The first 5 pages, shown as text so search engines and readers can find this document.
Page 1
Nature | www.nature.com | 1
Article
Ancient DNA reveals pervasive directional
selection across West Eurasia
Ali Akbari1,2,3 ✉, Annabel Perry2,3, Alison R. Barton2,3, Mohammadreza Kariminejad4,
Steven Gazal5,6,7, Zheng Li8, Yating Zeng5,9, Alissa Mittnik2,10, Nick Patterson2,3,
Matthew Mah1,11, Xiang Zhou12, Alkes L. Price3,13,14, Eric S. Lander3,15,16, Ron Pinhasi17,18,
Nadin Rohland1,2,3, Swapan Mallick1,2,3,11 & David Reich1,2,3,11 ✉
Ancient DNA has transformed our understanding of population history1, but its
potential to reveal as much about human evolutionary biology has not been realized
because of limited sample sizes and the difficulty of distinguishing sustained rises
in allele frequency increasing fitness—directional selection—from shifts due to
migrations, population structure, or non-adaptive purifying or stabilizing
selection2–7. Here we present a method for detecting directional selection in ancient
DNA time-series data that tests for consistent trends in allele frequency change over
time, and apply it to 15,836 West Eurasians (10,016 with new data). Previous work has
shown that classic hard sweeps driving advantageous mutations to fixation have been
rare over the broad span of human evolution8,9. By contrast, in the past ten millennia,
we find that many hundreds of alleles have been affected by strong directional
selection. We also document one-standard-deviation changes on the scale of modern
variation in combinations of alleles that today predict complex traits. This includes
decreases in predicted body fat and schizophrenia, and increases in measures of
cognitive performance. These effects were measured in industrialized societies,
and it remains unclear how these relate to phenotypes that were adaptive in the
past. We estimate selection coefficients at 9.7 million variants, enabling study of
how Darwinian forces couple to allelic effects and shape the genetic architecture
of complex traits.
Ancient DNA data hold extraordinary promise for revealing adapta-
tion, making it possible to track effects across time and obtain direct
measurements of selection coefficients10,11. Rather than being trapped
in the present and studying the scars left by selection on the genomes
of descendants12,13, ancient DNA makes it possible to test directly
whether frequencies of variants shifted more than could be expected
by chance2–5. Such data also make it easier to measure selection on
variants not of recent mutational origin (standing variation), which is
challenging to detect using retrospective methods14. Previous ancient
DNA selection studies in West Eurasia2–5 (Europe and its neighbours
in the Near East) have identified dozens of alleles influenced by selec-
tion, but despite growth in reported ancient individuals to more than
10,000 today15, the number of genome-wide significant loci reported
in a single study grew only from 12 in the first genome scan in 2015
(ref. 2) to 21 in 2024 (ref. 4), raising the concern that this approach might
not deliver broad insights into human adaptation. Here we increase
the yield of discoveries by 20-fold by adding more statistical power
(due to a qualitatively new method and larger sample size) and reduc-
ing artefactual signals (through intensive data cleaning).
First, we increased power by testing for a consistent trend in allele
frequency change over time. Several past studies have dealt with the
challenge posed by population mixture by treating more recent popula-
tions as linear combinations of more ancient ones, then searching for
alleles whose frequencies were outliers compared with what would
be expected from this history2,3. However, changes in frequency due
to selection are often less than what can be expected from gene flow
and random genetic drift6, and in this context, increasing sample
size helps little. We used a qualitatively different approach, using the
genetic similarity of each individual to every other, and testing whether
the date when they lived provides additional predictive power for the
allele frequencies of their population beyond what is expected from
the empirical population structure. Our test is simple: at each variant
https://doi.org/10.1038/s41586-026-10358-1
Received: 14 September 2024
Accepted: 4 March 2026
Published online: xx xx xxxx
Check for updates
1Department of Genetics, Harvard Medical School, Boston, MA, USA. 2Department of Human Evolutionary Biology, Harvard University, Cambridge, MA, USA. 3Broad Institute of MIT and Harvard,
Cambridge, MA, USA. 4Shamsipour Technical and Vocational College, Tehran, Iran. 5Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern
California, Los Angeles, CA, USA. 6Center for Genetic Epidemiology, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA. 7Department of Quantitative and
Computational Biology, University of Southern California, Los Angeles, CA, USA. 8Department of Biostatistics, School of Public Health, University of Michigan, Ann Arbor, MI, USA. 9Department
of Biostatistics and Data Science, School of Public Health, The University of Texas Health Science Center at Houston, Houston, TX, USA. 10Department of Archaeogenetics, Max Planck Institute
for Evolutionary Anthropology, Leipzig, Germany. 11Howard Hughes Medical Institute, Harvard Medical School, Boston, MA, USA. 12Department of Statistics and Data Science, Yale University,
New Haven, CT, USA. 13Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, MA, USA. 14Department of Biostatistics, Harvard T.H. Chan School of Public Health,
Boston, MA, USA. 15Department of Systems Biology, Harvard Medical School, Boston, MA, USA. 16Department of Biology, Massachusetts Institute of Technology (MIT), Cambridge, MA, USA.
17Department of Evolutionary Anthropology, University of Vienna, Vienna, Austria. 18Human Evolution and Archaeological Sciences, University of Vienna, Vienna, Austria. ✉e-mail: Ali_Akbari@
fas.harvard.edu; reich@genetics.med.harvard.edu
Page 2
2 | Nature | www.nature.com
Article
we asked whether hypothesizing a non-zero selection coefficient s—
causing allele frequency to trend in the same direction over all times
and places—predicts frequency differences across populations signifi-
cantly better than empirically measured population structure alone.
Second, we increased power through a 14-fold increase in sample
size, driven by 10,016 ancient individuals for whom we report new
data, which combined with previously reported data yields a dataset
of 15,836 people spanning 18,000 years (Supplementary Tables 1–3).
We co-analysed with 6,438 modern people, sub-sampled so that their
countries of origin were spread evenly over West Eurasia (Extended
Data Fig. 1a,b).
Third, we increased power by imputing diploid genotypes16, leverag-
ing known patterns of allelic correlation in a modern reference panel to
fill in missing data in all individuals. We also carried out intensive data
cleaning (Supplementary Section 1), filtering out single-nucleotide
polymorphisms (SNPs) susceptible to false positives. The final dataset
included 8,074,573 SNPs and 1,665,051 insertions or deletions (indels)
on chromosomes 1–22.
Test of selection on single variants
For each SNP in the genome, we estimated a selection coefficient, s,
which we found has a standard error of typically 0.17% for common
variants (Extended Data Fig. 1c,d). In theory, a valid test for selection
should be Z, the number of standard errors this quantity is from zero,
and we should be able to use a normal distribution to identify scores that
pass the standard threshold of genome-wide significance (P%<%5%×%10−8).
In practice, the median χ2 statistic (squared Z-score for the number of
standard errors s is from zero) is inflated by 4.75-fold relative to a χ2
distribution with 1%d.f. In human genetics studies, such inflation can
arise due to uncorrected population structure or non-normality of the
null distribution due to case–control sample imbalance17. Inflation like
this is often addressed by rescaling χ2 statistics by the median inflation
across the genome18,19. But if the null is not normally distributed, or a
substantial fraction of the genome carries real signal18, random sites
will not provide an appropriate neutral baseline. These concerns apply
here: the null distribution is not normal due to frequency-dependent
biases from imputation and the non-linear transformation of allele
frequencies (Supplementary Section 2), and most of the genome is
in linkage disequilibrium with sites showing evidence of directional
selection (Extended Data Fig.%2b and Supplementary Section 3).
Instead, we calibrated our test by taking advantage of a finding
about the connection between selection coefficients and associa-
tions to phenotypes in living people. The proportion of SNPs showing
significant association to a phenotype in genome-wide association
studies (GWAS) increases with our selection statistic, plateauing at
around 5.2 times the rate for random SNPs (Fig. 1a; this analysis relies
on 1,317,022 SNPs with a genome-wide significant association to at
least one phenotype for 452 traits in the UK Biobank, and conditions
on minor allele frequency (MAF) to remove bias due to signals being
easier to detect for higher MAFs). The plateau occurs at the same place
when we control for ‘background selection’, the loss of neutral varia-
tion linked to deleterious mutations removed by purifying selection,
although the enrichment in GWAS variants is reduced (approximately
2.4 times; Extended Data Fig. 3a). This is the pattern expected for a true
threshold for genome-wide significance: if SNPs beyond this threshold
reflect a combination of true signal and false discoveries, we would
expect enrichment to continue beyond it. This threshold occurs at
Z%=%8.23, larger than the standard threshold (5.45) for genome-wide
significance for a normal distribution, so we rescaled the naive score
by this quantity to obtain an X-statistic (Z/1.51) whose significance
threshold matches the standard threshold.
To validate our procedure, we conducted realistic forward-in-time
simulations (Fig. 1b, Extended Data Fig. 4 and Supplementary Sec-
tion 2). The simulations specify a demographic history inspired by
that of Europe and produce patterns of structure matching impor-
tant features of real data. The simulations include coding, functional
non-coding and neutral regions, recapitulating varying degrees of
background selection. The simulations also permit exploration of
the relationship between selected alleles and traits, in particular the
phenomenon of ‘stabilizing’ selection to remove genetic variation
when the average genetic predictor of a trait is at a fitness optimum.
We verified that these simulations produce a deficiency of common
polymorphisms influencing traits, matching patterns in real data20.
Our test for directional selection is robust to every confounding fac-
tor that we considered, including population structure, background
selection due to purifying or stabilizing selection, sampling bias and
selection in the ancestral population but not during the time transect
itself (Supplementary Section 2). None of these produces even a hint
of the enrichment in GWAS signals that we observe in data. By contrast,
in simulations with directional selection, including where we shift the
trait optimum, leading to deterministic changes in allele frequency to
bring the population closer to that optimum21, we observed exactly this
signal and could use it to provide a valid calibration of genome-wide sig-
nificance (Fig. 1b, Extended Data Fig. 4 and Supplementary Section 2).
To translate X to a posterior probability of selection π, we used a false
discovery rate (FDR) approach (Fig. 1a,b and Supplementary Section 3).
We fit a smooth function to the enrichment curve for GWAS signals and
estimate that at X greater than our threshold of 5.45, π >%99%.
A second line of evidence that these signals are real comes from com-
puting the mean haplotype allele frequency (HAF) score, the sum of
squared derived allele frequencies scaled by sample size in 200-kb win-
dows surrounding each tested variant. Previous work22,23 has shown that
directional positive selection on derived alleles is expected to increase
HAF scores, whereas background selection is expected to decrease it
(Extended Data Fig. 3b and Supplementary Section 4). After comput-
ing the residual HAF score for each variant controlling for background
selection24,25, we found that it increases with the X-statistic and begins to
rise before 5.45, the standard threshold for genome-wide significance
in GWAS (Fig. 1c and Extended Data Fig. 3c). Our forward-in-time simu-
lations (Supplementary Section 2) confirm that the rise in HAF score
occurs only for directional selection, with no hint of this pattern for
other scenarios (Fig. 1d and Extended Data Fig. 4).
We obtained a third line of evidence that we are detecting real sig-
nals of adaptation by showing that |X|%>%5.45 variants are associated
with some classes of traits much more than others. We found enrich-
ment for SNPs contributing to blood–immune–inflammatory traits
(95% confidence interval (CI) 3.12–8.24)3,5, compared with random
SNPs with matched characteristics defining the baseline. For mental–
psychiatric–nervous and behavioural traits, we did not detect enrich-
ment (95% CIs of 0.90–1.08 and 0.99–1.46, respectively; Fig. 1e and
Extended Data Fig. 5a). This cannot be explained by differences in allele
frequencies or background selection as we controlled for these factors.
The intensity of selection on both blood–immune–inflammatory and
cardiometabolic traits increased significantly in the Bronze Age relative
to the pre-farming period (Fig. 1e and Extended Data Fig. 5b), plausibly
reflecting adaptation to new diets, larger population or living closer
to domesticated animals26.
Hundreds of cases of directional selection
We found evidence of 479 independent loci (410 excluding the HLA
region) with |X|%>%5.45, corresponding to a π%>%99% probability of selec-
tion (Fig. 2a). To produce this list, we identified the strongest signal in
the genome and considered all SNPs in linkage disequilibrium with it
in unrelated Western European individuals (r2%>%0.05) to potentially
reflect the same signal. We then found the second-strongest signal
excluding these positions, and so on until no more variants passed this
threshold (Extended Data Fig. 2b). Visualizations of the trajectories
for these loci (Supplementary Section 5) and summary statistics for
Page 3
Nature | www.nature.com | 3
9.7 million variants (Supplementary Table 4) can be cross-referenced
with GWAS and with their frequency trajectories at the AGES browser
(https://reich-ages.rc.hms.harvard.edu).
The actual number of loci under selection is likely to be much larger.
Using a threshold of |X|%>%3.61 (FDR%=%50%), we identified 7,689 non-HLA
loci, implying more than 3,800 independent episodes of selection. The
exact number of distinct signals is difficult to count: residual linkage dis-
equilibrium exists among candidate loci, which at some loci may cause
some overestimation of signals (Supplementary Section 5), whereas
truly selected alleles in linkage disequilibrium with nearby stronger
alleles will be undercounted. Downsampling analyses showed that
further increases in sample size are expected to increase the number
of detected loci further, with people living more than 8,000 years ago
providing the most added power (Extended Data Fig. 1d,e).
To obtain insight into the phenotypes shaped by directional selec-
tion, we took advantage of the fact that a high proportion (62%) of the
variants with genome-wide evidence of selection are independently
associated to a phenotype in at least one UK Biobank GWAS. However,
biological interpretation is complicated as the allele that was the target
of selection may differ from the tag SNP that we are using to represent
the locus (and may even be in a neighbouring gene); because some
alleles affect multiple phenotypes, because the relevant modern trait
a b
c d
e
Category Number
of traits Heritability
enrichment
(95% CI)
Z-score for selection stronger in
Bronze Age than
Neolithic or Mesolithic
Historical period
than Bronze Age
Biomarkers 9 5.21 (3.17–7.26) 3.07 0.31
Blood–immune–inflammatory 20 5.68 (3.12–8.24) 3.96 1.92
Cardiometabolic 20 2.07 (1.27–2.87) 3.58 0.14
Life history–reproduction 7 1.89 (0.92–2.86) 2.54 0.00
Behavioural 13 1.23 (0.99–1.46) 1.96 0.95
Mental–psychiatric–nervous 13 0.99 (0.90–1.08) –0.61 0.33
Other 25 1.86 (1.23–2.49) 2.83 0.46
All 107 2.43 (1.86–3.00) 6.97 2.10
5.0 0
40
36
32
28
24
20
16
12
8
4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
4.5
4.0
3.5
3.0
2.5
Enrichment
Residual HAF/n
Enrichment
FDR
2.0
1.5
1.0
6 10
8
6
4
2
0
–2
5
4
3
2
1
0
0 1 2 3 4
|X|
5 6 7
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
FDR
0 1 2 3 4
|X|
5 6 7
0 1 2 3 4
|X|
5 6 7
0 1 2 3 4
|X|
5 6 7
Residual HAF/n
FDR
With directional selection
No directional selection
With directional selection
No directional selection
Fig. 1 | Multiple lines of evidence show that we are detecting genuine
directional selection. a, Enrichment of SNPs significant in at least one of 452
UK Biobank GWAS, for SNPs with |X| above the value on the x axis, controlling
for allele frequency. b, Simulations reveal enrichment of GWAS hits (blue; left
axis) closely aligns with 1%−%FDR (orange; right axis) until a plateau as a function
of |X|: 800 simulations with directional selection (200 each for selection
coefficients of 0.01, 0.02 or 0.03 for model 2.1, soft sweep; and 0.05 for model
2.2, hard sweep) and 800 simulations without it. c, Residual mean HAF score
((HAF)/n) for SNPs with |X| above the value on the x axis, computed as observed
minus expected, with n haploid sample size, from a linear regression correcting
for background selection using as explanatory variables McVicker-B,
Murphy-phastCons, Murphy-CADD, number of SNPs and heterozygosity in a
200-kb window. d, Residual mean HAF score in simulation for SNPs with |X|
above the value on the x axis, computed as observed minus expected, with n
haploid sample size, from a linear regression correcting for background
selection using as explanatory variables McVicker-B, coding region annotation,
functional non-coding region annotation, number of SNPs and heterozygosity
in a 200-kb window. The shaded areas show 95% CIs (a–d). e, The heritability
enrichment is a meta-analysis for annotations based on a binary selection
annotation, with FDR either below 1% (1) or above 1% (0). The Z-score for change
in selection intensity over time is based on a meta-analysis of heritability
enrichment comparing key cultural transitions: Mesolithic–Neolithic to
Bronze Age, and Bronze Age to historical era. We annotated each SNP
according to whether it is among the top 5% with the highest probability of a
stronger magnitude of selection coefficient in one time transect versus
another. Cells shaded grey denote P < 0.05.
Page 4
4 | Nature | www.nature.com
Article
may not be measured in one of the GWAS that we are analysing, or
because the phenotype in modern societies may not have existed in
the ancient societies where selection acted. The median selection mag-
nitude |s| at the tag SNPs for non-HLA loci is 0.86% (range of 0.43–4.3%),
and the median MAF is 13%. Standard errors in our estimates of |s| for
common alleles are 0.17% (Fig. 2b and Extended Data Fig. 1c,d).
We compared our results to those of five previous scans for selection
in Holocene West Eurasia (four based on ancient DNA)2–5,13 (Extended
Data Table 1 and Supplementary Section 6). Of 38 unique non-HLA loci
that met the formal threshold for genome-wide significance in at least
one previous study, 18 passed our π%>%99% threshold. The other 20 did
not replicate, in most cases due to what appears to be incompletely
controlled population structure driven by mixtures of populations
with different allele frequencies before they came together and in
a few cases due to failing quality control or fluctuating selection in
space or time.
We present a gallery of 36 single-allele trajectories of particular
interest (Fig. 3) as well as estimates of how their selection coefficients
changed over time (Extended Data Fig. 6). These loci are not necessar-
ily those with the largest X scores, but are highlighted as they address
long-standing debates. They include 27 passing the π%>%99% threshold,
5 with probable evidence of selection (57%%<%π%<%93%) and 4 with sur-
prising negative findings.
HLA-DQB1: selection in favour of the major risk factor for coeliac
disease
At the HLA region of chromosome 6, densely packed genes have key
roles in microorganism recognition. rs3891176 (C>A, meaning that the
ancestral allele is C and the newly arising mutation is A) is an excellent
tag for HLA-DBQ1*02/DQ2 HLA-DQB1*02/DQ2, with individuals carrying
two A alleles having a 19-fold higher susceptibility for coeliac disease
or gluten sensitivity (Extended Data Fig. 7a,b). The A allele has a selec-
tion coefficient of s%=%4.4% (π%>%99%), rising from approximately 0% to
approximately 20% in the past 4,000 years (Fig. 3a). These findings
suggest that the pathogenic exposures that drove its rise were not a
phenomenon only or largely of the rise of agriculture27.
ABO: positive selection for B at the expense of A
The ABO gene modifies oligosaccharides in glycoproteins on the surface
of red blood cells; its A, B and null (O) alleles encode distinct glycopro-
tein alterations that modulate resistance and susceptibility to diverse
pathogens28. The B allele rose from approximately 0% to approximately
a
b c
0
4.5
ABO
SLC39A8
NSF
KITLG
DUOX2
MUC4 TCHH
OAS1
SLC45A2
LCT
TYR
ATXN2
SLC22A4
FADS1
TLR1
CYP1A2
ADH1CDHCR7
FUT3 MUC5AC
106
90
70
50
30
10
–0.02 –0.01 0 0.01 0.02
105
104
104 105
Derived allele age from Relate
(years before present)
Derived allele
selection coefficient (s)
Simulation power (%)
106
4.0
Selection coefficient s (%)
Favoured allele age given a selective sweep
(years before present)
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0 0.05 0.10 0.15 0.20
MAF
0.25 0.30 0.35 0.40 0.45 0.50
2
4
6
8
10
PLA2G2A
SLC44A5
FLG TCHH
FCGR2A/3A/2B
RBBP5, CNTN2, DSTYK, TMCC2
MARC1
SDC1
PCBP1-AS1
IL1
GCA, IFIH1 LCT, MCM6
CCR5
DNAH1, NISCH, TNNC1, SEMA3G
RDUR
MED12L
MUC4
TLR1, TLR6
ADH1B/CABCG2
CENPE, SLC39A8
GYPA, GYPB
TPPP
ELOVL7
SLC45A2
SLC22A4/5
MHC
CDHR3
CFTR, ASZ1
8p23.1 inv. locus
SLC7A13, WWP1, CPNE3
NSMCE2
BNC2
9p21.1
9q31.2, RP11-240E2.2
ABO
FADS1/2
DHCR7, NADSYN1
TYR, GRMS
POU2F3
ESPL1
KITLG SH2B3, ATXN2
CRYL1
PCDH17
TNFSF13B
PLEKHG3
IGHG2
SORD, DUOX1/2/A1
CSK, CYP1A1/2
ITGAM
DEF8, MC1R, ZNF276, CDH15, ANKRD11
BLMH, KSR1 17q21.31 inv. locus
GRB2, MRPS7, NUP85, GGA3
FUT3, FUT6
FUT2
ASXL1
CEACAM19
B3GALT5
SREBF2, CENPM, TNFRSF13C (BAFF-R)
MUC2, MUC5AC
RSPH4A, ZUFSP, KPNA5, RWDD1
|X|
12
14
16
18
20
1 2 3 4 5 6 7
Chromosome
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Fig. 2 | Genome scan for directional selection. a, The x axis is the chromosomal
position, and the y axis is the selection signal for each variant. The dotted line
indicates our genome-wide significance threshold of |X|%=%5.45. For clarity, only
a subset of loci are annotated. b, Selection coefficient (s) estimated from our
scan plotted against MAF of tagging SNPs at independent loci with ( > 99%.
Overlaid grids are simulation-based power estimates (90%, 70%, 50%, 30% and
10% probability of detection). The error bars indicate standard errors from
20,374 unrelated individuals. c, The estimated age of the favoured allele in a
selective sweep versus the date of origin of the mutation inferred from Relate58,
for tagging SNPs at independent loci with ( > 99%. The age of the sweep is
defined as the time before present when the frequency of the favoured allele is
expected to have been 0.0001 given the present-day frequency in 1000
Genomes Project European populations and assuming that the selection
coefficient has been constant over time.
Page 5
Nature | www.nature.com | 5
10% over the past approximately 6,000 years (s%=%2.7%, π%>%99%), and
was matched by a concomitant decrease in A frequency (Fig. 3b). The
two alleles are associated with opposite effects on many phenotypes,
suggesting that with changing pathogenic exposures, the optimal
balance changed (Extended Data Fig. 7c,d).
TCHH: selection for an allele that reduced male-pattern baldness
An allele at missense SNP rs11803731 (A>T) in TCHH is a strong predictor
of straight hair and male-pattern baldness in European individuals. The
derived allele T is rare in African and East Asian individuals, and has been
0.25
a
0.16 0.08 0.06
0.12
0.12 0.05
0.04
0.03
0.02
0.01
0
0.10
0.08
0.06
0.04
0.02
0
0.08
0.04
0
0.05
0.04
0.03
0.02
0.01
0
0.06
0.04
0.02
0
0.15
0.6
1.0 0.5 0.30 1.0 0.05
0.04
0.03
0.02
0.01
0
0.8
0.6
0.4
0.2
0
0.25
0.20
0.15
0.10
0.05
0
0.4
0.3
0.2
0.1
0
0.8
0.6
0.4
0.2
0
0.4
0.2
0.10 0.8
0.6
0.4
0.2
0
0.8 0.9 1.0 0.5
0.4
0.3
0.2
0.1
0
0.9
0.8
0.7
0.6
0.5
0.8
0.7
0.6
0.5
0.6
0.4
0.2
0
0.08
0.06
0.04
0.02
0
0
0.08
0.04
0
0.12 0.55 0.12
0.15
0.06
0.05
0.04
0.03
0.02
0.01
0
0.10
0.05
0
0.10
0.08
0.06
0.04
0.02
0
0.45
0.35
0.25
0.15
0.10
0.08
0.06
0.04
0.02
0
HLA-DQB1 rs3891176 (C>A)
Coeliac disease susceptibility
CCR5-Δ32 rs333 (+>Δ32)
HIV resistance
BNC2 rs12350739 (G>A)
Light hair or skin tone
TSBP1 rs7775397 (T>G)
Coeliac disease susceptibility
*TLR1 rs5743618 (A>C)
Leprosy resistance
GCKR rs1260326 (C>T)
Hyperlipidaemia or gout susceptibility
SERPINA1 rs17580 (T>A)
α1-Antitrypsin deficiency
SERPINA1 rs28929474 (C>T)
α1-Antitrypsin deficiency
PTPN22 rs2476601 (G>A)
Hypothyroidism susceptibility
HFE rs1799945 (C>G)
Haemochromatosis
*IL23R rs11209026 (G>A)
Crohn’s disease resistance
*CYP1A2 rs2472304 (G>A)
Blood pressure (decrease)
*NADSYN1/DHCR7 rs2276360 (G>C)
Vitamin D levels (increase)
*ADH1B rs1229984 (C>T)
Lower risk for alcoholism
ABCG2 rs2231142 (G>T)
Gout susceptibility
APOE rs7412 (C>T)
Hyperlipidaemia
HLA-DQB1 rs1130398 (T>C)
Coeliac disease susceptibility
HLA-DRB1 rs660895 (A>G)
Rheumatoid arthritis susceptibility
GYPA rs7658293 (C>A)
MNS blood group N (increase)
DUOX2 rs57659670 (C>T)
Ferritin level (increase)
*SLC22A4 rs1050152 (C>T)
Crohn’s disease susceptibility
*SLC45A2 rs16891982 (C>G)
Light hair or skin tone
*TYR rs1042602 (C>A)
Light hair or skin tone
*TYR rs1126809 (G>A)
Light hair or skin tone
SLC24A5 rs1426654 (G>A)
Light hair or skin tone
CFTR-ΔF508 rs113993960 (+>Δ3)
Cystic fibrosis
MC1R rs1805008 (C>T)
Light hair or skin tone (red hair)
MC1R rs1805007 (C>T)
Light hair or skin tone (red hair)
DSTYK rs3851294 (G>A)
Light hair or skin tone
KITLG rs12821256 (T>C)
Light hair or skin tone
RALY/ASIP rs62209647 (G>C)
Light hair or skin tone
*ABO rs8176746 (G>T)
Blood type B (increase)
*TYK2 rs34536443 (G>C)
Tuberculosis susceptibility
*HLA-DRB1 rs3135388 (G>A)
Multiple sclerosis susceptibility
HFE rs1800562 (G>A)
Haemochromatosis
TCHH rs11803731 (A>T)
Straight hair (increase) or baldness (increase)
0.20
0.15
0.10
0.05
0
8,000 4,000 0
Time before present (years)
Derived allele frequencyDerived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
8,000 4,000 0
Time before present (years)
0.8 0.7 0.75 0.08 0.12 0.10
0.08
0.06
0.04
0.02
0
0.10
0.08
0.06
0.04
0.02
0
0.06
0.04
0.02
0
0.12
0.35 0.30
0.25
0.20
0.15
0.10
0.05
0
0.30
0.25
0.20
0.15
0.10
0.05
0.08
0.04
0
0.65
0.55
0.45
0.35
0.6
0.5
0.4
0.3
0.2
0.1
0.6
0.4
0.2
0.6 0.06
0.05
0.04
0.03
0.02
0.01
0
0.5
0.4
0.3
0.2
0.1
0
0
Derived allele frequencyDerived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
Derived allele frequency
0.06
0.05
0.04
0.03
0.02
0.01
0
s = 4.4%
|X| = 12.37
> 0.99
s = 2.7%
|X| = 10.44
> 0.99
s = –0.9%
|X| = 9.60
> 0.99
s = 2.3%
|X| = 5.62
> 0.99
s = –1.9%
|X| = 7.15
> 0.99
s = 4.2%
|X| = 12.57
> 0.99
s = –2.1%
|X| = 6.60
> 0.99
s = 2.7%
|X| = 6.23
> 0.99
s = –1.3%
|X| = 2.54
> 0.11
s = 2.3%
|X| = 6.63
> 0.99
s = 1.6%
|X| = 6.49
> 0.99
s = 1.3%
|X| = 7.43
> 0.99
s = 1.9%
|X| = 6.54
> 0.99
s = 1.9%
|X| = 6.90
> 0.99
s = 1.2%
|X| = 6.43
> 0.99
s = 0.6%
|X| = 6.14
> 0.99
s = 5.6%
|X| = 6.66
> 0.99
s = 2.4%
|X| = 13.15
> 0.99
s = 1.7%
|X| = 13.37
> 0.99
s = 1.0%
|X| = 9.62
> 0.99
s = 1.4%
|X| = 4.88
= 0.78
s = 0.0%
|X| = 0.07
< 0.01
s = 1.0%
|X| = 7.68
> 0.99
s = –0.9%
|X| = 6.48
> 0.99
s = –0.9%
|X| = 10.54
> 0.99
s = 1.3%
|X| = 10.74
> 0.99
s = 1.8%
|X| = 15.99
> 0.99
s = 0.9%
|X| = 5.08
= 0.87
s = 0.9%
|X| = 5.87
> 0.99
s = 2.3%
|X| = 8.84
> 0.99
s = 0.9%
|X| = 9.81
> 0.99
s = 1.0%
|X| = 10.41
> 0.99
s = 2.0%
|X| = 11.64
> 0.99
s = 0.5%
|X| = 5.03
= 0.84
s = 1.6%
|X| = 4.40
= 0.57
s = 2.5%
|X| = 5.16
= 0.90
s = –2.3%
|X| = 5.25
= 0.93
s = –0.0%
|X| = 0.31
= 0.01
s = 0.4%
|X| = 1.27
= 0.03
s = –0.5%
|X| = 3.52
= 0.27
s = 2.1%
|X| = 2.25
= 0.08
s = –0.6%
|X| = 3.72
= 0.33
b c d e f
g h i j k l
m n o p q r
s t u v w x
y z aa ab ac ad
ae af ag ah ai aj
Fig. 3 | Gallery of single-variant allele frequency trajectories. a–aj, Each
panel displays the derived allele frequency trajectory (solid line) over time
for a variant (uncorrected for structure), along with selection coefficient (s),
selection statistic (X) and posterior probability of selection (π). The shaded
regions represent 95% CIs around the estimated trajectory. The circles
represent frequencies in Western hunter–gatherers (green; n%=%131), early
European farmers (orange; n%=%452) and steppe pastoralists (pink; n%=%293); the
error bars indicate 95% CIs. The highlighted variants are not necessarily those
with the strongest signals and may include negative results or variants in
partial linkage disequilibrium. We highlight them here because of their
biological interest and because they speak to long-standing debates. For
panels d–f,ag,ai,aj, separate analyses are shown for transects before (blue)
and after (red) a manually selected peak (marked by a black dashed line). In
cases where π >%99%, the CI is shaded blue (or blue before and red after the
split); otherwise, the shading is grey. Variants reported in other ancient DNA
studies are marked with an asterisk.