Critical evaluation of the rs2076205 association with COVID-19 susceptibility in UK Biobank

In this post I critically evaluate the evidence for an association between rs2076205 and COVID-19 susceptibility, as previously reported here, here and here. There is a need to take stock because the sample sizes are as yet modest, and small differences in the details of the analysis of the UK Biobank data appear to produce different interpretations of the strength of evidence.

Statistical considerations that question the association include the observations that:

• The significance of the association goes down if individuals with close relatives in UK Biobank are included in the analysis.
• The signal appears specific to UK Biobank participants of white European ancestry.
• Power to detect a true association at a particular variant, if one were to exist, would be low because the sample size is modest.

Statistical considerations that support the association include the observations that:

• The direction and magnitude of the effect are compatible whether individuals with close relatives in UK Biobank are included in the analysis or not.
• The direction and magnitude of effect are compatible when comparing COVID-19 +ve individuals to either (i) COVID-19 -ve individuals, as in the discovery analysis, and (ii) untested individuals, as in the quasi-replication analysis.
• In individuals of white European ancestry, the variant is not strongly stratified geographically nor does it appear confounded with any of a large number of variables recorded by UK Biobank.

Robustness to the inclusion of close relatives

Comparison of results to other groups in the COVID-19 Host Genetics Initiative revealed that the signal in XPNPEP2 varies depending on the details of the analysis. Tomoko Nakanishi and Brent Richards have been helpful picking this apart. The differences are driven by whether one excludes participants with other close relatives (cousins or closer) in the analysis (as I do) or not (as many others do).

The table compares the effect of the minor (rarer) allele on the risk of a PCR positive for COVID-19, the standard error (uncertainty), and the -log₁₀ p-value (bigger is more statistically interesting) in a logistic regression analysis controlling for sex, age, age*age, sex*age and 40 genetic principal components:

The reasons for including individuals with close relatives in UKB are:

• Increased sample size should reduce uncertainty and improve the ability to discover signals.
• There are no pairs of close relatives who have both received a PCR test result among English UKB participants, so far.

The reasons for excluding individuals with close relatives in UKB are:

• Pairs of close relatives may appear in the PCR test positive vs negative comparison in future and do currently feature in other comparisons (e.g. PCR test positive vs no test).
• Determining inclusion/exclusion criteria depending on outcome variables (e.g. do they currently have a PCR test result) potentially introduces undesired ascertainment effects.
• Close relatives are more likely to share exposure to unmeasured environmental risk factors, notably exposure to SARS-CoV-2 which is highly variable and prerequisite for disease. If this non-independence is not fully controlled (difficult even with sophisticated tools) the inclusion of outcomes among a group of genetically similar individuals can unduly influence results.

On balance one would think excluding close relatives is the cautious approach, but with modest sample sizes the argument that more is better has merit. What is unusual is that including more individuals reduces signal at this gene.

Signal in other ancestries

The signal at SNP rs2076205 in XPNPEP2 is not apparent in UKB participants of non-white European ancestry. This means that it does not appear to be able to account for the strong increased risk of COVID-19 among individuals of non-white European ancestry in England.

Geographical stratification

In considering whether the signal could be an artefact, it is important to look at geographical stratification of the allele because risk factors including exposure to SARS-CoV-2 vary geographically. The figure shows the frequency of the minor allele across England, which was the allele estimated to be protective, among those of white European ancestry only. The map suggests that the allele has broadly similar frequency across England, although there are some local pockets of higher allele frequency scattered around. The map does not show a geographical trend (for example increasing allele frequency moving from North to South). (Note the density of points reflects UKB's sampling distribution, which was focused on particular urban centres.)

Robustness to measured confounders

Supposing rs2076205 had no effect on susceptibility to COVID-19, then what other explanations are there for the strong signal of association? To address this question Nicolas Arning and I have investigated the large number of variables recorded in UKB. In particular, Nick has helped to sift through the variables by applying machine learning methods, mainly XGBoost.

In the table, we have investigated the effect of including potential confounders on the association between COVID-19 PCR positive vs negative, both in terms of the statistical significance (measured by the -log₁₀ p-value) and the direction and magnitude of the effect itself. The focus is on individuals of white European ancestry, since that was where the signal was found.

The short story is we have failed to make the signal go away by including any of a large number of covariates. Moreover, the direction and magnitude of the effect is remarkably consistent, indicating that rs2076205 is not very correlated with any of the >18,000 covariates available. We have included potential confounders directly in the logistic regression, including those we might expect to be important beforehand (geographical location, genotyping array), known risk factors (body mass index), variables that came out in early analyses (Townsend deprivation index) and variables suggested by the XGBoost analysis. We have also directly included composite predictors produced by XGBoost.

Analysis of individuals with close relatives in UK Biobank

We investigated why inclusion or exclusion of individuals with close relatives in UK Biobank affects the signal, building on an earlier subgroup analysis.

Individuals with other close relatives in UK Biobank are not a random subset of all UKB participants. They are stratified geographically, because they are more likely to be found in the most densely-sampled urban areas and they differ in other respects. Nick ran an XGBoost analysis to identify predictors of having a close relative in UKB. Prediction had a low specificity and sensitivity of roughly 60%. Nonetheless, the predictors with highest feature importance were:

• Number of full brothers (UKB code 1873)
• Number of full sisters (1883)
• Distance Euclidean to coast (24508)
• Time urine sample collected (20035)
• Country of birth UK elsewhere (1647)
• Length of time at current address (699)
• Frequency of friend family visits (1031)

I included Nick's XGBoost predictor of having UKB relatives in the logistic regression (see table above) but the predictor did not alter the signal of association at rs2076205, whether relatives were excluded (as shown) or not.

The take-home is we don't understand why actually including versus excluding relatives in UKB affects the signal of association, but given the lack of evidence for stratification and confounding of rs2076205, it may be a combination of (i) re-weighting the contribution to the analysis of different regions or test centres and (ii) differences in interpretation or ascertainment of the PCR positive-versus-negative outcome by region or test centre. The latter I have tried to investigate by subgroup analysis but the numbers are as yet too small.

Statistical evaluation of the signal of association at rs2076205

Here I summarize the statistical evidence favouring or disfavouring an association between rs2076205 and susceptibility to COVID-19. The figure shows the relative strength of evidence for the association (positive and bigger is stronger) as a function of the effect size in four logistic regression models (discovery vs quasi-replication, excluding vs including relatives) and two combined models (excluding vs including relatives). The idea is that one can weight the strength of evidence by one's prior expectation of observing an effect of a given magnitude in a common SNP (~25% frequency). To assist, known effect sizes discovered in other genome-wide association studies for susceptibility to infection are marked from Tian and colleagues (2017).

Briefly, the result of combining evidence across the discovery and quasi-replication analyses indicates that

• The statistical evidence for an association between rs2076205 and susceptibility to COVID-19 just meets the stringent criterion for 'genome-wide' significance if one includes close relatives (orange line).
• The evidence substantially surpasses the 'genome-wide' significance criterion if one excludes close relatives (red line).
• The effect size of the SNP is broadly similar whether one excludes or includes relatives, and between discovery and quasi-replication analyses.
• The effect size is not outlandishly large compared to previous studies of infection.

In more detail, the y-axis shows the log₁₀ likelihood ratio between the alternative hypothesis with the stated effect size and the null hypothesis with effect size zero. The log-likelihoods are maximized with respect to all other parameters, so they have the same degrees of freedom, making the log-likelihoods directly comparable. The log-likelihood is essentially acting as a Bayes factor between the alternative and null hypotheses.

If one were testing a single candidate SNP, Jeffreys' interpretation of the y-axis would mean above 0.5 represents substantial evidence in favour of the association. However, there are generally considered around 10⁶ effective tests in a genome-wide association study, so the threshold becomes 6.5. This is equivalent to the 'genome-wide' significance threshold for p-values of 5×10^-8, indicated on the figure after transforming to the log-likelihood scale.

Since the logistic regression is not quite identical to the SAIGE analysis, which is preferred because of a more sophisticated model of genetic covariance, I have indicated with stars the maximized likelihood and corresponding effect size estimate from SAIGE. The values at the grey and blue stars (analyses including relatives) were provided by Tomoko Nakanishi.

Biological considerations

As discussed previously, XPNPEP2 appears to be a very plausible candidate for a role in susceptibility to COVID-19. It belongs to a pathway in common with ACE2, the receptor by which SARS-CoV-2 gains entry to the cell. Moreover, its involvement in bradykinin signalling and its previous implication in angioedema lends support to a new and - from the perspective of treatment options - potentially important hypothesis that the kallikrein-kinin system, rather than the renin-angiotensin (RAS) system, may mediate the life-threatening pathophysiology of COVID-19:

Kallikrein-kinin blockade in patients with COVID-19 to prevent acute respiratory distress syndrome
F. L. van de Veerdonk, M. G. Netea, M. van Deuren, J. W. M van der Meer, Q. de Mast, R. J. Brüggemann and H. van der Hoeven (2020)
eLife 9: e57555

Conclusions

In summary, the statistical genetic evidence from the UK Biobank cohort is 'substantial' or 'decisive' depending on whether one includes or excludes close relatives, but the signal appears restricted to those of white European ancestry. The >1000-fold difference in strength of evidence is a concern for the robustness of the association. Yet detailed analyses of the SNP in question do not reveal stratification or confounding. The phenomenon may therefore reflect differences in the interpretation or ascertainment of the COVID-19 PCR positive and negative outcomes by region or testing centre. There is a desire to replicate the signal in an independent population, taking care to consider the potential for COVID-19 test results and take-up to vary from place-to-place. There is a need to urgently find improved interventions for the ongoing pandemic, and the analyses here lend support to the kallikrein-kinin blockade hypothesis.

XPNPEP2 is genome-wide significant in a COVID-19 +ve/-ve analysis of UK Biobank participants

On April 19 I wrote about a biologically interesting signal in XPNPEP2 in a comparison of COVID-19 positive and negative individuals who have participated in UK Biobank. In the latest data release, which represents a doubling in sample size for this analysis, the hit is genome-wide significant in an analysis of white Europeans. However, it is not genome-wide significant when analysing individuals with all ancestries.

The top hit is rs2076205, a single nucleotide polymorphism whose common allele occurs in 68% of COVID-19 negative individuals and 79% of COVID-19 positive individuals. It has a p-value of 1.1 in 10 million, which is below the widely-used threshold of 5 in 10 million that is often deemed statistically interesting.

In the analysis of individuals of all ancestries, the signal is substantially muted, and no variants are genome-wide significant. The reasons for the difference in analysis are still unclear, but the SNP does appear to be strongly population stratified.

XPNPEP2 is of interest because its product's normal function (Aminopeptidase P) includes degrading bradykinin, also degraded by ACE2. The involvement of XPNPEP2 variants in ACE inhibitor-associated angioedema has been proposed as combining with the drug to produce higher circulating bradykinin.

Bradykinin has already been suggested as an important mediator of the COVID-19 because of clinical similarities between COVID-19 and ACE inhibitor induced angioedema. Since SARS-CoV-2 gains entry to the cell via ACE2, the virus may inhibit its normal role and contribute to elevated bradykinin. The dry cough associated with COVID-19 has been likened to the 'bradykinin cough' associated with the use of ACE inhibitors.

Interestingly, bradykinin inhibitors such as icatibant already exist and are used in ACE inhibitor induced angioedema. However, to make the jump that such drugs might be useful against COVID-19 on this evidence alone is speculative, and there are several caveats to the association itself, including the requirement to replicate the effect in an independent population. There is a need to understand the robustness of the association to population stratification and other possible confounders.

Update

This post was updated a second time on 8 May to include inpatient status in the subgroup analysis.

To dig a little deeper into the result I have looked at effect sizes in different groups defined by

• Ethnicity: British (self reported), euranc (white European genetically) or any (these are the only groupings so far with sufficient sample size)
• Sex: female, male or any
• Close relatives: whether they are excluded (as above) or included in the analysis (larger sample size but potential for family clusters to over-influence the results)
• Inpatients: whether individuals were in (ever inpatients when tested) or nin (never inpatients when tested)

The results suggest that the effect is both larger in magnitude and stronger in statistical significance in males. But this difference depends on inpatient status: the signal appears strong and similar in magnitude among male and female inpatients. In non-inpatients, the signal is non-significant in almost every subgroup. Including individuals not identified genetically as white European slightly dilutes the signal, as does including close relatives.

The graph shows the effect (and 95% confidence interval) of the rarer allele: negative coefficients (on a log odds scale) indicate that the rare allele reduces the risk of returning a positive test. To illustrate, an effect size of -0.35 would imply that a male with the rare allele is 30% less likely to test positive than a male with the common allele.

Some technical details: the effect size analysis is based on a logistic regression, whereas the GWAS above uses the SAIGE tool. The covariates included are: sex, age, age*age, sex*age and genetic principal components (40 in the logistic regression, 20 in the GWAS). The results are being contributed to the COVID-19 Host Genetics Initiative.