MWII01, a low-coverage femur from the Abbey Church of Middelburg: ancestry fit against medieval, Roman and Iron Age references

MWII01 is a low-coverage femur (ENA PRJEB113845) from the Abbey Church of Middelburg, traditionally attributed to William II, Count of Holland and King of the Romans (r. 1247–1256); historians have also named his brother Floris the Guardian, and the attribution is not settled. We describe which reference groups the genome resembles, using 106 qpAdm models fixed before they were run, f₃ and f₄ statistics, Fst among reference groups, a distal model and rescaled PCA distances. A single North Sea medieval source, Netherlands Medieval Saxon, is not rejected (p = 0.108), and no second medieval source is supported when added to it. In a Roman-period frame the only precise model is Denmark Roman Iron Age 90 ± 4 % plus Imperial Rome (Lazio) 10 ± 4 % (p = 0.068); the single-source version is rejected (p = 0.019). Only 4 of 104 two-source models are both feasible and precise. On the rescaled maps MWII01 lies with Low Countries and English medieval groups, and modern French, English and GBR cannot be separated. Fst among 32 reference groups puts Netherlands Medieval Saxon 9–11 × 10⁻³ from ten other North Sea medieval and Viking groups, which differ by at most 3 × 10⁻³ from one another. One low-coverage genome supports a Northwest European medieval profile with no detectable Slavic input and at most a small Mediterranean component; it does not resolve which North Sea group, and it says nothing about whose bones these are.

Sample and reference. MWII01 (femur) was converted from public sequence data¹ to the 1240K SNP set: 388,541 SNPs in the converted file, 369,415 overlapping the AADR v66 reference² (1,233,013 SNPs); between 221,000 and 235,000 SNPs entered each qpAdm model on the AADR frame (70,000 to 214,000 in the distal models). Comparison groups come from the AADR v66 extended frame in a Nexju custom dataset; Belgium Medieval is used without 11 genomes that duplicate libraries of other Belgian groups or are flagged for contamination (32 of 43 kept). The identity of the skeleton is not tested by any analysis here.

qpAdm. Models follow the qpAdm method³ with alpha 0.05, maximum weight SE 10 percentage points (pp) and all-SNPs off. Right panel D (14 populations): Australian, Papuan, Mbuti, Karitiana, Turkey_PPN, Israel_Natufian, Iran_WezmehCave_N, Italy_Sicily_Epigravettian, Turkey_N, Russia_Samara_EBA_Yamnaya, Iran_GanjDareh_N, Georgia_KotiasKlde_Mesolithic, Serbia_IronGates_Mesolithic, Russia_Karelia_Mesolithic; in rotations the candidates not used on the left join the rights. A model is retained when p ≥ 0.05, every weight lies in [0, 1] and every weight SE is below 10 pp. Retention is not a best-model claim: Table 1 shows every model that passes, not a ranking. Seven fixed models (W1–W7), six exhaustive two-source rotations (A, B, Iron Age, C, D, E; six candidates, 15 models each) and an add-one rotation (F; Netherlands Medieval Saxon locked, nine candidates) were specified in writing before each batch was run. In total 106 models were run on this one genome.

Distal model. On the distal frame (Anatolian farmer, Yamnaya, CHG-Zagros, WHG as lefts; 9 to 13 rights) rank 4 and rank 3 models were fitted.

f-statistics. Outgroup f₃(Mbuti; MWII01, X) for 29 groups and f₄(Mbuti, MWII01; X, Y) for 18 contrasts⁴, with block-jackknife SEs.

Fst. Hudson Fst⁵ among 32 reference groups, on the 970,494 SNPs that all of them share, with block-jackknife SEs. Fst is undefined for a single pseudohaploid genome, which has no within-group variance, so MWII01 is not one of the 32 groups; the analysis shows how well the reference groups can be told apart.

Rescaled PCA distances. Two reference panels build the axes (P1: Basque, Sardinian, Russian, Finnish, Adygei, Druze; P2: Palestinian instead of Druze; six components, no shrinkage, no LD pruning). MWII01 and 27 comparison groups are projected only; none is a reference. PC1 and PC2 explain 1.90 % and 1.60 % (P1, 311,543 SNPs) and 1.94 % and 1.49 % (P2, 321,506 SNPs). Distances follow Taleb et al.⁶ as implemented in the Nexju PCA-distance tool: entropy-rescaled maps A and B only (no classical map), four origin shifts and a bootstrap of individuals (2,000 resamples). A statement is stable if it holds in at least 80 % of settings and shifts and in the bootstrap, otherwise it depends. These are descriptions of a two-dimensional projection, not ancestry results, and a low-coverage genome is shrunk toward the centre of the cloud.

Of 104 two-source models, 44 are rejected (p < 0.05), 37 fit but need a weight outside 0–100 %, 19 are feasible but imprecise (SE ≥ 10 pp) and 4 are feasible and precise (Table 1 shows the models that pass, Table 2 gives every rotation). Of the two single-source models, W2 (Netherlands Medieval Saxon, 100 %) is retained (p = 0.108) and W1 (Belgium High Medieval) is rejected (p = 0.0016).

Rotation E gives the only precise Roman-period survivor: Denmark Roman Iron Age pool 89.7 ± 4.4 % with Imperial Rome (Lazio) 10.3 ± 4.4 % (p = 0.068, 230,854 SNPs). Twelve of its 15 models are rejected by p-value, and two fit only with weights far outside 0–100 % (SE 80 and 1,238 pp). The single-source model, Denmark Roman IA alone on the same rights, is rejected (p = 0.019, 231,419 SNPs), so inside this frame the Lazio-like component is needed; it lies 2.3 SE above zero and the model clears alpha by only 0.018, so it is a candidate, not a result.

In the medieval frame, rotations C and D produce survivors only with SE of 24–68 pp (C) and 23–36 pp (D). The two survivors of D put 83–89 % on Iceland Viking with 11–17 % Belgium Early Medieval or Belgium Medieval. In rotation F, which adds one source to Netherlands Medieval Saxon, the precisely estimated added weights are negative or near zero: Niederwunsch Slavic −4 ± 7 %, Basilicata Late Antiquity −4 ± 5 %, Croatia Velim Slavic −11 ± 7 % (rejected, p = 0.049) and Lazio Late Antiquity +4 ± 5 % (rejected, p = 0.026). The two survivors, with Belgium High Medieval (31 ± 29 %) and Estonia Early Medieval (3 ± 40 %) added, are imprecise. The Iron Age rotation has two precise survivors, Denmark Roman IA 73 ± 8 % with Magdalenenberg Hallstatt 27 ± 8 %, and Magdalenenberg 29 ± 7 % with Sweden Iron Age 71 ± 7 % (that group has 8 individuals, 6 flagged for quality); they share one source and are not separated by these data.

Three distal models are retained (Table 3). Natufian as an extra source gives 5 ± 9 % and a rejected model (p = 0.022). The decomposition is roughly one third Anatolian farmer, half Yamnaya and one eighth WHG; CHG-Zagros is not needed. Yamnaya and CHG-Zagros overlap, and their separation was not tested.

Table 4 gives 12 of the 18 f₄ contrasts f₄(Mbuti, MWII01; first, second). A positive Z means that MWII01 is closer to the second group; a difference is read at |Z| ≥ 3.

The Saxon group beats Imperial Rome by a wide margin (Z +13.7). Belgium High Medieval is the outlier: MWII01 is closer to Saxon, Danish and even Niederwunsch Slavic groups than to it, in line with the rejection of W1 (p = 0.0016). Between the North Sea groups themselves f₄ shows no difference.

Outgroup f₃ is flat: for 23 of 29 groups it runs from 0.06814 (Finnish) to 0.06878 (Wielbark), a spread of about 1.1 SE (SE about 0.0006). Lower are Imperial Rome (0.06617), Spanish, N. Italian, Sardinian and Russian, and England Viking, whose value rests on 55,603 SNPs (SE 0.0012).

Fig. 1 shows MWII01 among the 27 projected groups of panel P1 on the rescaled maps. Table 5 lists the groups nearest to MWII01 in each class and Table 6 the contrasts between pairs of groups, each over two panels, maps A and B and four origin shifts.

Fst and f₂ are undefined for one pseudohaploid genome, so MWII01 is not in this analysis; Table 7 shows how well the reference groups can be told apart. Belgium High Medieval, Denmark Viking, England Saxon and Denmark Roman IA lie within 0.8–3.1 of one another (ten North Sea medieval and Viking groups within 2.9). Netherlands Medieval Saxon, the best single-source fit, is 9.5–10.6 from the four, as far as Roman Italy (6.9–12.8) and Niederwunsch Slavic (10.1–12.9).

What the data support. MWII01 has a Northwest European medieval profile: a single Netherlands Medieval Saxon source fits, no Slavic source adds anything, f₄ puts it far from Imperial Rome, and the rescaled maps place it with Low Countries and English medieval groups and with the Iron Age group of Bucy-le-Long.

What they do not support. They do not choose between Saxon, Frisian, Flemish and Danish-like sources, they do not separate modern French, English and GBR, and the Roman-period model with a 10 % Lazio component is a candidate that clears alpha by 0.018 within a set of 15 tests. The single low-coverage genome limits resolution: Netherlands Medieval Saxon and Denmark Viking differ by Fst 9.5 × 10⁻³ (Table 7), yet f₄ does not separate them for MWII01 (Z −0.9). With 106 models on one genome and low power, a model that is not rejected is weak evidence.

Where methods disagree. The rescaled maps put MWII01 nearer to Belgium High Medieval than to Netherlands Medieval Saxon (Q 0.16), while f₄ puts it closer to the Saxon group (Z +3.5) and qpAdm rejects Belgium High Medieval alone. The measures answer different questions: position in a two-dimensional projection of a genome that is shrunk toward the centre of the cloud, against shared drift with an outgroup. Where they disagree, the maps describe position and qpAdm and f₄ test shared drift; the disagreement is not a test of the rescaling.

Identity. The remains in the Abbey Church of Middelburg have been assigned to William II or to his brother Floris the Guardian: an osteological examination in Leiden allowed either, and a reading of the tomb favours Floris⁷. Population-level ancestry cannot separate the two brothers or another male of the House of Holland, who would share the same ancestry. Only kinship, Y-lineage or documented provenance can speak to identity; none of these was tested.

Reproducibility. Each batch of models was specified in writing before it was run; all runs are stored on the Nexju platform (nexju.com).

Magnus Albertus Papers · @AlbertusMvagnus