henhaobio
Sign in

Science AdvancesEvolutionary BiologyPeer-reviewed journal article

Ancient regulatory evolution shapes individual language abilities in present-day humans

Lucas G. CasteniD, Tanner KoomariD, Taylor ThomasiD, Jin‐Young KohiD, Dabney Hofammann, Savantha Thenuwara, Allison M. MomanyiD, Marlea O’Brien, Jeffrey Clark MurrayiD, J. Bruce TombliniD, Jacob J. MichaelsoniD

PDFDOIPreprintCodeData

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 4 (14 Feb 2026) · CC BY 4.0

Polygenic score variants in HAQERs (Human Ancestor Quickly Evolved Regions, <0.1% of the genome) predict spoken language but not nonverbal IQ; language associations replicate in SPARK, ABCD and UK Biobank.

Rare FOXP2 mutations cause speech and language disorders, yet common FOXP2 variants do not explain typical language differences, and polygenic GWAS models leave open when in human evolution language-relevant regulatory variation arose. An evolutionary stratified polygenic score (ES-PGS) addresses this by splitting individual-level polygenic scores by the evolutionary origin of the sequence and testing each slice against matched control regions, not only the rest of the genome.

Key results

  1. ShownIn 350 sequenced children, the HAQER portion of a cognitive performance polygenic score (CP-PGS) was linked to four of seven language factors, most strongly sentence repetition, but not to nonverbal IQ; Human Accelerated Regions (HARs) gave no comparable signal.

    Sentence repetition (F1): r = 0.23, ES-PGS β = 0.18, p = 1.2 × 10^−4, FDR 0.004; nonverbal IQ (F3): β = 0.06, p = 0.19. Background plus matched-control scores (about 300,000 SNPs) explained 3.7% of F1 variance; adding 1,763 HAQER SNPs raised it to 7.7%, stated as 188 times more predictive power per SNP. HARs on F1: β = -0.05, p = 0.27. · Human-specific genomic regions predict…

  2. ShownThe HAQER language signal replicated, at modest significance, in SPARK, ABCD and UK Biobank, with no association to nonverbal measures in SPARK (nonverbal IQ) or UK Biobank (matrix reasoning); UK Biobank used a word reading polygenic score (WR-PGS) because the cognitive GWAS included its participants.

    SPARK: speaking in phrases β = 0.05, p = 0.008 (N = 29,266); language diagnoses in non-autistic parents p = 7.9 × 10^−5 vs psychiatric p = 0.58 (N = 713); verbal IQ β = 2.08, p = 0.022 vs nonverbal β = 0.59, p = 0.49 (N = 620). ABCD word recall β = 0.24, p = 0.048 (N = 5,625). UK Biobank digit recall β = 0.01, p = 0.002 (N = 84,433); matrix reasoning p = 0.44 (N = 26,051). · HAQERs influence language ability across…; ES-PGS replication in the UK Biobank

  3. SuggestedRare variants that restore the human-chimpanzee ancestral allele in HAQERs are linked to poorer language, and predicted hominin-gained binding of Homeobox and Forkhead transcription factors in HAQERs tracks sentence repetition scores.

    SPARK WGS: more rare HAQER reversions, more developmental language disorder (β = 0.16, p = 6.5 × 10^−4), no link to walking age or intellectual disability. EpiSLI (N = 350) motif integrity vs F1: HAQERs β = 0.14, p = 5.6 × 10^−4; HARs β = 0.01, p = 1; random regions β = 0, p = 1. Concordant motifs enriched for Homeobox (OR 11.58, p = 2.6 × 10^−34) and Forkhead (OR 7.28, p = 5.3 × 10^−6). · HAQERs influence language ability across…; HAQERs evolved stronger binding affinity for…; Joint Selection-Language Enrichment Analysis

  4. SuggestedIn ancient west Eurasians the background CP-PGS rose with time but the HAQER CP-PGS stayed flat; with excess heterozygosity in HAQERs and an ABCD link to both cognition and birth complications, the authors propose balancing selection via the obstetric dilemma.

    3,244 genomes dated 18,775 to 150 years ago: background selection coefficient 0.088 (p = 2.1 × 10^−12) vs HAQER -0.004 (p = 0.71). Heterozygosity in EpiSLI: HAQERs vs HARs t = 110, vs matched controls t = 147. ABCD canonical correlation: HAQER CP-PGS vs birth complication composite r = 0.69, vs cognitive composite r = 0.66, flagged by the authors as possibly inflated. · Selective pressures acting on language and…; Evidence of balancing selection from modern…; HAQERs link language evolution to the obstetric…

  5. SuggestedAmong 170 non-primate mammals, vocal learners have sequence more similar to human HAQERs than non-learners, whereas HAR-like similarity does not differ; HAQER-like similarity also tracks brain mass and relative birth weight.

    Phylogenetic logistic regression, 49 vocal learners vs 121 non-learners: HAQER-like β = 1.41, p = 1 × 10^−4; HAR-like β = -0.15, p = 0.67. Brain size β = 0.42, p = 0.006 (116 species); birth:adult weight ratio β = 0.44, p = 0.001 (115 species). HAQERs enriched near 50 vocal learning enhancers, p = 0.028 (HARs p = 0.04). · HAQER-like sequences show convergent evolution…; Defining mammalian vocal learning enhancer…; Convergent evolution of vocal learning analysis

Methods, data and reproducibility

ES-PGS splits a polygenic score into annotation, genome-background and matched-control parts (built with PRSet) and uses nested-model ANOVA to test whether the annotation adds prediction. It was run on five primate and human annotations against seven factor-analysed language and IQ scores, then backed by rare-variant, motif, ancient DNA, enrichment and cross-species analyses.

  • EpiSLI: community-based Iowa cohort of children with 17 language and cognitive assessments (kindergarten to 4th grade) and high-coverage (at least 20X) whole-genome sequencing · human · n = 350 unrelated Europeans (390 sequenced) · dbGaP phs002255.v1.p1
  • SPARK autism cohort: imputed SNP arrays, questionnaire language and psychiatric diagnoses, clinical IQ, and whole-genome sequencing for rare reversion counts · human · n = N > 30,000 (arrays); WGS N > 2,000 (max N = 11,545)
  • ABCD adolescent cohort (v4.0 release): imputed arrays with cognitive, MRI and birth phenotypes · human · n = 5,625 unrelated
  • UK Biobank imputed v3 genotypes, unrelated white British participants (verbal working memory, educational attainment, matrix reasoning) · human · n = 84,433; 370,830; 26,051 by trait
  • GWAS summary statistics for cognitive performance (primary PGS) and word reading (UK Biobank PGS) · human · n = N ≈ 250,000 and N ≈ 30,000
  • Allen Ancient DNA Resource v54 ancient west Eurasians, compared with 1000 Genomes Europeans · human · n = 3,244 ancient; 503 present-day
  • Archaic hominin genomes (AADR release for PGS; high-coverage VCFs from the Max Planck Institute for Evolutionary Anthropology for motif analysis) · Neanderthal and Denisovan · n = 10 in AADR (8 Neanderthals, 2 Denisovans); 4 high-coverage
  • Whole-genome alignment of > 400 species to hg38 (UCSC cactus447way), with PanTHERIA brain mass and birth weight data · multiple mammalian species · n = 170 non-primate mammals (49 vocal learners, 121 non-learners)
Methods & tools
ES-PGS (authors' custom code, GitHub), PRSet, LDpred2, nullranges, bedtools, PLINK, GCTA, bcbio with GATK, FreeBayes and Platypus, Ensembl VEP v109, JASPAR2020, gaston (lmm.aireml), phylolm
Software & environment
R 4.3.1; bcbio v1.1.6, BWA-mem v0.7.17, GATK v4.1.2.0, FreeBayes v1.1.0.46, Platypus v0.8.1.2, FastQC v0.11.8, VEP v109, bedtools v2.26.0. No container or requirements file is described.
Reproducibility gaps
Individual-level EpiSLI, SPARK, ABCD and UK Biobank data are available only to approved researchers. The methods point to the repository for annotation coordinates and matched regions through an unfilled placeholder ('LINK XX').
Limitations (6)
  • Replication effects were modest (UK Biobank p = 0.002, ABCD p = 0.048), which the authors attribute to briefer language measures in the replication cohorts than the standardized EpiSLI battery. · stated by the authors
  • Polygenic scores transfer poorly across populations, so the higher HAQER CP-PGS in Neanderthals and Denisovans (mean z-score 0.91, N = 10) needs cautious interpretation and further validation. · stated by the authors
  • The ABCD canonical correlations may be inflated by the optimization procedure; the authors base their inference on the qualitative pattern rather than the effect sizes. · stated by the authors
  • The excess of intermediate-frequency variants in HAQERs could also come from elevated mutation rates rather than balancing selection. · stated by the authors
  • The 3.7% to 7.7% variance gain and the 188-fold per-SNP figure come only from the 350-child discovery cohort, where HAQERs were picked out among five annotations; they are not reported for any replication cohort. · noted in this summary
  • Changes in transcription factor binding are predicted from JASPAR2020 position weight matrices; no binding or reporter assay tests them. · noted in this summary

Read this if you study the genetics or evolution of speech and language, or want a template for partitioning polygenic scores by evolutionary annotation against matched control regions.

Citations
5
FWCI
136.07
Field percentile
top 1%

Abstract

Language is a defining feature of our species, yet the genomic changes enabling it remain poorly understood. Despite decades of work since FOXP2s discovery, we still lack a clear picture of which regions shaped language evolution and how variation contributes to present-day phenotypic differences. Using a novel evolutionary stratified polygenic score approach in nearly 40,000 individuals, we find that Human Ancestor Quickly Evolved Regions (HAQERs) are specifically associated with language but not general cognition. HAQERs evolved before the human-Neanderthal split, giving hominins increased binding of Forkhead and Homeobox transcription factors, and show balancing selection across the past 20,000 years. Remarkably, language variants in HAQERs appear more prevalent in Neanderthals and have convergently evolved across vocal-learning mammals. Our results reveal how ancient innovations continue shaping human language.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMartins '18Goh '26Xia '22Gallardo '23Gokhman '20Benítez-Bur '17
60% semantic matchGenetics
Pedro Tiago Martins, Maties Marí, Cedric A. Boeckx·Journal of Language Evolution·Jan 1, 2018· 16 citations