CoChat

An ancestral recombination graph of human, Neanderthal, and Denisovan genomes

Open access

The DOI matches its Crossref record; 95 of the 95 references it cites were found in a registry or an index; 42 of the 67 sources its claims cite were checked for support.

Abstract

Many humans carry genes from Neanderthals, a legacy of past admixture. Existing methods detect this archaic hominin ancestry within human genomes using patterns of linkage disequilibrium or direct comparison to Neanderthal genomes. Each of these methods is limited in sensitivity and scalability. We describe a new ancestral recombination graph inference algorithm that scales to large genome-wide datasets and demonstrate its accuracy on real and simulated data. We then generate a genome-wide ancestral recombination graph including human and archaic hominin genomes. From this, we generate a map within human genomes of archaic ancestry and of genomic regions not shared with archaic hominins either by admixture or incomplete lineage sorting. We find that only 1.5 to 7% of the modern human genome is uniquely human. We also find evidence of multiple bursts of adaptive changes specific to modern humans within the past 600,000 years involving genes related to brain development and function.

Overview

CoChat Summary

## TL;DR

SARGE is a heuristic, parsimony-based method for inferring ancestral recombination graphs from phased modern and archaic genomes, and the authors use it to separate archaic admixture from incomplete lineage sorting (ILS). Applying it to 279 modern human genomes plus two Neanderthals and one Denisovan, they estimate that only about 1.5% to 7% of the autosomal genome is uniquely human and find human-specific regions enriched for genes involved in neural and RNA-splicing functions.

## Key findings

  • On simulated data, SARGE achieved 78.93% specificity (95% CI, 78.09 to 78.95%) but lower 25.36% sensitivity (95% CI, 25.32 to 25.40%), reflecting its design choice to leave uncertainty as polytomies.
  • In the human data, the authors estimate that about 7% of the autosomal genome is free of both archaic admixture and ILS, while only 1.5% of the assayed genome qualifies as “human-specific regions” after requiring a high-frequency, human-specific derived allele.
  • The human-unique regions are enriched for genes, exons, and regulatory element binding sites; deserts are enriched for genes and regulatory element binding sites, while admixed regions are depleted of genes.
  • They infer a recent mean TMRCA to Neanderthal of about 74 ka ago across Neanderthal-introgressed segments in non-African populations, and they identify two apparent bursts of human-specific mutations centered around 600 ka ago and 200 ka ago.

## Methods

  • Developed SARGE (Speedy Ancestral Recombination Graph Estimator), a heuristic, parsimony-guided ARG inference algorithm based on the four-gamete test, shared derived alleles, and inferred ancestral recombination events.
  • Analyzed 279 phased human genomes from the SGDP, plus two high-coverage Neanderthal genomes and one high-coverage Denisovan genome; ancient variants were recalled with snpAD, and the dataset was filtered and polarized using the chimpanzee genome.
  • Evaluated SARGE with simulation-based QC (panmictic, structured, bottleneck, and demographic/admixture simulations) and then used the inferred ARG to map archaic admixture, ILS, and human-specific regions, plus to estimate mutation ages and functional enrichment.

## Limitations

  • The paper states that SARGE systematically underestimates the true number of ancestral recombination events because the four-gamete test undercounts breakpoints; therefore it is not well suited for fine-grained recombination maps.
  • The authors note that their branch-length and mutation-age estimates rely on an assumption of constant local mutation rate around each locus since human/chimp divergence; they explicitly say this could bias TMRCA estimates downward if mutation rates differ across lineages.
  • Evident from the text: the method has low sensitivity and produces polytomies, so some true relationships and admixed segments are likely missed; the authors also warn that phasing limitations in archaic genomes and some populations may affect haplotype block lengths.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1126/sciadv.abc0776

Cites

95 of 95 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Data structure and performance of SARGE on simulated data. (A) Schematic of data structure. Top: Rectangles are “tree nodes” representing clades in trees. Each clade has member haplotypes (shown with letters A to G) and a start and end coordinate (blue numbers in brackets) determined by coordinates of single-nucleotide polymorphism (SNP) sites tagging the clade (yellow numbers in braces), along with a propagation distance parameter (100 in this example). Parent/child edges (vertical arrows) also have start and end coordinates determined by the nodes. Ovals are candidates for clades sharing an ancestral recombination event that can explain four-gamete test failures; colored edges indicate potential paths between tree nodes through candidate nodes that could explain four-gamete test failures (colors indicate types of paths). The candidate node with the most edges (here, AB) is eventually chosen as the most parsimonious branch movement, allowing for the inference of new nodes. The two trees at the bottom show the “solved” ancestral recombination event with the branch movement marked in red and all clades inferred without SNP data marked with yellow stars (haplotypes A and B share an ancestral recombination event; their ancestry is shared with haplotypes C, D, and G upstream of the recombination event and haplotype E downstream of it). The coordinates of the recombination event (blue numbers in brackets) are taken to be midway between the highest-coordinate upstream site (left side) and the lowest-coordinate downstream site (right side) involved in recombination. For a more detailed overview of the data structure, see figs. S3 to S5. (B) Accuracy of SARGE on simulated data (defined as percent of all clades correct according to the true ARG in the simulation), with increasing numbers of human-like haplotypes from an unstructured population. Error bars are one SD across five replicates. (C) Number of nodes per tree with increasing number of haplotypes in simulated data.
Fig. 1 Data structure and performance of SARGE on simulated data. (A) Schematic of data structure. Top: Rectangles are “tree nodes” representing clades in trees. Each clade has member haplotypes (shown with letters A to G) and a start and end coordinate (blue numbers in brackets) determined by coordinates of single-nucleotide polymorphism (SNP) sites tagging the clade (yellow numbers in braces), along with a propagation distance parameter (100 in this example). Parent/child edges (vertical arrows) also have start and end coordinates determined by the nodes. Ovals are candidates for clades sharing an ancestral recombination event that can explain four-gamete test failures; colored edges indicate potential paths between tree nodes through candidate nodes that could explain four-gamete test failures (colors indicate types of paths). The candidate node with the most edges (here, AB) is eventually chosen as the most parsimonious branch movement, allowing for the inference of new nodes. The two trees at the bottom show the “solved” ancestral recombination event with the branch movement marked in red and all clades inferred without SNP data marked with yellow stars (haplotypes A and B share an ancestral recombination event; their ancestry is shared with haplotypes C, D, and G upstream of the recombination event and haplotype E downstream of it). The coordinates of the recombination event (blue numbers in brackets) are taken to be midway between the highest-coordinate upstream site (left side) and the lowest-coordinate downstream site (right side) involved in recombination. For a more detailed overview of the data structure, see figs. S3 to S5. (B) Accuracy of SARGE on simulated data (defined as percent of all clades correct according to the true ARG in the simulation), with increasing numbers of human-like haplotypes from an unstructured population. Error bars are one SD across five replicates. (C) Number of nodes per tree with increasing number of haplotypes in simulated data.CC BY-NC · PubMed Central
Fig. 2 Performance of SARGE on SGDP and archaic hominin dataset. (A) Pairwise coalescence times for randomly sampled sets of up to 10 pairs of phased genome haplotypes per population (every possible pair was considered for archaic hominins, since fewer genomes were available). Values are calibrated using a 13-Ma human-chimp divergence time (see Supplementary Methods) and averaged across every variable site in the dataset, error bars show one SD, and branch shortening values for archaic samples were incorporated into calculations using mean values reported in (1). The lower value for humans comes from removing archaic-admixed clades from trees. (B) UPGMA (unweighted pair group method with arithmetic mean) trees computed using nucleotide diversity from SNP data (top and left) against similarity matrix from shared recombination events inferred by SARGE. Light yellow boxes (similar groups) are Native Americans and Papuans. (C) Average similarity between Orcadian haplotypes in the SGDP panel and other European haplotypes calculated on the basis of the number of shared ancestral recombination events. The best matches are in England, Iceland, and Norway, as expected.
Fig. 2 Performance of SARGE on SGDP and archaic hominin dataset. (A) Pairwise coalescence times for randomly sampled sets of up to 10 pairs of phased genome haplotypes per population (every possible pair was considered for archaic hominins, since fewer genomes were available). Values are calibrated using a 13-Ma human-chimp divergence time (see Supplementary Methods) and averaged across every variable site in the dataset, error bars show one SD, and branch shortening values for archaic samples were incorporated into calculations using mean values reported in (1). The lower value for humans comes from removing archaic-admixed clades from trees. (B) UPGMA (unweighted pair group method with arithmetic mean) trees computed using nucleotide diversity from SNP data (top and left) against similarity matrix from shared recombination events inferred by SARGE. Light yellow boxes (similar groups) are Native Americans and Papuans. (C) Average similarity between Orcadian haplotypes in the SGDP panel and other European haplotypes calculated on the basis of the number of shared ancestral recombination events. The best matches are in England, Iceland, and Norway, as expected.CC BY-NC · PubMed Central
Fig. 3 Neanderthal ancestry inferred in modern humans. (A) Genome-wide percent Neanderthal, Denisovan, and ambiguous (either Neanderthal or Denisovan) across SGDP populations, using the ARG and an estimator based on the D statistic. D statistic calculations considered only one archaic population at a time as introgressor and thus do not detect ambiguous ancestry and also might count some Denisovan ancestry as Neanderthal and vice versa. (B) For individual phased human genome haplotypes (points), mean TMRCA with Neanderthal in Neanderthal haplotype blocks (y axis) and mean Neanderthal haplotype block length (x axis) for all Neanderthal-introgressed haplotypes. TMRCA calculations assume a total of 6.5-Ma human-chimpanzee divergence and branch shortening values from (1), with a mutation rate of 1 × 10−9 per site per year. bp, base pair. (C) Overall number Neanderthal haplotype blocks versus geographically restricted (unique to a 3000-km radius) Neanderthal haplotype blocks. (D) Same as (B), but limited to geographically restricted (unique to a 3000-km radius) Neanderthal haplotype blocks. Only haplotypes with more than 10 geographically restricted segments are shown.
Fig. 3 Neanderthal ancestry inferred in modern humans. (A) Genome-wide percent Neanderthal, Denisovan, and ambiguous (either Neanderthal or Denisovan) across SGDP populations, using the ARG and an estimator based on the D statistic. D statistic calculations considered only one archaic population at a time as introgressor and thus do not detect ambiguous ancestry and also might count some Denisovan ancestry as Neanderthal and vice versa. (B) For individual phased human genome haplotypes (points), mean TMRCA with Neanderthal in Neanderthal haplotype blocks (y axis) and mean Neanderthal haplotype block length (x axis) for all Neanderthal-introgressed haplotypes. TMRCA calculations assume a total of 6.5-Ma human-chimpanzee divergence and branch shortening values from (1), with a mutation rate of 1 × 10−9 per site per year. bp, base pair. (C) Overall number Neanderthal haplotype blocks versus geographically restricted (unique to a 3000-km radius) Neanderthal haplotype blocks. (D) Same as (B), but limited to geographically restricted (unique to a 3000-km radius) Neanderthal haplotype blocks. Only haplotypes with more than 10 geographically restricted segments are shown.CC BY-NC · PubMed Central
Fig. 4 Extent of deserts of archaic hominin ancestry. (A) Left: In the SGDP dataset, fractions of the genome where any individual has archaic ancestry (admixture), where any individual is incompletely sorted with archaic hominin lineages but where there is no archaic admixture (ILS), where there is no evidence of either admixture or ILS with archaic hominins (deserts), and where there is a fixed derived allele private to and shared by all humans (human-specific regions). Right: The same values from a simulated dataset with single pulses of Neanderthal and Denisovan admixture, both with an admixture proportion (2%) that produced reasonable amounts of archaic ancestry per individual genome (fig. S47C). (B) For random subsamples of the SGDP dataset, along with the full dataset, fractions of the genome comprising deserts, deserts filtered for candidate archaic alleles using another dataset, and human-specific regions are shown. The points on the far right (full dataset) correspond to the desert and ILS bars in the left panel of (A).
Fig. 4 Extent of deserts of archaic hominin ancestry. (A) Left: In the SGDP dataset, fractions of the genome where any individual has archaic ancestry (admixture), where any individual is incompletely sorted with archaic hominin lineages but where there is no archaic admixture (ILS), where there is no evidence of either admixture or ILS with archaic hominins (deserts), and where there is a fixed derived allele private to and shared by all humans (human-specific regions). Right: The same values from a simulated dataset with single pulses of Neanderthal and Denisovan admixture, both with an admixture proportion (2%) that produced reasonable amounts of archaic ancestry per individual genome (fig. S47C). (B) For random subsamples of the SGDP dataset, along with the full dataset, fractions of the genome comprising deserts, deserts filtered for candidate archaic alleles using another dataset, and human-specific regions are shown. The points on the far right (full dataset) correspond to the desert and ILS bars in the left panel of (A).CC BY-NC · PubMed Central
Fig. 5 Approximate ages and consequences of fixed human-specific derived mutations. (A) Ages of candidate human-specific functional mutations (nonsynonymous substitutions and mutations within regulatory element binding sites) inferred by SARGE within desert regions (free of ILS and admixture between archaic hominins and modern humans). (B) For interacting pairs of genes in the STRING database (35) for which candidate human-specific functional mutations affect both genes, the ages of the two mutations are shown. (C) For each candidate human-specific functional mutation, the length of the surrounding desert region (x axis) and inferred mutation age (y axis) are shown. Mutations for which SARGE did not infer a date (mutations within CpG sites or for which the ancestral allele was unknown) are shown in the bottom panel. Mutations were scored on the basis of length of desert and age; genes with regulatory mutations in the top 99.5th percentile of this distribution, or nonsynonymous mutations in the top 95th percentile, are shown, with nonsynonymous changes in bold.
Fig. 5 Approximate ages and consequences of fixed human-specific derived mutations. (A) Ages of candidate human-specific functional mutations (nonsynonymous substitutions and mutations within regulatory element binding sites) inferred by SARGE within desert regions (free of ILS and admixture between archaic hominins and modern humans). (B) For interacting pairs of genes in the STRING database (35) for which candidate human-specific functional mutations affect both genes, the ages of the two mutations are shown. (C) For each candidate human-specific functional mutation, the length of the surrounding desert region (x axis) and inferred mutation age (y axis) are shown. Mutations for which SARGE did not infer a date (mutations within CpG sites or for which the ancestral allele was unknown) are shown in the bottom panel. Mutations were scored on the basis of length of desert and age; genes with regulatory mutations in the top 99.5th percentile of this distribution, or nonsynonymous mutations in the top 95th percentile, are shown, with nonsynonymous changes in bold.CC BY-NC · PubMed Central

Verification

Claims the paper makes

31 Valid · 10 Not valid · 14 Not checked · 139 Own results and interpretations

94 more claims · 9 with their cited sources checked

Create a free account to see all 194 claims and to verify the 8 references still open.

See all claims — free account
  • It may state two separate claims as one. 67
  • Probably not a substantial claim. 48

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 95 references

Cited by

Items 101–104 of 104

Metadata

Authors
Schaefer, N. K., Shapiro, B., & Green, R. E.
Affiliations
  • Howard Hughes Medical Institute (US)
  • University of California, Santa Cruz (US)
Corresponding author
Green, Richard E.
Venue
Science Advances
Published
2021-07-16
Volume
7
Issue
29
ISSN
2375-2548 (ISSN-L)
OpenAlex
W3182837943
Cited by
106
References
95
Citations by year
2026: 11 · 2025: 9 · 2024: 57 · 2023: 18 · 2022: 7 · 2021: 4
FWCI
81.49
Type
journal article
Language
en
Open access
Locations
  • Science Advances (publishedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
  • eScholarship (California Digital Library) (submittedVersion, cc-by-nc)
  • eScholarship (California Digital Library) (submittedVersion)
Funders
  • National Science Foundation (DEB-1754451)
  • National Institutes of Health (T32 HG00834, HG00834)
  • Howard Hughes Medical Institute
Topics
  • Social Sciences › Anthropology › Pleistocene-Era Hominins and Archaeology
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Forensic and Genetic Research
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Phylogenetic Studies
Keywords
ancestral recombination graph, Neanderthal admixture, Denisovan admixture, human genome, incomplete lineage sorting, linkage disequilibrium, adaptive evolution, brain development genes, modern human genomes
MeSH terms
Animals, Humans, Recombination, Genetic, Hominidae, Genome, Human, Genomics, Neanderthals

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.