A preliminary analysis of human DNA found a surplus of closely linked variant pairs where both changes appeared in many sampled chromosome copies, compared with a widely used model of population history. Researchers Matthew D. Schwartz and Michael M. Desai report that adding gene conversion, a short, nonreciprocal DNA-copying process, explained much of the mismatch in their model.
The team measured a two-site frequency spectrum: how often nearby DNA sites carry variants in each combination of sample counts. They analyzed 5.7 billion pairs from 113 Gambian participants in the 1000 Genomes Project. The strongest mismatch involved a pair of high-count variants present in six and seven of eight sampled chromosome copies; that combination appeared 4.8 times as often in the spectrum as the model predicted.
The comparison used the Kingman coalescent, a common model for how sample lineages trace back to shared ancestors, across a range of tested population histories. None fit the observed two-site spectrum, and models allowing multiple ancestral lineages to merge at once also failed. Adding gene conversion brought most of the spectrum closer to the data, though some differences remained.
Anthropic's guest post says Claude helped adapt and extend computational methods, while Desai directed the work toward mutation correlations. The paper also reports checks using Yoruba data and an independently sequenced and processed HGDP dataset; these are analyses within the same study, not independent replications. Marked preliminary, the paper leaves unresolved differences, including an estimated short-tract conversion rate below direct pedigree measurements.
Sources
- Schwartz and Desai, “Signatures of gene conversion in the human two-site frequency spectrum” (preliminary manuscript)
- Matthew D. Schwartz, Anthropic guest post on Claude-assisted scientific work
- Peer-reviewed review of gene conversion and crossover
- Image: Mikael Häggström, “Conversion and crossover,” Wikimedia Commons, CC BY-SA 3.0 (illustrative diagram, not study data).
