Optimized Use of Low-Depth Genotyping-by-Sequencing for Genomic Prediction Among Multi-Parental Family Pools and Single Plants in Perennial Ryegrass (Lolium perenne L.)

Fabio Cericola, Ingo Lenk, Dario Fè, Stephen Byrne, Christian S. Jensen, Morten Greve Pedersen, Torben Asp, Just Jensen, Luc JanssFrontiers in Plant Science 2018 · DOI

TL;DR

This paper tackles the challenge of using low-cost, low-coverage genotyping-by-sequencing (GBS) data for genomic prediction in perennial ryegrass. They developed a novel method to correct for biases in genomic relationship matrices (GRMs) caused by low sequencing depth, significantly improving predictive accuracy and enabling the integration of diverse breeding materials.

What Problem Does This Paper Address?

Genomic prediction (GP) holds immense promise for accelerating plant breeding by predicting an individual's genetic merit from its DNA. However, a major bottleneck is the cost of high-density genotyping. Genotyping-by-sequencing (GBS) offers a cost-effective solution, especially when performed at low coverage depths. The problem is that low coverage leads to increased missing data and, crucially, introduces biases into the estimation of allele frequencies. These biased allele frequencies, when used to construct genomic relationship matrices (GRMs), inflate estimates of inbreeding and lead to inaccurate variance estimates, ultimately reducing the accuracy of genomic prediction models like GBLUP. This paper aims to address this by developing methods to accurately use low-coverage GBS data for GP, specifically in perennial ryegrass, which involves various breeding materials like single plants and family pools.

Methods

The authors developed a two-pronged approach:

  1. Bias Correction for GRMs: They derived theoretical equations to quantify the bias in GRM diagonal elements (which reflect inbreeding) introduced by binomial sampling of sequence reads at low coverage depths. This method accounts for any ploidy level, making it applicable to different breeding materials. They then developed a method to correct for this bias by scaling down the inflated diagonal elements.
  2. Imputation of Missing Data: Recognizing that low coverage also increases missing data, they evaluated three imputation methods suitable for allele-frequency data: mean imputation (MNi), k-nearest neighbor imputation (kNNi), and random forest imputation (RFi). These methods were chosen for their map-independent nature, crucial given the lack of a high-quality reference genome for ryegrass.

These methods were first tested on simulated data to validate the bias correction and imputation strategies, and then applied to real-world perennial ryegrass breeding data.

Datasets

The study utilized two main sources of data:

  1. Simulated Data: Genomic data was simulated for three types of perennial ryegrass breeding material: single plants (SPs), F2 families (pools), and synthetic varieties (SYNs). This simulated data was crucial for understanding the magnitude of GRM bias at different coverage depths and for verifying the bias-correction method.
  2. Real Data: Phenotypic and genotypic data were obtained from a commercial perennial ryegrass breeding program. This dataset included:
  3. 1,225 single plants (SPs)
    • 1,791 bi-parental F2 families (genotyped as pools)
    • 127 multi-parental synthetic varieties (SYNs, genotyped as pools)

These real data were genotyped using three different GBS assays with varying coverage depths and missing data fractions, allowing for a comprehensive evaluation of the proposed methods. Phenotypic traits evaluated included heading date, crown rust resistance, and seed yield.

Key Insights and Ideas

  • Low-coverage GBS is inherently biased: The study provides strong theoretical and empirical evidence that low sequencing depth in GBS introduces a systematic bias in GRMs, specifically inflating the diagonal elements (overestimating inbreeding). This is a critical realization for anyone using low-cost GBS data for genomic analyses.
  • Bias correction is more impactful than imputation: While imputation methods improved predictive accuracy, the bias correction for GRMs had a more significant positive effect, highlighting the fundamental importance of accurate relationship estimation.
  • A unified GRM is feasible: The developed bias-correction method allows for the creation of a single, unbiased GRM that can effectively combine diverse breeding materials (single plants and family pools) for genomic prediction. This is a major step towards more integrated breeding programs.
  • Imputation matters, but with caveats: While random forest imputation (RFi) showed the highest accuracy among the tested methods, the gain from imputation was less pronounced than the gain from bias correction. This suggests prioritizing bias correction when dealing with low-coverage data.

Key Results

  • Bias Correction: The bias correction method effectively reduced the inflation of GRM diagonal elements across all breeding materials and coverage depths, bringing them closer to expected values. For instance, in real data, the average diagonal value for F2 families decreased from 1.81 to 1.35 after correction.
  • Imputation Accuracy: RFi achieved the highest imputation accuracy (R² = 0.77), followed by kNNi (R² = 0.73) and MNi (R² = 0.5).
  • Predictive Ability (PA):
    • Bias-corrected GRMs consistently led to higher PAs compared to standard GRMs across all breeding materials and traits.
  • The combination of RFi imputation and bias correction yielded the highest PAs, with improvements of up to 0.09 in some scenarios.
  • Predicting across different breeding materials (e.g., predicting SPs using F2 family data) was demonstrated to be feasible and effective.

Limitations and Future Work

  • Computational Cost of RFi: While RFi provided the best imputation accuracy, it was significantly more computationally intensive than MNi and kNNi. The authors note that kNNi offers a good balance between accuracy and computational efficiency.
  • Reference Genome Availability: The lack of a high-quality, chromosome-scale reference genome for ryegrass limited the imputation methods to haplotype-independent approaches. Future work could explore map-based imputation if such a resource becomes available.
  • Cost vs. Accuracy Trade-off: The paper highlights the ongoing challenge of finding the optimal balance between the cost of genotyping (lower coverage) and the accuracy of genomic prediction. This will be a key consideration for the widespread adoption of these methods in breeding programs.

The authors suggest that their methods provide a pathway for breeders to utilize low-cost GBS data more effectively, potentially reducing genotyping costs and increasing the number of individuals that can be included in training datasets, thereby accelerating breeding progress.

Summary updated

Abstract

Ryegrass single plants, bi-parental family pools, and multi-parental family pools are often genotyped, based on allele-frequencies using genotyping-by-sequencing (GBS) assays. GBS assays can be performed at low-coverage depth to reduce costs. However, reducing the coverage depth leads to a higher proportion of missing data, and leads to a reduction in accuracy when identifying the allele-frequency at each locus. As a consequence of the latter, genomic relationship matrices (GRMs) will be biased. This bias in GRMs affects variance estimates and the accuracy of GBLUP for genomic prediction (GBLUP-GP). We derived equations that describe the bias from low-coverage sequencing as an effect of binomial sampling of sequence reads, and allowed for any ploidy level of the sample considered. This allowed us to combine individual and pool genotypes in one GRM, treating pool-genotypes as a polyploid genotype, equal to the total ploidy-level of the parents of the pool. Using simulated data, we verified the magnitude of the GRM bias at different coverage depths for three different kinds of ryegrass breeding material: individual genotypes from single plants, pool-genotypes from F<sub>2</sub> families, and pool-genotypes from synthetic varieties. To better handle missing data, we also tested imputation procedures, which are suited for analyzing allele-frequency genomic data. The relative advantages of the bias-correction and the imputation of missing data were evaluated using real data. We examined a large dataset, including single plants, F<sub>2</sub> families, and synthetic varieties genotyped in three GBS assays, each with a different coverage depth, and evaluated them for heading date, crown rust resistance, and seed yield. Cross validations were used to test the accuracy using GBLUP approaches, demonstrating the feasibility of predicting among different breeding material. Bias-corrected GRMs proved to increase predictive accuracies when compared with standard approaches to construct GRMs. Among the imputation methods we tested, the random forest method yielded the highest predictive accuracy. The combinations of these two methods resulted in a meaningful increase of predictive ability (up to 0.09). The possibility of predicting across individuals and pools provides new opportunities for improving ryegrass breeding schemes.
BibTeX
@inproceedings{cericola2018optimized,
  author = {Fabio Cericola and Ingo Lenk and Dario Fè and Stephen Byrne and Christian S. Jensen and Morten Greve Pedersen and Torben Asp and Just Jensen and Luc Janss},
  title = {{Optimized Use of Low-Depth Genotyping-by-Sequencing for Genomic Prediction Among Multi-Parental Family Pools and Single Plants in Perennial Ryegrass (Lolium perenne L.)}},
  booktitle = {Frontiers in Plant Science},
  year = {2018},
  doi = {10.3389/fpls.2018.00369},
}
Figure