Comparison of additive, additive-dominant and machine learning (LightGBM) models for genomic prediction of corn hybrids

doi:10.21203/rs.3.rs-3588755/v1

Download PDF

Research Article

Comparison of additive, additive-dominant and machine learning (LightGBM) models for genomic prediction of corn hybrids

https://doi.org/10.21203/rs.3.rs-3588755/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The choice of the statistical method for estimating genomic breeding values in genome-wide selection (GWS) studies is essential to obtain high predictive accuracies. The goal of this study was to evaluate and compare the performance of parametric models containing Additive (rrBLUP) and Additive-Dominant (BL) effects, in addition to a non-parametric model based on Machine Learning (LightGBM). Such models were applied to the genomic selection of corn hybrids evaluated in three locations (Jataí-GO, Rolândia-PR and Sorriso-MT) for two important traits, grain yield and moisture. Through the results it was demonstrated that the BL model presents excellent stability in terms of predictive capacity, however its computational performance in model training is considerably lower than the rrBLUP and LightGBM methods. If computational time is a bottleneck for the development of the genomic selection study, the LightGBM model, which has high computational efficiency, can be used, however this use may imply a significant loss of predictive accuracy. Furthermore, it was observed that the heritability of traits affects model prediction accuracy, and that genetic effects arising from smaller heritabilities can be captured more efficiently with the use of models that incorporate additive-dominant effects or Machine learning models.

Accuracy

additive-dominant

heritability

LightGBM

machine learning

genome-wide selection

With the development of molecular markers, it was possible not only to study the genome of several species, but also to identify genomic regions associated with the expression of important traits. In plant breeding, molecular information can contribute significantly to obtaining greater genetic gains than when using only phenotypic data (Meuwissen et al. 2001). More recently, with the advancement of genotyping technologies and the emergence of SNP markers (single nucleotide polymorphism - SNP), thousands of markers can be genotyped in each corn inbred, allowing the application of genomic selection studies in this crop (Bernardo 2008).

Predicting the performance of corn hybrids without the need to cross in the field, makes it possible to evaluate hundreds of thousands of hybrids virtually, based on their predictions, and select only those with the best predictions to be crossed and evaluated in the field. As a result, the tendency is for more high-performance hybrids to be evaluated in the field (Windhausen et al. 2012). The usefulness of the model applied in genomic selection depends on the prediction accuracy, which is obtained through the correlation between the predicted and observed phenotypes in the validation population (Kaler et al. 2022). There are numerous factors that affect the prediction accuracy of a genomic selection model, among which we can highlight: the size of the training population, markers density, heritability of the trait, number of QTls involved for the trait expression and the statistical model used (Grattapaglia and Resende 2011).

Genomic selection can be carried out using different statistical approaches, ranging from linear regression, non-linear regression, and Bayesian methods to machine learning (ML) methods (Tong and Nikoloski 2021). In general, statistical methods provide different assumptions about the number and distribution of the effects of QTLs, presenting different approaches to the genetic model associated with the trait studied (Neves et al. 2012). The ideal method for GWS must accommodate the genetic architecture of the trait, in addition to considering problems of multicollinearity and dimensionality of the markers used in the model (Neves et al. 2012). One of the most used methods in genomic selection is Ridge Regression–Best Linear Unbiased Prediction (rrBLUP), which has a simpler assumption and assumes that the markers explain the same proportion of genetic variance, generally its application is adopted using only additive effects (Endelman 2011; Wang et al. 2018). Another parametric linear model that has been applied in genomic selection is the Bayesian Lasso (BL). This methodology allows fitting a high-dimensional linear regression model, including dense molecular markers, pedigree information and other covariates (Pérez et al. 2010). BL can incorporate additive and dominance effects in its implementation (Wellman and Bennewitz 2012), which can increase prediction accuracy, especially for species that present heterosis, such as hybrids (Ishimori et al. 2020).

Non-parametric ML methods have been applied quite successfully in genomic selection (Wang et al., 2018). These methods are capable of automatically learning data patterns and optimizing model parameters (Ma et al. 2014). Unlike conventional genomic selection methods, artificial intelligence-based methodologies do not require a priori assumptions about the relationship between inputs (markers) and output (phenotypic values), allowing great flexibility to deal with different genetic architectures, including additive effects, dominant and epistasis (Yan et al. 2021). Another advantage of ML algorithms is their excellent ability to deal with large data sets through automatic variable selection (Yan et al. 2021). One of the ML algorithms that has been explored with very promising results in genomic selection studies is Gradient Boosting (GB) (Ghafouri-Kesbi et al. 2017; Abdollahi-Arpanahi et al. 2020). GB is a cluster learning method that consists of bringing together several weak models to establish a more robust and complex model (Montesinos-López et al. 2022). One of the GB variants – LightGBM (Light Gradient Boosting Machine) has demonstrated superior performance in dealing with large data sets, offering high training speed and predictive accuracy (Yan et al. 2021). LightGBM is an open-source framework that was developed by Microsoft and is based on decision tree algorithms. According to Yan et al. (2021) this model presents excellent performance in terms of accuracy, stability, and computational efficiency for GWS studies. Despite the predictive and computational potential of this model, there are few studies on the performance of this algorithm in plant genomic selection, especially for quantitative traits in corn hybrids.

For hybrids, the effect of dominance deviation is expected to be an important component for the performance of the predictive model. In this sense, comparing the accuracy between additive, additive-dominant and machine learning models can be fundamental to define the best methodology for studying genome-wide selection. Choosing the best prediction model for each situation will promote greater efficiency in breeding programs and save resources through the right selection on truly superior hybrids.

In this study, 11,197 single-cross corn hybrids were evaluated from crosses between inbred lines from different heterotic groups from the germplasm bank of the company LongPing High-Tech, Brazil. The hybrids were evaluated in the 2022 harvest in three different locations: 3,985 hybrids in Jataí-GO, 1,811 in Rolândia-PR and 5,401 in Sorriso-MT. The experimental design was Federer Augmented Blocks, with 32 treatments in each block. Each block had 4 check hybrids, which were repeated in all blocks. The experimental plots were designed with 4 rows of 5m with a spacing of 50cm between the rows, and three plants per linear meter, corresponding to a density of sixty thousand plants per ha. The grain yield (GY) and grain moisture (GM) traits were evaluated at the harvest. As all plots were harvested on the same date, harvest moisture is related to the maturity cycle of the hybrids.

The heritability estimation for the two traits (GY and GM) were obtained from a subpopulation of hybrids evaluated at each location, and this subpopulation was composed of two replications of each hybrid. Broad-sense heritability was estimated using the mean square estimators of variance analysis, using the GENES statistical program (Cruz 2013). DNA extraction from parental lines (5682 Lines) was performed from leaf tissues, using the Fast ID Genomic DNA Extraction® Kit (GENETIC ID., Fairfield, IA, USA), according to the manufacturer's instructions. Sample genotyping was performed on the Ion GeneStudio S5 Prime platform (Thermo Fisher Scientific, Waltham, MA, USA) following the manufacturer's instructions. The parental lines were genotyped with a panel of 5076 SNP markers. The genotypes of the hybrids were obtained by concatenating the genotypes of the parents of each hybrid. Three models were used to estimate the effects of markers and predict the hybrids phenotypic performance:

rrBLUP (Ridge Regression-Best Linear Unbiased Prediction – Additive Model), which calculates the effects of markers by the following regression: y = Xb + Zm + e

Where, y is the vector of phenotypic observations; b is the fixed effects vector; m is the vector of random effects of the brands; e is the vector of random residues; X and Z are the incidence matrices for good respectively. The rrBLUP model used in this work only considers additive effects when predicting VGG. Thus, the Z matrix was coded by values − 1, 0 and 1, with − 1 and 1 being homozygotes and 0 being heterozygotes. This model was implemented using the R package “rrBLUP” (Endelman 2011).

BL (Bayesian Lasso - Additive-Dominant Model), which can be described as follows: y = 1n µ + Zµ + ε

Where, y is the vector of observations, Z is the incidence matrix for random effects, and µ is the vector of random effects (Liu et al., 2018). The BL model considers additive and dominance effects when predicting VGG. Thus, the Z matrix was coded as -1, 0 and 1 to consider additive effects, with values − 1 and 1 being homozygous and 0 being heterozygous. For dominance effects, the values were 1 for heterozygotes and 0 for homozygotes. This model was implemented using the R package “BLR” (De los Campos et al. 2013).

LightGBM (Light Gradient Boosting Machine): The construction of the LightGBM model was carried out using the lightgbm package version 3.3.4 using the R language (Shi et al. 2022). Genotype coding for this model was carried out using the values − 1, 0 and 1, with − 1 and 1 being homozygous and 0 being heterozygous.

The predictive ability of the three models was evaluated using the 3-fold cross-validation procedure (Montesinos-López et al. 2022). In this procedure, the phenotypic data set from each evaluated site was randomly divided into 3 groups of similar sizes, with one group representing the validation population and two groups representing the training population. This process was repeated three times, so that each group was the validation group once. In practice, 27 prediction models were generated (3 replications x 3 evaluation sites x 3 methods).

The prediction accuracy of each model was measured by the average of the simple correlation (Pearson correlation) between the observed and predicted phenotypes obtained from the three cross-validations for the three evaluated locations (Jataí-GO, Rolândia-PR and Sorriso-MT). To determine whether there were statistical differences between the mean predictive accuracy of the models evaluated, analyzes of variance were performed using the Tukey test at 5% probability, considering each location and traits evaluated.

The second parameter used to evaluate the predictive capacity of the models was T10Recall. This parameter indicates the proportion of hybrids that were simultaneously among the 10% of the best predictions and 10% of the best results in the field evaluation. Estimates of this parameter were obtained by averaging the percentages calculated from the three cross-validations for the three locations evaluated.

The computational performance of the models was evaluated through the time spent (seconds) to estimate the effects of markers for the two traits evaluated. All models were generated from an Intel Core i7-8550 computer, containing @1.80GHz and 8GB of RAM.

The descriptive statistics and heritability of the two traits evaluated (GY and GM) for the three locations (Jataí-GO, Rolândia-PR and Sorriso-MT) are described in Table 1. The heritability values for GY varied between 31.23% and 73.39%, and for GU between 63.37% and 87.15%. The place that presented the highest heritability estimates was Rolândia, with 73.39% for GY and 87.15% for GM. Jataí-GO and Sorriso-MT presented heritabilities below 50% for GY (31.23% and 34.62% respectively) and were equivalent for GM (64.89% and 63.37 respectively). Figure 1 presents the results obtained from cross-validation (3-fold) for the two traits in each evaluation environment. Each boxplot graph illustrates the correlation values between the predicted and observed phenotypes in each of the three models evaluated (BL, LightGBM and rrBLUP). The predictive ability of each model was evaluated through the average of the prediction accuracies obtained from the cross-validation process. The average prediction accuracy of the BL, LightGBM and rrBLUP models for the two traits evaluated in the three locations is described in Table 2. The average predictive accuracy for the evaluated locations ranged from 0.404 to 0.685 for GY, and 0.653 to 0.781 for GM, with the highest averages being obtained for Rolândia-PR.

Table 1

Descriptive statistics and heritability estimates for the Traits of Grain Yield (GY) and Grain Moisture (GM) in the three evaluation sites.
	Jataí-GO		Rolândia-PR		Sorriso-MT
Descriptive statistics	GY	GM	GY	GM	GY	GM
Average	9.65	21.19	8.10	24.40	11.81	20.54
Variance	1.34	3.39	2.36	9.17	2.52	4.54
Minimum	3.20	13.40	2.72	14.10	5.45	10.20
Maximum	16.41	30.30	12.57	38.90	17.24	29.20
Heritability (%)	31.23	64.89	73.39	87.15	34.62	63.37

Table 2

Average of prediction accuracies obtained from K = 3 cross-validation for the three models in this study (BL, LightGBM and rrBLUP) in the three evaluation locations. Means followed by the same letter in the column do not differ statistically at a 5% probability level using the Tukey test.
	Jataí-GO		Rolândia-PR		Sorriso-MT
Models	GY	GM	GY	GM	GY	GM
BL	0.519a	0.694a	0.680a	0.781a	0.420a	0.694a
LightGBM	0.487b	0.653b	0.685a	0.759a	0.404ab	0.681a
rrBLUP	0.477b	0.676ab	0.605b	0.762a	0.386b	0.677a

For Jataí-GO, the BL model presented the highest prediction accuracy for the two traits evaluated, having values of 0.514 for GY and 0.694 for GM. These accuracies were significantly higher than the accuracy values of the other two methods, using the Tukey test at the 5% probability level. Still there, it was observed that the LightGBM and rrBLUP models had equivalent predictive accuracy values, both for GY and GM. For Rolândia-PR, the BL and LightGBM models had the highest GY predictive accuracies (0.680 and 0.685 respectively) and differed statistically from the rrBLUP method. For GM, the models had equivalent predictive performance. For Sorriso-MT, there was no significant difference between the models evaluated for GM, however for GY, the BL and LightGBM models presented the highest predictive accuracy values, being 0.420 and 0.404 respectively. Regarding the traits evaluated, it was observed that the prediction accuracy for GM was greater than the prediction accuracy for GY in the three models, and for the three locations evaluated.

The T10Recall parameter indicates the percentage of correspondence between the data obtained in the prediction and the data observed in the field for the 10% of the most productive hybrids for GY, and the 10% earliest for GM. For this parameter, it was observed that the BL model presented the highest percentages of GY in the three locations evaluated, being 34%, 38% and 28% for Jataí-GO, Rolândia-PR and Sorriso-MT respectively, and in the latter location its percentage was the same as that obtained by the LightGBM model. For GM, the BL model had the highest percentage for Jataí-GO (46%) and Rolândia-PR (68%). The LightGBM model presented the lowest percentages of GY and GM for Jataí-GO (32% and 40% respectively) and for Rolândia-PR (36% and 62% respectively), but this model had the highest T10Recall values for Sorriso-MT, with 28% for GY and 43% for GM.

The rrBLUP model showed 33% correspondence for GY and 45% for GM in Jataí-GO and had the lowest GY and GM values for Sorriso-MT (25% and 39% respectively). For Rolândia-PR, the rrBLUP model presented the lowest percentage for GY (36%) and was the model with the highest percentage obtained for GM, together with the BL model (66%).

Figure 2 describes the computational time (seconds) spent to train the models for the two traits (GY and GM) at each evaluation location. The values were obtained from the average time spent in each cross-validation process (3-fold). The computational time of the BL model was considerably greater than the other two methods in all evaluated locations. This model took 8,214, 3,186 and 9,282 seconds to train the GY and GM models for Jataí-GO, Rolândia-PR and Sorriso-MT respectively. The Machine Learning model (LightGBM) presented the shortest computational time for all evaluated locations, being 12.42, 10.04 and 17.64 seconds for Jataí-GO, Rolândia-PR and Sorriso-MT respectively. The computational performance of the rrBLUP model in model training was intermediate to the BL and LightGBM models, presenting values of 319.53, 298.19 and 362.41 seconds for Jataí-GO, Rolândia-PR and Sorriso-MT respectively.

The computational time spent on genetic breeding value (GBV) prediction, and not only in the model training phase, is also a very important factor in choosing the method to be used in genomic selection. This is because this step is constantly carried out in the routine of corn genetic breeding programs for early selection of the best hybrid combinations. In this sense, the time spent (seconds) was estimated to obtain the predicted phenotypes values from the models built for GY and GM in the Jataí-GO cross-validation.

In this estimate, the time spent predicting phenotypic values ranged from 1.04 to 53.48 seconds. The longest computational time for GBV prediction was for the rrBLUP model (53.48) and the shortest time was obtained for the LightGBM model (1.04), the BL model was intermediate to the two models (22.40 seconds). The estimated time for predicting GBV in each model is shown in Fig. 3.

In the present study, the genomic prediction performance of two linear parametric methods (BL and rrBLUP) and a non-parametric Machine learning method (LightGBM) were compared. The comparative analysis was carried out using the genomic selection approach in corn hybrids from crosses between inbred lines from the company LongPing High-Tech in the locations of Jataí-GO, Rolândia-PR and Sorriso-MT in Safrinha 2022. Two important traits of corn crops were evaluated in this study: Grain Yield (GY) and Grain Moisture (GM). It is worth mentioning that moisture at harvest is used as an estimator of physiological maturity (Sala et al., 2007), since all plots were harvested on the same date.

The criteria used to compare the methods were: Prediction accuracy, defined as the correlation between observed and predicted phenotypic data; T10Recall, which indicates the proportion of hybrids from the validation population that were among the 10% most productive for GY and 10% precocious for GM, considering both prediction and field evaluation; in addition to computational performance in training GY and GM models.

The prediction accuracies obtained from the BL, LightGBM and rrBLUP models for the two traits evaluated in the three locations (Jataí-GO, Rolândia-PR and Sorriso-MT) were medium-high, ranging from 0.386 to 0.680 for GY, and 0.653 to 0.781 for GM (Table 2). Similar results were found by Terraillon et al. (2023), which using data from corn hybrids evaluated within the same year obtained predictive accuracies that ranged from 0.300 to 0.710 for GY and 0.490 to 0.830 for GM.

The highest estimates of predictive accuracy for GY and GM were observed for Rolândia-PR. Sorriso-MT and Jataí-GO had the lowest average accuracy for GY and GM respectively. In general, for the two traits evaluated in the three locations, it was observed that with the highest heritabilities, the predictive accuracies were also higher. This behavior is expected and is in line with the work reported by Liu et al. (2018), in which, working with genomic selection for grain yield in corn, they found that heritability has a positive correlation with prediction accuracy. Also, in agreement with the results found here, Hayes et al. (2010) showed that by increasing heritability from 10–90%, predictive accuracy was also increased from 0.35 to 0.75.

Considering the three locations evaluated, it was observed that the highest prediction accuracies for GY were obtained by the additive-dominant model (BL), and the lowest accuracies were obtained by the additive model (rrBLUP). The average range between these two models in terms of predictive ability for GY was 0.050.

For GM, the BL model also had the highest predictive accuracies compared to the additive model (rrBLUP), but the difference was not statistically significant. These results indicate that the genetic effect of dominance considered in the BL model contributed more effectively to the prediction of GY. Very similar results were found in the study carried out by Ferrão et al. (2020), who comparing additive and additive-dominant models in the genomic selection of corn hybrids, found greater prediction accuracies from models incorporating the dominance effect, especially for the grain yield trait.

Because of heterosis in maize hybrids, the incorporation of non-additive effects in genomic prediction models can contribute considerably to increase prediction accuracy (Fritsche-Neto et al. 2021), but this superior performance also depends on the genetic architecture and complexity of the trait of interest. This was demonstrated by Bernardo (1996), who, when evaluating the prediction accuracy for different traits in corn, found that the inclusion of the dominance effect in the modeling was not consistent for all traits studied.

The influence of each model's genetic effects on genomic prediction is related to the heritability of the trait (Zhang et al. 2017). In this study, GM heritability values were higher than GY values, for all sites evaluated. This indicates that the proportion of additivity effect on the phenotypic variance of GM is greater than that in GY. This proportion may explain the smaller differences found between the predictive accuracies of the additive (rrBLUP) and additive-dominant (BL) models for GM, but greater differences for GY.

Another approach that has been applied in genomic selection studies is based on Machine Learning (ML). The algorithms of this approach can capture non-additive effects without requiring any assumptions about the genetic architecture of the evaluated trait. This is a result of its high power in capturing complex and non-linear patterns (Abdollahi‑Arpanahi et al. 2020; Wang et al. 2018). In this study, the genomic prediction performance of the ML method (LightGBM) was evaluated and compared to the two conventional statistical methods BL and rrBLUP. In terms of prediction accuracy, the LightGBM model was equivalent to the BL and rrBLUP models in predicting GM for Rolândia-PR (0.759) and Sorriso-MT (0.681). There was a statistical difference only for Jataí-GO (0.653), where LightGBM had a predictive performance equivalent to rrBLUP, but was inferior to the performance of the BL model (Table 2). For GY, its prediction accuracy was equivalent to the BL model for Rolândia-PR (0.685) and Sorriso-MT (0.404), differing only in the prediction of Jataí-GO (0.487). In this location, its performance was like the rrBLUP additive model. These results confirm that for traits containing complex gene action, such as GY, ML models can perform equivalently or even better than conventional statistical methods (Abdollahi-Arpanahi et al. 2020). In agreement with the results found here, Perez et al. (2022) also observed the superiority of ML models when compared to additive linear models for genomic prediction of traits with lower heritability.

Another very useful parameter in evaluating the predictive capacity of genomic prediction models is T10Recall, which indicates the percentage of correspondence between the data obtained in the prediction and the data observed in the field for the 10% of the most productive hybrids for GY, and the 10% earlier for GM. In practice, through this parameter we were able to estimate, for the models evaluated, the success rate of those hybrids that are truly superior in the field. For this parameter the results indicated that, in general, the LightGBM model was inferior to the parametric models (BL and rrBLUP) for Jataí-GO and Rolândia-PR, but this model was superior for Sorriso-MT (Table 3). The BL and rrBLUP models had very similar percentages for Jataí-GO and Rolândia-PR, in both traits (GY and GM). Given the above, there is no clear association between the results obtained for prediction accuracy and T10Recall, but it is worth highlighting the performance stability of the BL model compared to the other two methods for these two parameters.

Table 3

Percentages of hybrids that were above 10% most productive for GY, and above 10% earliest for GM, both in the field result (Observed) and in the prediction (Predicted). The higher the percentage, the greater the correspondence between Predicted and Observed values.
	Jataí-GO		Rolândia-PR		Sorriso-MT
Models	GY	GM	GY	GM	GY	GM
BL	34%	46%	38%	68%	28%	41%
LightGBM	32%	40%	36%	62%	28%	43%
rrBLUP	33%	45%	37%	66%	25%	39%

In terms of computational efficiency, the LightGBM model stood out as the most efficient and fastest for the three locations evaluated (Jataí-GO, Rolândia-PR and Sorriso-MT), and in both traits. For the location with the largest amount of training data (Sorriso-MT), this model took less than 20 seconds to estimate the effects of markers for the two traits evaluated (Fig. 2). This excellent computational performance of LightGBM was expected and is justified by the way trees are built in this algorithm, which, unlike other gradient boosting algorithms, is based on leaves instead of a level-based structure (Shi et al. 2022).

Among the two parametric models, the additive model (rrBLUP) presented the best computational performance for training the GY and GM models, being considerably faster than the additive-dominant model (BL). Considering the location with the largest amount of training data (Sorriso-MT), the time difference between the two models was greater than 8,000 seconds (Fig. 2). The low computational performance of the BL model is associated with the method used in statistical inferences, which in this case is based on a Monte Carlo algorithm via Markov Chains (MCMC) (Yang et al. 2020).

Although BL has been largely surpassed by the other two methodologies (LightGBM and rrBLUP) in terms of computational performance for training GY and GM models, this inferiority is mitigated with the faster execution of GBV estimates. Once the model training and validation stage is completed, the time spent predicting the GBV of new selection populations becomes the biggest bottleneck. In this scenario, the BL model becomes the second fastest, taking around 31 seconds less than the rrBLUP model and 21 seconds more than the LightGBM model (Fig. 2). As genomic prediction projects in genetic breeding programs may require the prediction of a few million hybrids, computational time is an important parameter when choosing prediction models.

The BL model showed greater stability in terms of predictive ability. If computational time is not a limiting factor, this is the most suitable method to obtain greater prediction accuracies for situations like those found in this study. In a scenario with a high amount of data for model training, the additive method (rrBLUP) and Machine Learning (LightGBM) can be used to achieve better computational performance, however this use can imply a significant loss of predictive accuracy.

The heritability of traits influences the prediction accuracy of the models. Higher accuracies are obtained with higher heritabilities. The genetic effects resulting from smaller heritabilities can be better captured using additive-dominant and machine Learning models.

Funding: LongPing High Tech Seed and Biotechnology company, the data analyzed are part of the company's routine research.

Conflicts of interest/Competing interest: Not applicable

Availability of data and material: Data is available, but material is not available

Code availability: Not applicable

Abdollahi-Arpanahi R, Gianola D, Peñagaricano F (2020) Deep learning versus parametric and ensemble methods for genomic prediction of complex phenotypes. Genet Sel Evol, 52, 12. https://doi.org/10.1186/s12711-020-00531-z
Bernardo R (1996) Testcross additive and dominance effects in best linear unbiased prediction of maize single-cross performance. Theoretical and Applied Genetics, Berlin, v. 93, no. 7, p. 1098-1102. https://doi.org/10.1007/BF00230131
Bernardo R (2008) Molecular markers, and selection for complex traits in plants: learning from the last 20 years. Crop Science, vol. 48, no. 5, p. 1649-1664. https://doi.org/10.2135/cropsci2008.03.0131
Cruz CD (2013) GENES, A software package for analysis in experimental statistics and quantitative genetics. Acta Scientiarum. v.35, n.3, p.271-276.
de los Campos G, Pérez P, Vazquez AI, Crossa J (2013). Genome-Enabled Prediction Using the BLR (Bayesian Linear Regression) R-Package. In: Gondro, C., van der Werf, J., Hayes, B. (eds) Genome-Wide Association Studies and Genomic Prediction. Methods in Molecular Biology, vol 1019. Humana Press, Totowa, NJ. https://doi.org/10.1007/978-1-62703-447-0_12
Endelman JB (2011) Ridge regression and other kernels for genomic selection with R package rrBLUP. Plant Genome, 4, 250–255. https://doi.org/10.3835/plantgenome2011.08.0024
Ferrão LFV, Marinho CD, Munoz PR, Resende Jr MFR (2020) Improvement of predictive ability in maize hybrids by including dominance effects and marker × environment models. Crop Sci. https://doi.org/10.1002/csc2.20096
Fritsche-Neto R, Galli G, Borges KLR, Costa-Neto G, Alves FC, Sabadin F, Lyra DH, Morais PPP, Braatz de Andrade LR, Granato I, Crossa J (2021) Optimizing Genomic-Enabled Prediction in Small-Scale Maize Hybrid Breeding Programs: A Roadmap Review. Front. Plant Sci, 12, 658267. https://doi.org/10.3389/fpls.2021.658267
Ghafouri-Kesbi F, Rahimi-Mianji G, Honarvar M, Nejati-Javaremi A (2017) Predictive ability of random forests, boosting, support vector machines and genomic best linear unbiased prediction in different scenarios of genomic evaluation. Anim. Prod. Sci, 57, 229–236
Grattapaglia D, Resende MDV (2011) Genomic selection in forest tree breeding. Tree Genetics & Genomes, vol. 7, p. 241-255. https://doi.org/10.1007/s11295-010-0328-4
Hayes BJ, Daetwyler HD, Bowman P, Moser G, Tier B, Crump R, Khatkar M, Raadsma HW, Goddard ME (2010) Accuracy of genomic selection: comparing theory and results. In Proceedings of the 18th Conference of the Association for the Advancement of Animal Breeding and Genetics. Barossa Valley, Australia. https://hdl.handle.net/1959.11/5252
Ishimori M, Hattori T, Yamazaki K, Takanashi H, Fujimoto M, Kajiya-Kanegae H, Yoneda J, Tokunaga T, Fujiwara T, Tsutsumi N, Iwata H (2020) Impacts of dominance effects on genomic prediction of sorghum hybrid performance. Breed Sci, 70(5), 605-616
Kaler AS, Purcell LC, Beissinger T, Gillman JD (2022) Genomic prediction models for traits differing in heritability for soybean, rice, and maize. BMC Plant Biol, 22, 87. https://doi.org/10.1186/s12870-022-03479-y
Liu X, Wang H, Wang H, Guo Z, Xu X, Liu J, Wang S, Li WX, Zou C, Prasanna BM, Olsen MS, Huang C, Xu Y (2018) Factors affecting genomic selection revealed by empirical evidence in maize. Crop J, 6, 341–352. https://doi.org/10.1016/j.cj.2018.03.005
Ma C, Zhang HH, Wang X (2014) Machine learning for Big Data analytics in plants. Trends Plant Sci, 19(12), 798-808. https://doi.org/10.1016/j.tplants.2014.08.004
Meuwissen TH, Hayes BJ, Goddard ME (2001) Prediction of total genetic value using genome-wide dense marker maps. Genetics, 157: 1819–1829.https://doi.org/10.1093/genetics/157.4.1819
Montesinos-López OA, Montesinos-López A, Crossa J (2022) Multivariate Statistical Machine Learning Methods for Genomic Prediction. International Publishing: Cham, Switzerland, 115-117. https://doi.org/10.1007/978-3-030-89010-0
Neves HH, Carvalheiro R, Queiroz AS (2012) A comparison of statistical methods for genomic selection in a mice population. BMC Genet 13, 100. https://doi.org/10.1186/1471-2156-13-100
Pérez P, de los Campos G, Crossa J, Gianola D (2010) Genomic-enabled prediction based on molecular markers and pedigree using the Bayesian linear regression package in R. plant. Genome, 3, 106–116. https://doi.org/10.3835/plantgenome2010.04.0005
Perez BC, Bink MCAM, Svenson KL, Churchill GA, Calus MPL (2022) Prediction performance of linear models and gradient boosting machine on complex phenotypes in outbred mice. G3 (Bethesda), 4, 12(4):jkac039. https://doi.org/10.1093/g3journal/jkac039
Shi Y, Ke G, Soukhavong D, Lamb J, Men, Q, Finley T, Wang T, Chen W, Ma W, Ye Q, Liu T, Titov N, Yan Y, Microsoft Corporation, Dropbox Inc, Loden J,Daeschler D, Rodola G, Ferreira A, Lemire D,Zverovich V, IBM Corporation, Cortes D (2022) Lightgbm: Light Gradient Boosting Machine. R package version 3.3.4.
Terraillon J, Roeber FK, Flachenecker C, Frisch M (2023) Training set designs for prediction of yield and moisture of maize test cross hybrids with unreplicated trials. Front. Plant Sci. 14:1080087. doi: 10.3389/fpls.2023.1080087
Tong H, Nikoloski Z (2021). Machine learning approaches for crop improvement: leveraging phenotypic and genotypic big data. J. Plant Physiol. 257, 153354. https://doi.org/10.1016/j.jplph.2020.153354
Wang X, Xu Y, Hu Z, Xu C (2018) Genomic selection methods for crop improvement: Status and prospects. The Crop Journal, 6, 330-340. https://doi.org/10.1016/j.cj.2018.03.001
Wellmann R, Bennewitz J (2012) Bayesian models with dominance effects for genomic evaluation of quantitative traits. Genet. Res. (Camb), 94, 21–37. https://doi.org/10.1017/S0016672312000018
Windhausen VS, Atlin GN, Hickey JM, Crossa J, Jannink JL, Sorrells ME, Raman B, Cairns JE, Tarekegne A, Semagn K, Beyene Y, Grudloyma P, Technow F, Riedelsheimer C, Melchinger AE (2012) Effectiveness of genomics prediction of maize hybrid performance in different breeding populations and environments. G3 Genes Genomes Genetics, 2, 1427–1436. https://doi.org/10.1534/g3.112.003699
Yan J, Xu Y, Cheng Q, Jiang S, Wang Q, Xiao Y, Ma C, Yan J, Wang X (2021) LightGBM: accelerated genomically designed crop breeding through ensemble learning. Genome Biol 22, 271. https://doi.org/10.1186/s13059-021-02492-y
Yang Y, Wainwright MJ, Jordan MI (2016) On the computational complexity of high-dimensional Bayesian variable selection. The Annals of Statistics 44 (6), 2497–2532. https://doi.org/10.1214/15-AOS1417
Zhang A, Wang H, Beyene Y, Semagn K, Liu Y, Cao S, Cui Z, Ruan Y, Burgueño, J, San Vicente F, Olsen M, Prasanna BM, Crossa J, Yu H, Zhang X (2017) Effect of Trait Heritability, Training Population Size and Marker Density on Genomic Prediction Accuracy Estimation in 22 bi-parental Tropical Maize Populations. Front Plant Sci, 8, 8:1916. https://doi.org/10.3389/fpls.2017.01916

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Comparison of additive, additive-dominant and machine learning (LightGBM) models for genomic prediction of corn hybrids

Status:

Version 1

Abstract

Figures

Introduction

Material and methods

Results

Discussion

Declarations

References

Additional Declarations

Status:

Version 1