicons from individual isolates were pooled and sent to the Wadsworth Center Advanced Genomic Technologies Core for library preparation and indexing using the Nextera XT kit (Illumina, San Diego, CA) according to manufacturer’s protocols.

Sequencing was performed on the Illumina MiSeq platform (San Diego, CA). Paired-end reads were assembled to a 2014 Connecticut isolate from Cs. melanura (KX029260) deploying Geneious Prime’s reference mapping tool with high sensitivity and free end gaps using 10 iterations of fine tuning and trimming paired read overhangs. Mean coverage/base ranged from 703–2132x. Resultant consensus sequences were used for downstream analyses. We generated complete genome consensus sequences for all 80 sequenced isolates.

In total, 80 new samples of EEEV were sequenced from 2015–2019. 451 additional whole genome sequences from prior to this study were downloaded from Genbank (see Table S6). Those not from the US and without location data were removed, and the remaining combined with the new sequences to give a dataset of 523 sequences.

We aligned the sequences using Mafft version 1.3.7³⁷, and removed the non-coding regions at either end of the genome, giving a final multiple sequence alignment with a length of 11,277 bases. A maximum-likelihood phylogenetic tree was generated using IQ-TREE 2.1.4³⁸ using an HKY substitution model and temporal signal was assessed in TempEst³⁹. A single molecular clock outlier (ID: AY722102) was removed, as described in⁴⁰.

We estimated introductions into the Northeast (defined as New York, Connecticut, Massachusetts, New Hampshire, Vermont, Rhode Island and Maine) by conducting a discrete trait phylogeographic analysis (DTA) at the state level in BEAST 1.10⁴¹, with an asymmetric CTMC model. We also used a non-parametric skygrid coalescent model¹⁶ estimated using Hamiltonian Monte Carlo sampling⁴², an HKY substitution model⁴³, and a strict clock model⁴⁴. We used tip-date sampling for those sequences without exact sampling dates, with a starting date given as 0.6 of the way through the appropriate year (i.e. August) with a standard deviation of three months. We estimate Markov jump histories for the full posterior to obtain estimates of location transitions between states in the DTA, and summarize them using TaxaMarkovJumpHistoryAnalyzer⁴⁵. We performed two independent replicates of this analysis, with each chain running for 100 million iterations, removing 10% for burn-in. Convergence and mixing were assessed in Tracer 1.7⁴⁶. Further, we found a strong temporal signal for this dataset (Figure S1) with an estimated evolutionary rate of 1.86×10⁻⁴ (95% HPD: 1.77×10⁻⁴-1.95×10⁻⁴) substitutions per site per year, in line with previous estimates¹².

An introduction node was considered to be the first node which was inferred to be in one of the Northeastern states and downstream tips were counted as part of the cluster. One introduction left Massachusetts and returned to Florida. As this is unlikely given bird migration patterns, and the confidence in the location of the node was low (52% Massachusetts, 46% Florida), we instead used the child node which was found in Massachusetts, thereby excluding the Florida sequence from this cluster.

As there are no sequences from Florida (which was estimated to be the backbone of the phylogeny) after 2014, and relatively few before, the DTA will underestimate the number of introductions in total as they will not be broken up by Florida sequences. The date of the introduction node (the first node in the cluster inferred to be in the Northeast) will therefore also be too far in the past, as the lack of Florida sequences in the cluster means that the location of the node is inferred incorrectly. Therefore, we set out to split up some introductions which contained long branches and likely represented unsampled Floridian diversity. We identified all of the clusters which contained tips only from the Northeast with more than three sequences and more than 50% of the sequences sampled in 2014 or earlier (i.e. when the last Floridian sequence was sampled); and calculated their average branch length (1.28 years). We then traversed the tree, and for any branches where both the parent and child node were inferred to be in the Northeast, but were more than twice the standard deviation (1.52 years) above the average branch length (threshold = 4.33 years), we assigned the introduction node as the child of the pair instead of the parent. This therefore moved the introduction node closer to the present, and sometimes excluded Northeastern sequences on the sister branch, making separate introductions. The final number of introductions into the Northeast before this procedure was 42 prior to 2014 and 19 after, and 49 and 26 respectively, adding 14 more introductions in total.

To test for nucleotide substitutions common to 2019 sequences, we compared the consensus sequences to the reference sequence used in⁴⁷, which has Genbank accession ID KJ469556.

We have also built a Nextstrain page to further explore the data, which can be found at nextstrain.org/community/grubaughlab/eeev-genomics.

Index P is an estimate of the transmission potential by a single female mosquito using Bayesian methods and mosquito-virus specific priors (Table S3), including transmission probability as a function of temperature and relative humidity. Unlike the basic reproduction number, R₀, index P does not need to be greater than one to cause sustained transmission, as index P is multiplied by the ratio of human to mosquitoes to derive R₀. We estimated the variable index P using the Mosquito Virus Suitability Estimator (MVSE), an R package to download the functions for modeling the suitability of a given environment for mosquito-borne virus transmission²³. Temperature and humidity data for each state were calculated by taking the average temperature for each day in the center of each county in the state. These county level data were then averaged to obtain a single temperature estimate for the entire state. All weather data were provided by visualcrossing.com. Index P was calculated from April to October, in line with the transmission season for EEEV. All negative temperatures were set to zero in the model to avoid erroneous results, as at zero or below, mosquitoes are in a hibernation state and so there is no mosquito activity that varies by temperature.

We used a 1 month lag time from index P to cases to account for delays caused by requiring transmission into humans and horses, followed by incubation times and time to diagnosis. Therefore the relevant mosquito activity will be before human and horse cases are reported.

While index P is seen as an important covariate in the negative binomial regression model, it is difficult to tease out the seasonality of the mosquito season with cases. Due to the sporadic nature of EEEV in the Northeast, index P is a necessary but not sufficient parameter to consider. To test whether temperature could be used as a proxy, an additional model was tested by replacing index P with temperature which resulted in a slightly higher Akaike information criterion (AIC) score (231.4 vs 232.3) and wider confidence interval. Given the widespread use of modeled mosquito viability parameters, we opted to keep index P. However, future research and surveillance programs may wish to utilize temperature, an easily obtained parameter with little uncertainty in its estimate.

We restricted the case data (described above) to Connecticut and Massachusetts for the model where we had sufficient mosquito data to calculate the Vector Index. Dispersion was detected with a dispersion parameter alpha = 0.18 (p-value = 0.019) using the dispersion test from the AER package in R ⁴⁸. This led us to fit a negative binomial regression model (estimated using ML) to predict human_equine with vector_index (formula: human_equine ~ vector_index + month_detected_july + indexP_lag1 + st_grp + year_index + month_f).

Given the vast differences in cases between Massachusetts and Connecticut we attempted to run stratified models of each state separately. However the model failed to converge for Connecticut, likely due to low availability of cases. Thus a model combining both was used. Months were included as categorical variables with the months May through July as reference groups. These months were combined due to May and June having no cases, making parameter estimates fail to converge when separate. The parameter ‘first month detected July’ was a binary variable determined by identifying the month when the first non-zero value for the vector index occurred for each year. All continuous data were normalized so interpretation of estimates are for a 1 standard deviation increase in the term.

Effect modification of vector index by month of detection was explored in addition to the original model without effect modification. AIC scores of the model with effect modification and without were nearly identical (231.4 vs 231.2 respectively) and an ANOVA test was conducted between the two models and found to have no difference (p-value = 0.14). While no difference was identified, this may be due in part to the small sample size of years where the month detected was after July.

In addition to effect modification, given some of the complexity of index P, we explored the terms temperature 1-month lag and mosquito abundance 1-month lag in replacement of index P. Utilizing the two point rule of thumb for AIC, the abundance model performed worse (231.4 vs 233.2) and its estimate was not significant (CI 95%: 0.57–1.62). The temperature performed slightly worse but similar to the index P model (AIC 231.4 vs 232.3) but had a larger standard error (CI 95%: 1.15–9.31). Given the desire to explore the utility of index P and its higher performance we opted to focus on this model. However, for simplicity sake future work may utilize temperature for its ease of use and similar performance (Figure S6 and Table S4).

Utilizing the DHARMa version 0.4.6 package in R (https://CRAN.R-project.org/package=DHARMa), we simulated the residuals 1,000 times to test for heteroskedasticity, zero inflation, and autocorrelation. Upon visual inspection no heteroskedasticity of the residuals was detected. The simulated ratio of expected to actual zeros was 1.01 (p-value = 0.776) The Durbin-Watson test for autocorrelation was conducted on a subset of the Massachusetts data to avoid duplicate time indexes. In addition to subsetting the data, index P with and without a 1-month lag was tested per the suggestion of the Durbin-Watson test to avoid lagging covariates. Autocorrelation was borderline but insignificant without the index P lag (DW = 1.59, p-value = 0.051).

It must be noted that our model was restricted to a monthly time scale to match monthly case data. Given the short duration of the EEEV transmission season, certain dynamics may be missed on the weekly or even daily timescale. There was not enough case data within each month to explore impacts of detection on a month by month basis, making the month detection piece of the analysis less specific.

The maximum-likelihood estimation (MLE) of the infection rate is a pooled infection rate of positive EEEV pools, and was estimated using the CDC R software package PooledInfRate (https://github.com/CDCgov/PooledInfRate). This estimation procedure takes into account that there may be different numbers of positive mosquitoes in each positive pool, and so estimates the likely number of positive mosquitoes in each pool based on the overall number of positive pools. The relative abundance was calculated as the average number of mosquitoes captured per trap per night. Vector index is these two metrics multiplied together ^21,22.

To identify any possible associations between effective population size and case counts, we applied the latter as covariates to the estimation of population size using the skygrid model¹⁷. We began by comparing 12 possible covariates (all human and animal cases, just human cases, just horse cases, and all cases in Massachusetts, Connecticut, and New York with 0, 1 and 2 year lags) to effective population estimated sizes using a skygrid model with no covariates¹⁶. On the basis of this preliminary analysis, we ran formal analyses with all cases, human cases, and horse cases and the relevant lags. We also only used sequence and case data from 2003 onwards (n=423), when the surveillance program began, in order to eliminate some noise from the data. BEAST runs were set up as above with tip-date sampling, HKY substitution models and a non-parametric skygrid coalescent model. Grid points were externally set to correspond to the start of each year.

Case counts for years between the inferred time of origin of the tree and 2003 were estimated independently using a normal prior, whose mean is the recorded case count. The standard deviation was calculated such that all national cases and horse case covariates were allowed to have ±5 cases in the 95% confidence interval; and human covariates were allowed to have ±1 cases in the same interval, as the recording of the latter is more precise. This was also undertaken for 2021 data for both analyses with the 2 year lag covariates, as finalized data has not been released by ArboNET at time of writing.

Starting values for unobserved years were taken from news reports and other publicly available sources. Most of the researched outbreaks were reported in Corrin et al, 2021⁴⁹, and data were supplemented by going to each of the references given in that paper. Data for outbreaks in Michigan were supplemented by information in⁵⁰, and Massachusetts data were supplemented using Stobierski et al 2022, Grady et al 1978 and Feemster et al 1938^51–53. Data for all states were supplemented using CDC Morbidity and Mortality Weekly reports in relevant years. Horse cases used are both confirmed and suspected, due to a lack of equine testing, especially in the earlier epidemics.

How much time people spent indoors or outdoors was taken from Susswein et al 2022¹⁸, who used anonymized GPS data from mobile phones. This data was only available for 2018 and 2019. We took the average indoor/outdoorness for each county which had a case in 2019, or had ever had a case and performed a dependent t-test in Python for paired samples to compare the metric between years. Significance was taken to be a p-value of less than 0.05.

Bird banding data to calculate abundance and proportion were obtained from the North American Bird Banding Program dataset⁵⁴. All ages of birds were used. Bird species were selected using data from Kramer et al 2015²⁰, who calculated the relative feeding preference of Cs. melanura in Connecticut using a blood meal analysis and normalizing it by abundance of the bird species in question. High-preference birds in this analysis were the 8 species identified in this paper, and low-preference birds were randomly selected from their list of birds for which Cs. melanura appeared to have no preference. We conducted Poisson regressions in Python for each combination for the abundance or proportion of a bird species in a given year against the human and horse cases or mosquito infection rate (both defined above). A 5% significance level was used.