henhao
Sign in
GenomicsbioRxiv
Preprint

Local ancestry inference identifies robust evidence of selection in Neolithic Europe

Georgia MiesiD, Iain MathiesoniD

Abstract

During the European Neolithic, migrating Anatolian farmers admixed with local hunter-gatherers, coinciding with major shifts in diet, environment, and lifestyle that imposed strong selective pressures. Local ancestry inference is widely used to detect selection following admixture, but most methods were developed and validated on present-day populations. Their performance in ancient DNA - where reference panels are smaller, data are sparser, and admixture is more ancient - remains unresolved. We benchmark eight local ancestry inference methods on 176 imputed Neolithic genomes. While individual-level ancestry estimates are highly correlated across methods, inferred tract lengths and admixture time estimates vary by an order of magnitude. Overall, we recommend Gnomix or RFMix for general use. We also investigated our ability to detect natural selection using LAI. Integrating results across methods and replicating across methods and in two independent datasets (n=378 and 1,121) we identify a robust ancestry deviation at FADS1/2, consistent with adaptation on metabolism. We also identify IRAK4 (innate immunity) as a candidate locus, but with less consistent signal across methods. Finally, we replicate previous reports of excess hunter-gatherer ancestry at the HLA, but these results are inconsistent across methods and suggest that they may be affected by bias in local ancestry inference. Our findings demonstrate that while local ancestry inference recovers biologically meaningful signals in ancient genomes, results can be sensitive to the methods used for inference, particularly in complex regions like the HLA. Method choice critically influences inferred ancestry patterns and selection signals, underscoring the importance of multi-method validation.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
2
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 2 (17 Aug 2026) · CC BY 4.0

Local ancestry methods on 176 imputed Neolithic genomes disagree on tract lengths; replicated meta-analysis supports selection at SLC24A5 and FADS1/2, while HLA signals look artefactual.

Local ancestry inference (LAI) has been used to scan ancient genomes for post-admixture selection, yet the methods were built and validated on present-day data with large, well-matched reference panels, unlike sparse, low-coverage ancient DNA. This work benchmarks LAI methods on a well-characterized Neolithic admixture and tests candidate loci across methods and two independent datasets, separating reproducible selection signals from method artefacts.

Key results

  1. ShownOf ten autosomal regions passing a multi-method meta-analysis, five replicated in an independent pseudohaploid dataset and two of them (HLA, SLC24A5) also in a second imputed dataset; the other three are FADS1/2 and regions containing PER3 and IRAK4.

    Discovery: 10 regions with |Z| > 3 from a covariance-adjusted meta-analysis of per-method Z-scores (51-SNP windows). Replication dataset 1 (n=378, Ancestry HMM): 5 replicated (P < 0.005, within 2 Mb of the discovery lead SNP). Replication dataset 2 (n=1,121, RFMix): 2 replicated. Neither of 2 X chromosome hits (|Z| > 3) replicated at P < 0.025. · Meta-analysis identifies five regions with…; Ancestry deviation analysis; Replication analyses

  2. SuggestedAncestry deviation at SLC24A5 and FADS1/2 is concordant across methods and smaller in early, low hunter-gatherer-ancestry individuals, as expected under post-admixture selection; PER3 and IRAK4 stay high in that subset and HLA is inconsistent, pointing to misclassification.

    Deviations compared between full samples and low-ancestry subsets: 27 discovery individuals with <10% hunter-gatherer ancestry (mean age 7,127 years before present), 30 in replication dataset 1 and 692 in replication dataset 2. HLA deviation also differed across methods and replication datasets. · Meta-analysis identifies five regions with…; Replication analyses; Discussion

  3. ShownInferred tract lengths, and hence admixture dates, differ by more than an order of magnitude between methods; RFMix gives the most plausible distribution, Mosaic is next, and Ancestry HMM, AncestralPaths and simpLAI give tracts that are too short.

    TRACTS single-pulse fits: 10.3 (RFMix), 41 (simpLAI) and 6.5 (Mosaic) generations, against a maximum possible admixture time of about 3,000 years (about 100 generations) and earlier linkage-disequilibrium estimates of 10-30 generations. Ancestry HMM and AncestralPaths tracts fit admixture over 100 generations ago; Mosaic gives near whole-chromosome tracts in many individuals. · Local ancestry methods disagree about the…; Tract length analysis

  4. ShownWith all 48 hunter-gatherer and 7 farmer references, every LAI method gave higher individual hunter-gatherer ancestry than qpAdm; subsampling hunter-gatherers to 3 or 7 and keeping calls with posterior > 0.9 improved agreement.

    qpAdm mean hunter-gatherer ancestry 0.20 (range -0.10 to 0.41); ADMIXTURE 0.158 (range 0 to 0.39). The overestimate was absent in 1000 Genomes African Americans (ASW, n=61) with 7 YRI and 48 CEU sources. Imputation left qpAdm and ADMIXTURE estimates unchanged but substantially shifted Ancestry HMM, the only LAI method run on unimputed data. · Local ancestry inference is sensitive to sample…; Local ancestry inference on autosomes; Sensitivity analyses (source size, imputation…

  5. ShownMethods agree only modestly on where hunter-gatherer tracts lie along the genome, with AncestralPaths least concordant with the others, although every method's individual global ancestry estimates correlate with qpAdm.

    Pairwise Pearson correlations of population-level hunter-gatherer ancestry in 100 kb bins ranged 0.08 to 0.5; per-individual genome-wide correlations ranged 0.15 to 0.43 (mean 0.32), even after source-size optimisation and posterior filtering. · Local ancestry methods disagree about the…; Benchmarking local ancestry inference methods

Methods, data and reproducibility

Four LAI methods (RFMix v1.9, Mosaic, simpLAI, Ancestry HMM) plus published AncestralPaths calls were applied to 176 imputed admixed Neolithic genomes with Western hunter-gatherer and Anatolian farmer references and compared with qpAdm, ADMIXTURE and TRACTS fits. Per-method Z-scores in 51-SNP windows were meta-analysed with a covariance adjustment and tested in two independent datasets.

  • Discovery set: imputed, phased admixed Neolithic genomes from the Allentoft et al. (2024) release (GLIMPSE, 1000 Genomes phase 3 panel), filtered to 572,223 autosomal SNPs · human · n = 176
  • Source references: Western hunter-gatherers and Anatolian farmers from the same release (hunter-gatherers subsampled to 7, 3 or 1 in tests) · human · n = 48 hunter-gatherers; 7 farmers
  • X chromosome (non-PAR) data re-imputed with GLIMPSE from published reads of the discovery and source individuals · human · n = 176 admixed; 7 farmers; 48 hunter-gatherers
  • Replication dataset 1: pseudohaploid admixed Neolithic individuals from the Allen Ancient DNA Resource (AADR) 1240k set, discovery samples removed; 98 hunter-gatherer and 48 farmer sources · human · n = 378
  • Replication dataset 2: imputed shotgun and capture admixed Neolithic individuals from Akbari et al. (2026); 32 farmer and 26 hunter-gatherer sources · human · n = 1,121 · doi:10.7910/DVN/7RVV9N
  • 1000 Genomes Project African American (ASW) individuals with Yoruba (YRI) and European (CEU) sources, used to test source-size imbalance · human · n = ASW 61; YRI 7; CEU 48
Methods & tools
RFMix V1.9, Mosaic, SimpLAI, Ancestry HMM, AncestralPaths, Flare (failed on this sample), qpAdm, ADMIXTURE, TRACTS, Tractor, GLIMPSE, PLINK 1.90
Code
https://github.com/gmies/LAI_neolithic/
Data & accessions
https://zenodo.org/records/19684580https://erda.ku.dk/archives/917f1ac64148c3800ab7baa29402d088/published-archive.htmlhttps://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/7RVV9N
Software & environment
Stated versions: RFMix V1.9, plink 1.90 Beta 6.18, bcftools 1.20; simpLAI run with documented parameters (-s 1e6, -I 5e5, -n 2000, -m 1000, -t 5); GRCh37 (hg19) coordinates.
Reproducibility gaps
No container, dependency file or compute requirements described; AncestralPaths results are reused published calls rather than rerun; the X chromosome SNP count after filtering is given as 160,150 in one section and 157,610 in another.
Limitations (6)
  • No ground truth exists for these ancient genomes, so methods are judged indirectly: against allele-frequency estimates (qpAdm, ADMIXTURE), known limits on admixture timing, and replication in other datasets. · stated by the authors
  • Every method is affected by unbalanced source sizes, the makeup of the reference panel and uncertain ancestry calls; the authors expect such biases to grow in sparsely sampled ancient populations with small, unbalanced panels. · stated by the authors
  • Local ancestry deviation can reflect differentiation already present between the source populations, not only selection after admixture. · stated by the authors
  • Deviations at PER3 and IRAK4 stay high in individuals with <10% hunter-gatherer ancestry and HLA results vary across methods, so these may be misclassification (technical artefact or mismatched sources); the authors judge HLA local ancestry too unreliable for selection or trait inference. · stated by the authors
  • The selection-versus-misclassification call rests on comparing deviations between full samples and small low-ancestry subsets (27 discovery individuals) with no reported statistical test, so the split between loci is qualitative. · noted in this summary
  • Each replication dataset uses a single method: Ancestry HMM on pseudohaploid data (which the paper finds gives implausibly short tracts and shifts with imputation) and RFMix, so replication does not itself test robustness across methods. · noted in this summary

Who should read:Read this if you use local ancestry inference on ancient or small, unbalanced-reference datasets to scan for post-admixture selection and want to know which methods and checks hold up.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISDavy '22Pearson '23Bougiouri '26Chintalapat '22Lerga-Jaso '23Colbran '26
71% semantic matchEvolutionary Biology
T. Davy, D. Ju, I. Mathieson+1 more·bioRxiv·Sep 5, 2022