Higher airborne pollen concentrations correlated with increased SARS-CoV-2 infection rates, as evidenced from 31 countries across the globe

To examine the potential effects of pollen−virus coexposure, a large cross-sectional and longitudinal study was set up, based on 248 airborne pollen monitoring sites, from 31 countries in all inhabited continents across the globe (Fig. 1). The initiative started when, during 10 to 14 March 2020, a warm weather episode brought about higher airborne pollen concentrations across the Northern Hemisphere (denoted as larger circles in Fig. 2), which was evident in mainland Europe mainly on 12 March. This coincided with high SARS-CoV-2 infection rates (denoted with darker color circles in Fig. 2) characteristic for the early exponential infection phase.

The median day of onset of COVID-19 exponential phase (for definition, see Materials and Methods) was 13 March 2020 (Fig. 3), which corresponds, on average, to a cumulative pollen concentration of 1,201 grains/m³ up to 4 d before (daily average: 240 pollen/m³). On a cross-sectional design for all 80 regions under study, it was found that the onset date of the exponential phase per region positively and significantly correlated with the cumulative amount of pollen up to 4 d before (P < 0.001, r = 0.25). Those regions mainly with lower pollen concentrations and high human contact because of the carnival events in late February, as well as with humid, colder continental climates (on 20–21 March), were categorized as outliers in Fig. 3.

On a cross-sectional approach, we investigated for differences during the exponential infection phase between the infection rates for all sites of the study, grouped into four categories: low vs. high population density and low vs. high pollen concentrations (Fig. 4). To isolate the genuine pollen effect, we elaborated only intervals for all countries without any lockdown. The mean and median of the infection rates were found to differ between low- and high-pollen sites by ∼0.1 (low population density) and 0.3 (high population density); that is, a more pronounced pollen effect was observed for the high-population density sites. The extreme values revealed an even stronger signal: Regardless of the population density, near-zero infection rates were observed only in regions with low pollen levels. Conversely, the absolute maximum infection rate was reached in the high-population vs. high-pollen case (P < 0.01).

On a longitudinal setup and focusing on the geographically large or climatically diverse countries, which contain the vast majority of regions under study, we investigated for spatial anomalies of the infection rates, which were correlated country-wise with spatial anomalies of pollen concentrations. To eliminate low-level statistical noise, very low pollen concentrations (<50 pollen/m³) and regions sparsely populated (<100 inhabitants/m²) were not included in the analysis. Only the before or no lockdown time intervals were included in the analyses. It was found that the anomaly correlation coefficient was positive for all countries and significantly positive in six out of eight (Fig. 5). The regression slopes show that the infection rate’s sensitivity to pollen, on average, is 0.04 per 100 pollen/m³ (range: 0.03 to 0.25) for the countries with significant correlations. Depending on the region (note the different x axes values in Fig. 5), this corresponds to 6 to 15% of the exceedance of the rate over zero. The R² values shown in Fig. 5 (including also nonsignificant relationships) illustrate that 10% of variability in the infection rate is explained by its sensitivity to pollen fluctuations.

The pollen effect was proven strong, sometimes regardless of the population density. Switzerland, as one of the countries with the highest pollen concentrations across the world during the exponential phase of the pandemic, serves as a case study to illustrate the relative importance of the pollen effect, by comparing three cities located close to each other and with comparable climates and population densities, but with different pollen exposure (SI Appendix, Fig. S1).

To test the influence of other cofactors, environmental but also human interaction related, we performed a per-country longitudinal analysis (Fig. 6). Complementing the analysis and results in Fig. 5, ridge regressions were conducted for all 31 countries and 130 regions under investigation. For those countries in which no lockdown had been implemented, or the lockdown had started almost in parallel with the onset of the exponential infection phase (<5 d difference), we could not possibly consider the lockdown variable in the analysis. Despite the significant and negative effect of lockdown in the majority of countries for which we included it as dummy variable (11 out of 14 countries, in the mixed design with no lockdown−lockdown regime), environmental cofactors were still significantly correlated with increases in daily infection rates in 12/14 of cases (P < 0.05) (Fig. 6). Regardless of the exposure conditions, either with or without a lockdown regime (Fig. 6), of the three environmental factors examined here, pollen was significant in 10/21 countries, air temperature in 14/23, and relative humidity in 10/23. All significant correlations of infection rates with environmental factors (pollen, temperature, humidity) were, by rule, positive, and those with lockdown and weekend, by rule, negative. The average lag effect of airborne pollen on daily infection rates was 4 d (using backward stepwise removal of independent variables), which is consistent with the cross-sectional analyses described above. Under an early lockdown design (lockdown before or <5 d after the onset of the exponential infection phase), pollen concentrations were still significantly and positively correlated with daily infection rates in 6/14 countries, and, in 5/14 pollen, was the primary factor. Under a mixed lockdown design (full exposure ≥ 5 d, then lockdown), lockdown was significantly and negatively correlated with daily infection rates in 11/14 of cases, in 9/14 as the primary factor. Strikingly, even under an early lockdown, the synergy of environmental factors could explain, on average, 44% of the infection rate variability in 9 out of 14 countries (Fig. 6). It is worth mentioning that, of the remaining countries with no significant relationships with airborne pollen abundances (or with other environmental factors as well), 7 countries exhibited very low pollen concentrations during the examined period, explicitly less than 5% of the averaged total pollen load of all countries. These countries, by rule in the Southern Hemisphere or in colder and humid continental climates (Fig. 6), most frequently did not correlate with any environmental parameter at all.

We further investigated the lockdown effect, longitudinally, among countries, and, cross-sectionally, in association with airborne pollen concentrations. Almost all countries had a lockdown of some type, mostly a partial one. Only nine countries adopted a strict lockdown from the beginning. Lockdown significantly decreased the infection rates as compared to no lockdown (P < 0.001) (Fig. 7A). A significant positive correlation between daily infection rates and daily pollen concentrations was observed under both lockdown and no-lockdown regimes (R² = 0.02; P < 0.001). However, the magnitude of the lockdown effect was such that, under comparable amounts of pollen, daily infection rates were reduced to approximately half during lockdown compared to full exposure: The association of infection rates with pollen concentrations was still positive and significant (note the different y axes in Fig. 7B).

Our large-scale retrospective data analysis based on 80 individual time series from 130 regions in 31 countries in all inhabited continents across the globe (8,019 data points) enabled us to reveal a robust and significant positive correlation between SARS-CoV-2 infection rates and airborne pollen concentrations, which was halved under lockdown. We managed to obtain pollen data from the majority of all pollen monitoring stations worldwide that were operative despite considerable spread of COVID-19 infection rates already by that time, resulting in the most comprehensive aerobiological dataset possible to conduct such a study.

In the current pandemic situation, SARS-CoV-2 infection spread is primarily and foremost dependent on person-to-person interaction, which is mirrored by the observed, significant effect of lockdown. The rapid kinetic of infection in the absence of herd immunity is prone to mask any potential effect of environmental cofactors that may exacerbate contact-dependent mechanisms. The example of Switzerland shown in SI Appendix, Fig. S1 highlights the major assumption made in the longitudinal study: The cities should have similar weather conditions and be similar from a sociodemographic standpoint. On the opposite side of this case study, in the United States, these very requirements were not upheld for the five sites tested (distance between them exceeded 2,000 km, some were in maritime and some in strongly continental climate, different states with different strategies regarding lockdown, mean income, and other factors). This lack of homogenous conditions may easily explain the strong scatter in the United States anomaly correlation chart.

The COVID-19 pandemic hit Europe and North America during springtime, when rising air temperatures are associated with increased social and outdoor activities, which, in turn, means increased environmental exposure—to bioaerosols, pollutants, or infected humans. Given the complexity of intertwined environmental, social, and political cofactors, it is anticipated that no clear signal may be observed unless it is tremendously robust. Moreover, environmental exposures, whether climatic factors, air pollutants, or pollen, often exert their effects at the same time, and many of these factors are collinear, which complicates the statistical analysis. Nonetheless, from all the countries that showed a significant correlation of the infection rate with pollen, this correlation was always positive, which suggests that the mechanism reported for pollen exposure on antiviral immunity to rhinovirus (15) could also be influencing innate immunity toward SARS-CoV-2. To verify this statement, we conducted multiple tests to check for bias, including bootstrapping and permutation tests. If, under this statistical noise, we can still see such a signal, we may safely consider the results robust enough, with our concerns being actually about whether we potentially underestimate the magnitude of this effect.

Infections with endemic coronaviruses (strains OC43, HKU1, 229E, and NL63), as well as other frequent respiratory viruses, such as respiratory syncytial virus and influenza A, peak in winter or early spring; a general negative trend of air temperature on these infections has been evidenced (18). Therefore, it is likely that parameters like air temperature act, in the long term, as confounding factors for the short-term positive effect of pollen on infection rates. Also, while the anomaly correlation between airborne pollen and infection rates was significantly positive, the effect size was small, indicating that pollen is only one of a number of environmental factors influencing SARS-CoV-2 infection. However, if one considers that the study was conducted marginally in the start of the pollen season in most regions, this statement may be under dispute. Extending this study deeper into the 2020 pollen season would not offer clearer information, as we would have an even wider variety of data, with ceased lockdown measures and opening borders and tourist activities taking place almost up to the end of 2020.

When checking for additional environmental cofactors, including human interaction indicators, an average of 4 d of lag effect was found in increases in pollen concentrations associated with increases in infection rates. This was connected with the temperature and/or humidity lag of the same or the previous day. A 4-d lag effect of pollen is in agreement with the proposed physiological mechanism of action, an interference of pollen with the innate antiviral immune system. A study based on infection data from Singapore and the Chinese provinces of Tianjin and Hubei estimated an incubation time for COVID-19 of between 4 and 5 d (19, 20), which is much shorter than original estimates (2) but close to our results. It is also in agreement with a hypothesis of environmental exposure factors acting by reducing the incubation period. Unfortunately, this assumption could not be supported by similar pollen data from China, as aerobiological monitoring there is not yet well established.

Respiratory and olfactory epithelium has been shown to express the viral entry receptors for SARS-CoV-2, ACE-2, and TMPRSS2 (21, 22), which makes the nasal cavity a potential early virus reservoir and stresses its importance in innate antiviral defense (23, 24). Since the upper airways are also the entry site for pollen grains, the previously shown immunosuppressive effect of pollen on respiratory epithelia (15) could influence the susceptibility to SARS-CoV-2 infection as well. Pollen grains act on the very site of virus entry, the nasal epithelium, by inhibiting antiviral λ-IFN responses (15). Early treatment with IFN-λ has recently been discussed as a first-line therapeutic option to prevent COVID-19−associated cytokine storm (25–27). This highlights the conclusiveness of our primary hypothesis, which is supported by the epidemiological results reported here.

The observed correlation of airborne pollen with infections did not depend on the allergenic nature of the pollen types present in the air during the study period. Although we analyzed the entire biodiversity spectrum of pollen taxa (SI Appendix, Fig. S2), when stratifying pollen by “allergenic” and “total” pollen, both showed similar correlations with COVID-19 cases (SI Appendix, Fig. S3). This agrees with our previous findings on immune modulatory effects of pollen, for example, inhibition of NF-κB (28), MyD88 (29), and antiviral IFNs (15), which do not depend on pollen-derived allergens and are effective in sensitized as well as in nonsensitized individuals (30, 31). Thus, although we do not (and could not possibly, to our knowledge) have any information on the allergy status of the COVID-19 cases on which our analysis was based, we assume that the pollen effect is relevant for the entire population. It might, however, be more pronounced in allergics, asthmatics, or chronic rhinosinusitis patients, due to an intrinsically weaker antiviral immune response (32–35).

Our results were not yet able to reveal the genuine magnitude of the pollen effect, as the entire springtime pollen peak of the Northern Hemisphere was not fully included, either in terms of abundance or in its whole seasonality. The data acquisition was stopped in early April due to lockdown restraints. An unavoidable major limitation of the longitudinal data analysis is, therefore, the shortness of some of the time series. During that time, only a few studied sites were subjected to the substantially varying pollen load similar to that shown for Switzerland; practically, we had to deal with two subsets of data, one with a mixed design of lockdown−exposure effects and another design of early enough lockdown to almost annihilate the pollen effect in some occasions.

The sites located in the Southern Hemisphere were mostly out of the pollen season during the study period, and most had not reached the exponential infection phase yet. Whether this is in support of our hypothesis cannot be conclusively answered at this stage, but it should become evident by examining the Southern Hemisphere's pollen season in October 2020 and thereafter.

Another limitation is the spatial resolution of the COVID-19 cases, as, for some sites, local COVID-19 data (SI Appendix, Table S1) were not yet available, data had gaps or were registered in a biased way, or the number of cases was too low. In such occasions, we had to access the COVID-19 cases per country, which might not be the best approximation and is reliant on testing strategies within each country. At this early stage of the pandemic, infection rates were based on documentation of numbers of cases presenting to public hospital services and may not have included mild or asymptomatic cases in the community.

To minimize bias of COVID-19 data due to registry lags and errors, we regularly updated our database (last update: 10 May 2020). In most countries, COVID-19 databases were updated within the time frame of a month and then did not change any more. Therefore, we consider our COVID-19 database curated up to 8 April as “reliable.” We were, however, unable, at this stage, to correct for every possible confounder, such as underreporting or changes in testing strategy. In our cross-sectional analysis, we controlled for population density, but we are aware that, still, a comparison across all countries is problematic due to the above limitations, and we attempted to overcome this by doing longitudinal analyses per country, and by two different approaches.

We specifically searched the data, per site and per country, for weekly cycles that might arise from gaps in weekend recordings. While recurrent accumulations of COVID-19 cases on some weekdays, mainly on Wednesdays and Thursdays, can be most likely attributed to weather events, we still included “weekend” as a dummy variable in the ridge regression, where it turned out to be less significant than the effects of lockdown and environmental factors, with the exception of three countries.

In the light of the present pandemic situation, our findings should be communicated with caution so as to avoid misunderstandings and panic. It has to be made very clear that 1) the demonstrated correlations suggest that pollen is a modulating factor to the overall progression of the SARS-CoV-2 infection, with the potential to add an extra 10 to 30% to the infection rate (Fig. 5), 2) there is no evidence for airborne pollen grains themselves being carriers of virus particles (36), and 3) without contact, there is no risk of infection.

Of note is that the effect of pollen on reported infection rates was shown to be less pronounced under lockdown regimes. It is also possible that high temperatures in summer would counteract infections to some extent, provided, of course, that social distancing will still be kept. Therefore, the infection-promoting effect of pollen could become evident only during spring, when air temperatures are not high enough yet to limit viral spread, but high concentrations of tree pollen occur. To avoid future waves of high virus transmission under “favorable” combinations of air temperature, humidity, and pollen, we recommend taking stricter protection measures, for example, wearing particle filtering masks during springtime higher pollen concentrations. The installation of reliable, real-time bioaerosol measurement networks and the use of pollen information and forecasting systems should be encouraged.

Looking to the future, it is yet unknown whether other air particles, like fungal spores, or complex interactions with pollen, other meteorological variables, and air pollutants may also play a role. Even though there is published evidence on the effects of various environmental parameters, like nitrogen dioxide (NO₂), particulate matter (PM_2.5), and ultraviolet radiation (37–41), these usually refer to preliminary results and investigation of only a single factor. If one takes into account the huge effect of ongoing climate change and urbanization on the long-term trends in airborne pollen levels (42, 43), as well as emerging viral infections, it is of utmost importance to forecast the associated risk for human health in future pandemics and take appropriate measures to reduce it as much as possible. Coexposure is certainly not the exception but the rule under natural conditions, and, hence, we strongly suggest that modeling and forecasting of ongoing and future pandemics ought to consider the whole “soup” of exposome.

Following the strictest publishing recommendations during the COVID-19 pandemic, we followed the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) protocol, as follows.

We thank Mr. Luis-Leopold Moelter and Mr. Mehmet Gökkaya for assistance in overall data curation. The study was partly implemented in the frame of the European Cooperation in Science and Technology (EU-COST) program, "New approaches in detection of pathogens and aeroallergens (ADOPT)," Grant CA18226 (EU Framework Program Horizon 2020). D.B. and C.T.-H. were supported by the Helmholtz Climate Initiative (HI-CAM), Mitigation and Adaptation. A.Ch. and D.V. were supported by the Municipality of Thessaloniki, Greece (Directorate for the Management of the Urban Environment, Department of Environment). This research has been partly supported by the European Social Fund (Project 09.3.3-LMT-K-712-01-0066) under grant agreement with the Research Council of Lithuania (LMTLT). The study was also partly conducted within the frame of the project of the European Community European Regional Development Fund (EC ERDF) and PostDoc Latvia (Grant 1.1.1.2/VIAA/2/18/283). M.S. acknowledges the Academy of Finland (Project PS4A, Grant 318194). We thank the Department of Health and Rehabilitation of Vinnytsia Regional Council, Ukraine, for providing the numbers of COVID-19 cases. A.H.A. acknowledges Angel Chaves and the Government of Navarra: Institute of Public and Labor Health of Navarra, within LIFE-IP NAdatpa-CC (LIFE16 IPC/ES/000001). A.P. was supported by a predoctoral grant financed by the Ministry of Education, Culture and Sports of Spain, in the Program for the Promotion of Talent and its Employability (Grant FPU15/01668). B.S. acknowledges the Ministry of Education, Science, and Technological Development of the Republic of Serbia (Grant 451-03-68/2020-14/200358). C.T.H. acknowledges the Christine Kühne–Center for Allergy Research and Education (CK-CARE), and The Initiative and Networking Fund of the Helmholtz Association (Immunology & Inflammation). We thank Jan Bumberger, Marcus Karsten, Paul Remmler, Jan C. Simon, and Regina Treudler for data provision and curation from Leipzig, Germany. We thank Claudia Langford Brown and Dana Flanders for statistical advice and scientific discussions. We thank Penelope Jones, Edith Bucher, Reyhan Gumusburun, Haydar Soydaner Karakus, Su Ozgur, and Asli Tetik Vardarli for data curation.