Rice plays a key role in food security of Bangladesh. Climate change impact and ever-increasing population are pushing tremendous pressure on agriculture for increasing food production. Under the scenario of decreasing arable land annually by 0.43% and diminishing natural resources, no other viable alternative to increase production per unit area (Salam et al., 2020). Rice production in Bangladesh has increased by around four-fold during the last 5 decades through the introduction and use of improved varieties (MVs) and the practice of optimum crop management solutions. Although in recent years Bangladesh has attained self-sufficiency in rice production, it is still not sustainable. Different natural calamities and human-created crises are endangering food security. A study shows that Bangladesh will require 45 million tons of rice in 2050 to feed its 250 million people (Kabir et al., 2015). It will be a great challenge to meet this demand with the current rate of genetic gain in the yield of rice as estimated by Rahman et al. (2022) as 0.24% for winter rice (Boro) and 0.15% for monsoon rice (RLR-T.Aman), respectively. Therefore, the improvement of breeding materials needs to be focused on top of everything. Genetic gain in crops for a particular trait can be enhanced by shortening the breeding cycle, the time span required for the selection of parents from the progenies of a mating between two grandparents, and the recycling of high-value parents in the breeding program. Application of different speed breeding techniques, such as rapid generation advance (RGA), double haploid, embryo rescue, etc. is the effective means of shortening breeding cycle time. Recycling of elite germplasm in the breeding crosses increases the frequency of favorable alleles of quantitative traits like yield. The genomic selection approach expedites the recycling process of parents; can thereby accelerate the rate of genetic gain for yield.

Genomic selection (GS) is a form of marker-assisted selection, which utilizes markers across the entire genome to estimate genomic estimated breeding values (GEBVs) taking additive genetic effects of all genes into account. The GEBVs are directly used for making the selection of individuals for specific trait. As GEBVs can be predicted with or without phenotyping, the selection at early generation is possible, thus reducing breeding cycle time greatly. GS uses a training population of known phenotypes and genotypes to construct a model of each marker’s effect on the trait. The model is then applied to predict the phenotypic performance of the untested individuals having only genotypes. However, the reliability of such predicted phenotype depends on the accuracy of the estimates. The prediction accuracy is estimated from the correlation between the GEBVs of the individuals and measured phenotype for which it is available. The GS has been reported to be more efficient than the phenotypic selection considering resources involvement (Heffner et al., 2009; Jannink et al., 2010; Lorenz et al., 2011; 2012; Rutkoski et al., 2011; Rutkoski et al., 2012; Wang et al., 2012; Onogi et al., 2015; Spindel and Iwata 2018). Since its first application in cattle breeding (Schaeffer, 2006; Hayes et al., 2009; Venot et al., 2016; Wiggans et al., 2017). GS is increasingly being used in both plant and animal breeding programs to accelerate genetic gain of the traits governed by minor genes (Juma et al., 2021). The application of GS in rice was first reported by Grenier et al. (2015) and its use in rice breeding is continuously increasing. The GS in rice has been used for selection against yield (Xu et al., 2014; Grenier et al., 2015; Spindel et al., 2016; Wang et al., 2017), heading date (Onogi et al., 2016), plant height, flowering time (Grenier et al., 2015; Spindel et al., 2016; Wang et al., 2017), panicle weight (Grenier et al., 2015), tiller number, grain number, thousand kernel weight (Xu et al., 2014; Wang et al., 2017), as well as panicle length, secondary branch number, and productive panicle number per plant (Wang et al., 2017). Iwata et al. (2015) suggested that GS could be useful for predicting rice grain shape, with average accuracy ranging from 0.40 to 0.64. The GS accuracies for grain yield ranged approximately from 0.09 to 0.40 across different studies (Xu et al., 2014; Grenier et al., 2015; Spindel et al., 2016; Wang et al., 2017). In a GS study for heading date with 174 backcrossing inbred lines together with its parental lines of rice using different models, Onogi et al. (2016) reported very high accuracy (r > 0.9) across all models. The accuracy for plant height and flowering time ranged approximately from 0.25 to 0.86 in different studies (Grenier et al., 2015; Spindel et al., 2016; Wang et al., 2017). The GS accuracy reported by Xu et al. (2014) for tiller number, grain number, and thousand kernel weight ranged from 0.67 to 0.69 depending on the models used. In general, the GS accuracy in rice studies varies by trait, population, and the models being used. The commonly used genomic prediction models are ridged regression best linear prediction (rrBLUP) (Whittaker et al., 2000; Meuwissen et al., 2001), Bayesian LASSO (BL) (de los Campos et al., 2009; Park and Casella, 2008), reproducing kernel Hilbert spaces (RKHS) (Gianola et al., 2006) regression, and random forest (RF) (Breiman, 2001). The ridge regression best linear unbiased prediction (rrBLUP) model performs adequately well compared to many other models (Spindel et al., 2015; Spindel et al., 2016; Spindel and Iwata 2018). However, prediction accuracy depends on many factors, including the model, crop, size of the reference population, extent of linkage disequilibrium (LD), marker set, and heritability of the trait of interest (Crossa et al., 2010). Accurate phenotyping of a large training population, preferably over multiple environments and years is required to derive accurate predictions due to the interactions between these factors (Rikkerink et al., 2007; Xu and Crouch, 2008; Resende et al., 2012; Desta and Ortiz, 2014). In this paper, we report the progress of testing training populations in multiple environments and further scope of applying GS in enhancing genetic gain in the breeding program aiming to develop rice varieties for the favorable irrigated ecosystem of Bangladesh.

Grain yield data of 1445 breeding lines tested in 64 historical trials during 2014–2019 under the irrigated breeding program of Bangladesh Rice Research Institute (BRRI) were used to estimate baseline genetic gain. The trials included only the elite breeding lines and released varieties as standard check varieties with up to a maximum of 8% common entries in the succeeding years. Performance BLUP for yield extracted for each the breeding lines and used to determine baseline gain, while genomic BLUPs for 3767 breeding lines evaluated at multi-locations under 183 trials during 2019–2022 were extracted and used to estimate the rate of changes in genetic improvement of rice yield.

Genetic gain is the amount of increased genetic improvement of a population over time due to intervention of selection for specific traits. It is usually estimated per unit of time and/or per unit area and/or per unit investment. Measuring the genetic gain of rice breeding programs is extremely important, as it is the staple food crop in Bangladesh. The analysis of baseline genetic grain for yield based on trial year shows that the irrigated breeding program of Bangladesh Rice Research Institute had a value of 0.0146 t ha⁻¹ year⁻¹ (Figure 1) from 2016 to 2019. This rate of genetic improvement is quite low and inadequate compared to the expected genetic gain of at least 0.044 t ha⁻¹ per year (approximately 1% annually) (Kabir et al., 2015) to meet Bangladesh’s requirements through 2050 for ensuring food security. In a study using historical series data of the released varieties over 50 years from 1970 to 2020, Rahman et al. (2022) reported the baseline genetic gain as 0.01 t ha⁻¹ year⁻¹ for both rainfed lowland (monsoon rice) and irrigated rice (winter rice). However, these rates are consistent with those observed for other South Asian rice breeding programs serving favorable environments (Kumar et al., 2021). In general, low rates of genetic gain in South and Southeast Asian rice breeding are likely due to long breeding cycles caused by repeated use of older, popular varieties as parents, and by limited selection intensity for yield in multi-location trials. Post facto analysis of BRRI crosses showed that in its irrigated breeding program, the most popular varieties BRRI dhan28 and BRRI dhan29 were repeatedly used as parents. Also, frequent use of landrace varieties in the crossing programs (Supplementary Table S2) without proper pre-breeding activities has resulted in limited improvement in additive breeding value for grain yield of rice in the irrigated breeding program.

Breeder’s equation (Lush 1937) suggests that by increasing the selection differential per unit of time or cost, the genetic gain can be enhanced. Increased selection differential depends on the trial heritability/accuracy, selection intensity, the genetic variance of the trait, and re-cycling time or length of the breeding cycle. Multi-location trials improve trial heritability. The inclusion of high-breeding-value parents in the breeding program increases genetic variance and selection intensity. Cutting-edge speed breeding techniques, such as RGA, double haploid, embryo rescue, etc. have shown promise in reducing breeding cycle time (Cobb et al., 2019; Ahmar et al., 2020; Shanmugavel et al., 2022). GS provides an opportunity to hasten the cycle of selection. It also showed the potential to select high-breeding value individuals from early-generation populations without extensive field testing. GS has been shown effective in wheat (Bonnett et al., 2022), maize (Beyene et al., 2021), barley (Sallam et al., 2015), and even rice (Xu et al., 2021). In our study, we also found that genetic gain per unit of time is much faster in the GS strategy than in the conventional selection methods. Since 2018-19, the GS approach is routinely practiced in selecting high-breeding value parents for recycling in the breeding program. One hundred Six superior lines with high GEBVs comprising 27 lines from the breeding trials conducted in 2018–19, 23 lines from 2019–20’s trial 31 from 2020–21’s trial and 25 from 2021–22’s trial were isolated and recycled in the crossing program (Tables 2, 3) and thereby, frequency of favourable alleles for yield has been increased in the breeding population. GS strategy helped grab the high GEBV lines as it accounts for the marker effect with the phenotypic performance (Contaldi et al., 2021).

Prediction accuracy is a very important factor for applying GS in filtering selection candidates. The prediction accuracy depends on various factors including the model used in the GS scheme. The rrBLUP is the frequently used GS model in the field of plant breeding. However, Rutkosky et al. (2012) reported Reproducing Kernel Hilbert Spaces (RHKS) regression and Random Forest (RF) regression as the most accurate models for genomic prediction. The simplicity of rrBLUP to extract marker effect made it popular among plant breeders. Thus, we used this model in our study for genomic prediction of the untested breeding lines. Another factor is the size of the training population which significantly affects the accuracy of genomic prediction. In a study of optimization of the training population size, we found that prediction accuracy gradually increases with the increase of breeding lines and sharply increases when TP contains more than 80% of the breeding lines of a validation population (Figure 4). Data quality of the training population is another important factor for GS accuracy. The heritability of a trial is an ideal indicator of data quality. Heritability for a quantitative trait like yield between 0.4 to 0.6 is considered to be optimum for the best quality data.

The GS technology not only can capture high-value parents but can be used to predict the performances of the untested (validation population) lines together with the tested lines, thus it saves resources required for the field testing of the whole population. The sparse testing approach of GS (Jarquin et al., 2020), in which the breeding population is subsetted into training populations with different but genetically related lines by pedigree for field testing, saves resources further by reducing duplication of the lines across the locations. Applying the sparse testing model, we evaluated 650 breeding lines in 2020–21 and 548 breeding lines in 2021–22 at four locations without testing all of them in all locations. The level of prediction accuracy observed at Cumilla (0.456 and 0.396), Gazipur (0.715 and 0.603), Habiganj (0.499 and 0.378) and Rangpur (0.456 and 0.329) during 2020–21 and 2021–22 (Figure 3), respectively suggests the reliability of the predicted performance of the untested breeding lines. The GS accuracy for grain yield of rice was reported with a range from 0.09 to 0.40 in many studies conducted by Xu et al. (2014), Grenier et al. (2015), Spindel et al. (2016), Wang et al. (2017). In addition, GS accuracy in the multiple advanced yield trials conducted at different regional research stations with different sets of breeding lines in Boro 2018–19 (431 lines 26 trials), Boro 2019–20 (816 lines in 43 trials), Boro 2020–21 (1491 lines in 71 trials), Boro 2021–22 (1029 lines in 43 trials) season were 0.672, 0.701, 0.454, and 0.509 with high (1.933, 2.623, 0.989, and 1.58, respectively) relative efficiency of prediction (Table 2; Supplementary Figure S1). These results indicated that the sparse testing model of GS was as effective for capturing expected selection candidates as the GS models with the same set of training population tested across the sites.

The rate of genetic gain can be improved by increasing selection differential with sufficient accuracy and decreasing cycle time (Cobb et al., 2019). Before adopting the RGA technique in advancing segregating populations, the breeding cycle length was roughly 8–10 years in the breeding program of BRRI and IRRI (Collard et al., 2017; Cobb et al., 2019). The cycle time of BRRI’s breeding programs has been cut down from 8–10 years to 4–5 years by the use of RGA techniques (Rahman et al., 2019) as shown in Supplementary Table S3. For further reduction in cycle time, in this study, the GS technique was used for predicting the performance without yield testing of a portion of the total breeding lines at the initial yield trial called OYT and found reliable prediction accuracy in the trials of Boro 2020–21 and Boro 2021–22 (Figure 3). Also, marker-assisted selection was performed for different target traits viz. cold tolerance, disease and insect resistance, grain quality, etc. using trait makers embedded within the 1K-RiCA panel for filtering the superior selection candidates. Thus, GS has cut down another 0.5–1.0 years that would be required for yield testing in the advanced yield trials and phenotyping for grain quality and pest reaction before selecting parents for recycling (Figure 7). Applying GS together with MAS for key target traits at the line fixation stage (F₄-F₅) has further reduced cycle time by at least half a year and thereby increased the rate of genetic gain as indicated in Figure 6. However, the accuracy of prediction was compromised greatly (Figure 5). Careful selection of breeding lines in the training population, recycling only the elite lines with adequate genetic variance for the traits as parents in the crossing program and good quality phenotyping data could improve prediction accuracy. Non-etheless, practicing GS for the consecutive 4 years from 2019 to 2022, genetic improvement for yield has been recorded at the rate of 117 kg ha⁻¹ year⁻¹, which is around 6.77 fold higher than the baseline gain.

Based on the above findings it can be concluded that by applying GS, superior lines with high breeding value can be reliably captured with and without extensive field phenotyping. GS approach particularly sparse testing of the training population saved resources required for the phenotyping without sacrificing prediction accuracy. Moreover, results show that by practicing GS at OYT level, breeding cycle time could be reduced to 3.5 years from the existing 4.5 years. If GS is performed at the LST stage, cycle time can be further reduced by another half a year.

The original contributions presented in the study are publicly available. This data can be found here: https://www.ebi.ac.uk/eva/?eva-study=PRJEB59909.