TL;DR
As human genetics studies increasingly analyze massive sample sizes, the standard coalescent model, a simplification of population genetics, may become inaccurate. This paper shows that for large samples, the discrete-time Wright-Fisher model reveals significant deviations in genealogical properties, particularly concerning rare variants. The authors propose a hybrid approach combining the Wright-Fisher model for recent history and the coalescent for deep history to achieve accurate and efficient analysis.
What Problem Does This Paper Address?
The field of human genetics is experiencing an explosion in sample sizes, with studies now routinely analyzing hundreds of thousands to millions of individuals. This scale presents computational challenges and, more fundamentally, questions the validity of existing theoretical models. The coalescent model, a cornerstone of population genetics for studying ancestry, is derived as a mathematical limit that assumes the population size is much larger than the sample size. When this assumption is violated, as is increasingly common with large samples, the coalescent may no longer accurately reflect key genealogical properties, potentially leading to misinterpretations of population history, especially concerning the abundance of rare genetic variants. This paper aims to quantify these deviations and propose a more robust modeling approach.
Methods
The core of this work involves comparing two models: the standard coalescent and the discrete-time Wright-Fisher (DTWF) model. The DTWF model is a more direct, albeit computationally intensive, representation of a randomly mating population. The authors developed exact deterministic computation methods for the DTWF model, leveraging dynamic programming to calculate various genealogical quantities. To address the computational burden of the full DTWF model, they introduced a hybrid algorithm. This hybrid approach uses the DTWF model for a specified number of recent generations and then switches to the coalescent model for the more distant past. This strategy aims to capture the deviations caused by large sample sizes in recent history while benefiting from the computational efficiency of the coalescent for deep history.
Datasets
Not discussed in the paper. The study uses simulated data based on four demographic models inferred from real human population data.
Key Insights and Ideas
- The "large sample" problem: The coalescent's accuracy hinges on the population size being significantly larger than the sample size. As sample sizes grow, this assumption breaks down, leading to discrepancies.
- Multiple and simultaneous mergers: Unlike the coalescent, which only allows for pairwise mergers of lineages at any given time, the DTWF model, especially with large samples, exhibits a significant number of multiple and simultaneous mergers. These events are crucial for understanding recent population history and the generation of rare variants.
- Rare variants are key: The paper highlights that the most noticeable differences between the coalescent and DTWF models appear in the expected number of rare variants. This is particularly relevant because recent studies have observed an excess of rare variants in human populations, which the coalescent might not fully explain.
- Hybrid approach is effective: The proposed hybrid method, using DTWF for recent generations and coalescent for older ones, demonstrates a remarkable ability to approximate the DTWF model's results with much greater computational efficiency. Even a small number of DTWF generations can significantly improve accuracy.
Key Results
- For large sample sizes (e.g., $n = 2 \times 10^4$) and realistic demographic scenarios (especially those involving recent population growth), the DTWF model shows a substantial number of multiple and simultaneous lineage mergers, which are absent in the coalescent.
- Noticeable differences exist in the expected number of rare variants between the coalescent and DTWF models, with the DTWF model predicting a higher number of singletons (rare variants with frequency < 0.01%) by up to 11% when sample size approaches population size.
- The hybrid algorithm, with as few as 5 recent DTWF generations, can achieve frequency spectrum predictions very close to those of the full DTWF model, reducing the deviation from the coalescent by a substantial margin (e.g., from 11% to < 1% for singletons).
- For models with rapid recent population growth (Models 3 and 4), the deviation between coalescent and DTWF predictions becomes more pronounced as sample size increases, underscoring the need for the DTWF model or similar approaches.
Limitations and Future Work
- Computational cost: While the DTWF model provides a more accurate representation, its exact computation is computationally expensive, especially for large sample sizes and long time scales. The hybrid method is a step towards mitigating this.
- Generalizability of hybrid approach: The paper demonstrates the effectiveness of the hybrid method for computing the frequency spectrum. Future work could explore its application to other demographic inference methods and its performance across a wider range of demographic models and other random mating models beyond DTWF.
- Demographic inference: The authors suggest that their hybrid method could be a valuable tool for demographic inference, allowing for more accurate estimates of population history, especially concerning recent rapid growth, by better accounting for the dynamics of large samples. This is identified as a key area for future research.
Summary updated