Assessment of genetic diversity in a crop germplasm is a vital part of plant breeding. DNA markers such as microsatellite or simple sequence repeat markers have been widely used to estimate the genetic diversity in rice. The present study was carried out to decipher the pattern of genetic diversity in terms of both phenotypic and genotypic variability, and to assess the efficiency of random vis-à-vis QTL linked/gene based simple sequence repeat markers in diversity estimation. A set of 88 rice accessions that included landraces, farmer’s varieties and popular Basmati lines were evaluated for agronomic traits and molecular diversity. The random set of SSR markers included 50 diversity panel markers developed under IRRI’s Generation Challenge Programme (GCP) and the trait-linked/gene based markers comprised of 50 SSR markers reportedly linked to yield and related components. For agronomic traits, significant variability was observed, ranging between the maximum for grains/panicle and the minimum for panicle length. The molecular diversity based grouping indicated that varieties from a common centre were genetically similar, with few exceptions. The trait-linked markers gave an average genetic dissimilarity of 0.45 as against that of 0.37 by random markers, along with an average polymorphic information constant value of 0.48 and 0.41 respectively. The correlation between the kinship matrix generated by trait-linked markers and the phenotype based distance matrix (0.29) was higher than that of random markers (0.19). This establishes the robustness of trait-linked markers over random markers in estimating genetic diversity of rice germplasm.

Genomic prediction is meant for estimating the breeding value using molecular marker data which has turned out to be a powerful tool for efficient utilization of germplasm resources and rapid improvement of cultivars. Model-based techniques have been widely used for prediction of breeding values of genotypes from genomewide association studies. However, application of the random forest (RF), a model-free ensemble learning method, is not widely used for prediction. In this study, the optimum values of tuning parameters of RF have been identified and applied to predict the breeding value of genotypes based on genomewide single-nucleotide polymorphisms (SNPs), where the number of SNPs ($P$ variables) is much higher than the number of genotypes ($n$ observations) ($P >> n$). Further, a comparison was made with the model-based genomic prediction methods, namely, least absolute shrinkage and selection operator (LASSO), ridge regression (RR) and elastic net (EN) under $P >> n$. It was found that the correlations between the predicted and observed trait response were 0.591, 0.539, 0.431 and 0.587 for RF, LASSO, RR and EN, respectively, which implies superiority of the RF over the model-based techniques in genomic prediction. Hence, we suggest that the RF methodology can be used as an alternative to the model-based techniques for the prediction of breeding value at genome level with higher accuracy.