TL;DR
This paper investigates how biased sampling of individuals in genetic datasets can significantly impact the accuracy of population structure and ancestry inference. The authors develop a mathematical framework to model this bias and propose a practical resampling-based correction method that improves accuracy, particularly when sample sizes are not representative of the underlying populations.
What Problem Does This Paper Address?
In genetic studies, understanding population stratification—how genetic variation is structured across different populations—is crucial for both inferring evolutionary history and correcting for confounding factors in association studies. While large public datasets exist, practical constraints often mean that only a subset of individuals from a given population is genotyped. This creates a "sample selection bias" if the sampled individuals don't accurately reflect the true population's genetic diversity. This bias can lead to inaccurate inferences about an individual's ancestry and the overall population structure, potentially undermining the reliability of genetic analyses. The paper aims to quantify this effect and provide a solution.
Methods
The authors tackle this problem by:
- Developing a Mathematical Framework: They propose a probabilistic model to describe sample selection bias, where an individual's inclusion in a dataset is dependent on auxiliary features (like geography) rather than solely on their genotype. This leads to a formulation for the biased distribution of genotypes.
- Investigating Existing Methods: They analyze the impact of this bias on two widely used population structure inference tools:
- ADMIXTURE: A model-based method that uses an explicit admixture model.
- EIGENSOFT: An eigenanalysis-based method that projects genetic data into a lower-dimensional space.
- Proposing a Correction Method: They introduce a resampling-based correction technique. This method involves re-weighting individuals in the selected sample based on their estimated probability of being selected, effectively creating a more representative sample for analysis. This correction can be implemented using auxiliary information like geographical location.
Datasets
- Simulated Data: The primary evaluation was performed using simulated genetic data. This involved creating a two-population admixture scenario (CEU and YRI from HapMap Phase III) and then generating various biased subsets of this data by varying the number of individuals from ancestral and admixed populations. They also simulated demographic scenarios with varying effective population sizes to test the robustness of their findings.
- Real Cattle Genotype Data: To validate their findings on a real-world dataset, they used genotype data from 1461 cattle individuals across different breeds, focusing on admixture between African and European cattle.
Key Insights and Ideas
- Sample Selection Bias is a Real Problem: The study unequivocally demonstrates that biased sampling can significantly degrade the accuracy of both population structure and individual ancestry inference, even with well-established methods like ADMIXTURE and EIGENSOFT.
- Beyond Simple Counts: The accuracy of ancestry inference isn't solely determined by the absolute number of unmixed individuals. The ratio of admixed individuals to unmixed individuals (termed $T_{sample}$) plays a critical role. When this ratio is high ($T_{sample} > 5$), accuracy noticeably drops.
- Effective Population Size Matters: The impact of sampling bias is influenced by a population's effective population size, not just its census size. Populations with smaller effective population sizes (due to bottlenecks, for example) can be more accurately captured even with smaller sample sizes.
- Auxiliary Information is Key for Correction: The proposed correction method relies on auxiliary information (like geography) that is correlated with selection criteria. This highlights the importance of understanding the sampling process itself.
- ADMIXTURE is More Robust to Bias: ADMIXTURE generally showed more stable and predictable results compared to EIGENSOFT when dealing with biased samples, likely due to its underlying likelihood-based model.
Key Results
- Accuracy Degradation: In simulations, when $T_{sample}$ exceeded 5, the squared correlation between true and inferred ancestry dropped significantly. For cattle data, a ratio greater than 1 was sufficient to affect ancestry estimates.
- Correction Effectiveness: The proposed resampling correction method consistently improved accuracy in simulated datasets, often achieving squared correlations above 0.95.
- SNP Count Impact: Increasing the number of SNPs in the analysis (from 52,040 to 116,415) in simulations could mitigate the effects of sample selection bias, suggesting that more genetic information can help overcome sampling issues.
- Effective Population Size Impact: In demographic simulations, a stronger population bottleneck (leading to a lower effective population size) resulted in higher accuracy of ancestry inference even with fewer individuals from that population, demonstrating the importance of effective population size.
Limitations and Future Work
- Requirement for Auxiliary Information: The proposed correction method requires knowledge of auxiliary information correlated with the selection criteria. The accuracy of the correction is dependent on the strength of this relationship.
- Dataset-Specific Correction Factors: Correction factors may be dataset-specific, requiring careful consideration for different datasets.
- Modeling Complex Demographics: The simulations primarily used simple two-population admixture scenarios. Real-world populations often have more complex demographic histories, which could make stratification harder to resolve and exacerbate the effects of sample selection bias.
- Future Directions:
- Develop models that can jointly model population structure and auxiliary factors like geography or language to better capture selection processes.
- Explore subsampling/resampling strategies to assess the robustness of ancestry inference results to variations in sampling when bias is not fully understood.
- Investigate the impact of sample selection bias on local ancestry inference methods.
- Develop weighted eigenanalysis methods to address potential collinearity issues arising from resampling in EIGENSOFT.
Summary updated