Razib Khan's Unsupervised Learning 10,000 years of selection (in Western Eurasia)
May 1, 2026
A deep look at a Nature paper that mapped directional selection across Western Eurasia over 10,000 years. Discussion covers why the paper provoked strong media reaction and controversy. The methods are explained, including models that control population structure and massive variant testing. Results highlight many selection signals, with emphasis on immune, blood, and metabolic shifts and notable allele trajectories over time.
AI Snips
Chapters
Transcript
Episode notes
Individual-Level Models Reveal Hidden Selection
- Ancient DNA can detect selection by modelling allele-frequency change over time while accounting for individual relatedness rather than just population branches.
- Akbari et al. used a GLMM on ~15,836 West Eurasian samples and ~9.7M variants to infer selection across 10,000+ years with greater sensitivity than classical sweep tests.
Long Road From Preprint To Nature Publication
- Razib recounts knowing about the Akbari et al. work years earlier and discussing it with Ali Akbari at conferences, highlighting the paper's long development.
- He mentions the preprint (Sept 2024), lengthy review (accepted Mar 2026), and Robert-level delays like the project spanning many years.
Massive Temporal Panel Powers Allele Trajectories
- Akbari et al. combined ~15,836 ancient samples (≈10,000 new) with 6,438 modern genomes to scan selection on ~8M SNPs and 1,665 indels across autosomes.
- This dense temporal panel gives power to estimate allele trajectories over the last 10,000 years and detect selection episodes missed by older methods.