Comparative Analysis of GWAS Models for Cation Exchange Capacity in Oat (Avena sativa L.)

Genome-wide association studies play a key role in the selection of agricultural feed crops across many parameters, including nutrient uptake and stress resistance. The cation exchange capacity (CEC) is an important indicator for soil fertility and nutrient uptake, but its genetic nature in the oat (Avena sativa L.) studied in this paper is currently poorly understood. Open phenotypic and genotypic oat data were used in the study, and four GWAS models (MLM, FarmCPU, BLINK, and BayesCπ) were applied to evaluate their effectiveness in detecting SNP associated with CEC. When evaluating the results, it can be stated that in this work FarmCPU showed the highest sensitivity, MLM - better control of false finds, BLINK - high analysis speed, and BayesCπ gave accurate evaluations of SNP effects. These results can help with the selection of the GWAS model depending on the research goals and resources for subsequent work.

Oats (Avena sativa L.) are an important cereal crop that plays a significant role in the production of feed and food due to their high nutritional value, resistance to adverse conditions, and ability to improve soil structure [1]. In the current conditions of climate change and resource depletion, the selection of oats with improved agronomic characteristics is of particular importance. One of these characteristics is stress resistance and increased absorption of nutrients from the soil [2,3].

One of the key traits associated with plant fertility and the efficiency of element absorption from the soil is the cation exchange capacity (CEC) - a measure of the ability to retain ionic nutrients in the rhizosphere [4]. A high CEC value promotes better absorption of macro- and microelements, such as potassium, calcium, and magnesium, and can be considered a promising indicator for the selection of more productive oat varieties [5,6].

CEC reflects the ability of the plant and the surrounding soil environment to retain cations, and thereby maintain stable and balanced plant nutrition. This is especially important under stressful conditions (e.g., drought, soil acidity), where the availability of nutrients can decrease drastically. Plants with high CEC retain elements in the rhizosphere more efficiently, reduce their leaching, and improve interactions with the microbiota involved in the mobilization of phosphorus and nitrogen [5,6]. In addition, CEC is associated with the morphology of the ion-transport systems of roots, as well as with the general adaptive capacity of the crop to poor soils and extreme conditions [7,8]. Despite such importance, CEC is rarely studied as a target trait in breeding, including due to the complexity of phenotyping and the lack of genetic studies [9]. Given the multifactorial nature of CEC and its dependence on complex genetic interactions, genome-wide association analysis (GWAS) methods are a powerful tool for identifying markers associated with this trait [10]. In recent years, GWAS has been widely applied to various traits in oats, from yield to disease resistance, but research on traits related to ion exchange and interactions with the soil environment remains limited [11,12]. There are many GWAS methods, differing in their approaches to accounting for population structure, power, and resistance to false positives. In this paper, four methods were compared: Mixed Linear Model, FarmCPU, BLINK, and BayesCπ. MLM (Mixed Linear Model) is a classic model that takes into account both genetic relatedness and population structure [13]. FarmCPU is an improved method that combines fixed and random effects to increase the power of the analysis [14]. BLINK is a method that is designed for rapid analysis and increased accuracy by excluding irrelevant markers [15]. BayesCπ is an approach based on Bayesian regression that provides accurate identification of markers with a strong effect [16].

Despite significant progress in the development of GWAS models, questions about which methods are most effective in analyzing CEC-like traits in the context of crop plants remain unanswered. This study aims to compare statistical power and robustness to false positives between methods, assess the impact of accounting for population structure, and identify optimal approaches for further use of GWAS in oat breeding programs.

A publicly available oat (Avena sativa L.) dataset provided by the Global Landrace Collection Project was used for the analysis. Genotypic and phenotypic data were downloaded from the Agricultural Research Service repository. This dataset is linked to the study by Maughan et al. (2019) published in Nature Communications [17].

The genotypic data is in VCF format and contains ~394,000 SNP markers after basic filtering. The phenotype was the cation exchange capacity (CEC) extracted from the soil trait table. This continuous trait reflects the ability of a plant to interact with the mineral part of the soil and assimilate cations.

To control population stratification, principal component analysis (PCA) was performed on the genotypic data. The first two principal components (PC1 and PC2) were included in the GAPIT models as covariates. This allows us to correct the influence of population structure and reduce the probability of false positive associations.

Four methods of association analysis were tested and compared: MLM, FarmCPU, BLINK, and BayesCπ. For the GAPIT models, a common set of SNPs and a single number of PCs (2) were used. In the case of BayesCπ, due to the high computational load, the analysis was performed on a subset of SNPs, while the following filters were observed: minimum minor allele frequency (MAF) ≥ 0.05; exclusion of SNPs with missing values.

Comparison of the efficiency of the models was carried out by the number of significant SNPs, the overlap of the top 100 SNPs between the methods, visualization of the results using Manhattan and QQ plots, and for BayesCπ, histograms of SNP effects and plots of the relationship between the effect, MAF and the proportion of explained phenotypic variance were additionally constructed. The analysis was performed in R 4.x using the following packages: GAPIT v3 for MLM, FarmCPU, and BLINK models; BGLR for BayesCπ analysis; ggplot2, VennDiagram, prcomp() and other standard visualization tools.

A comparison of the number of significant SNPs showed different sensitivity of each of the methods. FarmCPU identified the largest number of significant markers (11), MLM — 7, BLINK — 4. The BayesCπ method, applied to a subset of SNPs, also identified 10 significant positions. These results reflect differences in approaches to controlling type I errors and model sensitivity. Manhattan plots (Figure 1, Figure 2 and Figure 3) show that FarmCPU identified clear peaks, including those with p-values above -log10(p) > 13, BLINK showed single high signals, but in smaller numbers, MLM gave a more uniform distribution of signals, without sharp spikes.

False positive results and False discovery rate were assessed based on the degree of overlap of the top 100 SNPs between three methods. To estimate possible false positives, a Venn diagram of the top 100 SNPs of each method was constructed (Figure 4). The overlap between the methods is extremely limited. MLM and FarmCPU do not share any SNPs, MLM and BLINK share 3 SNPs, FarmCPU and BLINK share 6 SNPs. All three methods together identified only 6 SNPs.

The QQ plot analysis allows us to assess the conformity of the p-value distribution with the expected one in the absence of associations, and thus to judge the presence of signals and possible inflation of Type I errors. FarmCPU (Figure 5a) shows a noticeable deviation from the diagonal in the right part of the plot, which indicates the presence of strong signals and increased sensitivity of the method. Despite the small inflation, the deviation starts quite late, which indicates an acceptable level of false positives. BLINK (Figure 5b) also shows a pronounced deviation from the null hypothesis in the tail of the distribution, but less extensive than that of FarmCPU. This indicates the presence of significant associations, but the number of strong signals is smaller, which is consistent with the number of detected SNPs. MLM (Figure 5c) gives a distribution close to the expected one, with a minimal deviation from the diagonal. This indicates high stringency of the model and a low level of inflation, but may also indicate a loss of sensitivity to weak and moderate effects.

For a more detailed analysis of significant SNPs identified by the FarmCPU method, plots of the minor allele frequency (MAF) were constructed (Figure 6a–c). Most of the highly significant and large-effect SNPs had low MAF, which is typical for rare variants with potential functional significance. One of the SNPs explained up to 8% of the phenotypic variance, indicating its possible role in the control of CEC.

The BayesCπ method allowed us to identify the most informative markers potentially involved in the formation of the CEC trait. 10 SNPs with the highest contribution were identified, for which the effects and the proportion of explained phenotypic variance were calculated (Table 1). The SNP effects ranged from 0.0029 to 0.0032, indicating a moderate influence of each individual locus. At the same time, each of these markers explained from 10.9% to 12.3% of the phenotypic variance, which can be considered highly informative for a complex agronomic factor.

To account for the population structure, Principal Component Analysis was performed. The scree plot (Figure 7) shows that the first two components explain most of the genetic variance (~23% and ~8%, respectively). Based on the results, 2 PCs were included in the model as covariates.

A limitation in this work was the use of BayesCπ only on a subset of SNPs due to the method’s high computational load. Because of this, the analysis was not performed on the entire SNP set (~394K), but only on a limited subset. This does not allow for a direct, full comparison with other methods on the entire dataset. In addition, the analysis was performed only for one trait (CEC), without testing the results on other phenotypes or samples. In the future, it will be possible to test the models on other traits and populations.

In this study, four GWAS methods were compared to identify SNPs associated with cation exchange capacity (CEC) in oats (Avena sativa L.).

The results confirm that the choice of method significantly affects the number, reliability, and interpretability of the associations found. FarmCPU and BayesCπ showed high sensitivity, while MLM provided strict control of false positive results. BLINK, in turn, showed exceptional speed in obtaining results.

The results of this work will help researchers select suitable GWAS models for the analysis of complex traits and contribute to the development of strategies for selecting optimal GWAS models for the analysis of agronomically significant traits.