Regularized selection indices for breeding value prediction using hyper-spectral image data

The BLUP method in evaluation of breeding values of Russian spring wheat lines using micro- and macroelements in seeds

Genomic selection is a technology that allows for the determination of the genetic value of varieties of agricultural plants and animal breeds, based on information about genotypes and phenotypes. The measured breeding value (BV) for varieties and breeds in relation to the target trait allows breeding stages to be thoroughly planned and the parent forms suitable for crossing to be chosen. In this work, the BLUP method was used to assess the breeding value of 149 Russian varieties and introgression lines (4 measurements for each variety or line, 596 phenotypic points) of spring wheat according to the content of seven chemical elements in the grain - K, Ca, Mg, Mn, Fe, Zn, Cu. The quality of the evaluation of breeding values was assessed using cross-validation, when the sample was randomly divided into five parts, one of which was chosen as a test population. The following average values of the Pearson correlation were obtained for predicting the concentration of trace elements: K - 0.67, Ca - 0.61, Mg - 0.4, Mn - 0.5, Fe - 0.38, Zn - 0.46, Cu - 0.48. Out of the 35 models studied, the p-value was below the nominal significant threshold (p-value < 0.05) for 28 models. For 11 models, the p-value was significant after correction for multiple testing (p-value < 0.001). For Ca and K, four out of five models and for Mn two out of five models had a p-value below the threshold adjusted for multiple testing. For 30 varieties that showed the best varietal values for Ca, K and Mn, the average breeding value was 296.43, 785.11 and 4.87 mg/kg higher, respectively, than the average breeding value of the population. The results obtained show the relevance of the application of genomic selection models even in such limited-size samples. The models for K, Ca and Mn are suitable for assessing the breeding value of Russian wheat varieties based on these characteristics.

Conserved patterns and locomotor-related evolutionary constraints in the hominoid vertebral column

The evolution of the hominoid lineage is characterized by pervasive homoplasy, notably in regions such as the vertebral column, which plays a central role in body support and locomotion. Few isolated and fewer associated vertebrae are known for most fossil hominoid taxa, but identified specimens indicate potentially high levels of convergence in terms of both form and number. Homoplasy thus complicates attempts to identify the anatomy of the last common ancestor of hominins and other taxa and stymies reconstructions of evolutionary scenarios. One way to clarify the role of homoplasy is by investigating constraints via phenotypic integration, which assesses covariation among traits, shapes evolutionary pathways, and itself evolves in response to selection. We assessed phenotypic integration and evolvability across the subaxial (cervical, thoracic, lumbar, sacral) vertebral column of macaques (n = 96), gibbons (n = 77), chimpanzees (n = 92), and modern humans (n = 151). We found a mid-cervical cluster that may have shifted cranially in hominoids, a persistent thoracic cluster that is most marked in chimpanzees, and an expanded lumbosacral cluster in hominoids that is most expanded in gibbons. Our results highlight the highly conserved nature of the vertebral column. Taxa appear to exploit existing patterns of integration and ontogenetic processes to shift, expand, or reduce cluster boundaries. Gibbons appear to be the most highly derived taxon in our sample, possibly in response to their highly specialized locomotion.

Genome-wide association and genomic prediction for iron and zinc concentration and iron bioavailability in a collection of yellow dry beans

Dry bean is a nutrient-dense food targeted in biofortification programs to increase seed iron and zinc levels. The underlying assumption of breeding for higher mineral content is that enhanced iron and zinc levels will deliver health benefits to the consumers of these biofortified foods. This study characterized a diversity panel of 275 genotypes comprising the Yellow Bean Collection (YBC) for seed Fe and Zn concentration, Fe bioavailability (FeBio), and seed yield across 2 years in two field locations. The genetic architecture of each trait was elucidated via genome-wide association studies (GWAS) and the efficacy of genomic prediction (GP) was assessed. Moreover, 82 yellow breeding lines were evaluated for seed Fe and Zn concentrations as well as seed yield, serving as a prediction set for GP models. Large phenotypic variability was identified in all traits evaluated, and variations of up to 2.8 and 13.7-fold were observed for Fe concentration and FeBio, respectively. Prediction accuracies in the YBC ranged from a low of 0.12 for Fe concentration, to a high of 0.72 for FeBio, and an accuracy improvement of 0.03 was observed when a QTN, identified through GWAS, was used as a fixed effect for FeBio. This study provides evidence of the lack of correlation between FeBio estimated in vitro and Fe concentration and highlights the potential of GP in accurately predicting FeBio in yellow beans, offering a cost-effective alternative to the traditional assessment of using Caco2 cell methodologies.

Robotized indoor phenotyping allows genomic prediction of adaptive traits in the field

Breeding for resilience to climate change requires considering adaptive traits such as plant architecture, stomatal conductance and growth, beyond the current selection for yield. Robotized indoor phenotyping allows measuring such traits at high throughput for speed breeding, but is often considered as non-relevant for field conditions. Here, we show that maize adaptive traits can be inferred in different fields, based on genotypic values obtained indoor and on environmental conditions in each considered field. The modelling of environmental effects allows translation from indoor to fields, but also from one field to another field. Furthermore, genotypic values of considered traits match between indoor and field conditions. Genomic prediction results in adequate ranking of genotypes for the tested traits, although with lesser precision for elite varieties presenting reduced phenotypic variability. Hence, it distinguishes genotypes with high or low values for adaptive traits, conferring either spender or conservative strategies for water use under future climates.

(Quasi) multitask support vector regression with heuristic hyperparameter optimization for whole-genome prediction of complex traits: a case study with carcass traits in broilers

This study investigates nonlinear kernels for multitrait (MT) genomic prediction using support vector regression (SVR) models. We assessed the predictive ability delivered by single-trait (ST) and MT models for 2 carcass traits (CT1 and CT2) measured in purebred broiler chickens. The MT models also included information on indicator traits measured in vivo [Growth and feed efficiency trait (FE)]. We proposed an approach termed (quasi) multitask SVR (QMTSVR), with hyperparameter optimization performed via genetic algorithm. ST and MT Bayesian shrinkage and variable selection models [genomic best linear unbiased predictor (GBLUP), BayesC (BC), and reproducing kernel Hilbert space (RKHS) regression] were employed as benchmarks. MT models were trained using 2 validation designs (CV1 and CV2), which differ if the information on secondary traits is available in the testing set. Models' predictive ability was assessed with prediction accuracy (ACC; i.e. the correlation between predicted and observed values, divided by the square root of phenotype accuracy), standardized root-mean-squared error (RMSE*), and inflation factor (b). To account for potential bias in CV2-style predictions, we also computed a parametric estimate of accuracy (ACCpar). Predictive ability metrics varied according to trait, model, and validation design (CV1 or CV2), ranging from 0.71 to 0.84 for ACC, 0.78 to 0.92 for RMSE*, and between 0.82 and 1.34 for b. The highest ACC and smallest RMSE* were achieved with QMTSVR-CV2 in both traits. We observed that for CT1, model/validation design selection was sensitive to the choice of accuracy metric (ACC or ACCpar). Nonetheless, the higher predictive accuracy of QMTSVR over MTGBLUP and MTBC was replicated across accuracy metrics, besides the similar performance between the proposed method and the MTRKHS model. Results showed that the proposed approach is competitive with conventional MT Bayesian regression models using either Gaussian or spike-slab multivariate priors.

Genomics combined with UAS data enhances prediction of grain yield in winter wheat

With the human population continuing to increase worldwide, there is pressure to employ novel technologies to increase genetic gain in plant breeding programs that contribute to nutrition and food security. Genomic selection (GS) has the potential to increase genetic gain because it can accelerate the breeding cycle, increase the accuracy of estimated breeding values, and improve selection accuracy. However, with recent advances in high throughput phenotyping in plant breeding programs, the opportunity to integrate genomic and phenotypic data to increase prediction accuracy is present. In this paper, we applied GS to winter wheat data integrating two types of inputs: genomic and phenotypic. We observed the best accuracy of grain yield when combining both genomic and phenotypic inputs, while only using genomic information fared poorly. In general, the predictions with only phenotypic information were very competitive to using both sources of information, and in many cases using only phenotypic information provided the best accuracy. Our results are encouraging because it is clear we can enhance the prediction accuracy of GS by integrating high quality phenotypic inputs in the models.

Enviromic-based kernels may optimize resource allocation with multi-trait multi-environment genomic prediction for tropical Maize

BACKGROUND

Success in any genomic prediction platform is directly dependent on establishing a representative training set. This is a complex task, even in single-trait single-environment conditions and tends to be even more intricated wherein additional information from envirotyping and correlated traits are considered. Here, we aimed to design optimized training sets focused on genomic prediction, considering multi-trait multi-environment trials, and how those methods may increase accuracy reducing phenotyping costs. For that, we considered single-trait multi-environment trials and multi-trait multi-environment trials for three traits: grain yield, plant height, and ear height, two datasets, and two cross-validation schemes. Next, two strategies for designing optimized training sets were conceived, first considering only the genomic by environment by trait interaction (GET), while a second including large-scale environmental data (W, enviromics) as genomic by enviromic by trait interaction (GWT). The effective number of individuals (genotypes × environments × traits) was assumed as those that represent at least 98% of each kernel (GET or GWT) variation, in which those individuals were then selected by a genetic algorithm based on prediction error variance criteria to compose an optimized training set for genomic prediction purposes.

RESULTS

The combined use of genomic and enviromic data efficiently designs optimized training sets for genomic prediction, improving the response to selection per dollar invested by up to 145% when compared to the model without enviromic data, and even more when compared to cross validation scheme with 70% of training set or pure phenotypic selection. Prediction models that include G × E or enviromic data + G × E yielded better prediction ability.

CONCLUSIONS

Our findings indicate that a genomic by enviromic by trait interaction kernel associated with genetic algorithms is efficient and can be proposed as a promising approach to designing optimized training sets for genomic prediction when the variance-covariance matrix of traits is available. Additionally, great improvements in the genetic gains per dollar invested were observed, suggesting that a good allocation of resources can be deployed by using the proposed approach.

Genomic selection for morphological and yield-related traits using genome-wide SNPs in oil palm

Oil palm is the most important oil crop worldwide. Colombia is the fourth largest producer, primarily relying on production from interspecific hybrids, derived from crosses between Elaeis oleifera and Elaeis guineensis (OxG). However, conventional breeding can take up to 20 years to generate a new variety. Therefore, reducing the breeding cycle while improving the genetic gain for complex traits is desirable. Genomic selection (GS) is an approach with the potential to achieve this goal. In this study, we evaluated 431 F1 interspecific hybrids (OxG) and 444 backcrosses (BC1) for morphological and yield-related traits. Genomic predictions were performed with the G-BLUP model using three different population datasets for training the model: the same population (TRN1), the other population (TRN2), and both populations (TRN1+2). Higher multi-family prediction accuracies were obtained for foliar area (0.3 in OxG) and trunk height (0.47 in BC1) when the model was trained with TRN1. Single-family prediction accuracies were lower in the OxG compared to BC1 families for traits such as trunk diameter, trunk height, bunch number, and yield using TRN1. Conversely, lower prediction accuracies were obtained for most traits when the model was trained using TRN2 (< 0.1). Multi-trait models showed a substantial increase of the predictions for traits such as yield (0.22 for OxG and 0.44 for BC1), because of the genetic correlations between traits. The results herein highlighted the potential of GS for parental selection in OxG and BC1 populations, but further studies are required to improve the models to select individuals by their genetic value.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11032-022-01341-5.

Combining novel feature selection strategy and hyperspectral vegetation indices to predict crop yield

BACKGROUND

Wheat is an important food crop globally, and timely prediction of wheat yield in breeding efforts can improve selection efficiency. Traditional yield prediction method based on secondary traits is time-consuming, costly, and destructive. It is urgent to develop innovative methods to improve selection efficiency and accelerate genetic gains in the breeding cycle.

RESULTS

Crop yield prediction using remote sensing has gained popularity in recent years. This paper proposed a novel ensemble feature selection (EFS) method to improve yield prediction from hyperspectral data. For this, 207 wheat cultivars and breeding lines were grown under full and limited irrigation treatments respectively, and their canopy hyperspectral reflectance was measured at the flowering, early grain filling (EGF), mid grain filling (MGF), and late grain filling (LGF) stages. Then, 115 vegetation indices were extracted from the hyperspectral reflectance and combined with four feature selection methods, i.e., mean decrease impurity (MDI), Boruta, FeaLect, and RReliefF to train deep neural network (DNN) models for yield prediction. Next, a learning framework was developed by combining the predicted values of the selected and the full features using multiple linear regression (MLR). The results show that the selected features contributed to higher yield prediction accuracy than the full features, and the MDI method performed well across growth stages, with a mean R² ranging from 0.634 to 0.666 (mean RMSE = 0.926-0.967 t ha^-1). Also, the proposed EFS method outperformed all the individual feature selection methods across growth stages, with a mean R² ranging from 0.648 to 0.679 (mean RMSE = 0.911-0.950 t ha^-1).

CONCLUSIONS

The proposed EFS method can improve grain yield prediction from hyperspectral data and can be used to assist wheat breeders in earlier decision-making.

Identification of Spring Wheat with Superior Agronomic Performance under Contrasting Nitrogen Managements Using Linear Phenotypic Selection Indices

Both the Linear Phenotypic Selection Index (LPSI) and the Restrictive Linear Phenotypic Selection Index (RLPSI) have been widely used to select parents and progenies, but the effect of economic weights on the selection parameters (the expected genetic gain, response to selection, and the correlation between the indices and genetic merits) have not been investigated in detail. Here, we (i) assessed combinations of 2304 economic weights using four traits (maturity, plant height, grain yield and grain protein content) recorded under four organically (low nitrogen) and five conventionally (high nitrogen) managed environments, (ii) compared single-trait and multi-trait selection indices (LPSI vs. RLPSI by imposing restrictions to the expected genetic gain of either yield or grain protein content), and (iii) selected a subset of about 10% spring wheat cultivars that performed very well under organic and/or conventional management systems. The multi-trait selection indices, with and without imposing restrictions, were superior to single trait selection. However, the selection parameters differed quite a lot depending on the economic weights, which suggests the need for optimizing the weights. Twenty-two of the 196 cultivars that showed superior performance under organic and/or conventional management systems were consistently selected using all five of the selected economic weights, and at least two of the selection scenarios. The selected cultivars belonged to the Canada Western Red Spring (16 cultivars), the Canada Northern Hard Red (3), and the Canada Prairie Spring Red (3), and required 83–93 days to maturity, were 72–100 cm tall, and produced from 4.0 to 6.2 t ha⁻¹ grain yield with 14.6–17.7% GPC. The selected cultivars would be highly useful, not only as potential trait donors for breeding under an organic management system, but also for other studies, including nitrogen use efficiency.

Association Studies and Genomic Prediction for Genetic Improvements in Agriculture

To feed the fast growing global population with sufficient food using limited global resources, it is urgent to develop and utilize cutting-edge technologies and improve efficiency of agricultural production. In this review, we specifically introduce the concepts, theories, methods, applications and future implications of association studies and predicting unknown genetic value or future phenotypic events using genomics in the area of breeding in agriculture. Genome wide association studies can identify the quantitative genetic loci associated with phenotypes of importance in agriculture, while genomic prediction utilizes individual genetic value to rank selection candidates to improve the next generation of plants or animals. These technologies and methods have improved the efficiency of genetic improvement programs for agricultural production via elite animal breeds and plant varieties. With the development of new data acquisition technologies, there will be more and more data collected from high-through-put technologies to assist agricultural breeding. It will be crucial to extract useful information among these large amounts of data and to face this challenge, more efficient algorithms need to be developed and utilized for analyzing these data. Such development will require knowledge from multiple disciplines of research.

The L-shaped Selection Algorithm for Multi-trait Genomic Selection

Selecting for multiple traits as opposed to a single trait has become increasingly important in genomic selection. As one of the most popular approaches to multi-trait genomic selection (MTGS), index selection uses a weighted average of all traits as a single breeding objective. Although intuitive and effective, index selection is not only numerically sensitive but also structurally incapable of finding certain optimal breeding parents. This paper proposes a new selection method for MTGS, the L-shaped selection, which addresses the limitations of index selection by normalizing the trait values and using an L-shaped objective function to find optimal breeding parents. This algorithm has been proven to be able to find any Pareto optimal solution with appropriate weights. Two performance metrics have also been defined to quantify MTGS algorithms with respect to their ability to accelerate genetic gain and preserve genetic diversity. Computational experiments were conducted to demonstrate the improved performance of L-shaped selection over index selection.

Integrating and optimizing genomic, weather, and secondary trait data for multiclass classification

Modern plant breeding programs collect several data types such as weather, images, and secondary or associated traits besides the main trait (e.g., grain yield). Genomic data is high-dimensional and often over-crowds smaller data types when naively combined to explain the response variable. There is a need to develop methods able to effectively combine different data types of differing sizes to improve predictions. Additionally, in the face of changing climate conditions, there is a need to develop methods able to effectively combine weather information with genotype data to predict the performance of lines better. In this work, we develop a novel three-stage classifier to predict multi-class traits by combining three data types-genomic, weather, and secondary trait. The method addressed various challenges in this problem, such as confounding, differing sizes of data types, and threshold optimization. The method was examined in different settings, including binary and multi-class responses, various penalization schemes, and class balances. Then, our method was compared to standard machine learning methods such as random forests and support vector machines using various classification accuracy metrics and using model size to evaluate the sparsity of the model. The results showed that our method performed similarly to or better than machine learning methods across various settings. More importantly, the classifiers obtained were highly sparse, allowing for a straightforward interpretation of relationships between the response and the selected predictors.

Steady agronomic and genetic interventions are essential for sustaining productivity in intensive rice cropping

Steady agronomic and genetic interventions helped sustain high annual rice production in an intensive irrigated monoculture system under a changing climate. However, the system did not achieve the increases in yield required to keep pace with the growing global demand for rice because annual yield potential was stagnant, and apparent biotic constraints limited yield in the wet season.

Intensive systems with two or three rice (Oryza sativa L.) crops per year account for about 50% of the harvested area for irrigated rice in Asia. Any reduction in productivity or sustainability of these systems has serious implications for global food security. Rice yield trends in the world’s longest-running long-term continuous cropping experiment (LTCCE) were evaluated to investigate consequences of intensive cropping and to draw lessons for sustaining production in Asia. Annual production was sustained at a steady level over the 50-y period in the LTCCE through continuous adjustment of management practices and regular cultivar replacement. Within each of the three annual cropping seasons (dry, early wet, and late wet), yield decline was observed during the first phase, from 1968 to 1990. Agronomic improvements in 1991 to 1995 helped to reverse this yield decline, but yield increases did not continue thereafter from 1996 to 2017. Regular genetic and agronomic improvements were sufficient to maintain yields at steady levels in dry and early wet seasons despite a reduction in the yield potential due to changing climate. Yield declines resumed in the late wet season. Slower growth in genetic gain after the first 20 y was associated with slower breeding cycle advancement as indicated by pedigree depth. Our findings demonstrate that through adjustment of management practices and regular cultivar replacement, it is possible to sustain a high level of annual production in irrigated systems under a changing climate. However, the system was unable to achieve further increases in yield required to keep pace with the growing global rice demand.

Multi-generation genomic prediction of maize yield using parametric and non-parametric sparse selection indices

Genomic prediction models are often calibrated using multi-generation data. Over time, as data accumulates, training data sets become increasingly heterogeneous. Differences in allele frequency and linkage disequilibrium patterns between the training and prediction genotypes may limit prediction accuracy. This leads to the question of whether all available data or a subset of it should be used to calibrate genomic prediction models. Previous research on training set optimization has focused on identifying a subset of the available data that is optimal for a given prediction set. However, this approach does not contemplate the possibility that different training sets may be optimal for different prediction genotypes. To address this problem, we recently introduced a sparse selection index (SSI) that identifies an optimal training set for each individual in a prediction set. Using additive genomic relationships, the SSI can provide increased accuracy relative to genomic-BLUP (GBLUP). Non-parametric genomic models using Gaussian kernels (KBLUP) have, in some cases, yielded higher prediction accuracies than standard additive models. Therefore, here we studied whether combining SSIs and kernel methods could further improve prediction accuracy when training genomic models using multi-generation data. Using four years of doubled haploid maize data from the International Maize and Wheat Improvement Center (CIMMYT), we found that when predicting grain yield the KBLUP outperformed the GBLUP, and that using SSI with additive relationships (GSSI) lead to 5-17% increases in accuracy, relative to the GBLUP. However, differences in prediction accuracy between the KBLUP and the kernel-based SSI were smaller and not always significant.

MegaLMM: Mega-scale linear mixed models for genomic predictions with thousands of traits

Large-scale phenotype data can enhance the power of genomic prediction in plant and animal breeding, as well as human genetics. However, the statistical foundation of multi-trait genomic prediction is based on the multivariate linear mixed effect model, a tool notorious for its fragility when applied to more than a handful of traits. We present MegaLMM, a statistical framework and associated software package for mixed model analyses of a virtually unlimited number of traits. Using three examples with real plant data, we show that MegaLMM can leverage thousands of traits at once to significantly improve genetic value prediction accuracy.

Using Hybrid Artificial Intelligence and Evolutionary Optimization Algorithms for Estimating Soybean Yield and Fresh Biomass Using Hyperspectral Vegetation Indices

Recent advanced high-throughput field phenotyping combined with sophisticated big data analysis methods have provided plant breeders with unprecedented tools for a better prediction of important agronomic traits, such as yield and fresh biomass (FBIO), at early growth stages. This study aimed to demonstrate the potential use of 35 selected hyperspectral vegetation indices (HVI), collected at the R5 growth stage, for predicting soybean seed yield and FBIO. Two artificial intelligence algorithms, ensemble-bagging (EB) and deep neural network (DNN), were used to predict soybean seed yield and FBIO using HVI. Considering HVI as input variables, the coefficients of determination (R2) of 0.76 and 0.77 for yield and 0.91 and 0.89 for FBIO were obtained using DNN and EB, respectively. In this study, we also used hybrid DNN-SPEA2 to estimate the optimum HVI values in soybeans with maximized yield and FBIO productions. In addition, to identify the most informative HVI in predicting yield and FBIO, the feature recursive elimination wrapper method was used and the top ranking HVI were determined to be associated with red, 670 nm and near-infrared, 800 nm, regions. Overall, this study introduced hybrid DNN-SPEA2 as a robust mathematical tool for optimizing and using informative HVI for estimating soybean seed yield and FBIO at early growth stages, which can be employed by soybean breeders for discriminating superior genotypes in large breeding populations.

Improving Genomic Prediction Using High-Dimensional Secondary Phenotypes

In the past decades, genomic prediction has had a large impact on plant breeding. Given the current advances of high-throughput phenotyping and sequencing technologies, it is increasingly common to observe a large number of traits, in addition to the target trait of interest. This raises the important question whether these additional or “secondary” traits can be used to improve genomic prediction for the target trait. With only a small number of secondary traits, this is known to be the case, given sufficiently high heritabilities and genetic correlations. Here we focus on the more challenging situation with a large number of secondary traits, which is increasingly common since the arrival of high-throughput phenotyping. In this case, secondary traits are usually incorporated through additional relatedness matrices. This approach is however infeasible when secondary traits are not measured on the test set, and cannot distinguish between genetic and non-genetic correlations. An alternative direction is to extend the classical selection indices using penalized regression. So far, penalized selection indices have not been applied in a genomic prediction setting, and require plot-level data in order to reliably estimate genetic correlations. Here we aim to overcome these limitations, using two novel approaches. Our first approach relies on a dimension reduction of the secondary traits, using either penalized regression or random forests (LS-BLUP/RF-BLUP). We then compute the bivariate GBLUP with the dimension reduction as secondary trait. For simulated data (with available plot-level data), we also use bivariate GBLUP with the penalized selection index as secondary trait (SI-BLUP). In our second approach (GM-BLUP), we follow existing multi-kernel methods but replace secondary traits by their genomic predictions, with the advantage that genomic prediction is also possible when secondary traits are only measured on the training set. For most of our simulated data, SI-BLUP was most accurate, often closely followed by RF-BLUP or LS-BLUP. In real datasets, involving metabolites in Arabidopsis and transcriptomics in maize, no method could substantially improve over univariate prediction when secondary traits were only available on the training set. LS-BLUP and RF-BLUP were most accurate when secondary traits were available also for the test set.

Evaluating the performance of machine learning methods and variable selection methods for predicting difficult-to-measure traits in Holstein dairy cattle using milk infrared spectral data

Fourier-transform infrared (FTIR) spectroscopy is a powerful high-throughput phenotyping tool for predicting traits that are expensive and difficult to measure in dairy cattle. Calibration equations are often developed using standard methods, such as partial least squares (PLS) regression. Methods that employ penalization, rank-reduction, and variable selection, as well as being able to model the nonlinear relations between phenotype and FTIR, might offer improvements in predictive ability and model robustness. This study aimed to compare the predictive ability of 2 machine learning methods, namely random forest (RF) and gradient boosting machine (GBM), and penalized regression against PLS regression for predicting 3 phenotypes differing in terms of biological meaning and relationships with milk composition (i.e., phenotypes measurable directly and not directly in milk, reflecting different biological processes which can be captured using milk spectra) in Holstein-Friesian cattle under 2 cross-validation scenarios. The data set comprised phenotypic information from 471 Holstein-Friesian cows, and 3 target phenotypes were evaluated: (1) body condition score (BCS), (2) blood β-hydroxybutyrate (BHB, mmol/L), and (3) κ-casein expressed as a percentage of nitrogen (κ-CN, % N). The data set was split considering 2 cross-validation scenarios: samples-out random in which the population was randomly split into 10-folds (8-folds for training and 1-fold for validation and testing); and herd/date-out in which the population was randomly assigned to training (70% herd), validation (10%), and testing (20% herd) based on the herd and date in which the samples were collected. The random grid search was performed using the training subset for the hyperparameter optimization and the validation set was used for the generalization of prediction error. The trained model was then used to assess the final prediction in the testing subset. The grid search for penalized regression evidenced that the elastic net (EN) was the best regularization with increase in predictive ability of 5%. The performance of PLS (standard model) was compared against 2 machine learning techniques and penalized regression using 2 cross-validation scenarios. Machine learning methods showed a greater predictive ability for BCS (0.63 for GBM and 0.61 for RF), BHB (0.80 for GBM and 0.79 for RF), and κ-CN (0.81 for GBM and 0.80 for RF) in samples-out cross-validation. Considering a herd/date-out cross-validation these values were 0.58 (GBM and RF) for BCS, 0.73 (GBM and RF) for BHB, and 0.77 (GBM and RF) for κ-CN. The GBM model tended to outperform other methods in predictive ability around 4%, 1%, and 7% for EN, RF, and PLS, respectively. The prediction accuracies of the GBM and RF models were similar, and differed statistically from the PLS model in samples-out random cross-validation. Although, machine learning techniques outperformed PLS in herd/date-out cross-validation, no significant differences were observed in terms of predictive ability due to the large standard deviation observed for predictions. Overall, GBM achieved the highest accuracy of FTIR-based prediction of the different phenotypic traits across the cross-validation scenarios. These results indicate that GBM is a promising method for obtaining more accurate FTIR-based predictions for different phenotypes in dairy cattle.

Optimal breeding-value prediction using a sparse selection index

Genomic prediction uses DNA sequences and phenotypes to predict genetic values. In homogeneous populations, theory indicates that the accuracy of genomic prediction increases with sample size. However, differences in allele frequencies and linkage disequilibrium patterns can lead to heterogeneity in SNP effects. In this context, calibrating genomic predictions using a large, potentially heterogeneous, training data set may not lead to optimal prediction accuracy. Some studies tried to address this sample size/homogeneity trade-off using training set optimization algorithms; however, this approach assumes that a single training data set is optimum for all individuals in the prediction set. Here, we propose an approach that identifies, for each individual in the prediction set, a subset from the training data (i.e., a set of support points) from which predictions are derived. The methodology that we propose is a sparse selection index (SSI) that integrates selection index methodology with sparsity-inducing techniques commonly used for high-dimensional regression. The sparsity of the resulting index is controlled by a regularization parameter (λ); the G-Best Linear Unbiased Predictor (G-BLUP) (the prediction method most commonly used in plant and animal breeding) appears as a special case which happens when λ = 0. In this study, we present the methodology and demonstrate (using two wheat data sets with phenotypes collected in 10 different environments) that the SSI can achieve significant (anywhere between 5 and 10%) gains in prediction accuracy relative to the G-BLUP.

Machine learning in plant science and plant breeding

Technological developments have revolutionized measurements on plant genotypes and phenotypes, leading to routine production of large, complex data sets. This has led to increased efforts to extract meaning from these measurements and to integrate various data sets. Concurrently, machine learning has rapidly evolved and is now widely applied in science in general and in plant genotyping and phenotyping in particular. Here, we review the application of machine learning in the context of plant science and plant breeding. We focus on analyses at different phenotype levels, from biochemical to yield, and in connecting genotypes to these. In this way, we illustrate how machine learning offers a suite of methods that enable researchers to find meaningful patterns in relevant plant data.

Prior Biological Knowledge Improves Genomic Prediction of Growth-Related Traits in Arabidopsis thaliana

Prediction of growth-related complex traits is highly important for crop breeding. Photosynthesis efficiency and biomass are direct indicators of overall plant performance and therefore even minor improvements in these traits can result in significant breeding gains. Crop breeding for complex traits has been revolutionized by technological developments in genomics and phenomics. Capitalizing on the growing availability of genomics data, genome-wide marker-based prediction models allow for efficient selection of the best parents for the next generation without the need for phenotypic information. Until now such models mostly predict the phenotype directly from the genotype and fail to make use of relevant biological knowledge. It is an open question to what extent the use of such biological knowledge is beneficial for improving genomic prediction accuracy and reliability. In this study, we explored the use of publicly available biological information for genomic prediction of photosynthetic light use efficiency (Φ PSII ) and projected leaf area (PLA) in Arabidopsis thaliana. To explore the use of various types of knowledge, we mapped genomic polymorphisms to Gene Ontology (GO) terms and transcriptomics-based gene clusters, and applied these in a Genomic Feature Best Linear Unbiased Predictor (GFBLUP) model, which is an extension to the traditional Genomic BLUP (GBLUP) benchmark. Our results suggest that incorporation of prior biological knowledge can improve genomic prediction accuracy for both Φ PSII and PLA. The improvement achieved depends on the trait, type of knowledge and trait heritability. Moreover, transcriptomics offers complementary evidence to the Gene Ontology for improvement when used to define functional groups of genes. In conclusion, prior knowledge about trait-specific groups of genes can be directly translated into improved genomic prediction.

Accelerating forest tree breeding by integrating genomic selection and greenhouse phenotyping

Breeding forest species can be a costly and slow process because of the extensive areas needed for field trials and the long periods (e.g., five years) that are required to measure economically and environmentally relevant phenotypes (e.g., adult plant biomass or plant height). Genomic selection (GS) and indirect selection using early phenotypes (e.g., phenotypes collected in greenhouse conditions) are two ways by which tree breeding can be accelerated. These approaches can both reduce the costs of field-testing and the time required to make selection decisions. Moreover, these approaches can be highly synergistic. Therefore, in this study, we used a data set comprising DNA genotypes and longitudinal measurements of growth collected from a population of Populus deltoides W. Bartram ex Marshall (eastern cottonwood) in the greenhouse and the field, to evaluate the potential impact of integrating large-scale greenhouse phenotyping with conventional GS. We found that the integration of greenhouse phenotyping and GS can deliver very early selection decisions that are moderately accurate. Therefore, we conclude that the adoption of these approaches, in conjunction with reproductive techniques that shorten the generation interval, can lead to an unprecedented acceleration of selection gains in P. deltoides and, potentially, other commercially planted tree species.

Application of Machine Learning Algorithms in Plant Breeding: Predicting Yield From Hyperspectral Reflectance in Soybean

Recent substantial advances in high-throughput field phenotyping have provided plant breeders with affordable and efficient tools for evaluating a large number of genotypes for important agronomic traits at early growth stages. Nevertheless, the implementation of large datasets generated by high-throughput phenotyping tools such as hyperspectral reflectance in cultivar development programs is still challenging due to the essential need for intensive knowledge in computational and statistical analyses. In this study, the robustness of three common machine learning (ML) algorithms, multilayer perceptron (MLP), support vector machine (SVM), and random forest (RF), were evaluated for predicting soybean (Glycine max) seed yield using hyperspectral reflectance. For this aim, the hyperspectral reflectance data for the whole spectra ranged from 395 to 1005 nm, which were collected at the R4 and R5 growth stages on 250 soybean genotypes grown in four environments. The recursive feature elimination (RFE) approach was performed to reduce the dimensionality of the hyperspectral reflectance data and select variables with the largest importance values. The results indicated that R5 is more informative stage for measuring hyperspectral reflectance to predict seed yields. The 395 nm reflectance band was also identified as the high ranked band in predicting the soybean seed yield. By considering either full or selected variables as the input variables, the ML algorithms were evaluated individually and combined-version using the ensemble-stacking (E-S) method to predict the soybean yield. The RF algorithm had the highest performance with a value of 84% yield classification accuracy among all the individual tested algorithms. Therefore, by selecting RF as the metaClassifier for E-S method, the prediction accuracy increased to 0.93, using all variables, and 0.87, using selected variables showing the success of using E-S as one of the ensemble techniques. This study demonstrated that soybean breeders could implement E-S algorithm using either the full or selected spectra reflectance to select the high-yielding soybean genotypes, among a large number of genotypes, at early growth stages.