Reply to: Insufficient evidence for natural selection associated with the Black Death

Abstract

Barton et al.¹ raise concerns regarding the statistical analyses² of our recent work, highlighting challenges of inferring natural selection from ancient genomic data. We have collected new data and conducted additional analysis, all of which support our original conclusions. First, when applying a maximum likelihood approach to allele frequency estimation, filtering sites to known human variants, and downsampling our data to the same mean coverage across sites, we continue to recover a consistent enrichment of high F_ST values at immune loci relative to putatively neutral sites. Second, permutations show that rs2549794 near ERAP2 remains the strongest candidate for selection during the Black Death. Third, the evidence for selection on ERAP2 is supported by functional data demonstrating the impact of the ERAP2 genotype on the immune response to Y. pestis infection in immune cells. In particular, we show that the functional evidence linking the putatively selected ERAP2 allele to cytokine release and Y. pestis clearance is unusual relative to all immune loci SNPs tested. Finally, independent epidemiological data has recently emerged showing that the putatively selected ERAP2 allele does indeed protect against severe respiratory infection in contemporary populations. Together, all lines of evidence converge in support of selection on ERAP2 during the Black Death.

We thank Barton et al.¹ for their careful consideration of our work². They raise important concerns, which have led them to conclude that our data set is underpowered to be informative about human evolution during the Black Death. We agree that small sample sizes are an inherent limitation of ancient DNA studies. To address this, we incorporated multiple lines of evidence—including genomic data spanning before, during, and after the Black Death, replication across two distinct population cohorts, and experimental data on cellular responses to Yersinia pestis infection—to build a case for selection. Our functional analyses—a rarity in ancient DNA studies—are further supported by epidemiological evidence from an independent group³. Together, all lines of evidence continue to support selection on ERAP2.

Despite the importance of functional evidence, Barton et al.’s critique focuses entirely on the detection of candidate loci from ancient genomic data. They raise concerns about biases in our original allele frequency estimates and inclusion of false positive SNPs from aDNA damage, which have been addressed in an accompanying Correction⁴. Using a maximum likelihood-based allele frequency estimate as suggested by Barton et al. (which are not biased by sequencing depth: Extended Data Fig. 1), we recapitulate our original finding: high F_ST values are enriched for immune loci relative to putatively neutral sites (Figure 2A in the corrected article^2,4). We confirm that this enrichment is not an artifact of differences in coverage between neutral and immune loci—a concern raised by Barton et al.—in two ways. First, we recover a similar enrichment signature after down-sampling the sequencing data to the same mean coverage across neutral and immune loci (Figure 1A: London: 2.07-fold, p = 2.7 × 10⁻⁴; Denmark: 2.14-fold, p = 1.6 × 10⁻⁴). Second, we identify an excess of high-F_ST values in immune loci when calculating enrichments based on pseudo-haploid data, which eliminates differences in coverage at the cost of excluding a large proportion of the data (London: 3.1-fold, p = 7.8 × 10⁻¹⁵; Denmark: 2.4-fold, p = 1.9 × 10⁻⁸). These analyses indicate that coverage differences alone cannot account for the enrichment we detect. This conclusion is robust to the number of MAF-matched neutral variants used to identify highly differentiated sites (Table S1, Methods).

Figure 2: Candidate loci for positive selection.

(A) Loci ranked by evidence for positive selection, shown on the y-axis as the −log₁₀ p-value. This value was calculated using Fisher’s method from empirical p-values for the degree of genetic differentiation observed in London and Denmark separately, based on the proportion of permutations for which permuted F_ST values exceeded the one observed in the real data. Candidate immune loci (n=290) were selected as sites where the allele frequency change between pre- and post-Black Death was in the same direction for both London and Denmark and the allele frequency change between pre- and during-Black Death in London was in the opposite direction. These points are shown in green and values are jittered along the x-axis to limit overlapping points. Dashed lines correspond to p-values of 0.05 (orange) and 0.01 (red). Among our 4 original candidate loci, rs2549794 (ERAP2) and rs1052025 (NFATC1) are shown in red. The other two variants failed to meet the criteria where the changes in allele frequency between pre- vs post-Black Death and pre- vs during-Black Death are in the opposite direction. (B) Allele frequencies over time for rs2549794 and rs2548527, which are strongly linked and near ERAP2. Error bars represent the standard deviation based on bootstrapping individuals from that population and each time point 10,000 times. Allele frequencies for London are shown in red and for Denmark are shown in blue.

Figure 1: Enrichment of sites with high F_ST in candidate immune loci compared to putatively neutral loci is not an artifact of allele frequency estimation method or coverage.

(A) Down-sampling our data such that coverage was even across exon, immune, and neutral sites preserves the observed enrichments. The figure shows the degree of enrichment (odds-ratio) for candidate sites that exceed the 99^th percentile of neutral, MAF-matched F_ST values. Significance was assessed using a binomial test comparing the number of sites observed to pass the 99^th percentile to the number expected by chance. We note that down-sampling discards ~1/3 of the data, which likely reduces our power to detect a significant enrichment of high F_ST in immune loci relative to putatively neutral sites. (B) Heat map of enrichment statistics obtained by permuting sites across locus type (neutral versus candidate immune loci). Observed enrichments for the down-sampled (orange) and full (red) datasets were larger than observed in any of the 10,000 permutations.

Barton et al. propose a permutation-based method for generating a null distribution of allele frequency changes across time (randomly assigning individuals to pre- or post-Black Death). Applying this, we find only 0.86% of permutations yield a higher enrichment of high F_ST values among immune loci than observed in both the London and Denmark cohorts (Fisher’s combined p = 0.046). A limitation of this null is that it assumes that F_ST between time periods is uniformly 0, treating all non-zero values as noise. This differs from our original analysis, which tested whether immune genes showed greater divergence due to selection compared to neutral regions. Comparing neutral and selected regions of the same genome identifies candidates for selection relative to other sites that experienced similar demographic processes (e.g. population collapse from the Black Death). These sources of variance are entirely lost in Barton et al.’s approach, producing a null that is significantly different from what’s observed in the real data (Methods).

To better capture the intent of our original hypothesis, we propose randomizing “neutral” and “immune” variants instead. Across 10,000 such permutations, none produced enrichment scores for immune loci (relative to neutral loci) as large as those in the observed data (i.e., p < 0.0001; Figure 1B). While, in principle, these results could be confounded by an artifact between neutral and candidate loci, we show that the coverage-based explanations proposed by Barton et al. cannot explain our enrichment and are unable to find evidence of any other systematic differences in data quality (e.g., DNA damage or mapping bias) between immune and neutral loci. Across all methods — our original analysis, Barton et al.’s proposed permutations, and our neutral vs. immune variant approach — we consistently observe enrichment of high-F_ST variants in candidate regions compared to neutral regions, regardless of SNP matching scheme (Table S1 and Methods), though the strength of statistical support varies by method.

To identify individual sites that may have been the target of selection, our original approach focused on identifying F_ST outliers, a common practice in scans of natural selection in humans^5–7.

Barton et al. argue that, instead, we should have relied on explicit p-value estimates for individual SNPs. We now calculate site-specific p-values by comparing observed F_ST values with those derived from permuting individuals between time points (Methods). rs2549794, near ERAP2, remains the top hit among all variants (Figure 2A, p = 1.4×10⁻³; Table S2), although this p-value does not pass a conservative Bonferroni correction (p = 1.7×10⁻⁴ based on the 290 candidate sites tested). However, this approach does not capture two important features of our original prioritization criteria: that changes in allele frequency must be directionally concordant between Denmark and London; and that, in the London sample, selected variants must decrease in frequency among individuals who died during the Black Death and increase in frequency thereafter. When we include these criteria, only 24 of one million permutations (i.e., 0.0024%) produce F_ST levels greater than or equal to those observed in the real data from London and Denmark (Methods). rs2549794 was again the top candidate (Extended Data Fig. 2). Another variant (rs2548527) in strong LD with rs2549794 exhibits qualitatively similar changes in allele frequency between pre- and post-Black Death samples, indicating that our results for ERAP2 are unlikely to be due to DNA damage (Figure 2B). An independent study from Cambridgeshire, England, also shows an increase in frequency from 52% to 60%⁸ of ERAP2 following the Black Death⁸.

Barton et al. report that previous scans of selection^9,10 do not support selection on ERAP2. This does not contradict our findings, since those scans were designed to detect directional selection over many generations. Rather, we propose that ERAP2 variation has evolved under long-term balancing selection^3,11,12, with a short pulse of strong directional selection during the Black Death. Specifically, we suggest that the rs2549794-C allele is advantageous against Y. pestis and potentially other pathogens, but that it is also associated with a fitness cost that maintains a long-term balanced allele frequency. In support of this model, a recent study of 2,376 ancient genomes, covering a period of 10,000 years, also identified evidence for positive selection for the C allele in the rs2549794 variant of the ERAP2 gene¹³. They conclude that selection started between 1,000 and 2,000 years ago, a timeline that aligns with significant European pandemics, including both the first and second pandemics of plague.

We note that our estimated selection coefficient for rs2549794 cannot be directly compared to long-term selection cases like lactase persistence. In our study, we modeled directional selection during a short, 3-generation pulse, whereas, selection for lactase persistence occurred over at least 150 generations⁹. Nevertheless, we acknowledge that the selection coefficient may be over-estimated due to small sample sizes and ascertainment bias towards high F_ST variants. Larger cohorts are needed for more precise estimates, a challenge given our study’s narrow temporal and geographic focus.

Finally, Barton et al dismiss the functional data, arguing that it is unsurprising that the putatively selected ERAP2 variant influences the immune response to Y. pestis, suggesting that any immune variant would produce similar results. We disagree based on both the established literature (most genetic variants in immune genes are not associated with infectious disease phenotypes) and additional analyses. In our original analysis, we showed that the ERAP2 variant predicts the response to experimental Y. pestis inoculation in macrophages for four of 10 tested cytokines². Carriers of the putatively selected C allele also better restrict bacterial growth post-infection² and the C allele is protective against respiratory diseases (specifically pneumonia)³. To further emphasize the rarity of this finding, we tested 1,741 other immune variants for an association with cytokine levels and bacterial clearance. Only three showed similar levels of association with Y. pestis growth and host cytokine response (Table S3), and none were linked to pneumonia-related diseases (Table S4). Thus, rs2549794 is not only the strongest candidate based on genetic differentiation over time but the plausibility of selection on ERAP2 is further supported by functional evidence.

Concluding remarks

Our new analyses support our original conclusions: (i) changes in allele frequencies at immune genes during the Black Death are larger than expected based on neutral regions; (ii) the temporal patterns of genetic differentiation at the ERAP2 locus are the most extreme amongst all variants tested; and (iii) the ERAP2 genotype has strong and unusual effect—relative to other immune loci—on the immune response to Y. pestis.

As for classic cases of positive selection such as alleles for lactase persistence or malaria resistance, we argue that it is the combination of sequence-based evidence with functional, organismal, and epidemiological data that make these cases for selection compelling. Such convergent evidence is crucial in ancient DNA studies, where achieving the statistical power of modern GWAS is challenging. We look forward to additional ancient DNA studies with larger sample sizes and full-genome resolution to fully delineate the impact of the Black Death on the evolution of the human immune system.

Methods:

Site curation and allele frequency estimates

Barton et al. raise a valid concern that our original dataset included many sites not previously identified as polymorphic in humans, many of which may have been introduced by DNA damage. We correct for that oversight by limiting our analyses here to variants that are also reported in the 1000G project⁷. Doing so results in a transition to transversion ratio of 3.18 and a strong correlation in allele frequency between our population in London and the 1000 Genomes project GBR population⁷ (r² = 0.94, p < 10⁻³⁰⁰). As with all ancient DNA studies, we cannot completely rule out damage as a contributing factor to our genotype calls. However, as described in the original publication, we took several steps to minimize the impact of deamination and other forms of ancient DNA damage in our estimates of allele frequencies. Quality scores in each bam file were adjusted for DNA degradation based on their initial qualities, position in reads, and damage patterns. In addition, we systematically trimmed off the first and last 4 bases of each sequencing read where the bulk of damage accrued for our samples is concentrated (Figure S1 of original publication). Moreover, we note that the average coverage across our most highly differentiated sites is considerably high for ancient DNA data (mean of 4.5x across all variants initially reported in Table S4, and 9.4x rs2549794 near ERAP2), which minimizes the potential contribution of DNA damage to our genotype calls.

We also revise our allele frequency estimator to use the ML-based approach suggested by Barton et al. Specifically, we let the genotype likelihoods for individual i in 1:n be given by $x_i^j$ where j is the genotype denoted as 0, 1, or 2 alternate alleles. The ML estimate of the allele frequency p is then:

$$\widehat{p}=\mathit{\text{argmax}}\left\{\sum _{i=1}^n\text{log}(x_i^0{\left(1-p\right)}^2+{2x}_i^{1}p\left(1-p\right)+x_i^2{p}^2)\right\},\mathit{\text{with}}pϵ[\mathrm{0,1}]$$

Prior to estimating F_ST, we applied the numerical implementation of this estimator used by Barton et al. to estimate allele frequencies in each population and each time point, based on the genotype likelihoods reported in our original manuscript.

Test for enrichment at immune loci

We replicated our previous analysis demonstrating an enrichment for highly differentiated candidate sites using these ML-based allele frequency estimates. Specifically, we retained candidate loci with a minor allele frequency greater than 5% (n=1,758) and identified highly differentiated sites by comparing the observed F_ST values for candidate immune sites against 250 MAF-matched neutral sites. Matched neutral sites were selected as the sites with the closest MAF to the candidate variant such that 125 had higher MAF than the candidate variant and 125 had a lower MAF. This represents a departure from our initial approach using MAF bins, which was necessary due to the fact that we have now limited our analyses to a smaller set of variants known to be polymorphic in modern humans. Importantly, however, we found a strong correlation between the original F_ST percentiles for each of the candidate SNPs and those calculated using the allele frequencies estimated using the ML-based approach suggested by Barton et al. (Spearman’s rho > 0.88, p < 10⁻¹⁰⁰ in both London and Denmark). Thus, although some of the variants that met the original selection criteria may now fall below the original cutoffs, the evidence for selection remains qualitatively unchanged (e.g., the F_ST quantile in London for rs2549794 near ERAP2 changed from 0.045 (original) to 0.055 (revised).

While this approach was chosen to replicate our original analyses, we recognize that there is a tradeoff between controlling for MAF and having sufficient neutral sites to capture the F_ST distribution. Importantly, increasing or decreasing the number of neutral sites (down to 100 loci or up to all neutral variants) has little to no impact in the observed enrichment of high F_ST in candidate sites as compared to neutral variants (Table S1; log₂(OR) > 1.98, p < 10⁻²⁰ for all binomial enrichment tests in in London or Denmark). Whether using all neutral variants or 100 MAF-matched sites to define F_ST quantiles, rs2549794 near ERAP2 emerges as a candidate for selection when following our original prioritization criteria.

Coverage alone does not drive the observed enrichment

Barton et al. argue that differences in coverage between sets of sites drives the observed enrichment for high F_ST values among immune loci. To test this hypothesis, we down-sampled our dataset such that the coverage was the same between the three capture sets (exons, GWAS, and neutral loci). Specifically, we estimated the coverage of each sample at analyzed sites using “samtools depth” separately for sites captured in the immune GWAS, immune exons, and neutral capture sets¹⁴. We then calculated the minimum coverage in each individual, and used the “samtools view” with the -s option to down-sample the other two capture sets to the minimum coverage for that individual. For samples where one of the capture sets was not used, we instead down-sampled sequence data sets for that sample to the average coverage for all individuals. We then estimated allele frequencies and calculated the enrichment of highly differentiated loci, as described above. As expected, this procedure resulted in very similar coverage between the sets of immune and neutral loci (mean coverage at GWAS loci 5.416 ± 3.56x, at exonic loci 5.419 ± 3.57x, and at neutral loci 5.417 ± 3.58x). Based on further concerns that differences in the distribution in coverage were maintained following this downsampling approach, we also generated pseudo-haploid data using ANGSD¹⁵ with `-dohaplocall 1 -doCounts 1`. Allele frequencies were estimated using the maximum likelihood estimator for downsampled data and using the proportion of alternate reads for pseudo-haploid data. We then applied the identical filtering criteria and analysis pipeline as for the full data set.

Permutation test for enrichment at immune loci

We first used the permutation approach employed by Barton et al. to test whether the observed enrichment of large F_ST values at candidate loci exceeded random expectations. Specifically, within each population, we permuted individuals without replacement across time points. Whereas Barton et al. only performed this analysis for the London pre-Black Death and London post-Black Death samples, our approach also considered the London samples collected during the Black Death and the pre-Black Death and post-Black Death samples collected in Denmark.

A limitation of this null is that it assumes that F_ST between time periods is uniformly 0, and that all non-0 values reflect measurement noise—which differs a different null from our original analyses which tested the hypothesis that immune genes show greater divergence (as a result of selection) than putatively neutral regions of the same genomes. A comparison against neutral regions identifies candidates for selection relative to other sites that experienced the same effects of population structure, migration, and other demographic processes (including severe population contractions as the a result of the Black Death itself). These sources of variance are entirely lost in Barton et al.’s approach, producing significant differences in variance when compared to real data (Methods). Consequently, 57% and 66% of permuted distributions in London and Denmark, respectively, show a significant difference in variance compared to real data when using their approach (F-test, p < 0.05; 11.4 and 13.4-fold more than expected by chance).

In our proposed alternative approach, we permuted SNP assignment into the “putatively neutral” versus “immune” categories, while keeping the sizes of both categories identical to those in the original data. We repeated this process 10,000 times for each permutation approach, and for each iteration calculated the proportion of sites that exceeded 99% of neutral, MAF-matched loci. We then estimated the proportion of permutations in which more sites passed this threshold in the permuted data than in the observed data.

Permutation test for selection at immune loci

Inspired by the permutation approach proposed by Barton et al., we next used permutations to calculate a site-specific, empirical p-value for each site that is independent of MAF-matched neutral loci. Following the same reasoning as in the original publication, we filtered for sites where (i) in London, the direction of change from the pre-Black Death samples to those collected during the Black Death was opposite to the direction of change between the pre-Black Death to post-Black Death samples; and (ii) the pre to post-Black Death allele frequencies change in the same direction in London and Denmark. For each site, we created one million permutations and, after filtering for permutations where these conditions were also met, calculated empirical p-values for London and Denmark separately, based on the number of cases in which the permuted F_ST value exceeded the observed F_ST in that population. We then combined these two values using Fisher’s method to calculate a single p-value for each site. This produces a conservative estimate as these p-values do not reflect the change in allele frequency in London during the Black Death or replication between London and Denmark.

To include evidence based on the direction of allele frequency changes across time points and populations, we also calculated, for each site, the proportion of permutations for which the allele frequency change between pre- and post-Black Death was in the same direction for both London and Denmark, the allele frequency change between pre- and during-Black Death in London was in the opposite direction, and the permuted F_ST was greater than the observed F_ST in both London and Denmark. This approach asks about the probability of identifying, under the null, a pattern of genetic differentiation that matches the F_ST levels observed in the real data and that meets our original prioritization criteria. However, requiring greater F_ST in both London and Denmark results in poorly calibrated p values as it is not possible to incorporate the directional criteria alongside a proper multivariate permutation test. We therefore suggest treating these proportions as ranked evidence for selection, rather than p values.

Allele frequencies and differentiation at rs2549794 and rs2548527

rs2549794 and rs2548527 are linked variants near ERAP2 (r² = 0.79 in 1000Genomes’ GBR population). To show the change in allele frequency over time, we estimated the variance of the allele frequencies at each time point using ML allele frequency estimates by bootstrapping with 10,000 replicates. Mean coverage of rs2549794 was more than twice that of rs2548527 (9.4x vs 3.7x) and many more individuals were missing genotype calls for rs2548527 (56 samples vs 6 for rs2549794), reducing confidence in the estimated allele frequencies at rs2548527.

Functional consequences of immune loci

To ask whether the functional characteristics of the putatively selected ERAP2 variant and the closely linked rs2548524 variant were unusual in their functional association with the immune response to Y. pestis, we generated new genotype data for all the samples used in our functional assays using the Gencove genotyping platform, focusing on the set of immune loci with MAF >5% in our ancient genomic sampled also genotyped here (n=1,741). We successfully generated genotype data in 32 samples (out of the original 33). We proceeded to assess whether the genotype at each candidate locus correlated with the levels of the ten cytokines we previously assayed using Luminex technology, as detailed in our original manuscript, and with Y. pestis bacterial clearance, based on the functional data from our initial study. For each SNP, we established a threshold for a significant association at a nominal p-value of less than 0.05. This criterion was based on the premise that if any given SNP had emerged as our most extreme candidate, our analysis would have been exclusively focused on that particular SNP. Out of all 1,741 variants tested, only three other loci show similar levels of association with Y. pestis growth or the host cytokine response as seen for the ERAP2 variant functionally tested in the original study. As in Hamilton et al.³, we extracted GWAS summary statistics from the UK Biobank for the association between these four loci (ERAP2 plus the other three independent variants) and the risk of pneumonia, hospitalization from pneumonia, death from pneumonia, and death from pneumonia while in critical care from the IEU OpenGWAS Project (https://gwas.mrcieu.ac.uk/datasets/, accessed 8 August 2024). Only the ERAP2 variant showed any evidence of an association with pneumonia-related traits, notably an association with critical care (p = 0.0076) and death from pneumonia (p = 0.014). Confirming this finding, the other three variants are also not reported to be associated with infectious disease susceptibility in the NHGRI-EBI GWAS Catalog (https://www.ebi.ac.uk/gwas/, 8 August 2024).

Extended Data

Extended Data Fig. 1: Maximum likelihood allele frequency estimates are not biased by coverage.

The difference between simulated “true” allele frequencies and the allele frequency estimated using either the genotype likelihoods (GL) in Klunk et al. or the maximum likelihood (ML) approach outlined by Barton et al. The GL approach shows a bias towards overestimating the frequency of rare variants, which is exaggerated at lower coverages. No similar bias is apparent using the ML approach.

Extended Data Fig. 2: Candidate loci for positive selection, based on the proportion of permutations exceeding the observed patterns.

Loci ranked by evidence for positive selection, shown on the y-axis as the −log₁₀ proportion of permutations where the F_ST was greater than the observed F_ST in both London and Denmark, the allele frequency change between pre- and post-Black Death was in the same direction for both London and Denmark, and the allele frequency change between pre- and during-Black Death in London was in the opposite direction. Candidate immune loci (n=290) are shown in green, and values are jittered along the x-axis to limit overlapping points. Among our 4 original candidate loci, rs2549794 (ERAP2) and rs1052025 (NFATC1) are shown in red. The other two variants failed to meet the criteria where the changes in allele frequency between pre- vs post-Black Death and pre- vs during-Black Death should be in the opposite direction. This plot differs from Figure 2A in that it shows the proportion of permutations on the y-axis rather than the multivariate p-value used in Figure 2A.

Data availability

New genotype data were generated for 32 individuals for whom we had previously generated functional data and genotyped at the putatively selected ERAP2 variant. These genotype data are available via the GitHub repository for this project.

Code availability

Code to replicate the analyses described here is available at https://github.com/TaurVil/VilgalysKlunk_response_to_commentary