COVID-19 Susceptibility By Race, Blood Type

Geographic distribution of cases
The prevalence of infection, estimated as the number of reported test positives in a state
relative to the number of study participants in the state, varied across geographic regions. The
highest proportions of positive COVID-19 tests were reported in New York (4.4%) and New
Jersey (3.3%), and the lowest proportions were reported in Maine (0.4%) and West Virginia
(0.4%). As the majority of case data was collected between late April and early June, hotspots
that developed earlier in the pandemic are better represented compared to those that arose
later in the course of the pandemic (Figure 3a; Supplementary Figure 1). Nonetheless, the selfreported
prevalence of positive COVID-19 tests at the US state level was reasonably well
correlated with the number of positive tests reported per capita16 as of July (Figure 3b; Pearson
r = 0.85). However, the prevalence of self-reported COVID-19 test positive status was higher in
the 23andMe database than per-capita estimates, likely reflecting differences in the composition
of the 23andMe database and the general population, and potential selection bias arising from
individuals with a positive test potentially being more likely to choose to participate in the study.



GWAS
Across the 5 phenotypes, the trans-ethnic meta-analysis identified 11 genome-wide significant
associations from 7 distinct regions of the genome (Figure 4; Supplementary Figures 2 - 4;
Supplementary Table 5).



ABO: In our phenotype contrasting COVID-19 test positive individuals to test negative
individuals, we identified an association at chr9q34.2, with index SNP rs9411378 (p-value =
5.3e-20, C allele OR = 0.857; Figure 5). This index SNP is in LD with a functional variant in the
ABO gene, specifically rs8176719 (r2 = 0.57 in the European population), which is a well-known
single-nucleotide deletion that usually confers a type O blood group when present in the
homozygous form. While multiple rare variants elsewhere within the ABO gene can contribute to
blood group determination, individuals heterozygous for the deletion are most likely to have
blood groups A or B, whereas individuals without any copies of the deletion are most likely to
have blood groups A, B, or AB.

Discussion
The COVID-19 pandemic represents a unique emergency in recent human history, and has
dramatically accelerated the pace of scientific investigation into the effects of the virus on
human health. In this paper, we have utilized a direct-to-consumer research platform to collect
data regarding experiences of COVID-19 at large scale and over a compressed timeline.
Both Latino and African American groups reported a higher rate of COVID-19 infection and, as
reported elsewhere2,18,19, our data show an elevated risk of hospitalization in these populations.
For Latino respondents, the higher rate of hospitalization was broadly consistent with the higher
rate of infection. However, for African American respondents, the risk of hospitalization was
disproportionately high, and remained so after adjusting for socio-demographic characteristics,
age, sex, obesity, type 2 diabetes, hypertension, and fatty liver disease.

Our data also strengthens the evidence for a role for ABO in COVID-19 host genetics. ABO
blood group has been reported as a risk factor for both COVID-19 susceptibility20 and severity15,
and is notable given the reported links between COVID-19 and blood clotting complications21,22.
Our data supports a role in susceptibility to infection, suggesting that blood group O is protective
in contrast to non-O blood groups. Whereas previous reports suggested protection was limited
to the rhesus positive group23, our data does not support that conclusion.

The mechanism by which ABO is associated with COVID-19 is unclear, but ABO blood groups
can play a direct role in pathogen infection by serving as receptors and/or coreceptors24. SARSCoV-
2 is an enveloped virus that carries ABO antigens on the viral spike (S) glycoprotein and
host envelope glycolipids. Recent work has shown the SARS-CoV-2 S protein interacts with
multiple host C-type lectin receptors in a glycosylation-dependent manner25,26, similar to
previous work on the SARS-CoV virus from the earlier SARS outbreak27,28. Differential
glycosylation of the spike protein or the envelope glycolipids from expression of different ABO
glycosyltransferases may then impact the binding and internalization of SARS-CoV-2 viral
particles. Others have speculated that the lower susceptibility of blood group O could be linked
to anti-A blood antibodies inhibiting the adhesion of coronavirus to ACE2-expressing cells,
thereby providing protection29.

The ABO locus is also highly pleiotropic30 and exhibits complex population structure31.
Interestingly, while old literature regarding the association between ABO and influenza is
inconsistent32, our own data suggests that blood type O is actually a risk factor for seasonal flu.
This is notable because COVID-19 testing in the United States was largely restricted to
individuals with flu-like symptoms at the time we were collecting the majority of our data. As
such, it is possible that the population of individuals receiving COVID-19 tests was enriched for
influenza cases, and the apparent protective nature of ABO for COVID-19 could arise from a
subtle form of collider bias. However, our data also shows that blood type O is protective in the
severe-outcome population, which wouldn’t be expected under the collider bias hypothesis.

Likewise, our data strengthens the evidence of association at the chr3p21.31 gene cluster, first
identified by Ellinghaus et al.15. The locus contains multiple genes (SLC6A20, LZFTL1, CCR9,
CXCR6, XCR1, FYCO1) that could be functionally implicated in COVID-19 pathology. In
particular, SLC6A20 can be linked to the association via eQTLs in relevant tissue and has been
noted15 as potentially forming a complex with angiotensin converting enzyme 2 (ACE2), the cell
surface receptor for SARS-CoV-2 viral entry33,34. It is possible that increased SLC6A20
expression leads to increased ACE2 protein levels and greater viral uptake. LZFTL1 has been
implicated in ciliogenesis and intracellular trafficking of ciliary proteins, which may impact airway
epithelial cell function. As noted elsewhere15, CXCR6 promotes NKT cell and tissue resident
memory CD8+ T cells residence in the lung35 and plays a role in the trafficking of T lymphocytes
to the bronchial epithelia during respiratory infection and inflammatory lung disease. CCR9
predominantly regulates T cell homing to the gut, which may indirectly impact the response in
the lung, however it has also been shown to regulate eosinophil recruitment to the lung36.
Recent studies have identified elevated chemokines and eosinophilia as a hallmark of severe
disease37–39, but additional work will be required to define any functional contribution of these
genes to the genetic association with COVID-19.

In phenotypes contrasting individuals with severe COVID-19 symptoms to controls without a
COVID-19 diagnosis, the risk variants at the chr3p21.31 locus achieve odds ratios of
approximately 2.0 in our data, which is relatively large in the context of GWAS studies. Given
the risk alleles are also relatively common (~3 - 8% frequency, depending on population), it is
likely that this locus makes a meaningful contribution to determining why some individuals
experience severe COVID-19 outcomes. However, while the population sample sizes in our
study differ considerably, we found little evidence to suggest that allele frequency differences at
this locus could account for the higher rate of severe outcomes from COVID-19 for non-
European ancestry groups. In fact, the primary risk allele at the chr3p21.31 locus is most
common in European populations, and less common in Latino, and African American
populations.

This study is a testament to the power of the 23andMe research platform, which in less than
four months enabled over one million research participants to contribute to the study of a novel
disease. However, there are notable caveats to relying on self-reported data for a disease with
lethal outcomes. Namely, the cases identified in this study were healthy enough to respond to
the survey therefore are likely biased towards a healthier case population than otherwise exists.
In addition, 23andMe research participants are a self-selected group and may not reflect the
general population. Furthermore, the scarcity of testing has likely further obscured the true
picture of SARS-CoV-2 infections in the United States, leading to misclassification of true cases
as controls in this study if they did not receive a positive test result. The effect of these types of
error would bias the reported effect estimates towards the null, meaning that the true impact of
risk factors reported here may be expected to be larger if the sample were randomly drawn from
the broader population and had perfect case and control classification.