Machine Learning Reveals Genetic Divide Between Chinese Kirghiz and Kyrgyzstan Sheep

High in the eastern Pamir Mountains, where China meets Kyrgyzstan across some of the most forbidding terrain on Earth, flocks of sheep have been herded for centuries along ancient transhumance routes. A new genomic study published in BMC Genomics has now used a combination of classical population genetics and machine learning to ask a deceptively simple question: how genetically distinct are the Chinese Kirghiz sheep from the local breeds grazing just across the border in Kyrgyzstan? The answer, delivered through nearly 49,000 genetic markers from 722 animals, is nuanced. The sheep are clearly differentiated, yet they also carry the fingerprints of regional connectivity, a finding that carries real weight for conservation planners and livestock breeders alike.

The research team, led by Lingfeng Kong and Wanwan Peng of Tarim University in Xinjiang alongside collaborators at Osh State University in the Kyrgyz Republic, assembled one of the most extensive genomic datasets yet compiled for Central Asian sheep. After rigorous quality control and filtering to remove closely related individuals, the researchers retained 48,902 single nucleotide polymorphisms, the tiny DNA variations that serve as genomic landmarks, from 722 sheep representing 14 distinct populations. These included the focal Chinese Kirghiz breed, three Kyrgyzstan local populations known as Ala-Too, Gissar, and Kyrgyz native sheep, and a broader comparative panel spanning Kazakh, Dolang, Bashibai, Afshari, Chinese Merino, German Texel, Irish Suffolk, Moghani, Norduz, and Tibetan sheep.

The Chinese Kirghiz sheep, which inhabit the eastern Pamir region near the China–Kyrgyzstan border, presented a striking genetic profile. Their observed heterozygosity, a direct measure of the proportion of gene copies carrying different alleles at a given locus, stood at just 0.249, notably lower than their expected heterozygosity of 0.355. This gap between observed and expected values produced an elevated inbreeding coefficient, FIS, of 0.296, signaling a substantial deficit of heterozygotes that likely reflects mating among related individuals within geographically constrained herds. Yet despite this elevated inbreeding signal, the breed retained relatively high nucleotide diversity, measured at π = 0.356, indicating that the overall pool of genetic variation within the population remains far from depleted.

Runs of homozygosity, long stretches of DNA where both chromosomes carry identical variants inherited from a common ancestor, told a more reassuring story. The mean genomic inbreeding coefficient derived from these runs, FROH, was a modest 0.009, suggesting that while recent mating patterns may skew heterozygosity estimates at individual loci, the breed has not experienced the severe, sustained bottlenecks that leave dramatic homozygous tracts across the genome. This distinction between locus-level inbreeding signals and genome-wide runs of homozygosity illustrates why modern conservation genomics increasingly relies on multiple, complementary metrics rather than any single statistic.

Principal component analysis, a dimensionality-reduction technique that projects genetic variation onto orthogonal axes of greatest variance, clearly separated the Chinese Kirghiz sheep from the three Kyrgyzstan local populations. However, complementary ancestry and gene-flow analyses painted a more interconnected picture, indicating that despite the differentiation detected by PCA, genetic material has continued to move across the border region. This apparent tension between differentiation and connectivity is not contradictory; it reflects the reality that populations can maintain distinct identities while still exchanging migrants at low but meaningful rates, particularly in mountainous landscapes where valleys channel both animal movement and human trade.

To quantify the degree of separation, the researchers calculated pairwise fixation index values, or FST, between the Chinese Kirghiz and each Kyrgyzstan local breed using balanced resampling to correct for the severe sample-size imbalance in the original dataset. The balanced estimates came in at 0.099 for Ala-Too, 0.092 for Gissar, and 0.098 for Kyrgyz native sheep. These values fall within the moderate range of genetic differentiation, confirming that the populations are genetically distinguishable but not so divergent as to represent entirely separate gene pools. The consistency of the three estimates, all hovering just below 0.10, suggests a shared regional history with recent but limited gene flow.

Perhaps the most technically innovative aspect of the study was its use of machine learning classifiers, specifically LightGBM and XGBoost, two gradient-boosting frameworks that have become workhorses of modern computational biology. In the full, unbalanced dataset, these classifiers achieved near-perfect discrimination between the Chinese Kirghiz and Kyrgyzstan local sheep, with feature importance rankings generated through SHAP values and LASSO regularization highlighting specific SNPs as candidate markers of population identity. However, when the researchers applied balanced resampling to equalize sample sizes across groups, both classification performance and feature stability dropped noticeably. This sensitivity to sample composition serves as a cautionary tale for the growing number of studies that deploy machine learning on genomic data without adequately testing whether their results are artifacts of unbalanced training sets.

The researchers went further, intersecting regions of low nucleotide diversity in the Chinese Kirghiz genome with elevated pairwise FST values to identify 395 candidate genes potentially involved in local adaptation. Among these were CAMK2A, involved in calcium signaling and neural function; CACNA1C, a calcium channel gene; EPHB1 and ROBO1, both implicated in nervous system development; GRID1 and GRID2, glutamate receptor genes; and RXFP2, a gene well known in livestock genetics for its association with horn morphology. However, when the researchers subjected these candidate genes to Gene Ontology and KEGG pathway enrichment analyses with multiple-testing correction, no term remained statistically significant. This sobering result underscores the difficulty of moving from population-level genomic signals to confirmed functional adaptation.

The study’s most important methodological contribution may be its explicit treatment of the confounding effects of sample-size imbalance and population structure on machine learning results. When the researchers applied principal component adjustment to their association testing, none of the prioritized SNPs remained significant after false discovery rate correction at the threshold of FDR < 0.01. This finding demonstrates that what appears to be a robust machine-learning signal in an unbalanced dataset can evaporate entirely once population structure is properly accounted for. The authors are careful to frame their 395 candidate genes as population-informative markers requiring further validation rather than as confirmed adaptive variants, a level of methodological transparency that is still relatively rare in the application of machine learning to population genomics.

For the pastoralists who have tended Kirghiz sheep across the Pamir for generations, the study provides scientific grounding for what they have long known intuitively: their animals are distinct, shaped by the particular demands of high-altitude grazing, harsh winters, and centuries of selection within a geographically bounded landscape. For conservation biologists, the moderate FST values and evidence of ongoing gene flow suggest that the Chinese Kirghiz and Kyrgyzstan local breeds represent connected but distinguishable components of a broader Central Asian sheep metapopulation, a framing that has direct implications for how breeding programs and genetic resource conservation efforts should be designed across national borders. And for the growing community of researchers applying machine learning to genomic data, the study offers a methodological template: always test your classifiers under balanced conditions, always adjust for population structure, and never mistake a confident prediction for a confirmed biological mechanism.

Subject of Research: Population genomic differentiation between Chinese Kirghiz sheep and Kyrgyzstan local sheep breeds using SNP data and machine learning classifiers

Article Title: Machine learning-assisted genomic characterization of population differentiation between Chinese Kirghiz and Kyrgyzstan local sheep

Article References: Kong, L., Peng, W., Peng, Y., Li, Y., Zhu, L., Yang, R., Erkinbay, A., Kou, F., Liao, Y., Mutalifu, M., Kursantbek, A., Rakhima, M., Roza, A., Abdugani, A., Mukhatai, G., & Liu, S. (2026). Machine learning-assisted genomic characterization of population differentiation between Chinese Kirghiz and Kyrgyzstan local sheep. BMC Genomics. https://doi.org/10.1186/s12864-026-13459-3

Image Credits: AI Generated

DOI: 10.1186/s12864-026-13459-3

Keywords: Chinese Kirghiz sheep, Kyrgyzstan local sheep, population genomics, genetic differentiation, machine learning, SNP markers, FST, runs of homozygosity, LightGBM, XGBoost, conservation genomics, Central Asia