Why G25, Admixture Models and qpAdm All Struggle to Tell Celts and Germanics Apart, And What Actually Works

If you have Northwestern European ancestry and have tried modeling your heritage with G25 coordinates, you have probably noticed something frustrating: the tool often cannot clearly separate Celtic from Germanic ancestry. A Breton may model as partly "Germanic," a Dutch person as partly "Celtic," and an Englishman might get wildly different Celtic-to-Germanic ratios depending on which calculator they use. This is not a bug - it reflects a deep biological reality. Celtic and Germanic populations share so much of their genetic history that common allele frequencies alone struggle to tell them apart. Here is why, and what tools actually work.

1. A Shared Deep Ancestry: Bell Beakers to Bronze Age

To understand why modern genetic tools struggle with the Celtic-Germanic boundary, we need to look back approximately 4,500 years to the Bell Beaker phenomenon. Between 2800 and 2000 BCE, populations carrying substantial Steppe ancestry - ultimately derived from Yamnaya-related groups via the Corded Ware culture - swept across Northwestern Europe. This migration replaced approximately 90% of Britain's gene pool (Olalde et al., 2018) and caused massive Y-chromosome turnovers across the continent, with R1b-M269 replacing the pre-existing Neolithic paternal lineages almost entirely.

The crucial point is that both Celtic-speaking and Germanic-speaking populations of the Iron Age descended from these same Bell Beaker populations. The future "Celts" and "Germans" shared the same three ancestral components in broadly similar proportions: Western Hunter-Gatherer (WHG) ancestry, Early European Farmer (EEF) ancestry, and Steppe pastoralist ancestry. The populations that would later become the Hallstatt Celts, the Atlantic Celts of Britain and Armorica, the Belgae of northern Gaul, and the Germanic tribes of the North Sea littoral all emerged from a single post-Beaker continuum stretching from the Atlantic coast to southern Scandinavia.

The Atlantic Celtic populations - the ancestors of the Irish, Welsh, Bretons, and Galicians - were characterized by particularly high frequencies of R1b-L21 (a downstream subclade of R1b-P312), which likely expanded through founder effects during the Bronze Age. Meanwhile, the Germanic populations that emerged around the western Baltic and North Sea coasts were characterized by R1b-U106 (the "sister" branch of P312 under L11) alongside I1-DF29 and R1a-Z284, haplogroups that had been present in Scandinavia since the Corded Ware and Funnelbeaker periods. But autosomally, these groups remained remarkably similar because they drew from the same underlying Bell Beaker gene pool.

2. Why G25 Cannot Separate These Populations

The Global25 (G25) system, created by Davidski of the Eurogenes blog, is based on Smart PCA - a principal component analysis of genome-wide SNP data. It reduces each individual's autosomal profile to 25 principal components, which capture the major axes of human genetic variation worldwide. G25 has proven to be an extraordinarily powerful tool for distinguishing populations that differ substantially in their ancestral proportions: it can clearly separate Europeans from East Asians, Northern Europeans from Southern Europeans, and even distinguish Sardinians from mainland Italians.

However, G25 fundamentally relies on common allele frequency differences - that is, it captures variation at SNP positions where different populations carry different alleles at measurably different frequencies. And this is precisely where the Celtic-Germanic problem arises. Because both groups descend from the same Bell Beaker substrate, they share the vast majority of their common alleles at nearly identical frequencies. The genetic signal that differentiates an Irish person from a Danish person in G25 space is largely driven by:

• Slightly higher WHG (Western Hunter-Gatherer) ancestry in Atlantic Celtic populations, reflecting deeper Mesolithic substrate survival in the western fringes
• Slightly higher Scandinavian Hunter-Gatherer (SHG) ancestry in Germanic populations, inherited from the pre-Indo-European inhabitants of Scandinavia (associated with haplogroups I1, I2a-M223)
• Varying levels of additional EEF admixture through Hallstatt-era continental Celtic migrations into Britain during the Iron Age, which shifted insular populations slightly toward higher farmer ancestry

These differences are real but extremely subtle in the context of global human variation. On a standard G25 PCA plot, the Irish, English, Dutch, Danes, and Norwegians cluster so tightly together that the distances between them are dwarfed by the distances to, say, Greeks or Finns. When you feed these coordinates into an admixture calculator, the algorithm is forced to partition tiny differences into discrete "Celtic" and "Germanic" components, leading to unstable results that shift dramatically depending on which reference populations and source proxies are used.

3. G25 Admixture Modeling vs. qpAdm: Two Approaches, Same Wall

When people use G25 coordinates to explore their ancestry, they typically run them through an admixture calculator - a tool that decomposes their genetic profile into a weighted mix of reference populations. But how does this actually work, and why does it fail specifically at the Celtic-Germanic boundary? Understanding the mechanics of both G25 admixture modeling and the academic gold-standard tool qpAdm reveals why this problem is so stubborn.

How G25 Admixture Modeling Works

G25 admixture modeling (as implemented in Vahaduo and ExploreYourDNA calculators) uses a straightforward mathematical principle: it finds the combination of reference populations whose weighted average G25 coordinates best match your own coordinates. The algorithm minimizes the Euclidean distance (or sometimes Chebyshev distance) between the modeled mix and the target individual. For example, if a calculator uses "England_Iron_Age" and "Denmark_Iron_Age" as sources, it will find what percentage of each source produces coordinates closest to yours.

This works beautifully when the reference populations are genetically distinct. Modeling a Sicilian as a mix of "Anatolian Neolithic" + "Steppe Pastoralist" + "WHG" produces robust, stable results because these three sources occupy very different regions of G25 space, and the algorithm can confidently assign proportions. Each source contributes a unique "pull" in a distinct direction across the 25 dimensions.

The Celtic-Germanic case is fundamentally different. When "England_IA" (Celtic) and "Denmark_IA" (Germanic) are offered as sources, their G25 coordinates differ by only 0.005, 0.02 on most dimensions. The algorithm is trying to distinguish between two points that are virtually on top of each other in 25-dimensional space. Mathematically, this creates an ill-conditioned system: many different combinations of the two sources produce nearly identical fit distances. A model of 70% Celtic / 30% Germanic and one of 40% Celtic / 60% Germanic may differ in total fit by less than 0.001 - well within the margin of noise in G25 coordinates themselves.

How qpAdm Works - and Why It Hits the Same Wall

qpAdm is part of the ADMIXTOOLS suite developed by David Reich's laboratory at Harvard, and it is the gold-standard method used in academic ancient DNA papers for formal admixture modeling. Unlike G25's Euclidean distance approach, qpAdm uses f-statistics - specifically f4-statistics - to test whether a target population can be modeled as a mixture of specified source populations, relative to a set of "outgroup" or "right" populations.

The principle behind f4-statistics is elegant: they measure the excess of shared alleles between populations beyond what is expected by chance, using the formula f4(A, B; C, D) = E[(pA − pB)(pC − pD)], where p represents allele frequencies. If the f4-statistic deviates significantly from zero, it indicates that A shares more alleles with C than B does - evidence of admixture or shared ancestry beyond the tree-like model. qpAdm constructs a matrix of f4-statistics between the target population, the proposed source populations, and multiple outgroup populations, then tests whether the observed pattern is consistent with the proposed admixture model.

qpAdm has several advantages over G25 modeling. It uses the full genome-wide SNP data (typically 500,000, 1.2 million SNPs) rather than 25 compressed dimensions. It explicitly tests whether a model is statistically valid through a p-value (models with p < 0.05 are rejected as inadequate fits). And it uses outgroup populations to "triangulate" the ancestry sources, providing a more robust framework than simple distance minimization.

However, even qpAdm struggles with the Celtic-Germanic question, for a reason that goes deeper than methodology: the outgroup populations cannot "see" the difference either. f4-statistics work by detecting differential allele sharing with outgroups. If an English person shares slightly more alleles with an East Asian outgroup than a Dane does, this might signal different levels of WHG ancestry (which does show some ancient East Asian affinity). But these signals are extraordinarily weak between Celtic and Germanic populations precisely because their ancestral components are so similar. The f4-statistics approach zero, and the resulting models have wide confidence intervals.

Choosing the Right Outgroups: A Critical (and Often Overlooked) Step

One area where qpAdm offers a real advantage is in outgroup selection. The choice of "right" populations in a qpAdm model critically determines what the model can resolve. For the Celtic-Germanic question, the most informative outgroups are those that have differential affinity to the subtle ancestral differences between the two groups:

• Scandinavian Hunter-Gatherers (SHG) - more related to Germanic populations via the pre-Bronze Age Scandinavian substrate
• WHG (Western Hunter-Gatherers like Loschbour) - slightly more relevant to Atlantic populations where Mesolithic ancestry persisted longer
• Anatolian Neolithic farmers - captures the EEF gradient (higher in continental Celts via Hallstatt influence)
• Yamnaya / Steppe populations - captures the pastoralist component, which is broadly similar but can vary slightly
• East Asian populations (e.g., Han Chinese) - a "deep" outgroup that can detect subtle differences in ancient Eurasian admixture layers

The Gretzinger et al. (2022) Anglo-Saxon study succeeded partly because they carefully selected outgroups and used ancient reference populations from both sides of the North Sea, rather than modern ones (which have been further blurred by centuries of post-Iron Age migration). They also used multiple complementary approaches - not just qpAdm, but also supervised ADMIXTURE, PCA projection, and f4-ratio estimation - to cross-validate their findings. This multi-method approach is essential when dealing with closely related sources.

G25 vs. qpAdm: When to Use Which

Table: Comparison of G25 admixture modeling and qpAdm. Both methods are powerful but both struggle when source populations are genetically near-identical, as Celtic and Germanic Iron Age groups are.

4. Davidski's Celtic vs. Germanic PCA: A Specialized Solution

Recognizing this limitation, Davidski developed a specialized PCA view in 2018 specifically designed to maximize the separation between Celtic and Germanic populations. Unlike the standard G25 PCA, which captures global variation, the Celtic vs. Germanic PCA is built from a carefully selected set of Northwestern European populations - both modern and ancient - and the principal components are extracted specifically to highlight the axes of variation that distinguish these closely related groups.

The result is remarkably effective for a PCA-based tool. On this specialized plot, Ireland sits at one extreme, Scandinavia at the other, and populations like the English, Scottish, Dutch, and French distribute along a cline that reflects their historical mixture of Celtic and Germanic ancestry. Crucially, it also places ancient samples - Bell Beaker, Iron Age British, Anglo-Saxon, Viking - in positions that make historical sense.

What makes this PCA work where standard G25 does not? By restricting the input populations to Northwestern Europeans, Davidski effectively "zooms in" on the tiny fraction of genetic variation that differentiates these groups - variation that is swamped by larger signals (like the North-South European cline) in a global PCA. The tool captures recent genetic drift that has accumulated since the Bronze Age, including the distinct Scandinavian hunter-gatherer signature in Germanic populations and the higher EEF signal in populations influenced by Hallstatt-era expansions.

However, it is important to note that the Celtic vs. Germanic PCA still has limitations. As forum users on Anthrogenica observed, Northern Dutch individuals can plot very close to - or even within - Scottish clusters, reflecting their shared Bell Beaker deep ancestry rather than recent Celtic admixture. This underscores that even with a specialized tool, the autosomal signal separating these populations remains thin. The PCA captures tendencies and averages, but individual variation can easily cross the Celtic-Germanic "boundary."

5. Y-DNA Haplogroups: The Most Effective Differentiator

While autosomal DNA struggles with the Celtic-Germanic boundary, uniparental markers - particularly Y-chromosome haplogroups - offer far more discriminating power. This is because Y-DNA follows strict patrilineal inheritance with no recombination, accumulating unique mutations over time that act as permanent markers of paternal lineage. Different founder effects and demographic expansions within Celtic and Germanic societies concentrated specific Y-DNA lineages in each population, creating much clearer genetic signatures than autosomal admixture.

The power of Y-DNA for this specific question is illustrated dramatically by the case of England. While G25 coordinates for modern English individuals overlap heavily with both Irish and Dutch reference populations, Y-DNA haplogroup frequencies tell a clear story: eastern England carries approximately 20-35% R1b-U106 and 10-20% I1 (both Germanic markers), proportions that drop sharply in Wales, Scotland, and Ireland where R1b-L21 dominates at 60-80%+. This creates a much cleaner signal than autosomal admixture can provide.

Table 1: Approximate Y-DNA haplogroup frequencies in selected Northwestern European populations. Note how haplogroup frequencies provide much sharper differentiation than autosomal admixture. Sources: Eupedia compilations from multiple published studies; Capelli et al., 2003; Leslie et al., 2015.

6. Rare Alleles and IBD: The Future of Fine-Scale Differentiation

If common allele frequencies (G25) are too blunt and Y-DNA is limited to a single patrilineal thread, is there a middle ground that could offer autosomal-level resolution for closely related populations? The answer lies in rare alleles and Identity-by-Descent (IBD) segments - two related approaches that are now at the frontier of population genetics research.

Why Rare Alleles Matter

Common alleles (those with frequencies above 5%) are ancient and widely shared across populations. They were already segregating in the ancestral populations from which both Celts and Germans descend. Rare alleles, by contrast, arose more recently through new mutations and have had limited time to spread. Because they are young, they tend to be geographically restricted - confined to the specific population in which the mutation occurred and its immediate descendants. A rare variant that arose in Iron Age Denmark, for example, may be shared among modern Danes, Swedes, and English (via Anglo-Saxon migration) but largely absent in Ireland or Brittany. Conversely, a rare mutation originating in Bronze Age Ireland may be widespread in modern Irish and Highland Scots but almost absent in the Netherlands.

This makes rare alleles extraordinarily powerful for differentiating closely related populations that common allele frequencies cannot tell apart. Research by Baye et al. (2011) and Tennessen et al. (2012) has shown that rare variants delineate fine-scale population structure with much greater resolution than common variants. A 2020 study from bioRxiv demonstrated that IBD segments enriched for rare variants could detect relatedness and population divergence far beyond what standard IBD methods achieve.

IBD Segments: Shared Ancestry in Action

Identity-by-Descent analysis detects long stretches of DNA that two individuals share because they inherited them from a common ancestor. The length of these shared segments is inversely proportional to the age of the common ancestor: longer segments indicate more recent shared ancestry. Ralph and Coop (2013) used this approach to map the common ancestry of European populations, finding that virtually all Europeans share common ancestors within the last 1,000 years, but the density and length of shared IBD segments varies systematically by geography and history.

For the Celtic-Germanic question, IBD analysis holds particular promise because it can detect the specific population-level sharing patterns that resulted from historical migrations. English people, for instance, share significantly more long IBD segments with modern Dutch, Germans, and Danes than the Welsh or Irish do - a direct reflection of Anglo-Saxon and Viking-era migration. This signal is detectable even when standard autosomal admixture models produce ambiguous results.

Why These Methods Are Not Yet Widely Available

Despite their potential, rare allele and IBD-based approaches face significant practical barriers. Processing rare variants requires whole-genome sequencing data rather than the SNP arrays used by commercial testing companies and G25. The computational demands are enormous: detecting IBD segments across hundreds of thousands of individuals requires specialized algorithms (like hap-IBD or GERMLINE) and substantial computing infrastructure. Tools like HapFABIA, designed specifically to identify very short IBD segments characterized by rare variants, push these computational requirements even further.

For the amateur genetic genealogy community, this means that rare allele analysis remains largely inaccessible. G25 coordinates can be generated from standard 23andMe or AncestryDNA files and analyzed on a personal computer in seconds. A comprehensive rare variant analysis of the same individual against thousands of reference populations might take hours or days of computation on specialized hardware. Until sequencing costs drop further and efficient consumer-facing tools are developed, this powerful approach will remain in the domain of academic research.

7. What the Anglo-Saxon Study Teaches Us

The landmark 2022 study by Gretzinger et al. (Nature) on the Anglo-Saxon migration beautifully illustrates both the challenge and the solution. The researchers sequenced 460 medieval Northwestern Europeans, including 278 individuals from England, and attempted to disentangle local Celtic British ancestry from incoming Continental Northern European (Germanic) ancestry.

Crucially, the study explicitly acknowledged that separating these two sources was difficult because of their shared deep ancestry. On standard PCA plots, Iron Age British and Continental North Sea populations clustered closely together. The researchers overcame this by using a combination of approaches: F-statistics (f4 and FST) to identify subtle allele sharing differences, supervised ancestry modeling using ancient reference populations from both sides of the North Sea, and careful selection of populations that maximized the Celtic-Germanic contrast.

Their findings were striking: early medieval eastern England had approximately 76% Continental Northern European ancestry, meaning the Anglo-Saxon migration involved mass population movement, not merely elite dominance. But in modern England, this proportion has diluted to roughly 25-47%, reflecting subsequent centuries of mixing. The study also identified a third ancestry component - Continental Celtic / French Iron Age - that contributed independently to some regions of England, likely representing Hallstatt-era or Roman-era movements.

What the Gretzinger study demonstrates is that professional-grade tools can distinguish Celtic from Germanic ancestry, but it requires: carefully curated ancient reference populations, multiple complementary statistical methods, and access to full genome-wide data rather than PCA-compressed coordinates. These are resources that G25 calculators, powerful as they are for many questions, simply cannot replicate.

8. Practical Implications for Your DNA Results

If you are of Northwestern European descent and want to understand your Celtic vs. Germanic heritage, here is a practical framework for interpreting what different tools tell you:

G25 Admixture Calculators: Use these to understand your broad ancestral profile - your mix of WHG, EEF, and Steppe ancestry, and your position within the European genetic landscape. Do not take Celtic vs. Germanic percentage breakdowns literally. If a calculator tells you that you are "55% Gaul and 45% Saxon," treat this as an indication that you fall somewhere in the middle of the Northwestern European cline, not as a precise ancestry report.

Celtic vs. Germanic PCA (Eurogenes): This is currently the best PCA-based tool for this specific question. It will show you where you fall on the Celtic-Germanic spectrum relative to reference populations. If you cluster near the Irish, you probably have predominantly insular Celtic ancestry. If you cluster near the Danes, your ancestry is predominantly Germanic. Most English, Dutch, and northern French individuals will fall somewhere in between. Use the Vahaduo Celtic vs. Germanic PCA view to visualize your position.

Y-DNA Testing: If you are male and want the sharpest answer to the Celtic vs. Germanic question for your direct paternal line, a Y-DNA test (such as FTDNA's BigY-700) is invaluable. R1b-L21 subclades point to Celtic British/Atlantic ancestry. R1b-U106 subclades or I1-DF29 point to Germanic ancestry. R1b-U152 may suggest continental Celtic (Hallstatt/La Tène) origins. Of course, this tells you about only one ancestral line out of thousands, but it is the single most informative marker available.

Commercial DNA Tests (23andMe, AncestryDNA): These companies use proprietary haplotype-based methods that actually have better resolution than G25 for distinguishing recent Celtic vs. Germanic ancestry, because they compare long haplotype segments rather than just allele frequencies. If 23andMe separates your ancestry into "British & Irish" vs. "Scandinavian" or "French & German," these categories, while imperfect, do carry meaningful information. AncestryDNA's regional breakdowns (e.g., "Scotland" vs. "England & Northwestern Europe") reflect similar haplotype signals.

9. G25 Coordinates: Celtic and Germanic Iron Age Populations

Below are G25 scaled coordinates for key Celtic and Germanic reference populations from the Iron Age and early medieval period. These can be used with Vahaduo and ExploreYourDNA calculators to explore your own positioning. Note how close these coordinates are - this is precisely why admixture modeling between them is so challenging.

Iron Age & Early Medieval Reference Populations