About

Methods for individual traits are named systematically after their sources. The following descriptors indicate how they are used: Exploratory admits a broader but less certain evidence base; Strict uses the same sources and model with tighter quality gates; Minimal denotes a deliberately small comparison model; and PGS denotes a published polygenic score. All direct models use observed AADR markers without imputation. Only the Deep SNP scan explicitly evaluates positions outside the AADR panel.

Complex traits such as hair colour and skin pigmentation are analysed primarily through genome-wide association studies (GWAS). The relevant variants are usually single-nucleotide polymorphisms (SNPs): differences at individual DNA positions, each defined by one of the four bases A, T, C or G. Larger genomic structures can also affect a trait, but SNP effects are the principal basis used here and can be traced most transparently.

Genes and SNPs differ greatly in their association with a trait. In the strongly skewed effect-size distribution described here, the 5–20 largest-effect SNPs can carry more weight than thousands of individually weak associations combined. These direct models therefore evaluate a deliberately selected set of the strongest known signals rather than thousands of small effects, as a PGS does. Some traits, including height, are so polygenic that the weaker effects still make a material contribution; for those traits, PGS models are included as a benchmark.

The general procedure for these direct SNP models was as follows:

•The literature was reviewed for relevant genetic markers, and the reported effect sizes were extracted.

•Each marker was then checked for coverage in AADR. Markers absent from the panel were documented and re-evaluated against the raw sequencing data in the Deep SNP scan.

•Among the markers covered by AADR, the most robust, well-replicated and large-effect candidates were selected and combined into a transparent, trait-specific evidence index. The calculation uses published effect sizes from sources with the greatest available transferability; markers whose effect direction or weight could not be harmonised reliably were excluded.

•Unobserved loci are not treated as counter-evidence. Evidence that is too weak, non-independent or internally inconsistent is blocked so that it cannot distort the result.

Direct models, minimal models, and PGS

Genome-wide association studies (GWAS) and polygenic scores (PGS) are closely related but serve different purposes. A GWAS compares the genomes and measured traits of many modern participants to identify associated variants and estimate their effects. A PGS then combines effect estimates from such association studies into a model that assigns an aggregate score to an individual.

The Exploratory and Strict direct models use effect sizes derived in the same way. In a PGS method, by contrast, all SNPs and weights come from a jointly calibrated score. This improves internal consistency, but also introduces many weak markers whose calibration in modern populations may not transfer readily to populations that lived thousands of years ago. The source-based direct models focus on a small number of strong, transparently traceable markers; PGS models use a much broader evidence base that is more sensitive to temporal and population differences.

Lona-Durazo/Sulem (Minimal) and Morgan 2018 (Minimal) are deliberately not described as PGS models: both are small, GWAS-based direct-marker comparison models. The genuinely polygenic models are identified explicitly as Tanigawa 2022 (PGS), Privé 2022 (PGS) and Yengo 2022 (PGS).

Hysi/Morgan/Lona-Durazo (Strict)

•Selection and availability: The Strict mode uses exactly the same current source, marker, and score basis as Hysi/Morgan/Lona-Durazo (Exploratory). It must not subsequently alter markers, weights, or class boundaries.

•Quality gate: The Strict mode additionally requires greater observed effect-weight coverage, sufficient sequencing coverage, and a passing AADR quality assessment. Marker counts are diagnostic only. These requirements may block a result but cannot change the direction of an admitted score. The separate Hysi rank is reported as a consistency diagnostic and does not alter the class.

Hysi/Morgan/Lona-Durazo (Exploratory)

•Sources: Hysi et al. 2018, Morgan et al. 2018, and the raw plus COJO-filtered CanPath/Lona-Durazo basis from 2021. Hysi provides a European ordinal hair-colour axis; Morgan and CanPath provide categorical hair-colour effect sizes.

•Selection: The backbone uses only the shared Morgan/Lona-Durazo contrast of blonde versus brown+black. Effects are aligned to the same AADR allele and combined with their standard errors; significant heterogeneity triggers a random-effects weight. Brown versus black is outside the target and is fully excluded. 250 kb locus groups prevent LD double counting.

•Available: The audit covers 2,111 source rows: 258 from Hysi, 293 from Morgan, 1,543 raw Lona-Durazo rows, and 17 Lona-Durazo COJO rows. Eighty-seven unique AADR markers are active: 40 Morgan/Lona-Durazo blonde-axis markers, seven functional MC1R red markers, and 50 Hysi consistency markers; some markers belong to more than one separate axis. Every source row has a documented inclusion or exclusion reason.

•Calculation and interpretation: Blond and red are not treated as genetically identical. The MC1R red branch and Morgan/Lona-Durazo blonde axis remain separate internally; only their empirical evidence ranks are combined as alternative routes to the final Blond/Red class. Hysi standardised betas are never added to log odds because the values are not directly compatible. One symmetric dark-blond colour band marks the area near the boundary without changing binary classes or correlations.

Morgan/Lona-Durazo (Strict)

•Sources and SNPs: Identical to Morgan/Lona-Durazo (Exploratory): Morgan 2018 and CanPath/Lona-Durazo 2021, with 16 active effect rows across 15 unique SNPs. Of 62 reviewed hair markers, 28 are directly visible in AADR.

For Morgan/Lona-Durazo (strict) the source-weighted marker basis remains the same as for Morgan/Lona-Durazo (exploratory), but the release of results becomes narrower. Robust blond requires at least 15 per cent positive blond effect weight basis, a blond score of at least 0.50 and dark competition of at most 0.25. Robust red-blond follows the measured red-hair penetrance of the observed MC1R genotype (Morgan 2018) from 0.5. Brown/dark requires positive weighted dark evidence of at least 0.50; where a competing light basis is present, dark must lie at least 0.25 above the blond score. A dark statement additionally requires at least 25 per cent of the light effect weight basis to be observable. Marker and locus counts are diagnostics only. In the current run, 2,366 profiles are blond, 522 red-blond, 4,620 brown/dark, 1,675 dark blond tending dark, 2,773 dark blond tending light and 5,629 unknown.

Morgan/Lona-Durazo (Exploratory)

•Sources: Morgan et al. 2018 and CanPath/Lona-Durazo et al. 2021 provide the source-weighted hair-colour effect sizes. HIrisPlex, Guenther 2014, Zorina-Lichtenwalter 2019, Branicki 2011, Sulem 2007 and Wilde 2014 informed marker selection, allele orientation, functional interpretation and context.

•Selection: The active model contains 16 effect estimates across 15 unique SNPs. Blond/light: KITLG rs12821256 C, HERC2/OCA2 rs12913832 G, SLC24A4 rs12896399 T, TPCN2 rs3829241 A and rs35264875 T, TYR rs1042602 A, and OCA2 rs1800407 T. Red-blond: MC1R rs1805006 A, rs11547464 A, rs1805007 T, rs1805008 T, rs1805009 C and rs1805005 T. Brown/dark: SLC45A2 rs16891982 C, HERC2/OCA2 rs12913832 A, and IRF4 rs12203592 T. rs1110400 and rs2228479 remain documented as MC1R context markers but are not scored; rs28777 remains blocked because this small model cannot treat it as sufficiently independent of rs16891982. Other observed context markers are not scored unless their direction, independence and effect weight are sufficiently secure.

•Panel coverage: 62 hair-colour markers were reviewed. Of these, 29 are directly observable in AADR and 33 are currently absent. The score uses 16 effect estimates across 15 unique SNPs whose effect direction is secure and whose weights can be traced to their sources; rs1110400, rs2228479 and rs28777 are documented but blocked. The remaining observed markers are retained for documentation or context until their direction and weighting are sufficiently secure.

•Weighting: Only the strongest observed candidate within each locus group is used, preventing tightly linked markers from double-counting the same signal. Weights come from the master SNP list: Morgan odds ratios are represented as log(OR), while CanPath betas are treated as log-odds-like effect sizes; each scoring_allowed row contributes abs(effect_size_value). Each subscore is the source-weighted number of observed effect-allele copies divided by the source-weighted maximum observable evidence. Missing markers are not treated as counter-evidence. A robust blond call requires at least 15% coverage of the positive blond effect weight, a blond score of at least 0.50, and competing dark evidence no greater than 0.25. Red-blond follows the observed MC1R genotype's measured red-hair penetrance in Morgan 2018 and is assigned at a penetrance of 0.5 or above. Brown/dark requires positive weighted dark evidence and is never inferred from missing light evidence. Marker and locus counts are diagnostic only. In the current run, 2,667 profiles are blond, 522 red-blond, 2,991 brown/dark, 2,731 dark blond leaning light, 4,063 dark blond leaning dark, and 4,611 unknown.

Morgan 2018 (Minimal)

Morgan et al. 2018 on hair colour in the UK Biobank is used. This mode is not a fully fledged large PGS, but a small GWAS-oriented comparison mode built from directly observed markers.

10 unique markers from 11 effect rows are actively usable; their number is thus considerably smaller than in the Tanigawa PGS. Several markers are deliberately not used if they are not sufficiently independent, have an unclear direction, or act too weakly or contradictorily in the small model.

AADR coverage is correspondingly limited: this mode can use only the directly observed markers from the small GWAS model. In the current run, 11,286 profiles are calculated, 1,254 have a limited basis, and 5,045 are not evaluable.

Weighting follows the strength of the GWAS association, cautiously attenuated so that individual modern associations are not transferred too rigidly to ancient DNA. What is output is support for blond, dark and red, or a mixed status if no clear category dominates. The robustness is lower than with a large PGS and also lower than with robust individual markers. The Morgan mode is useful as an independent comparative indication, but not as a primary hair-colour method.

Tanigawa 2022 (PGS)

The hair-colour PGS family of Tanigawa et al. 2022 is used. It contains separate models for black, blond, brown, dark brown, light brown and red hair colour.

In the current runtime version, the Tanigawa PGS contains 6,896 unique active markers and 8,896 active score entries. By category, the following are active: 691 for black, 2,759 for blond, 1,642 for brown, 2,271 for dark brown, 866 for light brown and 667 for red.

Only candidate variants that are directly observable in AADR and can be harmonised reliably are usable. Many candidates are excluded because the AADR panel does not cover them or because their alleles cannot be matched with sufficient confidence.

The map output adopts the existing intermediate PGS categories according to fixed builder thresholds: red is only output clearly if red Z reaches at least 1.10 and lies at least 0.35 above blond Z and dark Z. Blond or dark are only output clearly if the respective Z value reaches at least 0.75 and lies at least 0.30 above the two competing axes. Dark blond appears if blond Z or dark Z reaches at least -0.15 but no clear category dominates. The lighter dark-blond side applies if blond Z minus dark Z plus 0.35 times max(light-brown Z, -1.0) yields at least 0; otherwise dark blond is displayed towards the darker side.

For a calculated result, an individual needs at least 3.5 per cent of the total absolute effect-weight basis and at least 2 per cent of the high-weight basis. For a limited result, 0.4 and 0.2 per cent respectively are sufficient. Marker count and unweighted coverage are only diagnostics. In the current run, 15,136 profiles are calculated, 1,186 limited and 1,263 not evaluable.

The Tanigawa PGS is broad and internally consistent, but it is not a forensic hair-colour verdict for ancient DNA. It is particularly useful as a comparison model. Over very long periods it has to be interpreted with caution, because many weak modern effects may be noisier or weighted differently in ancient populations.

What the genotype cannot tell us here

The light-dark axis of hair colour is largely additive: every additional light effect allele shifts the value a little, and observing individual alleles is enough for that. The red axis works differently. Red hair arises mostly when two function-reducing MC1R variants meet - including when they are two different ones, that is, as a compound pair. Whether someone had red hair therefore does not depend on the sum of their markers but on the combination of both gene copies.

That combination is precisely what cannot be established reliably in ancient genotypes. The 1240K data are pseudo-haploid: usually only a single base call is available per marker, which makes a heterozygote look like a homozygote. A compound pair can therefore neither be demonstrated nor ruled out. The GeoGens red values should accordingly be read as observed allele evidence, not as a determined phenotype: for the dominance- and recessiveness-dependent components of hair colour, the genotype permits only an imprecise inference about the actual hair colour. What remains defensible are the broad tendencies of the light-dark axis and group patterns, not a red-hair diagnosis for a single individual.

Lona-Durazo/Sulem (Strict)

•Sources and SNPs: Identical to Lona-Durazo/Sulem (Exploratory): CanPath/Lona-Durazo Data S1 and Sulem/GWAS Catalog, with five active effect rows across four unique SNPs. All 15 reviewed eye markers are visible in AADR, but only the cleanly oriented and weighted GWAS basis is scored.

With this method the GWAS-based marker basis remains the same as for Lona-Durazo/Sulem (exploratory), but proxy-like and contradictory profiles are blocked more strongly. Blue requires rs12913832 G/G as its structural main anchor, at least 3 per cent independent blue effect weight basis outside that anchor, a blue value of at least 0.50 and brown competition below 0.35. Brown/dark requires rs12913832 A/A, a brown/dark value of at least 0.50 and both blue competition and conflict competition below 0.35. Green/mixed is released only where blue and brown/dark clearly compete; direct intermediate markers contribute no mixed value of their own. In the current run, 1,696 profiles are blue, 6,150 brown/dark, 372 green/mixed and 9,367 unknown.

Lona-Durazo/Sulem (Exploratory)

•Sources: The SNP score uses a small, audited eye-colour GWAS model based on Lona-Durazo et al. 2022/CanPath Data S1 and Sulem et al. 2007/GWAS Catalog. IrisPlex/HIrisPlex-related eye markers, Chaitanya et al. 2014, and the more recent findings of Abbatangelo, Lona-Durazo, Edwards & Parra 2026 remain review and raw-signal context wherever the effect allele, effect direction, or suitability for an unstratified scoring model is not sufficiently secure.

•Selection: The SNP score contains five GWAS effect estimates across four unique SNPs: HERC2/OCA2 rs12913832 G forms the central blue-eye axis and rs12913832 A its central brown/dark counterpart; IRF4 rs12203592 T, SLC24A4 rs12896399 T and TYR rs1393350 A provide additional blue/light evidence. rs1129038, rs1800407 and rs16891982 are used by neither the Score nor the Strict model; rs1800407 also produces no direct green/mixed raw signal. Green/mixed is not an independent marker axis. It is a residual or conflict category derived from competing blue and brown/dark evidence.

•Panel coverage: All 15 reviewed eye-colour markers are observable in AADR. The score actively uses five GWAS effect estimates across four unique SNPs. Other observed effect estimates remain documented but are excluded from the score because they are proxies, depend on haplotype or stratified contexts, have an uncertain effect direction, or lack a sufficiently robust GWAS basis for this model.

•Weighting: Beta estimates are weighted as abs(beta) × 0.7; odds ratios as abs(log(OR)) × 0.7. The resulting weights are 0.885115 for each rs12913832 allele, 0.29309723 for rs1393350 A, 0.115157 for rs12203592 T, and 0.0844564 for rs12896399 T. Only the strongest observed contribution within each locus group is used. A profile is classified as blue when rs12913832 G/G is observed with a brown score below 0.45, or when the blue axis plus supporting loci reaches at least 0.55. Brown/dark is assigned when rs12913832 A/A is observed without strong competing blue or conflict evidence, or when the brown axis reaches at least 0.50. Green/mixed is assigned when competing or insufficiently broad GWAS evidence supports neither a robust blue nor a robust brown/dark call; its value is a derived conflict score without a dedicated marker axis. Profiles without sufficient positive eye-colour evidence are blocked. In the current run, 4,324 profiles are blue, 6,433 brown/dark, 880 green/mixed, and 5,948 unknown.

Simcoe/Lona-Durazo (Exploratory)

•Sources: Simcoe et al. 2021 and Lona-Durazo et al. 2022, both based on European-ancestry samples. The audit reviews 115 Simcoe rows, 401 Lona GWAS rows, and 35 additional Lona fine-mapping context rows.

•Selection: Eleven unique markers that can be harmonised directly in AADR are active, derived from 14 included source components. Reproducible Simcoe regional effects and independent Lona COJO signals are used; LD followers, unresolved palindromic variants, and findings without an accessible effect size remain excluded.

•Availability and calibration: The runtime model uses a rank-preserving internal reference calibration to Lona-Durazo/Sulem (Strict) for European samples from the first millennium CE as a geohistorical validation point. The green/middle band is deliberately very narrow. Calibration work is ongoing.

Simcoe/Lona-Durazo (Strict)

•Selection and availability: The Strict mode uses the same 11-marker score and the same internal calibration as Simcoe/Lona-Durazo (Exploratory). The genetic value and its category boundaries are not recalculated.

•Release: Exploratory and Strict use all 11 markers, the same effect weights, the same calibrated index, and the same category boundaries; Strict therefore does not alter any released value. The only difference is the evidence-release policy. Exploratory becomes evaluable from 8% observed absolute effect-weight basis, whereas Strict requires 18%. In both modes, a profile below 35% is classed as a limited basis and one at or above 35% as calculated. For AADR samples, Strict additionally requires a ‘Pass’ quality assessment, a reported sequencing coverage, and at least 0.1× coverage. If any of these conditions is missing, the profile is blocked as an insufficient basis despite observed markers. In the current run, Exploratory has 9,306 calculated, 3,312 limited, and 4,967 blocked profiles; Strict has 7,033 calculated, 1,512 limited, and 9,040 blocked profiles.

Lona-Durazo/Sulem (Minimal)

No large eye-colour PGS is currently active. This mode instead provides a small GWAS-based evidence model derived from Lona-Durazo et al. 2022, Sulem et al. 2007, and a more recent stratified source that has been reviewed for context but is not yet fully incorporated into the score.

4 unique markers from 5 effect rows are active; they are restricted to directly observable GWAS markers whose direction is unambiguous at row level. The mode is therefore called Lona-Durazo/Sulem (Minimal) and not PGS.

AADR coverage is sufficient for some individuals but remains narrow overall. In the current run, 10,846 profiles are calculated, 513 have a limited basis, and 6,226 are not evaluable.

Weighting follows the reported GWAS effect. The output is Blue or Brown/dark where the support is unambiguous enough; otherwise the profile remains unknown or limited. Its robustness is limited. Eye colour is already strongly shaped by a few main markers, above all the HERC2/OCA2 axis. The other methods are therefore usually more comprehensible and more robust than this small minimal mode.

SNP deep search

Because the AADR panel contains only the most common SNPs, a multi-day intensive computation examined the readily accessible raw genomes for additional SNPs from the known sources. The run is not yet complete and will be expanded over time as the more difficult genomes are processed.

•Source: A dedicated damage-aware scan of raw ENA BAM data at 1,250 target positions. The scan covers 8,043 ancient genomes and also reaches markers outside the AADR 1240K panel.

•Selection: The underlying eye-colour weight table contains 34 blue and brown/dark axis rows across 28 unique SNPs. After runtime harmonisation, the active artifact combines 27 weighted markers into one eye-colour axis.

•Available: In the current runtime build, the eye axis contains 6,371 evaluable profiles: 3,218 calculated, 1,273 with a limited basis, and 1,880 with an insufficient basis.

•Calculation and interpretation: Each position is counted both in full and with clipped read ends; transition markers are assessed using the less damage-prone trimmed counts. Genotypes are treated as diploid or pseudo-haploid depending on depth. The result is a weighted allele index.

What the genotype cannot tell us here

Eye colour is the least additive of the pigmentation traits shown here. It depends predominantly on the HERC2/OCA2 region, and the light variant acts largely recessively there: only two copies reliably produce blue eyes, whereas heterozygotes usually have brown eyes and merely pass the light variant on. The phenotype therefore does not follow the sum of the markers but the interplay of both gene copies at a few positions.

Ancient genotypes cannot resolve that interaction reliably. The 1240K data are pseudo-haploid: usually only one base call is available per marker, so a heterozygote cannot be distinguished from a homozygote. An observed light allele therefore demonstrates genetic potential, not blue eye colour itself. GeoGens eye values must be read as observed allele evidence rather than a determined phenotype. For traits governed by dominance and recessiveness, these genotypes permit only an imprecise inference about actual eye colour. Group means and temporal patterns are considerably more defensible than an attribution to a single individual. Homozygosity is generally assumed to have been higher in ancient populations, so the broad observation is more likely to hold, especially in earlier periods.

Liu/Jacobs (Strict)

•Sources and SNPs: Identical to Liu/Jacobs (Exploratory): six active R²-weighted markers from Liu 2015 and Jacobs 2013 within a registry of 30 skin-marker rows. Four requested direct markers are absent from the AADR 1240K panel; the remaining inactive rows stay as review context.

For Liu/Jacobs (strict) the marker basis remains the same as for Liu/Jacobs (exploratory). In addition, all low confidences, mixed dosage modes, missing quality information and sequencing coverage below 0.1× are blocked. The genetic minimum basis remains effect-weighted: 45 per cent total weight and 40 per cent anchor weight for calculated, 25 and 20 per cent respectively for limited. In the current run, 6,860 profiles are calculated and 10,725 blocked.

Liu/Jacobs (Exploratory)

•Sources: The model primarily draws on Liu et al. 2015 for direct European variation in skin pigmentation and Jacobs et al. 2013 for quantitative pigmentation axes.

•Selection: Six markers are active: HERC2/OCA2 rs12913832 A is the dark-pigmentation anchor (R² 5.376; weight 1.0); IRF4 rs12203592 T the light-pigmentation anchor (R² 3.627; weight 0.82138); TYR rs1393350 A a light-pigmentation marker from Liu (R² 0.573; weight 0.326473); OCA2 rs7495174 G a dark-pigmentation Jacobs SD marker (partial R² 0.48; weight 0.298807); ASIP rs4911442 G a light-pigmentation Jacobs saturation marker (partial R² 0.35; weight 0.255155); and SLC24A4 rs17128291 G a light-pigmentation marker from Liu (R² 0.241; weight 0.211728). Inactive markers include rs1805007, rs1805008, rs376397, rs16891982, rs10756819, rs1540771, rs4911414, rs17426596, rs1800407, rs6742078, rs12896399, rs12821256, rs1408799, rs35264875, rs1426654, rs2675345, rs2470102, rs11637235, rs8028919 and rs6602666. They are absent, belong to another axis, cannot be transferred to the current R² scale, are not sufficiently harmonised, or remain review context in this release.

•Panel coverage: The reviewed skin-pigmentation registry contains 30 entries. Six markers are directly observable and active. Four requested direct markers are absent from the AADR 1240K panel (rs183671, rs4268748, rs6059655 and rs2924567). Another 20 entries are observable or documented but not currently scored because they have not been approved for this model, or because their axis, effect direction or effect-size scale is not sufficiently compatible.

•Weighting: Weights follow the proportion of pigmentation variance explained: R² and partial R² values are square-root transformed and normalised to a weight of 1.0 for rs12913832. Lighter pigmentation is negative on the scale and darker pigmentation positive. Legacy AADR 0/2 calls are treated as half dosage for a single observed allele; missing markers are omitted from both numerator and denominator. A calculated result requires observation of at least 45% of the total effect weight and 40% of the anchor effect weight; a limited result requires 25% and 20%, respectively. Marker count and unweighted coverage are diagnostic only. Values up to −0.55 are classified as markedly paler, up to −0.2 as paler, below 0.25 as intermediate-light, below 0.6 as more tanned, and values above that as markedly more tanned. In the current run, 10,497 profiles are calculated, 1,355 limited, and 5,733 not evaluable.

Ju/Liu/Jacobs (Exploratory)

•Sources: The six active Liu/Jacobs markers are combined with independent European UK Biobank signals from Ju and Mathieson 2021 and one supplementary effect from Farré et al. 2023.

•Selection: A total of 171 source components are included — 164 from Ju/Mathieson, six from Liu/Jacobs, and one from Farré. After merging one overlap, 170 unique directly harmonised runtime markers remain. Findings that are not independent, are palindromically ambiguous, or cannot be transferred to a European light-to-dark axis are excluded.

•Available: In the current run, 11,105 profiles are calculated, 3,278 are limited, and 3,202 are blocked for insufficient basis. The six established Liu/Jacobs anchors remain part of the extended model.

•Calculation and interpretation: Effects are first normalised within each source; source blocks are then combined with equal weight. This prevents the numerically large Ju list from overwhelming the smaller Liu/Jacobs and Farré blocks merely through its row count. The value is an uncalibrated European evidence index, not a phenotype probability.

HIrisPlex-S

HIrisPlex-S is a forensic prediction model that uses approximately 40 SNPs to infer pigmentation in modern populations. It was calibrated for broad global applicability rather than specifically for European genetic variation, using data from roughly 30 population groups, including African, East Asian, Middle Eastern and admixed populations. This makes it highly error-prone, and in some settings nearly unusable, for ancient European DNA. In GeoGens it produces substantial statistical noise and can yield incorrect results.

The skin-pigmentation view used here is based on the Walsh line of global skin-colour prediction and the HIrisPlex-S system developed from it, which was validated for forensic use by Susan Walsh, Manfred Kayser and colleagues at Erasmus MC. GeoGens does not apply the full modern forensic probability calibration. Instead, it uses a direct-marker index that counts only observed, unambiguously harmonised SNPs and does not replace missing markers.

The active implementation uses the skin-colour component of the HIrisPlex-S/Walsh model with five directional categories: markedly paler, paler, intermediate-light, more tanned, and markedly more tanned. The result is a pigmentation index, not an individual forensic diagnosis or a modern calibrated probability.

The model requests 36 skin-pigmentation markers. Of these, 29 are observable and actively harmonised in the current AADR dataset; the other seven are not directly covered. Core markers include SLC24A5 rs1426654, SLC45A2 rs16891982, IRF4 rs12203592, OCA2 rs1800414 and rs1800407, HERC2/OCA2 rs12913832, HERC2 rs1129038 and rs1667394, ASIP rs6119471, and MC1R rs1805008.

The remaining active markers add further signals from MC1R, TYR, SLC24A4, OCA2, HERC2, KITLG, ASIP/PIGU, RALY, DEF8, TYRP1, ANKRD11 and BNC2. In the current run, 10,107 profiles are calculated, 1,997 have a limited basis, and 5,481 are not evaluable. Of all profiles, 9,554 reach moderate confidence; 8,031 remain at low confidence and are not rendered as a robust map colour.

Weights follow the ordinal coefficients of the Walsh/HIrisPlex-S skin model. For standard markers, dosage is centred: one copy of the effect allele is the midpoint, two copies shift the index in the coefficient's direction, and zero copies shift it in the opposite direction. Rare MC1R variants are handled differently. Only directly observed copies of the effect allele are counted, preventing the absence of a rare variant from being misread as evidence of darker pigmentation.

The raw score is capped at an index from −1 to +1. Negative values indicate paler pigmentation and positive values more tanned pigmentation. A calculated result requires observation of at least 45% of both the total and core-marker effect weight; a limited result requires 25% of each. Moderate confidence requires 50% coverage of each weight basis and at least 0.5 support for the leading category. Marker count and unweighted coverage are diagnostic only.

HIrisPlex-S has limited reliability over long time spans. It was developed for modern forensic use, not for 10,000-year-old genomes with extensive missing data, different population structure, and potentially shifted weak-effect associations. In ancient European samples, its signals may therefore provide broad pigmentation context but must not be interpreted as a reliable skin-colour diagnosis.

In GeoGens, HIrisPlex-S is therefore explicitly a supplementary direct-marker index. The more robust skin-pigmentation logic is covered by Liu/Jacobs (Exploratory) and in particular by Liu/Jacobs (Strict). HIrisPlex-S remains visible because it represents a well-known forensic reference, but its categories here are only cautious pigmentation indicators.

Privé 2022 (PGS)

This mode uses a modern skin-pigmentation PGS from the PGS Catalog/UK Biobank family, based on the self-reported modern skin-colour categories in the relevant UK Biobank field. Like HIrisPlex-S, it is retained here as a negative control. The score includes very weak associations that are not robust across long time spans, making it unsuitable for substantive claims about ancient European DNA. Its strong statistical noise and repeated validation failures make that limitation readily apparent.

The active skin-pigmentation PGS contains 9,330 usable markers out of 9,342 candidates. It is therefore broad but composed largely of weak effects. AADR provides sufficient coverage for many individuals, though not for all. In the current run, 11,643 profiles are calculated and 5,942 are not evaluable.

For a result, an individual needs at least 25 per cent of the observable absolute effect-weight basis and at least 20 per cent of the high-weight basis. Marker count and unweighted coverage are diagnostics only; missing markers are not counted as zero.

The Light complexion mode is based on the observation that light hair and light eyes are strongly correlated in present-day populations (approximately 0.8–0.9). GeoGens makes it possible to test whether the same relationship existed in past populations. The result is strikingly consistent: the correlation between blond hair and blue eyes remains high and stable, and increases slightly further back in time.

The map visualises this relationship, traditionally described in anthropology as a light complexion, and may help trace the historical distribution of the characteristic ‘Nordic phenotype’. Pale skin pigmentation is also substantially correlated with the other light traits, so the analysis was extended to three-way combinations. In this broader sense, light complexion can refer to:

blond hair x blue eyes

blond hair x pale skin

blue eyes x pale skin

blond hair x blue eyes x pale skin

The combinations differ in data coverage but produce similar patterns, supporting the interpretation that this is an old and substantial population-genetic relationship.

Light complexion continues to combine the three standard methods Morgan/Lona-Durazo (Strict), Lona-Durazo/Sulem (Strict), and Liu/Jacobs (Strict) with equal weight. Depending on the selected component, two or all three traits are combined; missing bases are not scored as zero.

Four component modes are available: hair × eyes, hair × skin, eyes × skin, and all three traits. Hair uses the blond score from the Strict hair model, eyes the blue score from the Strict eye model, and skin the paler-pigmentation score from the Strict skin model. Each component ranges from 0 to 1. A missing component is not treated as dark or low; it blocks the corresponding combination.

When all required components are present, their mean is calculated. This 0-to-1 mean is then mapped to the −1-to-+1 map axis: 0 becomes −1, 0.5 becomes 0, and 1 becomes +1. A high positive value therefore indicates strong light-complexion evidence; a value near 0 indicates an intermediate or partial pattern; and a negative value indicates low evidence. In band mode, values of +0.5 or above are very light, values from 0 are light, values from −0.5 are darker, and lower values are dark. The colours follow the same logic: green indicates low, yellow partial, orange stronger, and red high light-complexion evidence; grey indicates a missing component basis.

The five-factor model of personality (also called the Big Five or OCEAN model) comprises five main dimensions. GeoGens currently implements three of them: extraversion, agreeableness, and neuroticism. These three currently provide robustly validated genetic results. Conscientiousness and Openness are known to suffer from self-assessment biases. It is hoped that future studies will also validate these dimensions.

The basis consists of European-validated genome-wide significant associations from Gupta 2024 at p ≤ 5 × 10⁻⁸. Only SNPs whose effect and alternative alleles can be harmonised reliably in the raw data and directly observed in at least one sample are used. There is no imputation and sparse dimensions are not padded with markers from other studies. The current runtime model contains 181 active SNPs for the three implemented dimensions: 9 for extraversion, 3 for agreeableness, and 169 for neuroticism. This very uneven breadth must be kept in mind when comparing the dimensions.

How the value is calculated

For every observed SNP, GeoGens counts the fraction of damage-aware filtered reads carrying the effect allele. A fraction of 0 means that no read carries the effect allele, 0.5 means half, and 1 means all reads. This fraction is weighted by the beta direction reported in Gupta 2024. To keep betas of different magnitudes comparable within a dimension, each beta is divided by the median absolute beta for that dimension; an individual weight is capped at ±2.

The raw score is the sum of effect-allele fraction × signed weight, divided by the sum of the absolute weights of all markers actually observed for that individual. Missing SNPs therefore disappear from both numerator and denominator and are not treated as opposing alleles or neutral values. The observed weight basis is reported separately and states how much of the dimension's total possible effect weight is available for that individual.

At 35 per cent observed weight basis or more, the status is ‘calculated’; from 20 to below 35 per cent it is ‘limited basis’; below that the profile has no evaluable basis. At 50 per cent weight basis or more, computational confidence is at most moderate. Marker count, mean read depth, and single-read markers remain additional quality indicators and do not replace these weight thresholds.

To colour individuals comparably on the map, each dimension's raw score is centred relative to all evaluable GeoGens profiles. The median forms the midpoint and the median absolute deviation provides a robust measure of spread. The visible index is (raw score − median) / (3 × robust spread) and is clamped to −1 to +1 for colouring. Zero therefore means approximately the centre of the evaluable ancient sample, not the absence of the trait. Positive and negative values are relative deviations within GeoGens, not universal personality measurements.

The three implemented dimensions

Extraversion comprises sociability, outgoingness, activity, excitement seeking, and lively emotionality; lower values point more towards reserve and introversion. The active basis contains 9 SNPs.

Agreeableness comprises altruism, trust, modesty, cooperativeness, and sincerity; lower values point more towards competitiveness and lower cooperativeness. The active basis contains only 3 SNPs and is therefore especially weak.

Neuroticism comprises anxiety, irritability, emotional instability, and nervousness; lower values point more towards emotional stability and inner calm. The active basis contains 169 SNPs and is therefore much broader than the other two implemented dimensions.

Interpretation

The three genetic evidence indices are based on small effects from modern European samples and are applied to incomplete ancient genomes sequenced at uneven depth. Cautious group means, temporal patterns, and results that remain stable under bilateral inspection and different quality filters are therefore more defensible than interpreting a single individual.

Lifelong lactose tolerance is rare worldwide today and almost exclusively limited to (Northern) Europeans. In the past, this trait was far rarer still, and virtually non-existent during the Paleolithic era. Its spread can be traced so effectively because, in Europeans, it is linked to essentially a single SNP marker: rs4988235. The persistence allele acts dominantly; a single copy suffices—individuals carrying one or two copies are generally lactase-persistent, whereas only two copies of the alternative allele result in lactose intolerance. Consequently, detecting a persistence allele constitutes a reliable positive finding, whereas detecting the alternative allele does not rule out carrier status—meaning the map tends to underestimate persistence rather than overestimate it. Markers that are not detected remain omitted and do not count as a negative finding.

Metabolism is an exploratory mode based on two SNPs that previous studies have associated with adaptation to contrasting dietary patterns: a more hunter-like diet rich in animal foods and a more agricultural diet rich in plant foods.

The mode examines only these two loci and classifies the observed combination as two plant-diet variants, two animal-diet variants, or one of each—a mixed type. These labels are necessarily simplified; the underlying signals concern specific metabolic pathways:

The first locus, CLTCL1/CHC22, affects how the body handles high carbohydrate intake: whether glucose is made available quickly or retained in cellular stores for longer, helping sustain blood-glucose levels during fasting. Rapid handling of large carbohydrate loads is thought to be more adaptive in farming populations, whereas longer retention may better fit the lower-carbohydrate diets of hunting or pastoral populations.

The second locus, FADS1/FADS2, affects the conversion of plant-derived fatty-acid precursors into longer-chain polyunsaturated fatty acids. This conversion is less critical when meat and fish provide these fatty acids directly, but more important in a predominantly plant-based diet.

In detail, the genes act as follows:

CLTCL1 encodes CHC22, a clathrin heavy-chain variant involved in the control of GLUT4-containing vesicles. GLUT4 is the central insulin-dependent glucose transporter in muscle and adipose tissue. The variant observed here, rs1061325, marks M1316V: T/T corresponds to the ancestral M1316 profile, C/C to the V1316 profile. Functionally, this region is linked to differences in how quickly glucose is taken up from the blood into cells following an insulin signal, or retained in cellular stores.

FADS1/FADS2 control the desaturase steps by which the body can produce longer-chain polyunsaturated fatty acids from plant precursors. Such conversions are especially important when high-value animal or marine fatty acids are scarcer in the diet and have to be formed to a greater extent from plant precursors.

The active FADS component is represented by rs174546 and rs174570. At both markers, C is assigned to the plant-diet/farmer direction and T to the animal-diet/hunter direction; the two markers are weighted equally. A further FADS context marker is documented but not scored because it is not directly observable in the dataset used here.

The current implementation has two equally weighted components: CLTCL1/CHC22 and FADS1/FADS2. CLTCL1 is represented by rs1061325. The FADS component combines the observed active markers rs174546 and rs174570, each with a weight of 0.5. Every observed marker direction is mapped to a scale from −1 to +1: −1 denotes the animal-diet direction, +1 the plant-diet direction, and 0 an intermediate state.

If both components are observed, their mean is returned; if only one component is observed, a limited value from that single basis is used. From +0.35 upwards plant-food metabolism is shown, down to -0.35 animal-food metabolism, and mixed type in between. If every evaluable component is missing, the individual remains unknown. In the current run, 4,937 profiles are plant-food metabolism, 4,312 animal-food metabolism, 3,821 mixed type, and 4,515 unknown. CLTCL1 is observed in 10,100 profiles, FADS in 12,156 profiles.

Muscle type is an exploratory single-variant mode based on rs1815739, a polymorphism that affects muscle-fibre composition.

rs1815739 is the well-known ACTN3 R577X variant. The C allele corresponds to the R form and preserves functional alpha-actinin-3 in fast-twitch, power-oriented muscle fibres. GeoGens therefore labels C/C as Strength type. On average, this genotype is more compatible with sprinting, explosive power, rapid changes of direction, lifting, throwing and short bouts of intense exertion. This does not imply poor endurance, only less specialisation for highly efficient prolonged aerobic activity.

The T allele corresponds to the X form. T/T produces no functional alpha-actinin-3 and is associated with a shift towards more oxidative, endurance-oriented muscle properties. GeoGens therefore labels T/T as Stamina type. This may favour prolonged steady exertion, marching, running, repetitive work and energy efficiency, but it does not establish actual athletic performance. C/T is shown as mixed type. In the current run, 5,264 profiles are Strength type, 4,408 Stamina type, 594 mixed type, and 7,319 unknown.

Type I was accordingly called “Stamina type”, type II “Strength type”. Since the Strength type becomes slightly, but discernibly, rarer over time, an actual adaptation to changed living conditions appears to be observable here as well.

The COMT variant became widely known through the “warrior/worrier” label, although that description is misleading. The mode concerns a single SNP whose principal reported effect is on dopamine regulation in the prefrontal cortex.

rs4680 is the COMT Val158Met variant. The G allele corresponds to Val and is associated with higher COMT enzyme activity, causing faster dopamine breakdown in the prefrontal cortex. GeoGens labels G/G as fast-acting type: a tendency towards rapid responses and readiness to act under pressure. The possible trade-off is less support for fine-grained planning, working memory and calm deliberation under low-stress conditions.

The A allele corresponds to Met and is associated with lower COMT activity, leaving dopamine available for longer in the prefrontal cortex. GeoGens labels A/A as brooder type, reflecting greater support for working memory, anticipatory thought, deliberation and stable cognitive control. The possible trade-off is greater sensitivity to stress or overload under high dopaminergic load or pressure. A/G is shown as mixed type. In the current run, 4,864 profiles are fast-acting type, 4,643 brooder type, 530 mixed type, and 7,548 unknown.

We have therefore called the Met/Met variant “brooder type” and the Val/Val variant “fast-acting type”, because these terms capture the general effect of the gene more accurately. The brooder types appear to be associated to a small degree with hunter cultures, but more strongly still with northern origin. Even today, north-west Europeans show peak values here. In general, however, almost all populations display a relatively balanced ratio between brooder and fast-acting types, which is in itself an indication that the gene's real relevance for population-genetic adaptation must not be overestimated.

Shared genetic drift (f3-outgroup) measures the population-genetic drift shared by two individuals relative to a fixed outgroup. The current calculation explicitly uses four observed YRI reference individuals rather than Mbuti: NA18486, NA18488, NA18489 and NA18498. YRI denotes Yoruba in Ibadan, Nigeria, the West African reference population in the 1000 Genomes Project. It serves as the outgroup because it is approximately equally distant from the Eurasian individuals compared here. The calls come from the published 1000 Genomes reference in the hapROH 1240K resource and are harmonised to the same autosomal 1240K positions and allele orientations as the ancient individuals.

The materialised f3 source matrix contains 7,199 individuals and 25,909,201 pairs. After identity resolution, quality filters, and the current sample release, 6,317 individuals with 19,646,930 valid relationships were admitted to the GeoGens runtime. f3 counts allele states shared relative to the fixed outgroup.

How the value is calculated

For pseudo-haploid alleles encoded as 0 or 1, the implemented formula is f3(A,B;O) = (1/N) × ΣO Σl (xA,l − xO,l)(xB,l − xO,l). The sum runs over all four YRI references O and all positions l at which A, B, and the respective O jointly have a call; N is the total number of these valid A–B–O comparisons. With binary single-allele calls, a position contributes exactly 1 to the numerator when A and B carry the same allele and the respective YRI individual carries the other allele. If all three agree, or A and B differ, the contribution is 0. The raw value is the mean across all valid comparisons.

The reported f3 overlap is therefore not simply the number of unique SNP positions. Each position can enter the denominator with up to four YRI individuals, so the number of valid A–B–O comparisons can exceed the number of SNPs in the panel. A pair enters the runtime only with at least 50,000 SNPs jointly observed in A and B. To form the bilateral f3 rank, the f3 comparison basis must additionally contain at least 50,000 valid A–B–O observations. If it falls below this threshold, an available raw value can remain documented but receives no bilateral f3 score.

Larger raw f3 values indicate more drift shared relative to YRI. Every individual's valid partners are therefore ranked by descending f3 value; ties receive the midrank. The position is normalised from 0 to 1, with 1 for the nearest and 0 for the most distant partner. B's rank in A's complete profile and A's rank in B's complete profile are calculated separately. The displayed bilateral score is the conservative minimum of the two perspectives. This check prevents shifted individual value distributions or one-sided outliers from appearing as generally strong proximity.

Interpretation

For group and cumulative selections, each target uses the unweighted median of the available bilateral f3 pair scores. Missing relationships are not inserted as zero. Contributor count and median f3 overlap show how many selected individuals and how many valid A–B–O comparisons support the summary. The three colour scales change only the visual stretch of the bilateral rank; raw f3, percentile, and ordering remain unchanged.

For an ancient origin the bilateral f3 value is reported as a percentile and marked with “P”, not with a percent sign. It states how far forward an individual stands in the ranking against all others. The raw f3 value itself is not a percentage quantity but a drift covariance relative to the outgroup; it cannot be read as a share of common DNA.

The f3 value is a relative measure of shared genetic drift. The map method calculates f3 between individual pseudo-haploid individuals. SNP overlap and bilateral stability document the data basis. Absolute raw values and relative ranks should therefore only be compared within their respective scale and interpretation.

The cluster view divides the currently filtered map selection into two to twelve computational groups according to the selected features. It searches for similarities in a multidimensional feature space using a deterministic k-means procedure. The groups are derived entirely from the currently visible individuals and the selected settings. They are not predefined populations, ethnicities, or archaeological cultures and need not remain stable when the selection, period, features, weights, or number of clusters changes.

Hair colour uses the three raw axes blond, red, and dark from Morgan/Lona-Durazo (Strict); eye colour uses the blue, brown/dark, and green/mixed axes from Lona-Durazo/Sulem (Strict); skin colour uses the light and dark axes from Liu/Jacobs (Strict). Numerical model values are used rather than the colour labels derived from them. Shared genetic drift (f3-outgroup), by contrast, enters as a global profile of the complete measured relationship network.

How the value is calculated

For f3, a weighted graph is constructed: every evaluable individual is a node, every positive bilateral relationship is an edge, and its bilateral rank is the edge weight. In the current build, the graph contains 6,317 nodes and 18,741,188 positive measured edges. The adjacency matrix W is symmetrically normalised with weighted node degrees D as S = D⁻¹ᐟ²WD⁻¹ᐟ². The leading eigenvectors of S are determined, the trivial first eigenvector is discarded, and the following twelve axes are weighted by their eigenvalues. Each individual's resulting twelve-dimensional vector is normalised to length 1. This translates the complete bilateral relationship profile into comparable spectral coordinates without replacing individual missing pair values with invented raw values.

Within the current map selection, every included dimension is brought onto a 0-to-1 scale by min-max scaling. If all individuals have the same value in a dimension, a neutral value of 0.5 is assigned. The selected feature weight is distributed equally across that feature's dimensions and applied to the coordinates as a square-root factor. A three-axis colour model therefore does not automatically exert more influence than a one-dimensional trait merely because it contains three numbers. Multiple selected features then form one joint feature vector per individual.

In ‘Exclude incomplete’ mode, only individuals with every selected dimension available are clustered. In ‘Tolerate gaps’ mode, no imputation is performed: the distance between two vectors is the mean squared difference across only their jointly available dimensions, and centroids are averaged per dimension from the member values available there. If an individual lacks every usable dimension, it remains grey and unclustered. The effective number of clusters also cannot exceed the number of evaluable individuals.

Starting centroids are selected reproducibly: first the evaluable vector with the greatest magnitude, then at each step the point farthest from its nearest already selected centroid. K-means then assigns every individual to the nearest centroid and recomputes each centroid as the dimension-wise mean of its members. Assignment and recomputation repeat until stable, but no more than 40 times. Thus, with an identical selection and identical settings, there is no random initialisation that produces different groups on every run.

Interpretation

After clustering, groups are ordered along the first principal axis of their feature vectors and only then receive numbers and colours. For geographic naming, a two-by-two contingency table of membership and non-membership is formed for every group and every defined geographic region. GeoGens calculates the phi coefficient and uses the region with the strongest positive association and greatest support as a descriptive name. Geography therefore does not enter the cluster calculation retrospectively; the name only summarises the group already formed.

Cluster distances are calculated from the distances between their centroids. The dendrogram then joins the groups or branches with the smallest mean pairwise centroid distance using average linkage. It visualises proximity among the already calculated clusters but changes neither their members nor the k-means assignment. Because k-means always forms the requested number of groups, a cluster alone is not evidence of a natural population-historical boundary. The most defensible patterns are those that remain stable with appropriate features, sufficient data, and plausible changes to filters, weights, and cluster count.

Y-DNA traces the paternal line and is therefore generally available only for male individuals. Mitochondrial DNA traces the maternal line and may be recovered from either sex. Y-DNA and mtDNA haplogroups are consequently analysed as separate lineage systems.

The current haplogroup map is based on the merged GeoGen sample universe. After spatial exclusions and deduplication, the current public export contains 17,585 find and analysis points. Y-DNA is published with two separate current values: a hierarchical designation and a terminal SNP node. Imported AADR/Poseidon values and approved TruHaplo reanalyses remain available as auditable provenance; curated evidence decisions can replace both current values together. For mtDNA, the mtHaplogroup label is used.

TruHaplo extends this dataset through intensive reanalysis of aligned Y-chromosome read evidence. The run finaler_pathphynder_y_run_20260703-071029 used PathPhynder 1.2.4 and BigTree.Y_GRCh37.V2.250606. The 1,163 initially released Step 2 calls from lineages I and R were subsequently checked against the derived and ancestral markers at their actual final call. 1,064 calls have a derived-marker majority and remain approved; 99 ties or ancestral majorities are rejected and cannot override the imported call. Review, no-call, and Step 1 results remain excluded.

PathPhynder/BigTree and the YFull/YTree tree used for relationship and TMRCA functions do not consistently share the same nomenclature. The hierarchical designation, terminal SNP, and internal calculation node therefore remain separate. Terminal filters use only the current terminal designation; internal calculation nodes appear neither as terminal filters nor in parentheses after the hierarchical designation. YFull age estimates never rewrite a haplogroup call automatically.

Tree resolution is not based on string comparison alone. Y-DNA is checked against YFull/YTree 14.04.0 (43,261 nodes, 1,307,048 alias forms, CC BY 4.0); mtDNA against HaploGrep's PhyloTree 17 Forensic Update 1.3 (6,380 nodes, MIT). Its rCRS-oriented source topology is reproducibly rerooted at mt-MRCA for lineage calculations. Apostrophes and stars remain meaningful name components: composite apostrophe nodes denote the common ancestor of alternative daughter clades and are not treated as terminal lineages. Every single-letter mtDNA node and the reviewed macro clades are broad resolutions; question marks and tildes remain explicit source uncertainty.

The count is therefore stricter than a mere label count. Across the dataset there are 8,855 Y-DNA calls by calculation priority; 8,029 are usably resolved, 819 are broad root nodes only, 5 are invalid, and 2 cannot be resolved in the reviewed tree. 8,730 profiles have no Y call by calculation priority. There are 12,890 mtDNA calls: 12,147 are usably resolved, 604 are broad major nodes, 101 are composite clades, 12 are explicitly uncertain, and 26 are invalid multiple-call or free-text values; 4,695 profiles have no mtDNA call. Missing means only that no analysable haplogroup call is available for this axis. It is not a negative finding or a population-genetic zero.

The map surface for haplogroups is categorical. For Y-DNA, only sufficiently deep clades are admitted as a surface category, typically something like I1 or R1b rather than merely I or R; for mtDNA, subclades such as H1 or U5, and not merely H or U, are grouped accordingly. Every analysable point contributes its group colour with constant weight, within the current time window and after the active filters. Where several lineages overlap spatially, the weighted average colour of the competing groups results. The point colour can additionally be overridden by selected prefixes, so that sublineages such as I2a, R1b, or U5 can be traced specifically.

Clicking a data point switches the haplogroup display to lineage mode. The view shows which other individuals in the active time window could, in principle, share a paternal or maternal lineage relationship with the selection in the reviewed tree. These links are hypotheses rather than established genealogies, but they are derived from explicit phylogenetic and chronological criteria.

For both Y-DNA and mtDNA, the selected and candidate calls are first resolved against the corresponding reviewed tree index. A candidate is marked as a possible Y ancestor or mtDNA ancestral lineage when its node lies above the selected node and its date is earlier under the same dating rule. Identical resolved nodes—for example I2a1 and I2a1—may also be directional: a clearly older candidate can be shown as a possible ancestor and a clearly younger one as a possible descendant. Without a clear chronological direction, the pair remains visible as the same lineage without an arrow direction. A candidate is treated as a possible Y descendant or mtDNA sublineage when its node lies below the selected node and it is chronologically younger. Broad parent nodes and invalid or unresolved calls cannot produce directed arrows because they do not support a robust direct-lineage hypothesis, although they may still be colour-coded. The arrows therefore show the most plausible connections, not every conceivable one.

Chronological validation is strict by default and follows the active dating anchor. Earliest uses the older bound, Midpoint the mean, and Latest the younger bound. In Full range mode, the midpoint determines lineage direction, arrows and sorting, while the entire dated span still controls visibility. Under strict validation, a possible ancestor must be unequivocally older than the selection and a possible descendant unequivocally younger. When strict validation is disabled, different tree nodes with equal anchor dates may still be assigned a direction.

Undirected Y relationships are shown without separate close or distant TMRCA collateral-lineage classes and without a separate same-unresolved-root category. The remaining Y hits are consolidated as a broad shared Y major lineage and shown in orange: broad I, for example, remains visible for I2. Extremely basal cross-clade stem nodes such as CT are suppressed completely. For mtDNA, the view continues to distinguish direct ancestral lines, sublineages, the same mtDNA line, and the same but unresolved major clade.

Arrow mode can hide the display completely, limit it to the nearest links by space and time, by space only, or by time only, or show every arrow. By default, five ancestors and five descendants are selected by space and time: the calculated arrow width, which combines spatial and temporal proximity, determines the order. The quantity fields apply only to the three limited modes. The small label along an arrow estimates generations at 25 years per generation and rounds up. Dashed arrows are not guaranteed under every allowed haplogroup resolution or have low lineage confidence.

Competing ancestors are shown only in single-selection mode, and they are shown in green. For every visible descendant, other potential ancestors are searched for and ranked by geographic proximity, with the three geographically nearest ones being displayed. The condition is that they are not more than twice as far from the descendant as the selected point. Put loosely, these green arrows warn against assuming descent, because competing ancestors exist. If such a competing ancestor is too far away, it no longer carries an arrow.

In culture mode, a click first selects the sample's subgroup, where one exists. Clicking again within the same culture expands the selection to the entire culture group. The time window is then adjusted to this selection and widened to at least 3,000 years so that as many points as possible become visible. In cumulative mode, individual samples are added or removed by clicking. The point chosen most recently remains the primary anchor, again with a window width of at least 3,000 years. Both modes compute the lineage relationships separately for each selected anchor.

Y point colours follow the priority selection > ancestors/descendants > broad Y major lineage. If a point is both a possible ancestor of the selection and a competing ancestor of a visible descendant, it remains marked as a possible ancestor while its additional green arrow to the descendant stays visible. Legend counts in multiple-selection modes count relationships across the individual contexts.

AADR does not provide a consistent system of archaeological cultural attribution. This is a serious limitation because archaeological interpretation depends heavily on material culture and find context. More than a decade of archaeogenetic data collection has not been matched by a comparable effort to connect the samples with specific archaeological cultures. A few AADR labels contain broad terms such as “CordedWare” or “BellBeaker”, but most individuals carry only terse labels such as “Spain EBA” (Spanish Early Bronze Age) or “UkraineN” (Ukrainian Neolithic). People who lived within the same modern national borders thousands of years ago could belong to sharply different material and social contexts. Since such archaeological groups may in some cases approximate historical tribes or peoples, it matters whether an individual can be assigned securely or must remain unresolved.

Culture-group classifications must nevertheless be treated with caution. An archaeological culture does not necessarily correspond to a single community of customs, language and exchange, even though Migration Period cultural horizons show that it sometimes can. National research traditions also produce uneven nomenclature: the same group may have several names, while materially distinct cultures with only loose connections may be combined under one label. Chronological boundaries are equally inconsistent. A continuous material tradition may be divided sharply into several named phases, while proposed proto- or post-cultures may have little substantive connection to the culture whose name they borrow. Archaeogenetic interpretation should therefore neither equate every culture group with a “people” nor depend on cultural labels alone. Yet ignoring material context would be equally misleading—for example, by grouping all European hunter-gatherers from 20,000 to 7,000 BCE under “WHG” despite marked archaeological differences and anthropological clustering. Questions such as the Indo-Europeanisation and Neolithisation of Europe can ultimately be addressed only in dialogue with the archaeological cultural record.

I therefore undertook a record-by-record review of the archaeological contexts, working period by period and region by region to identify culture groups and subgroups. AI-assisted literature search was used selectively because the corpus contains more than 17,000 records, but every such lead was checked critically and repeatedly against the sources. The curation consequently remained a substantial manual research effort.

Now to the technical side:

•Not every culture group has subgroups. Where subgroups are defined, however, every individual in the parent group must be assigned to one of them.

•Every time stage contains individuals whose cultural affiliation is unclear and who therefore have no culture-group assignment. Further research support is needed for these cases. In some contexts—bog bodies, beach finds or burials without grave goods, for example—an archaeological assignment may be impossible in principle. Such records are labelled only, for example, “Uncertain assignment – Neolithic”.

•Some find contexts could be assigned to a particular culture with reasonable, but not complete, confidence. These are marked as “Uncertain assignment”, and a question mark appears after the cultural attribution in the live detail panel. A checkbox in the share panel on the right controls whether such uncertain assignments are included.

•Note: Culture-group assignments are under continuous review. Some are deliberately broad, and many finds in Central Asia and the Near East remain uncurated even though specialist knowledge of the sources could support a more precise attribution. Individual assignments may change as the evidence develops. Corrections, additions and research assistance are welcome.

Map colouring does not delineate territories or claim to reconstruct exact historical conditions. It is a visual aid for recognising broad tendencies in a dense field of sample points. Every evaluable point contributes to the surface at its geographic location through the same spatial influence kernel.

The result is a smoothed interpretive surface, not a population or territory map. When several evaluable individuals share a site, each remains an independent observation rather than being collapsed into a site-level majority. Conflicting values therefore compete within the surface, and a tie does not make an entire site disappear.

In addition, points surrounded by a large amount of land mass have a stronger effect than points located on islands such as Sicily or Gotland, which can therefore colour hardly any land mass. The haplogroup-arrow system is subject to a similar representational distortion, in that arrows spanning greater distances take up more space and catch the eye more quickly than small-scale relationships.

A further limitation is the conversion of model values into colour. A colour gradient cannot provide a uniform representation of every trait's underlying variance, so its intensity must be interpreted cautiously and comparatively—primarily against other finds from the same period and always within the same method. An identical visual difference may represent a much smaller model difference for trait A than for trait B or method C.

For the bilateral A-or-B values (muscle type, lactase persistence, eye colour, etc.), colour bands were used. For numerical value gradients (light complexion, height, skin pigmentation), by contrast, gradients were used. For light complexion, the colour-band mode was also built for the sake of comparability.

In principle the map generates a circle of influence from every evaluable data point. For hair colour, eye colour, light-complexion bands, metabolism, muscle type, COMT, culture groups and haplogroups the radius is 3.0 degrees of latitude; for height, skin values and continuous light complexion likewise 3.0; for lactase persistence 3.8. The influence falls off sharply with distance. Below a minimum strength no surface is drawn: 0.32 for discrete colour bands, 0.28 for continuous values and 0.38 for lactase persistence. The minimum strength is governed by the total local kernel strength and not by how extreme the trait value is. Lactase persistence additionally has a small close range of 0.28 degrees, so that a positive signal remains visible immediately at the site.

Hair colour, eye colour, metabolism, muscle type, COMT and the banded light-complexion view use discrete categories rather than continuous gradients. Every point that passes the active model's evidence gate receives the same spatial kernel. Score and quality determine whether the point is admitted and which category it carries, but not its geographic reach or opacity.

Hair colour: the Morgan/Lona-Durazo, Hysi/Morgan/Lona-Durazo, Morgan 2018 and Tanigawa 2022 methods each take the winning category of their active model. Blond and red-blond run towards the light colour direction, brown/dark towards the dark one, dark-blond intermediate bands towards intermediate colours. Unknown or blocked profiles do not contribute to the surface.

Eye colour: the Lona-Durazo/Sulem, Simcoe/Lona-Durazo and Minimal methods, together with the Deep SNP scan, use the same mapping logic. Blue contributes to the blue end of the scale, brown/dark to the brown end, and green/mixed to an intermediate colour. Only profiles admitted by the active model are coloured; unknown profiles do not contribute.

Skin pigmentation and height are continuous gradients. Skin values are limited to -1 to +1 and coloured from pale to tanned pigmentation. Height values are likewise limited to -1 to +1; lower values run into blue, higher ones into red, the middle stays neutral. Yengo/Marouli/Hawkes (exploratory) and (strict) use their own centring around the active middle range, so that small real differences do not disappear visually. Insufficient profiles are excluded; every admitted point receives the same spatial kernel.

Light complexion exists in two map forms. The continuous mode averages the available strict light values of the selected components and maps them onto -1 to +1: green means low light complexion, yellow/orange medium to stronger, red high light complexion. Every admitted point receives the same spatial kernel irrespective of how extreme its value is; the value determines only the colour. Only sufficiently well-supported strict eye profiles are used as a component. The band mode uses the same values but applies fixed classes: from +0.5 very light, from 0 light, from -0.5 darker, below that dark. If a required or sufficiently well-supported component is missing, the point remains grey, that is, without any contribution to the surface.

Lactase persistence is bilateral. An observed rs4988235 LP allele colours green, an observed marker without an LP allele blue. Missing markers are not counted as absent lactase persistence and do not contribute to the positive or negative decision. Where both signals compete within a region, the stronger weighted neighbourhood prevails.

Metabolism, muscle type and COMT are small discrete modes. Metabolism colours animal-food metabolism, mixed type or plant-food metabolism. Muscle type colours Strength type, mixed type or Stamina type. COMT colours Doer type, mixed type or Ponderer type. Missing individual markers do not contribute to the surface.

Culture groups, culture subgroups, Y haplogroups and mt haplogroups are categorical surfaces. Every point contributes the group colour assigned to it with a constant weight; in overlapping regions the weighted average colour of the competing groups emerges. Uncertain, missing or neutral assignments are not asserted as a dominant colour of their own.

One important aspect of our data visualisation was that of contemporaneity. Comparing genetic data across long spans of time has repeatedly led to errors and inconsistencies in the past, and for various reasons it is generally hazardous. Hence our approach of always viewing the map within a concrete time window. The width of the observed window is adjustable; it can be chosen from 100 to 3,000 years (for culture groups only up to 1,000 years, for performance and labelling reasons). For haplogroups a range between 500 and 10,000 years can be selected, because here it is above all ancestor-descendant relationships that are examined, and less so contemporaneity. The current year shown at the right of the timeline always determines the midpoint of the time window, so that with a 1,000-year window all data points are displayed that lie up to 500 years before and up to 500 years after this midpoint.

So where does a data point lie? That question is harder than it first appears. Prehistoric individuals rarely have a headstone giving the day of their death. They are radiocarbon-dated, and a considerable range is often possible. A skeleton may be dated to the period from 2400 to 1900 BCE, for example. Where does that data point lie then? For this there are four selection modes in the “Dating anchor” panel with which the question can be answered. Either the earliest point of the dating is chosen or the latest. The most pragmatic third option is to use the midpoint, which in our example would be 2150 BCE. Anyone unwilling to accept this arithmetic “trimming” chooses the fourth dating method, “Full range”, in which an individual is shown at both its earliest and its latest date and, of course, throughout the period in between. In extreme cases, however, there are Palaeolithic finds dated only very roughly to between 20,000 and 10,000 BCE. With the “Full range” dating anchor such a data point would therefore stay visible on the map for a very long time and might gain an “unfair” dominance and visibility merely because archaeology happens to be unable or unwilling to state anything more precise here.

The AADR data basis lists several analyses for some individuals. This is because individual finds were analysed at different times, in different laboratories, or with different degrees of stringency. An individual may therefore have been examined very strictly in analysis A and consequently show far fewer genes, but these genes come with far greater scientific robustness. In analysis B the search run may have been less strict, which permits more statements but gives those statements less weight.

The filter area therefore offers, under the basic filters, the option of displaying either the "most informative analysis" of each individual (relative to the trait currently under investigation), disregarding data quality. Alternatively, only the strictest analysis per individual can be displayed, which may cause data points to drop out but guarantees higher quality in return. The third option is "All analyses", with which such individuals show all of their data points on the map. This makes it possible to check where the analyses differ, but it may risk mistaking a single individual for several different ones if one is not paying close attention.

The right-hand share panel is a real-time ranking of the currently loaded map selection. It answers the question: in which culture groups, geographic regions, or haplogroups is the selected trait value highest? Only individuals for which the active map mode can supply a numerical value are used for this; unknown, grey, or blocked profiles are removed from this calculation and are not counted as zero.

Every evaluable point is first brought onto a uniform scale from 0 to 1. A value of 1 always denotes the positive direction of the heading, for example lactase persistence, animal-food metabolism, Strength type, worrier, blond/light hair, blue/light eyes, lighter skin, a higher height index, or high light complexion. For indices running from -1 to +1, (value + 1) / 2 is used; for skin values the direction is deliberately inverted, so that paler skin appears as the higher value: (1 - value) / 2.

For every possible group, the mean of its evaluable members is calculated. This mean appears in the bar as Ø in per cent. In addition, a point-biserial Pearson r is calculated: group membership is coded as 1, non-membership as 0, and the trait value remains the 0-1 value. Formally this is r = ((group mean - remainder mean) / standard deviation of all values) × sqrt(p × q), where p is the share of the group and q the share of the remainder. A group is only shown if it reaches the configured minimum sample size.

The two geographic gradients are calculated separately. Northing and westing are normalised coordinate values within the fixed map BBox; the current trait value is correlated against this coordinate using Pearson r.

The correlation page uses only individuals with a valid dateMidpointBce in the range 15,000 BCE to 1,000 CE, and only complete cases: both selected traits must be evaluable for the same individual. A missing marker basis is not counted as 0 but is removed from the pair count and the statistics.

All values are harmonised to a 0 to 1 scale. Blond hair and blue eyes use the respective strict evidence, skin pigmentation is read as the paler skin direction, and height is transferred from the index bounded to -1 to +1 via (value + 1) / 2. Lactase persistence is 1 for an observed LP allele and 0 for an observed non-LP; metabolism, muscle type, and COMT use 1/0.5/0 for the respective positive, mixed, and negative direction. Northing and westing are normalised coordinate values. Time stages, culture stages, and geographic regions are shown as group means in stage mode.

For two numerical traits, one pair (x, y) is formed per individual. GeoGens calculates Pearson r as the covariance divided by the product of the standard deviations; r lies between -1 and +1. Positive values mean trait expression in the same direction, negative values in opposite directions. Below n = 5, r remains locked; between n = 5 and n = 19 the block counts as limited, and from n = 20 as calculated. The time series divides the same complete cases into selectable blocks of 100, 200, 300, or 500 years; the overall r uses all cases, while the time-window r uses only the blocks visible in the graph.

The 95% confidence interval is calculated using the Fisher z approximation: z = 0.5 × ln((1 + r) / (1 - r)), standard error = 1 / sqrt(n - 3), back-transformation via tanh. p values and Benjamini-Hochberg q values are deliberately not reported. The data points are not independent: find spots are spatially autocorrelated, neighbouring time blocks temporally, individuals from the same burial ground frequently related, and one individual can enter the same calculation through several AADR analyses. The standard error 1 / sqrt(n - 3) assumes independent observations; here it understates the true error, a p value derived from it would be anti-conservative, and an FDR correction on top would inherit that defect while lending it the appearance of multiplicity control. The interval is kept because it is read as a width: it shows how little the sample pins r down.

This view asks two targeted questions: which traits are elevated or reduced within the selected group, and with which other groups does it overlap most strongly?

A group appears in the selector only after reaching the configured minimum size of 3, 5, 10, 30 or 100 individuals. Calculated values additionally require at least 3, 5 or 10 evaluable cases, depending on the selected threshold. By default, culture groups and subgroups are compared only with data points from their own date range. Enabling “Compare with the entire period” instead uses the full correlation window, from 15,000 BCE to 1,000 CE.

For each trait, group membership is coded as 1 and non-membership as 0; the trait value is again the harmonised 0-1 value. Pearson r is calculated from this. In addition, GeoGens shows the mean within the group, the mean outside the group, and the difference in percentage points. A positive delta means that the trait value is higher in the selected group than in the comparison basis. A missing trait basis, a missing culture or region basis, and a missing haplogroup resolution are excluded from the respective calculation.

The variance figure refers to the standard deviation of the analysable individual values within the selected group. It is not interpreted in absolute terms but ranked relative to other groups of the same category and the same trait: the lower roughly 35 per cent count as rather homogeneous, the upper roughly 35 per cent as rather heterogeneous; the middle values receive no strong classification. The intention is to show not only whether a group has, for example, high blond values, but also whether those values scatter widely within the group or occur strikingly tightly clustered.

Over time, the same calculations are repeated block by block. The blocks can be formed as thirds, fifths, millennia, or centuries; for thirds and fifths, the division can be made either by time span or by equal numbers of individuals. A time series requires a comparison window of at least 300 years. Each block must contain enough analysable overall cases and enough group cases; otherwise it remains locked. Y-DNA and mtDNA series additionally show the five most frequent haplogroups at the level selected above or, when “Terminal” is selected, the five most frequent terminal calls.

The top correlating groups are computed as an overlap analysis. For a candidate group, it is checked how often it occurs inside the selected group and how often outside it. Only positive enrichments are shown; sorting is by the difference of the rates, then by r, overlap, and candidate size. Here too only r, n, and the 95% confidence interval are reported; p and q values are omitted for the same reason as in the trait correlations.

To enable better filtering and classification, each individual was assigned a time stage and culture stage according to its dating and archaeological find circumstances. The time stage is absolute and based on exact calendar years (the midpoint of the individual's dating was used):

Palaeolithic (until 10,000 BCE)

Mesolithic (10,000-7,000 BCE)

Neolithic (7,000-3,500 BCE)

Chalcolithic (3,500-2,000 BCE)

Bronze Age (2,000-800 BCE)

Iron Age (800 BCE-300 CE)

Migration Period (300-700 CE)

Early Middle Ages (700-1,000 CE)

Middle Ages (1,000-1,500 CE)

Early Modern Period (1,500-1,800 CE)

Modern period (from 1,800 CE)

The culture stage, by contrast, is relative, meaning that it derives from the actual find circumstances. Mesolithic hunters still lived in Northern Europe when Neolithic farmers had long been living in Southern Europe; the same applies to the spread of bronze casting and ironworking. Culture stages shift gradually, and the question of the actual stage is therefore often more interesting than an absolute chronological classification, for example when comparing changes in metabolism or character traits between sedentary and non-sedentary peoples. It goes without saying that culture stages imply no value judgement regarding the people concerned.

To enable better filtering and to examine traits by geography, common geographic regions were introduced as categories. Some are very broad (Western Europe, Eastern Europe), others more localised (Baltic region, Italian islands). Each individual can therefore belong to several Geographic Regions. The geographic regions are defined as follows:

Italian Peninsula

Present-day Italy excluding the part north of Rimini.

Italian islands

Corsica, Sardinia, and Sicily.

British Isles

Present-day borders of Great Britain and Ireland.

Iberian Peninsula

Present-day borders of Spain, Portugal, and Andorra.

Gaul

France, Belgium, Luxembourg, Monaco, Channel Islands, the Federal Republic west of the Rhine, and the Dutch provinces of Limburg, North Brabant, and Zeeland.

Central Europe

Germany east of the Rhine, Lower Austria, Upper Austria, Vienna, Czechia, Slovakia, Poland, and Kaliningrad Oblast.

Scandinavia

Sweden, Denmark, and Norway up to the 65th parallel.

Finno-Karelia

Finland, European Russia north of the 60th parallel, and Norway/Sweden north of the 65th parallel; Siberian administrative units are excluded.

Baltic region

Estonia, Latvia, Lithuania, Kaliningrad Province, and Warmian-Masurian Voivodeship.

Northeastern Europe

Belarus, Ukraine excluding Odessa, Transcarpathia, and Crimea, the Central Federal District of Russia, and the Baltic region.

Balkan Peninsula

Bounded to the north by and including Slovenia, Hungary, Romania, and Moldova; includes Odessa and Transcarpathia in Ukraine.

Aegean

Greece, Eastern Thrace, and Turkish provinces 41, 77, 16, 10, 17, 45, 35, 9, 48, 7.

Adriatic

Slovenia, Croatia, Bosnia and Herzegovina, Montenegro, and Albania.

Anatolia

Turkey excluding Eastern Thrace and excluding the eastern provinces 8, 25, 75, 36, 76, 4, 49, 12, 21, 63, 47, 72, 56, 13, 30, and 73; including Van.

Pontic Steppe

Ukraine east of Kyiv and the Russian regions Krasnodar, Rostov, Volgograd, Adygea, and Crimea.

Caspian Steppe

Southern Russia excluding the Pontic Steppe, the North Caucasus, and the Kazakh regions West Kazakhstan, Atyrau, Mangystau, and Aktobe.

Pontic-Caspian Steppe

Pontic Steppe plus Caspian Steppe.

Carpathian Basin

Hungary, Serbia north of Belgrade, the Romanian planning regions Nord-Vest/Centru/Vest, southwestern Slovakia, and defined Croatian counties.

Lower Danube region

Northern Bulgaria, western/southern/eastern Romanian regions and Bucharest, and Serbia excluding Kosovo.

Alpine region

Everything within the polygon Linz-Munich-Zurich-Geneva-Grenoble-Grasse-Sanremo-Turin-Como-Verona-Udine-Maribor-Vienna-Linz.

Western Mediterranean

Land within 100 km of the western Mediterranean coast; west of Catanzaro and east of Gibraltar.

Eastern Mediterranean

Land within 100 km of the eastern Mediterranean coast; south of Istanbul and east of Catanzaro.

Mediterranean

Land within 100 km of the Mediterranean coast.

Levant

Lebanon, Syria, Israel, the West Bank, and Jordan.

Mesopotamia

Iraq excluding al-Anbar, selected Iranian provinces, the Syrian provinces ar-Raqqa, al-Hasakah, and Deir ez-Zor, and the Turkish provinces of Mardin, Diyarbakır, and Şanlıurfa.

Caucasus

Russia south of the Makhachkala-Krasnodar line, plus Georgia, Armenia, and Azerbaijan.

Central Asia

Kazakhstan, Turkmenistan, Uzbekistan, Kyrgyzstan, and Tajikistan.

Persian Highland

Iran, Afghanistan, and Balochistan.

India

India and Pakistan excluding Balochistan.

Tarim Basin

Xinjiang region.

Ural

Southern and western Ural: the Russian regions of Chelyabinsk, Bashkortostan, Orenburg, Udmurtia, Tatarstan, Kurgan, Sverdlovsk, Perm, and Samara.

Siberia

Siberian administrative units east of the Urals, including Western Siberia, Altai, Central Siberia, and Transbaikalia.

Himalaya

Nepal and the Tibet Autonomous Region.

Iceland

Iceland and the Faroe Islands.

Canary Islands

Canary Islands.

Arabian Peninsula

Saudi Arabia, Yemen, Oman, Kuwait, Bahrain, Qatar, the United Arab Emirates, and the Iraqi region al-Anbar.

Maghreb

Tunisia, Algeria, and Morocco.

Northeast Africa

Libya, Egypt, and Sudan.

Northern Europe

Europe north of the polygon line Nantes-Paris-Dortmund-Magdeburg-Warsaw-Vilnius-Saint Petersburg.

Southern Europe

Europe south of the line Bordeaux-Zurich-Budapest-Odessa, excluding Anatolia and Caucasus regions.

Western Europe

Europe west of the Trieste-Szczecin line, excluding Norway and Sweden.

Eastern Europe

Europe east of the Trieste-Szczecin line up to the longitude of Atyrau, excluding Finland, Norway, and Sweden.

North European Plain

Flanders, the Netherlands, Luxembourg, Lower Saxony, North Rhine-Westphalia, Schleswig-Holstein, Saxony-Anhalt, Mecklenburg-Western Pomerania, Brandenburg, Berlin, Hamburg, Bremen, and Poland excluding Opole, Silesia, Holy Cross, Lesser Poland, Subcarpathian, and Lublin.

At the outset, the GeoGens project had to define a spatial and temporal framework. Since the main focus was to be European prehistory, Europe and adjacent regions (North Africa and Western Asia), as well as the period from 45,000 BCE to 1,000 CE, were chosen. The exact map boundaries are 25° west longitude in the west, 85° east longitude in the east, 72.5° north latitude in the north, and 20° north latitude in the south. This area may be expanded in future, but curating individuals and cultural groups requires additional work that had to be weighed accordingly. In any event, the vast majority of AADR individuals are located in Europe.

With regard to the temporal framework, hardly any informative genomes survive from before 45,000 BCE. The timeline can be switched at the lower left to begin at 15,000 BCE, making the selection clearer. The limit of 1,000 CE was chosen to avoid having to curate the large mass of medieval and modern genomes. They are insignificant for questions of prehistory. However, extending the period to the present is under consideration: first, to allow better validation of the trait maps, and second, to enable future comparison with personal genomes and their individual origins.

The status box at the lower left is a snapshot of the current map view. The time window, date anchor, filters, analysis selection, and method immediately affect its numbers.

Basic information

Visible

Individuals remaining after the current time and content filters.

Window

The two boundaries of the active time window around the slider midpoint.

Imported

The size of the loaded AADR map core before current filters.

C14 provenance

Summary of detected direct, C14-like, calendar, and contextual dating evidence.

Active time window

The rows below depend on the selected method. “Calculated” passes the full quality gate; “Limited basis” is mappable but narrower; “No basis” is grey and is not counted as a biological negative.

Marker basis

Directly observed target-marker calls; this is a call sum, not automatically the number of distinct SNPs.

Diploid calls

Calls with enough depth for both alleles. Shallow pseudo-haploid calls do not count as true diploid calls.

Median index

The middle value of evaluable individuals; it is a relative model index, not a phenotype prediction.

Weight basis

The observed share of absolute effect weight. A strong marker therefore counts more than a weak one.

All status numbers describe the visible selection. They are neither a population-genetic estimate nor a smoothed prevalence estimate. Missing markers are not treated as counter-evidence in GeoGens.

A point represents a visible individual. The map uses the published find coordinates; point size, opacity, and colour are determined by the active method. Hovering shows a short summary; clicking opens the detail panel.

Multiple individuals at the same site

Individuals with identical coordinates are deterministically spread around a small ring. The order follows the sample ID; the ring grows with group size and is capped. It is formed across the whole dataset rather than the currently visible selection, so an individual keeps its place when the time window or the filters change, even when only part of a site's ring then remains. This is only a readability aid. The detail panel retains the published coordinates.

There is no general colour hierarchy. In most views points share the same drawing rank, so residual overlaps follow technical draw order. Exceptions are LP-positive points in the lactase view and selected or stronger relationships in the genetic-similarity view.

Display modes

All points

All individuals in the current selection, including grey or non-evaluable cases.

Relevant points

Only points with an evaluable basis for the active method. Hidden points are not historically irrelevant.

Surface only

Point circles are hidden while an active surface remains visible. In genetic similarity, the selection may remain as an orientation.

Colours and timeline

Categorical traits use fixed class colours; continuous indices use a gradient. Grey means missing, blocked, or non-mappable basis and is not a negative class. Optional colouring override can show another trait than the active analysis, so the legend must always be read with the selected method.

The timeline shows filtered, dated individuals across the full available period. “Earliest”, “midpoint”, and “latest” choose an endpoint or midpoint of a range; “full range” uses the full range for visibility and its midpoint for point position. Dense periods are only visually softened, not aggregated.

Data sources

•Allen Ancient DNA Resource (AADR) v66.0, Dataverse DOI 10.7910/DVN/FFIDCW, Release 2026-04-13T04:33:11Z

•AADR v66.1240K.aadr.PUB.anno and v66.1240K.aadr.PUB genotype dataset

•Poseidon Community Archive: 2025_Gretzinger_SlavicPeriod, 2020_Furtwaengler_Switzerland, 2025_SkourtaniotiJia_SCaucasus

•Poseidon Minotaur Archive: 2025_Schuetz_Sarmatians

•PGS Catalog: PGS001092, PGS001093, PGS001094, PGS001095, PGS001096, PGS001098, PGS001897, PGS002135, PGS002804

•UK Biobank Field 1717 Skin colour scale; Tran et al. 2026, UK Biobank Field 1717 distribution

Poseidon imports

•Gretzinger et al. 2025, Ancient DNA connects large-scale migration with the spread of Slavs - https://doi.org/10.1038/s41586-025-09437-6

•Furtwängler et al. 2020, Ancient genomes reveal social and genetic structure of Late Neolithic Switzerland - https://doi.org/10.1038/s41467-020-15560-x

•Skourtanioti et al. 2025, The genetic history of the Southern Caucasus from the Bronze Age to the Early Middle Ages - https://doi.org/10.1016/j.cell.2025.07.013

•Koptekin et al. 2023, Spatial and temporal heterogeneity in human mobility patterns in Holocene Southwest Asia and the East Mediterranean - https://doi.org/10.1016/j.cub.2022.11.034

•Schütz et al. 2025, Unveiling the origins and genetic makeup of the forgotten people: A study of the Sarmatian-period population in the Carpathian Basin - https://doi.org/10.1016/j.cell.2025.05.009

Haplogroups

•YFull/YTree 14.04.0, YFullTeam/YTree current_tree.json, CC BY 4.0

•HaploGrep/PhyloTree 17 Forensic Update 1.3 (rCRS), phylotree-fu-rcrs, MIT

•Dür, Huber & Parson 2021, Fine-Tuning Phylogenetic Alignment and Haplogrouping of mtDNA Sequences - https://doi.org/10.3390/ijms22115747

•van Oven & Kayser 2009, Updated comprehensive phylogenetic tree of global human mitochondrial DNA variation - https://doi.org/10.1002/humu.20921

•Weissensteiner et al. 2016, HaploGrep 2: mitochondrial haplogroup classification in the era of high-throughput sequencing - https://doi.org/10.1093/nar/gkw233

Ancestry components

•Lazaridis et al. 2014, Ancient human genomes suggest three ancestral populations for present-day Europeans - https://doi.org/10.1038/nature13673

•Haak et al. 2015, Massive migration from the steppe was a source for Indo-European languages in Europe - https://doi.org/10.1038/nature14317

•Mathieson et al. 2015, Genome-wide patterns of selection in 230 ancient Eurasians - https://doi.org/10.1126/science.aaa2773

•Hofmanová et al. 2016, Early farmers from across Europe directly descended from Neolithic Aegeans - https://doi.org/10.1073/pnas.1523951113

•González-Fortes et al. 2017, Paleogenomic evidence for multi-generational continuity of ancient European hunter-gatherers - https://doi.org/10.1016/j.cub.2017.05.023

•Mathieson et al. 2018, The genomic history of southeastern Europe - https://doi.org/10.1126/science.aau5085

•Olalde et al. 2018, The Beaker phenomenon and the genomic transformation of northwest Europe - https://doi.org/10.1038/nature25738

•Lazaridis et al. 2022, The genetic history of the Southern Arc - https://doi.org/10.1126/science.abm4247

•Olalde et al. 2022, Large-scale migration into Britain during the Middle to Late Bronze Age - https://doi.org/10.1038/s41586-021-04287-4

•Ghalichi et al. 2024, The rise and transformation of Bronze Age pastoralists in the Caucasus - https://doi.org/10.1038/s41586-024-08113-5

•Lazaridis et al. 2025, The genetic origin of the Indo-Europeans - https://doi.org/10.1038/s41586-024-08531-5

•Nikitin et al. 2025, A genomic history of the North Pontic Region from the Neolithic to the Bronze Age - https://doi.org/10.1038/s41586-024-08372-2

Hair colour

•Walsh et al. 2013, HIrisPlex system for simultaneous prediction of hair and eye colour - https://doi.org/10.1016/j.fsigen.2012.07.005

•Branicki et al. 2011, Model-based prediction of human hair color using DNA variants - https://doi.org/10.1007/s00439-010-0939-8

•Guenther et al. 2014, A molecular basis for classic blond hair color in Europeans - https://doi.org/10.1038/ng.2991

•Wilde et al. 2014, Direct evidence for positive selection of skin, hair, and eye pigmentation in Europeans - https://doi.org/10.1073/pnas.1316513111

•Sulem et al. 2007, Genetic determinants of hair, eye and skin pigmentation in Europeans - https://doi.org/10.1038/ng.2007.13

•Zorina-Lichtenwalter et al. 2019, A study in scarlet: MC1R as the main predictor of red hair - https://doi.org/10.1093/hmg/ddz018

•Hysi et al. 2018, Genome-wide association meta-analysis of individuals of European ancestry identifies new loci explaining a substantial fraction of hair color variation and heritability - https://doi.org/10.1038/s41588-018-0100-5

•Morgan et al. 2018, Genome-wide study of hair colour in UK Biobank explains most of the SNP heritability - https://doi.org/10.1038/s41467-018-07691-z

•Lona-Durazo et al. 2021, CanPath hair-colour GWAS meta-analysis - https://doi.org/10.1038/s42003-021-02764-0

•Tanigawa et al. 2022, Components of genetic associations across 2,138 phenotypes in the UK Biobank highlight adipocyte biology - https://doi.org/10.1371/journal.pgen.1010105

Eye colour

•Walsh et al. 2013, HIrisPlex system for simultaneous prediction of hair and eye colour - https://doi.org/10.1016/j.fsigen.2012.07.005

•Chaitanya et al. 2014, The HIrisPlex system for simultaneous prediction of hair and eye colour from DNA - https://doi.org/10.1016/j.fsigen.2014.04.006

•Sulem et al. 2007, Genetic determinants of hair, eye and skin pigmentation in Europeans - https://doi.org/10.1038/ng.2007.13

•Sulem et al. 2008, Two newly identified genetic determinants of pigmentation in Europeans - https://doi.org/10.1038/ng.160

•Liu et al. 2010, Digital quantification of human eye color highlights genetic association of three new loci - https://doi.org/10.1371/journal.pgen.1000934

•Lona-Durazo et al. 2022, Investigating the genetic architecture of eye colour in a Canadian cohort - https://doi.org/10.1016/j.isci.2022.104485

•Simcoe et al. 2021, Genome-wide association study in almost 195,000 individuals identifies 50 previously unidentified genetic loci for eye color - https://doi.org/10.1126/sciadv.abd1239

•Farré et al. 2023, Skin Phototype and Disease: A Comprehensive Genetic Approach to Pigmentary Traits Pleiotropy Using PRS in the GCAT Cohort - https://doi.org/10.3390/genes14010149

•Abbatangelo, Lona-Durazo, Edwards & Parra 2026, A comparative GWAS of eye colour in light and dark eye genetic backgrounds defined by HERC2 rs12913832 polymorphism in a Canadian cohort of European ancestry - https://doi.org/10.1038/s41598-026-44580-8

Skin pigmentation

•Walsh et al. 2017, Global skin colour prediction from DNA - https://doi.org/10.1007/s00439-017-1808-5

•Walsh et al. 2017, Global skin colour prediction from DNA, Erratum - https://doi.org/10.1007/s00439-017-1817-4

•Chaitanya et al. 2018, The HIrisPlex-S system for eye, hair and skin colour prediction from DNA: introduction and forensic developmental validation - https://doi.org/10.1016/j.fsigen.2018.04.004

•Liu et al. 2015, Genetics of skin color variation in Europeans - https://doi.org/10.1007/s00439-015-1559-0

•Jacobs et al. 2013, Comprehensive candidate gene study highlights UGT1A and BNC2 as new genes determining continuous skin color variation in Europeans - https://doi.org/10.1007/s00439-012-1232-9

•Ju & Mathieson 2021, The evolution of skin pigmentation-associated variation in West Eurasia - https://doi.org/10.1073/pnas.2009227118

•Farré et al. 2023, Skin Phototype and Disease: A Comprehensive Genetic Approach to Pigmentary Traits Pleiotropy Using PRS in the GCAT Cohort - https://doi.org/10.3390/genes14010149

•Kim et al. 2024, East Asian skin-pigmentation GWAS; audited but not used in the Ju/Liu/Jacobs model because no reproducible European effect-weight table was available - https://doi.org/10.1038/s41467-024-49031-4

•Privé et al. 2022, Portability of 245 polygenic scores when derived from the UK Biobank and applied to 9 ancestry groups from the same cohort - https://doi.org/10.1016/j.ajhg.2021.11.008

Lactase persistence, metabolism, muscle type and COMT

•Enattah et al. 2002, Identification of a variant associated with adult-type hypolactasia - https://doi.org/10.1038/ng826

•Fumagalli et al. 2019, Genetic diversity of CHC22 clathrin impacts its function in glucose metabolism - https://doi.org/10.7554/eLife.41517

•Ye et al. 2017, Dietary adaptation of FADS genes in Europe varied across time and geography - https://doi.org/10.1038/s41559-017-0167

•Irving-Pease et al. 2024, The selection landscape and genetic legacy of ancient Eurasians - https://doi.org/10.1038/s41586-023-06705-1

•Garcia-Retortillo et al. 2024, A systematic review and meta-analysis of the association between ACTN3 R577X genotypes and performance in endurance versus power athletes and non-athletes - https://doi.org/10.1186/s40798-024-00711-x

•Egan et al. 2001, Effect of COMT Val108/158 Met genotype on frontal lobe function and risk for schizophrenia - https://doi.org/10.1073/pnas.111134598

•Intelligence (the evidence base is poor, and all analyses attempted to date have failed validation)

•The 30 facets of the Big Five/OCEAN model

•Testosterone

•Risk propensity

•Psychopathy

•Field independence

•Spatial versus verbal intelligence

•Slow life-history strategy

•Aspects of K-selection (bone density, growth rate, lifespan/ageing, sexual activity, parental care, cooperation and altruism)

•Corruption

•Willpower and impulse control

•Agency

•Conformity and collectivism

•Creativity

•Autism

•Aggression

•Novelty seeking

•Neanderthal ancestry proportions

•Blood groups

•Craniometrics (cranial measurements)

•3D facial reconstruction

•Succession trees for archaeological cultures

•Haplogroup trees and charts by culture

•Major events (the flooding of Doggerland, the founding of Rome, Attila, and others)

•Map cultural relationships by identifying culture groups as related where there is an unambiguous succession of material forms

•Integrate GeoGens with archaeological distribution maps

•Enrich GeoGens with anthropological evidence

If you have an idea or would like to contribute, please get in touch: gggeogens.eu

Contributions are especially welcome for culture-group curation, research and integration of new genomes and publications, and the identification of additional SNPs or ideas for new traits.

Long-term access to substantial computing capacity and high-bandwidth infrastructure would also be a major help.