Michele Jorge da Silva

This study focused on incorporating dimensionality reduction based on marker significance to better harness the potential of machine learning for genomic prediction in different trait-genomic structures. The aim was to show that outcomes achieved with reduced data would improve predictive accuracy ( ) and precision (root-mean-square error: RMSE) while reducing computational time. Distinct subsets of markers, in simulated data, were chosen by prioritizing importance via the Bagging technique. Predictive modelling was subsequently conducted using both Bagging and the diverse architectures of a Multilayer Perceptron (MLP) neural network. This study was carried out with six traits of an F2 simulated population (derived from contrasting homozygotes) with 1,000 individuals. Three traits had three different heritabilities (0.4, 0.6, and 0.8) and were controlled by a set of 40 quantitative trait loci (QTLs). Additionally, four QTLs with more pronounced heritability effects (set at unity) were introduced in three other traits while preserving the same genetic control structure as the earlier traits. In our investigation, as the number of markers increased, both techniques gradually increased training time; however, the time needed for computation notably extended beyond the threshold of 100 markers for Bagging. In comparison to the MLP model, the Bagging model generally obtained better accuracy (higher ) and precision (lower RMSE) values regardless of heritability and added QTLs. Most importantly, results highlight that for traits subject to robust genetic control of additional QTLs, MLP networks experienced a decline in prediction performance from a few markers (~10). In contrast, Bagging kept constant or subtly improved predication performance. Finally, the dimensionality reduction procedure effectively improves genomic prediction, and Bagging captures complex genetic control structures for prediction better than MLP networks.

The prediction of traits allows the breeder to guide strategies to select and accelerate the progress of genetic improvement. The objective of this work was to determine the best prediction approach and establish a network with better predictive power for white oat using methodologies based on artificial intelligence, and machine learning. Seventy-eight white oat genotypes were evaluated. The design was randomized blocks with three replications. The models were evaluated with and without fungicide, and prediction models were established using four sets of experiments. The grain yield was used as a response trait the others as explanatory traits. The coefficient of determination was considered to evaluate the proposed methodologies. The importance of the traits was assessed through the impact of destructuring or disturbing the information of a given input on the estimation of R². For machine learning, decision trees, bagging, random forest, and boosting were used. The traits indicated to assist in decision-making are plant height, leaf rust severity, and lodging percentage. The R² ranged from 30.14% - 96.45% and 10.57% - 94.61% for computational intelligence and machine learning, respectively. A high estimate of the coefficient of determination, which was larger than the other estimates, was obtained using the bagging technique.

The development of superior cultivars involves parents with superiority for the traits of interest and wide genetic variability. Efficient plant breeding and selection strategies that allow for the identification of superior genotypes are essential in breeding programs. This work aims to carry out a comparative study between several strategies for choosing parents, for hybridization purposes, based on phenotypic analysis and molecular information. To obtain the phenotypic and genotypic information of the parents, data simulation was used. For genotyping, 2000 single nucleotide polymorphism markers were used, and from all possible gametes to be formed (22000), 5000 were randomly sampled to form each of the 100 individuals of the population of recombinant inbred strains. To obtain the phenotypic information, five characteristics with different levels of complexity were simulated. The comparative study was carried out using data referring to simulated genotypic values of hybrids and parents. Then, aiming to choose the parents destined for hybridization, different traditional selection strategies based on phenotypic analysis and the genome-wide selection methodology were approached. The genomic information resulted in the choice of the best lines and in obtaining superior hybrids when compared with traditional methodologies. The inclusion of the genomic genetic values of the parents in determining the crosses to be carried out increases the probability of generating phenotypically superior hybrids. Thus, the traditional methods of choosing parents for hybridization purposes are effective, but when incorporating the information from genome-wide selection, the choice of parents provides superior and promising results.

The objective of this work was to evaluate the genetic variability and associations of yield traits of pigeon pea, for indication of this crop for the breeding program in Mozambique. Eleven pigeon pea traits were evaluated in a experimental design with three randomized complete blocks, in the environments of Namapa (I) and Montepuez (II), in 2017/2018. To evaluate the traits, analyses of variance, genetic parameters, correlations, correlation network, and path analysis were used. In both environments, there is a high genetic and phenotypic variability for primary branches, secondary branches, pods per plant, number of seed per pod, pod width, pod length, and yield, as well as a high heritability for all traits except for plant height. The primary and secondary branches have a high correlation with yield and a positive direct effect on it. However, plant height and pods per plant show a high phenotypic variability and a positive correlation with yield in environment I. Pods per plant, weight of 100 seed, and primary branches show high direct positive effects in environment I, and secondary branches and number of seed per pod, in environment II. The primary and secondary branches show high variability and associations with yield in both environments and are indicated for the pigeon pea breeding program in Mozambique.

The GxE interaction is one of the major difficulties of plant breeding programs, both in the selection phase and in the recommendation of cultivars. To assess adaptability and stability, various statistical methods are used. The simultaneous use of some methodologies, using multi-information criteria for cultivar’s recommendation, can extract information that cannot be observed using each methodology separately. The aim of this work was to perform a large description of the behavior of flooded-irrigated rice genotypes, responding to environmental variations, using methods already established in the literature, but exploring the particularities of each methodology that together establish an information criterion for cultivar recommendation. To this end, 18 rice genotypes belonging to flood-irrigated rice breeding program were evaluated over four agricultural years, 2012/2013 to 2015/2016, totaling 12 environments (3 sites × 4 years). Multi-information estimates were performed to adaptability and stability analysis. There was no sign for the effect of the genotypes, and there was the significance of the effects of environment and GxE interaction. The aggregation of information and the large description of the behavior of the flooded rice genotypes demonstrated to be an efficient tool for studies of adaptability and stability.

The performance of genotypes in a wide range of environments can be affected by extensive genotype × environment (G × E) interactions, making the subdivision of the testing environments into relatively more homogeneous groups of locations (mega‐environments) a necessary strategy. The genotype main effects + genotype × environment interaction biplot method (GGE) allows identification of mega‐environments and selection of stable genotypes adapted to specific environments and mega‐environments. The objectives of this study were to identify mega‐environments regarding sorghum [Sorghum bicolor (L.) Moench] grain yield and demonstrate that the GGE biplot method can identify essential locations for conducting tests in different mega‐environments. A total of 22 competition trials of grain sorghum genotypes were conducted over three crop seasons across several production locations in Brazil. A total of 25, 22, and 30 genotypes were evaluated during the first, second, and third crop seasons, respectively. After identifying the presence of G × E interactions, the data were subjected to adaptability and stability analyses using the GGE biplot method. A phenotypic correlation network was used to express functional relationships between environments. The GGE biplot was found to be an efficient approach for identifying three mega‐environments in grain sorghum in Brazil, selecting representative and discriminative environments, and recommending more adaptive and stable grain sorghum genotypes.

Genomic‐wide selection (GWS) consists of the use of a large number of molecular markers for the prediction of genetic values and has been shown to be highly relevant for genetic improvement. The objective of this work was to evaluate and compare the predictive performance of statistical (ridge regression‐best linear unbiased predictor [RR‐BLUP] and BayesB) and machine learning methods through GWS in simulated populations with traits presenting different levels of heritability and quantitative trait loci (QTL) numbers in the presence of dominant and epistatic effects. The simulated genome of population F₂ was formed by 1,000 individuals and genotyped with 2,010 single nucleotide polymorphism (SNP) markers. Twenty‐six traits were simulated considering QTL numbers ranging from two to 88 and heritabilities of .3 and .6. The selective and predictive performances were evaluated using the multilayer perceptron (MLP), radial basis function (RBF), decision trees (DT), bagging (BA), random forest (RF), and boosting (BO) machine learning models and the classical RR‐BLUP and BayesB methods. A high effect of heritability was observed for the results of selective accuracy when compared to the increased QTL number. In addition, the selective accuracy based on the number of QTL demonstrates that the application of alternative machine learning models, such as RBF, BA, BO, and RF, can be suitable for the analysis according to QTL number. Machine learning methods are powerful tools for predicting genetic values with epistatic gene control in traits with different degrees of heritability and different numbers of controlling genes.