Europe teaches the standard recipe; South Asia teaches the proxy problem; the Middle East teaches collinearity — what happens when the candidate sources are themselves relatives. The region's four founding streams share deep structure, exchange gene flow from the Neolithic onward, and produce the classic pathology: wild weights, huge errors, models that pass while meaning little. Modelling here is a discipline of making the streams distinguishable before asking about them.
The four streams#
The Epipalaeolithic-to-Neolithic Near East resolves into four deeply divergent but interrelated ancestries:
- Natufian / Levant_N — the Epipalaeolithic Levant and its early farmers.
- Anatolia_N (Barcın) — northwest Anatolian farmers, Europe's EEF source.
- Iran_N (Ganj Dareh) — early Zagros farmers.
- CHG (Kotias/Satsurblia) — Caucasus hunter-gatherers, Iran_N's closest deep relative.
Iran_N and CHG are the notorious pair — close enough that naive models cannot split them — with Natufian↔Anatolia_N the second-tightest edge. Later prehistory then stirs the pot continuously, which is why this region, more than any other, runs on the distal discipline: all four sources predate the mixing they explain.
The standard recipe#
Left: target + two to four of Levant_N (or Natufian), Anatolia_N, Iran_GanjDareh_N
and Georgia_CHG — chosen by geography, and never Iran_N and CHG together in a first model:
lowest rank first means starting two-source (Levantine targets:
Levant_N + Iran_N; Anatolian/Mesopotamian: Anatolia_N + Iran_N or + CHG) and letting
the rank test decide, relative to the right set, whether a further
source is needed. Testing the lowest rank first is statistical parsimony: the simplest model that
fits, relative to the right set used; it is not a head count of the migrations in anyone's history.
Right: the region's models live or die on the o9a-style extension of
the O9 spine: Mbuti.DG, Ust_Ishim.DG,
Mota.DG, MA1, Ami.DG, Onge.DG, plus the discriminators — Russia_EHG, WHG,
Morocco_Iberomaurusian (the North African edge, essential for Levantine and Egyptian-adjacent
targets), and where the source list allows it, whichever of the four streams sits out of the
left set. This is the published case where the machinery visibly rewards craft: the Bronze Age
Levant work found one added differentially-related outgroup cut standard errors roughly
threefold — the exact anti-collinearity lever, applied.
Later-era targets (Bronze Age onward, and all moderns) usually need a steppe-related fourth
column (Yamnaya/Sintashta-related, for Anatolian, Iranian-plateau and Levantine-diaspora
targets) and, for the Arabian peninsula and Horn-adjacent targets, an African stream
(Mota.DG-related or Egypt-adjacent proxies) — each addition paid for by
a nested-model rejection, never added on vibes.
The named failure modes#
- The Iran/CHG seesaw. Both in the left set with a generic right set: weights trade against each other run to run, SEs balloon, sometimes one goes negative. Fix from the right (EHG, steppe-related and Anatolia contrasts split them best), or accept the honest merged reading — "Iranian/Caucasus-related" as one stream — when the data sit below the resolution floor.
- Missing North Africa. Levantine and coastal targets modelled without an Iberomaurusian-related term fail — or worse, pass with distorted Natufian weights (Natufians themselves relate to North African lineages; the right set must be able to see that edge).
- Umbrella labels.
Iran_Npooling Ganj Dareh with later Chalcolithic plateau samples, or Levant labels pooling PPNB with Bronze Age — check the .anno; era-pure source labels are worth more here than anywhere. - Modern reference temptation. Using present-day populations as sources for other moderns, the proximal shortcut, gives up temporal order: on simulated admixture-graph histories (Flegontova et al. 2025), proximal screens on the rotating protocol had median false-discovery rates of 72.5 to 100%. On the same study's simulated stepping-stone landscapes the protocols compared did not differ significantly, and neither result is an error rate for real studies. The region's dense ancient record makes the distal frame available; use it.
A worked reading#
A Lebanese-ancestry target: p = 0.24, Levant_N 0.58 ± 0.027, Iran_N 0.28 ± 0.031, Anatolia_N 0.09 ± 0.025, Yamnaya-related 0.05 ± 0.014; nested three-source models rejected; ~680k SNPs; Iran_N/CHG swap re-run shifts weights within one SE (stability check passed). Reading: compatible and consonant with the published Bronze-Age-Levant-plus-steppe-trickle picture; the Anatolia term sits barely two SEs from zero, so the honest report flags it as the model's softest claim; and the swap test is the region's signature move — a result that survives proxy rotation is the only kind worth publishing here.
Terms used here are defined in the glossary.
References#
- Lazaridis, I. et al. (2016). Genomic insights into the origin of farming in the ancient Near East. Nature, 536, 419–424. (The four-stream framework and O9.)
- Agranat-Tamir, L. et al. (2020). The genomic history of the Bronze Age Southern Levant. Cell, 181(5), 1146–1157. (o9a and the threefold SE reduction.)
- Feldman, M. et al. (2019). Late Pleistocene human genome suggests a local origin for the first farmers of central Anatolia. Nature Communications, 10, 1218.
- Haber, M. et al. (2017). Continuity and admixture in the last five millennia of Levantine history. American Journal of Human Genetics, 101(2), 274–282.
- Flegontova, O. et al. (2025). Performance of qpAdm-based screens for genetic admixture on graph-shaped histories and stepping stone landscapes. Genetics, 230(1), iyaf047.
