Efficiency of indirect selection in sorghum breeding for target production environments in moisture deficit areas of Ethiopia Authors: Teklay Abebe Teferi, Taye T Mindaye, Gemechu Keneni, Gurja Belay, Amare Seyoum, Tokuma Legesse, Gebreslassie Gebru Berhe Published: 2026 ABSTRACT AbstractThe success of crop breeding program ultimately hinges on developing, disseminating, and integrating suitable varieties into the production system. A critical challenge in variety development process is whether indirect selection in one environment is more effective than direct selection in the target environment. Hence, the efficiency of direct vis-à-vis indirect selection was compared, and recommendations are provided to support more objective breeding strategies. Thirty-six series of variety trials conducted between 1994 and 2020 at the pre-national and national levels, and evaluated under selection and target environments were used for analysis. Grain yield, heritability, predicted genetic gain, and genetic correlations varied significantly across environments. The mean yield was 3565.6 kg ha− 1 in the selection environment (SE) Melkassa, and 3627.6, 5079.1, 5059, and 4200.9 kg ha− 1, respectively in target environments/TEs/ Miesso, Kobo, Srinka, and Zema. Similarly, for SE (Miesso), the mean yield in the TE (Kobo, Zema, Erer, Shewarobit, and Sheraro) were 3802.1, 1450.1, 2819.2, 6172.5, and 2366.5 kg ha− 1, respectively, with the SE itself yielding 3165.8 kg ha− 1. Heritability estimates ranged from 5% to 96% with the majority (84.7%) of the trials exhibited medium to very high, while genetic correlations ranged from 0.002 to 0.62 indicating substantial genotype-by-environment interaction. The correlation suggests that the alleles controlling physiological mechanisms in the SEs are at least partially different from those controlling in the TEs. When Melkassa was used as SE the efficiency of indirect selection relative to direct selection was 0.23, 0.21, 0.33, and 56 for Miesso, Kobo, Srinka, and Zema, respectively. When Miesso was used as SE, the efficiency of indirect selection was 0.14, 0.08, 0.16, 0.17, and 0.06 for Kobo, Zema, Erer, Shewarobit, and Sheraro, respectively. This indicated that indirect selection under SE can identify superior genotypes for TEs to some extent, however, the relative efficiency of indirect selection was consistently lower than direct selection under the TEs themselves. These results highlight limitations in the current sorghum breeding approach and suggest that the selection strategy requires refinement to better predict performance across diverse growing conditions and to maximize genetic gains.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-52825-9.Keywords: Broad-sense heritability, Direct selection, Genetic correlation, Indirect selection, Predicted genetic gain, Selection environment, Target environmentSubject terms: Biotechnology, Genetics, Plant sciences Introduction The global food production system, particularly in the developing world, face increasing pressure from rapidly burgeoning population, limited arable land, climate change, and environmental degradation1–3. At the same time, the global population projected to reach between 9 and 10 billion by 20504,5. Meeting this demand will require food production to increase by 60–110%6, while the expansion of cultivated land is no more possible indefinitely7–10. This demands paradigm shift and transformative technological advances in agriculture to substantially enhance food output in a sustainable manner1,6,11. Crop production plays a critical role to food security as well as economic and social development7,10,12. Crop production has undergone dramatic increase in recent decades, driven by both the expansion of agricultural land and, more significantly, by rising crop yields10. Nevertheless, crop productivity has generally remained below expectations13 primarily due to lack of exploiting the full genetic potential of crop14, biophysical challenges, and insufficient technological interventions15,16. The important ways of narrowing the gap between the actual and potential crop productivity involves genetic manipulation of the crop and better management of the growing environment15,17. However, genetic manipulation is preferred over continual manipulation of the growing environments because of its sustainability and affordability. In Ethiopia, the inception of formal crop breeding can be traced back to the mid- 20th century with the establishment of higher learning institutions18. Since then, Ethiopia’s agricultural research system has evolved significantly, progressing through various stages and advancements. As a result of these efforts, 1,603 varieties from 125 crops and 28 varieties from forage and pasture species have been released and registered19. Sorghum breeding in Ethiopia was first conceived in 195320. The establishment of the Ethiopian Sorghum Improvement Project (ESIP) in 1972 at Adama (the then Nazareth) represents a landmark for the establishment of formal sorghum research in Ethiopia. As a result of sustained efforts, 66 improved sorghum varieties have been developed and promoted for production19. It is evident that a mismatch exists between the varieties released and those actually adopted and cultivated by farmers in Ethiopia21 as well as in other parts of the world22–24. The main reasons, among others, may often been related to the discrepancies between the selection and target environments. Such differences can result in the development of varieties that are poorly adapted or irrelevant to the diverse agro-ecological conditions of the target production systems24. When genotypes perform consistently across environments, breeders can effectively evaluate germplasm in a limited number of environments at lower cost and deploy varieties across wider geographic areas. However, varietal development aimed at wide adaptation without adequately considering the diverse biophysical conditions under different agro-ecological zones, often results in underperformance and low adoption by the farmers. Ethiopia is characterized by remarkable geographical diversity, resulting in substantial environmental variability across both temporal and spatial scales25. In such heterogeneous environment, genotype by environment interaction (GEI) is expected to be substantial, leading differential genotype responses across environments26–28. Consequently, genotypes selected for superior performance under a particular set of conditions may perform poorly under different environmental conditions29,30. High levels of GEI hinder progress in breeding programs and complicate genotype selection31,32. GEI become particularly important when target breeding environments differ from the environments used for selection and evaluation32,33. Changes in genotype rank across environments and years, including crossover interactions, provide strong evidence of significant GEI34. Whether varieties centrally selected at a single or a limited number of primary breeding centers can consistently perform across diverse target agro-ecologies conditions therefore remains a major concern32,35. The environment for which breeding is undertaken should be clearly defined and classified into relatively homogenous groups to reduce the magnitude of GEI, and thereby maximize gains from breeding efforts36. In this context, optimizing primary selection environments using meta-analysis of long-term, multi-environment trial data (already available in Ethiopia) is critically important. Among the available approaches, evaluating the efficiency of selection in one environment as a primary breeding center for improving performance in other environments have been proposed as a means of optimizing selection environments across different recommendation domains26. The expected response in target evaluation environments to selection conducted in selection environments can be predicted from genetic variances in the target environments, heritability in the selection environments, selection intensity, and the genetic correlation between genotype performance in selection and target environments26,37. Despite apparent differences among test environments, systematic efforts to optimize sorghum breeding programs by clearly defining and classifying primary selection and target environments into homogeneous groups, so as to minimize GEI and maximize genetic gain, have not been adequately undertaken in Ethiopia. Therefore, this study is designed to determine the efficiency of indirect selection in the selection environments for improving productivity in target environments as compared to direct selection under the target environments; and reassess and redirect towards a more objective and perceptible breeding approaches. Materials and methods Description of the study areas The study has been conducted in three regions of Ethiopia. These include the Oromia (Melkassa, Miesso, and Erer), Amhara (Kobo, Shewarobit, Zema, and Sirinka), and Tigray (Sheraro). A series of experiments have been conducted in various agro-ecological zones simultaneously in the selection and target environments for decades (Fig. 1). Melkassa Agricultural Research Center (MARC) of Ethiopian Institute of Agricultural Research (EIAR) is a central coordinator of sorghum research at the national level and also the primary breeding center for a large mass of agro-ecologies considered as indirect selection site. Varieties initially developed at MARC have been tested at a number of collaborative centers/sites. Still MARC is coordinating the national breeding program, Miesso also considered both as the target (1994–2004) and selection (2010–2020) environment. Generally, these research sites (selection and target) represent arid and semi-arid Agro-Ecological Zone (AEZ) (Fig. 1), typically affected by severe moisture stress. Map of Ethiopia showing sorghum selection and target environments (A, Major Agricultural Zones; B, Elevation; C, Annual Temperature; D, Annual Precipitation). Source: In this figure A—38 and B to D—39. All maps were created by the authors using40 (version 4.5.1) with spatial packages including dplyr, ggplot2, raster, sp, sf, and tmap. A: Major Agricultural Zones data were obtained from the Ethiopian Ministry of Agriculture38. Climate data (B–D) were derived from the WorldClim database39(Version 2; https://www.worldclim.org/). The two direct selection environments, Melkassa and Miesso, exhibited comparable temperature-related climatic variables. Mean annual temperature was 21 and 22 °C, maximum temperature of the warmest month was 30.7 and 31.2 °C, minimum temperature of the coldest month was 10.9 and 12.1 °C, and mean temperature of the coldest quarter was 18.9 and 20.1 °C, for Melkassa and Miesso, respectively (Fig. 2A). The remining temperature-related climatic characteristics of these two sites are presented in same figure. In contrast, the two sites differ markedly in precipitation-related variables. Annual precipitation was 767 mm at Melkassa and 892 mm at Miesso, precipitation of the wettest month was 188 and 136 mm, precipitation seasonality was 93 and 59, precipitation of the wettest quarter was 466 and 376 mm, and precipitation of the coldest quarter was 30 and 56 mm, respectively (Fig. 2B). Climatic variables of the sampling sites. Source: Fick and Hijmans39, https://www.worldclim.org/. (A) Temperature variables; (B) Precipitation variables; Soil Variables; Vegetation and land cover variables. Comparing the soil characteristics revealed that Melkassa had higher soil cation exchange capacity 36 vs. 34.1 cmol kg-1, higher soil pH (7.59 vs. 7.27), and greater soil silt content (33.9% vs. 24%) than Miesso (Fig. 2C). Conversely, Miesso exhibited a higher proportion of soil clay (37.1% vs. 31%), higher soil organic matter content (82.7% vs. 77.3%), and higher sand content (39.9% vs. 35.1%). With respect to vegetation and land cover characteristics, the two sites showed comparable proportion of forest and grassland cover (Fig. 2D). However, cultivated land differs between the sites, accounting for 25.35% at Melkassa and 35.85 at Miesso. Among the target environments, Sheraro was identified as the warmest agro-ecology, recording the highest annual temperature (26.4 °C), with maximum temperature of the warmest and coldest months reaching 39.3 °C and 14.6 °C, respectively. The mean temperature of the wettest, driest, warmest, and coldest quarters were 24.7 °C, 25 °C, 29.3 °C, and 24.7 °C, respectively. In contrast, Srinka represent the coldest agro-ecology, characterized by the lowest annual temperature (11.7 °C). The maximum temperature of the warmest and coldest months were 21.4 and 1.2 °C, respectively. Mean temperature for the wettest, driest, warmest and coldest quarters were 12.2 °C, 10.2 °C, 12.9 °C, and 10 °C, respectively. The remaining target environments, Erer, Kobo, Zema, and Shiwarobit, displayed broadly comparable temperature regimes across the evaluated bioclimatic variables (Fig. 2A). Lower annual precipitation was recorded in the target environment Sheraro (686 mm) and Kobo (815 mm), whereas the remaining sites received more than 1086 mm annual rainfall (Fig. 2B). Higher precipitation seasonality was observed in Sirinka (117), Sheraro (131), and Zema (116). Comparatively, lower seasonality values were recorded in Kobo (84), Erer (84), and Shiwarobit (65). Precipitation during the wettest quarter was highest in Zema (861 mm), followed by Sirinka (746 mm), and Erer (690 mm). Conversely, lower wettest quarter precipitation was obsereved in Kobo (396 mm), Sheraro (517 mm), and Shiwarobit (561 mm). Some of the target environments exhibited heterogenous soil characteristics (Fig. 2C). For instance, soil cation exchange capacity was lowest in Erer (27 cmol kg-1) and highest in Zema (49.7 cmol kg-1), while the remining sites displayed comparable values ranging from 38.9 to 42.9 cmol kg-1. Soil clay content was relatively lower in Sheraro (30.1%), Shiwarobit (32.3%), and Kobo (33.1%), whereas higher clay proportions were observed in Sirinka (45.9%), Zema (42.8%), and Erer (42.2%). Higher soil organic carbon content was observed in Sirinka (121.9 g kg-1), Erer (118.1 g kg-1), and Shiwarobit (116.1 g kg-1). In contrast, Kobo exhibited the lowest value (97.5 g kg-1), while Sheraro (105.3 g kg-1) and Zema (105.3 g kg-1) displayed intermediate levels. Soil pH across most target environments fell within the tolerable range for sorghum production (6.1 to 7.51). Exceptions included Erer, which exhibited moderately acidic conditions (pH 5.8), and Sheraro, which showed slightly alkaline conditions (pH 7.89). Sand content was highest in Sheraro (47.8%) and lowest in Sirinka (25%), while the remaining target environments showed comparable sand proportions ranging from 29.9% to 35.5%. Almost all target environments exhibited a high proportion of cultivated land, ranging from 34% to 44.99%, except for Erer, which had only 17.01% cultivated area (Fig. 2D). This lower level of cultivated land at Erer is associated with its extensive grassland (73.21%) and forest cover (5.76%). In contrast, Sheraro recorded the lowest the forest cover among the target environments (0.81%). The remaining sites, Kobo, Sirinka, Zema, and Shiwarobit, showed comparable forest cover, ranging from 2.36% to 2.67%. Similarly, most target environments (Kobo, Sirinka, Sheraro, Zema, and Shiwarobit) exhibited comparable grassland coverage, varying between 48.06% and 57.6%. Experimental design and agronomic practices Thirty-six series of pre-national and national variety trials were conducted simultaneously under selection and target environments for one to two years. The analysis considered Melkassa (1994–2004) and Miesso (2010–2020) as selection environments and many target environments representing the major sorghum-growing regions of Ethiopia. The historical dataset consisted of 36 experiments involving 818 genotypes tested across 98 environments or locations. (Table 1). These experiments were part of the Ethiopian multi-location varietal trial program. The genotypes evaluated included breeding populations, elite lines, released varieties, local cultivars, and exotic introductions. Trials were arranged using either lattice designs (LD) or randomized complete block designs (RCBD), typically with two to four replications. Table 1Overview of experiments under the selection environments (Melkassa, 1994–2004 and Miesso: 2010–2020) and target environments.ExperimentTrial nameYearEntryReplicationLocations1Early White Sorghum NationalVariety Trial Set-I1994144Melkassa, Kobo2Early White Sorghum NationalVariety Trial Set-III1994144Melkassa, Kobo3Early White Sorghum NationalVariety Trial Set-I1995124Melkassa, Kobo4Advanced Sorghum HybridTrial1996183Melkassa, Kobo5Early Red brown SorghumNational Variety Trial199684Melkassa, Kobo6Elite Sorghum hybrid Trial199684Melkassa, Kobo7Elite Sorghum hybrid Trial199884Melkassa, Kobo8Early white sorghum nationalvariety trial1999123Melkassa, Miesso, Kobo9Lowland Sorghum PreliminaryVariety Trial Set-II1999153Melkassa, Miesso, Kobo10Lowland Sorghum PreliminaryVariety Trial Set-III1999243Melkassa, Miesso, Kobo11Advanced Sorghum HybridTrial1999263Melkassa, Miesso, Kobo12Elite Sorghum Hybrid Trial Set II1999184Melkassa, Miesso, Kobo13Early White Sorghum NationalVariety Trial2000144Melkassa, Miesso, Sirinka14Elite Sorghum Hybrid TrialSet-II2000144Melkassa, Miesso, Adet15Early White SorghumPreliminary Variety Trial-I2000143Melkassa, Miesso, Srinka16Elite Sorghum Hybrid Trial2001164Melkassa, Miesso, Srinka17Early White Sorghum NationalVariety Trial2002184Melkassa, Srinka18Advanced Sorghum HybridTrial2002223Melkassa, Srinka19Early White Sorghum NationalVariety Trial2003184Melkassa, Miesso, Kobo20Early Maturing SorghumPreliminary Trial2003183Melkassa, Miesso, Kobo21Advanced Sorghum HybridTrial Set-II2003243Melkassa, Miesso, Kobo22Advanced Sorghum HybridTrial Set-III2003233Melkassa, Miesso, Kobo23Advanced Sorghum HybridTrial Set-I2004243Melkassa, Miesso24Advanced Sorghum HybridTrial Set-II2004223Melkassa, Miesso25Early Maturing SorghumPreliminary Trial2004183Melkassa, Miesso26Early Maturing NationalVariety Trial2010254Miesso, Kobo27Elite Sorghum Hybrid Trial2010204Miesso, Kobo28Stay Green Introgressedimproved lines2010313Miesso, Kobo29Advanced Tan Hybrid Trial2011343Miesso, Kobo30Elite Sorghum Hybrid Trial2011264Miesso, Kobo31Regional Sorghum Hybrids Trial2013163Miesso, Erer, Kobo32Regional Sorghum Hybrids Trial2013173Miesso, Erer, Kobo33Early Maturing National Variety2014173Miesso, Erer, Shewarobit, Kobo, Sheraro34Evaluation of Red SorghumHybrids2014382Miesso, Erer, Shiwarobit, Kobo, Sheraro35Stay Green ObservationNational Variety Trial20191002Miesso, Erer, Zema, Shiwarobit, Kobo, Sheraro36Stay Green National VarietyTrial2020722Miesso, Erer, Kobo Overview of experiments under the selection environments (Melkassa, 1994–2004 and Miesso: 2010–2020) and target environments. Early White Sorghum National Variety Trial Set-I Early White Sorghum National Variety Trial Set-III Early White Sorghum National Variety Trial Set-I Advanced Sorghum Hybrid Trial Early Red brown Sorghum National Variety Trial Early white sorghum national variety trial Lowland Sorghum Preliminary Variety Trial Set-II Lowland Sorghum Preliminary Variety Trial Set-III Advanced Sorghum Hybrid Trial Early White Sorghum National Variety Trial Elite Sorghum Hybrid Trial Set-II Early White Sorghum Preliminary Variety Trial-I Early White Sorghum National Variety Trial Advanced Sorghum Hybrid Trial Early White Sorghum National Variety Trial Early Maturing Sorghum Preliminary Trial Advanced Sorghum Hybrid Trial Set-II Advanced Sorghum Hybrid Trial Set-III Advanced Sorghum Hybrid Trial Set-I Advanced Sorghum Hybrid Trial Set-II Early Maturing Sorghum Preliminary Trial Early Maturing National Variety Trial Stay Green Introgressed improved lines Evaluation of Red Sorghum Hybrids Stay Green Observation National Variety Trial Stay Green National Variety Trial Experimental plots of either 11.25 m2 (2.25 m × 5 m) or 18.75 m2 (3.75 m × 5 m) were used, depending on the stage of the trials within the variety development hierarchy, corresponding to preliminary and national trials, respectively. Planting was conducted during the main growing season, from May to early July, depending on the onset of rainfall at each location and the relative physiological maturity of the genotypes. Inter-row and intra-row spacing were maintained at 75 cm and 25 cm, respectively. Fertilizer was applied at rates of 100 kg ha− 1 of diammonium phosphate (DAP) 18% N and 46% P2O5 and 50 kg ha− 1 of urea (46% N). Standard agronomic practices, including land preparation, weed control, pest management, and other pertinent practices were applied uniformly across all locations. Grain yield data were collected on a plot basis and subsequently converted to kg ha− 1 for analysis. Estimation of selection efficiency The efficiency of selection at one environment (primary breeding center or selection) for improving performance in another environment (target) was assessed following the methodology described by Falconer26. The expected response in the ultimate evaluation environments to selection in the selection environments was estimated using genetic variances in the evaluation environments, heritability in the selection environments, selection intensity, and the genetic correlation between genotype performance across environments. For this purpose, historical datasets collected over the last three decades from a series of variety trials conducted simultaneously at MARC and its sub-center Miesso (primary selection environment), as well as several collaborating testing sites, were compiled. Only datasets of good quality from preliminary and national variety trials conducted in common environments were retained and used for the meta-analysis. All statistical analysis and figures were performed using R software version 4.5.140 within the RStudio integrated development environment. The R packages dplyr, lmer4, and emmeans were employed for the meta-analysis41–43. Spatial delineation of agrological zones and visualization of climatic variables were conducted using the R packages raster, sf, ggspatial, ggplot2, and tmap44–48. Each experiment was analyzed independently to avoid biases associated with pooling trials differing in design, genetic composition, and environmental conditions. Although some of the trials were originally laid out in lattice designs and the others in RCBD, analyses were conducted using linear mixed models to ensure consistency across experiments. Genotype was treated as a random effect, while replication was included as a blocking factor were supported by the data. Variance components were estimated using restricted maximum likelihood (REML), and genetic parameters were derived from these components. Phenotypic variance was calculated at the entry-mean level by accounting for replication effects when applicable. Therefore, heritability estimates therefore correspond to entry-mean repeatability within each experiment. In addition, a pooled analysis of variance was conducted for each variety trial across years within each location using the following model: Where is the grand mean; ρijk and Φijk is the observed value of the trait for the ith genotype in the kth replication during the jth year; is the effect of the ith genotype; is the effect of the jth year; is the genotype × year interaction; is the effect of replication nested within year; and is the residual error. The total phenotypic variance for yield was partitioned into components attributable to genotype (), year () and genotype by year interaction (). Variance components were estimated using expected mean squares following the procedure of Singh and Chaudhary49 (Table 2). Table 2Combined analysis of variance under selection and target environments, assuming randomized complete block designs.Source of variationDegree offreedomMean Square (MS)Expected mean square (EMS)Component variancesYear (Y) Y–1 MS1 -Replications/Year Y (R-1) MS2 -Genotypes (G) G–1 MS3 G × Y (G–1) (Y-1) MS4 Error Y (G-1) (R–1) MS5 Y= number of years, R = number of replications and G = number of genotypes. Combined analysis of variance under selection and target environments, assuming randomized complete block designs. Y= number of years, R = number of replications and G = number of genotypes. Broad-sense heritability (h2) was calculated as: Genotypic correlations coefficients (rg) between environments were estimated using only genotypes common to the environments being compared. Within each experiment, the same set of genotypes was evaluated across locations using a balanced design with comparable replication, ensuring strong genetic connectivity between the selection and target environments. To ensure robustness, rg was calculated only when at least five genotypes were shared between environments, thereby avoiding unreliable estimates from weakly connected genotype sets. All genetic parameters, including variance components, heritability, and genotypic correlations, were estimated independently within each experiment. Therefore, rg between grain yield performances at the primary breeding centers (ρ) and target evaluation centers (ϕ) was estimated as: Where σg(xy) is the genotypic covariance between genotype effects expressed in the primary breeding centers (ρ) and the target evaluation center (ϕ). The terms x and y stand grain yield performance in the primary breeding and collaborative environments. The predicted response of grain yield to direct selection (DR) under the primary breeding (selection) environment (SE) and the target environments (TE) was determined as a function of selection intensity assuming a standardized value of 1.026, genotypic variance ( and broad-sense heritability (in the TE. The indirect response (IR) to selection in the SE to improve grain yield in the TE was determined as a function of selection intensity (i), the genetic correlation between environments (rg), genotypic variance ( in the TE, and broad-sense heritability of the SE. The relative efficiency (RE) of indirect selection (IR) practiced in the SE, compared with direct selection (DR) in the TE, was calculated as follows50–52. Where h(TE) and h(SE) are the square roots of heritability at the TE and SE, respectively. The relative reductions (RR) in yield, heritability, and genetic gain when moving from SE to TE was calculated as: . Results Grain yield performance under the selection and target environments varied substantially among trials and locations (Table 3 and Supplementary Fig. S1 and S2). Overall, mean grain yield tended to be higher in the target environments than in the selection environments, reflecting differences in environmental productivity. In the first scenario (1994–2004), mean yield at Melkassa (3565.6 kg ha–1) was lower than yields recorded at the target environments of Miesso (3627.59 kg ha–1), Kobo (5079.14 kg ha–1), Srinka (5059 kg ha–1), and Zema (4200.9 kg ha–1) (Table 3). In the second scenario (2010–2020), mean yield at Miesso (3186.2 kg ha–1) exceeded that of Erer (2762.9 kg ha–1) and Sheraro (2446.3 kg ha–1), but was lower than yields at Kobo (3699.3 kg ha–1) and Shewarobit (5987.5 kg ha–1). These contrasts reflect environmental heterogeneity among test locations, while evidence of genotype-by-environment interactions is supported by the significant genotype x environment effects observed in the analysis (Supplementary Tables S1, S2 and S3). Comparisons of mean yields suggest that the trials included promising candidate genotypes, as their performance was comparable to or exceeded that of released varieties, some of which were included in the evaluated genetic population (Supplementary Figs. S1 and S2). Table 3The average yield performance in the selection (Melkassa:1994–2004; Miesso: 2010–2020) and target environments in Ethiopia.YearExperimentGrain yield (kg ha− 1 ±SE)MelkasaMiessoKoboSrinkaZemaErerShewarobitSheraro199412121.7 ± 87.2-3923.1 ± 145-----199422691.9 ± 97-4309.6 ± 129-----199531405.1 ± 101-2921.7 ± 149-----199644490.4 ± 141-2281.8 ± 92.8-----199653045.5 ± 160-1475.5 ± 79.9-----199664276.2 ± 154-2362.5 ± 108-----199873294.1 ± 230-7151.6 ± 203-----199983220.7 ± 1533806.4 ± 1637341.6 ± 120-----199993717.2 ± 1913710 ± 1426784.1 ± 244-----1999103191.3 ± 993942.8 ± 1144572.1 ± 148-----1999113174.5 ± 964192.4 ± 1248348.1 ± 222-----1999122658.3 ± 93.85203.1 ± 1618089.3 ± 215-----2000133077.7 ± 1013006.3 ± 178-4271.2 ± 1673078.5 ± 101---2000145274.1 ± 1953294.1 ± 130--5323.3 ± 190---2000154391.4 ± 1732937.5 ± 105-5446.2 ± 180----2001164697 ± 107-4247.7 ± 1392002171220.5 ± 108--5529 ± 209----2002184058.2 ± 178--5800.9 ± 256----2003194122.4 ± 1284829.6 ± 90.44853.6 ± 146-----2003204541.9 ± 2084842.8 ± 1045094.6 ± 189-----2003213498.2 ± 3603081.5 ± 855662.6 ± 181-----2003223239.8 ± 1174474.1 ± 74.76094.3 ± 147-----2004234876.1 ± 1831790.1 ± 97.5------2004244922.1 ± 942462 ± 77.2------2004253934.1 ± 1232841.2 ± 92.7------Mean3565.6 ± 1473627.6 ± 115.95079.1 ± 157.45059 ± 190.24200.9 ± 145.5---2010262814.4 ± 453226.2 ± 87---201027-1953.9 ± 57.93012.9 ± 99-----201028-1983.5 ± 543425.9 ± 184-----201129-5465.7 ± 1325735.2 ± 158-----201130-4349.7 ± 52.54199.3 ± 111-----201331-3400.6 ± 91.84957.3 ± 104--3426.6 ± 167--201332-2806.4 ± 1323676.6 ± 142--2743.9 ± 138--201433-2745 ± 962660.1 ± 50.6--3105.8 ± 79.44923.7 ± 73.91798.7 ± 50.5201434-3280.1 ± 79.74753.7 ± 127--4101.3 ± 1797134.5 ± 1773258.8 ± 50.2201935-3477.8 ± 39.83893.7 ± 120-1450.1 ± 34.91025.8 ± 27.66459.2 ± 1062041.9 ± 45.3202036-2547.1 ± 1182281.7 ± 94.9--2511.8 ± 85.3--Mean-3165.8 ± 81.73802.1 ± 116.1-1450.13 ± 34.92819.2 ± 112.76172.5 ± 1192366.5 ± 48.7 The average yield performance in the selection (Melkassa:1994–2004; Miesso: 2010–2020) and target environments in Ethiopia. In the first scenario, with Melkassa as the SE, the relative yield reduction (RYR) indicates that, except for Meisso, most TEs produced higher yield than the SE Melkassa, as reflected by negative RYR values (Table 4). Negative RYR values indicate higher yield in the TE relative to the SE, whereas positive values indicate yield reduction in the TE compared to the SE. Relative yield reduction value close to zero or with non-significant difference from zero imply minimal yield differences between environments. Accordingly, except for Miesso (5.27%), negative mean RYR values for Kobo, Srinka, and Zema were negative − 61.2%, -93.12%, and − 0.48%, respectively, indicate that grain yield was higher in these TEs than in Melkassa (Table 4). Similarly, in the second scenario with Miesso as the SE, negative mean RYR values for Kobo (-23.15%) and Shewarobit (-67.7%) indicate that grain yield exceeded that of the SE Miesso. In contrast, positive mean RYR values observed for Zema, Erer, and Sheraro 58.3%, 9.65%, and 10.6%, respectively, indicate yield reductions in these TEs relative to SE Miesso, suggesting that the SE provided more favorable growing conditions than these environments. Table 4Relative yield reduction (RYR) between the selection and target environments in Ethiopia.YearExperimentRelative yield reduction (%)MiessoKoboSrinkaZemaErerShewarobitSheraro19941--84.9-----19942--60.1-----19953--107.9-----19964-49.2-----19965-51.6-----19966-44.8-----19987--117.1-----19998-18.2-128.0-----199990.1-82.5-----199910-23.6-43.3-----199911-32.1-162.9-----199912-95.7-200.4-----2000132.3--38.8-0.03---20001437.5---0.93---20001533.1--24.0----200116---6.9-20021789.8--353.0----200218---42.9----200319-17.2--17.74-----200320-6.6--12.17-----20032111.9-61.87-----200322-38.1-88.10-----20042363.3 - -----20042450.0 - -----20042527.8------Mean5.27-61.20-93.12-0.48---201026-14.6-----201027--54.2-----201028--72.7-----201129--4.9-----201130-3.5-----201331--45.8---0.8--201332--31.0-2.2--201433-11.6--9.60.1-10.1201434--44.9---25.0-117.50.6201935--12.0-58.370.5-85.741.3202036-10.4--1.4--Mean--23.1558.309.65-67.7010.60 Relative yield reduction (RYR) between the selection and target environments in Ethiopia. The magnitude of broad sense heritability (h2b) varied across years, trials and environments (Table 5), explain substantial environmental heterogeneity (Fig. 1) and genotype by environment interaction (Supplementary Table S1). In the first scenario, the mean h2b across trials was 69.4% in the SE Melkassa and 69.3%, 70.8%, 66.4%, and 74% in the TEs Miesso, Kobo, Srinka, and Zema, respectively. Across experiments, heritability estimates for selection and target environments were generally moderately high, indicating that a considerable proportion of the observed phenotypic variation was attributable to genetic differences among genotypes in most trial conditions. For instance, of the 15, 16, 5 and 2 trials shared between Melkassa (SE) and Miesso, Kobo, Srinka, and Zema (TEs), respectively, 73.3%, 75%, 80%, and 100% of trials showed moderately high to very high heritability (Table 5). Table 5The broad sense heritability (H2b) and relative broad sense heritability reduction (RHR) in the primary breeding centers and simulated target environments.YearExperimentHeritability (%)RHR (%)MelkasaMiessoKoboSrinkaZemaErerShewarobitSheraroMiessoKoboSrinkaZemaErerShewarobitSheraro1994188-86------2.5-----1994293-87------7.1-----1995375-91-------20.6-----1996482-65------20.8-----1996587-79------8.3-----1996686-87-------1.1-----1998722-93-------323.1-----19998929669------5.124.7-----19999878844------1.049.1-----199910778263------6.018.0-----199911748437------14.049.9-----199912709691------36.7-30.1-----2000137913-1864---83.4-77.620.1---2000148586--84----1.4-1.3---2000158280-80----3.2-2.6----20011674--72------3.3----20021756--74-------33.0----20021880--88-------11.0----200319727956------9.622.5-----200320924181-----55.611.9-----200321156272------306.5-371.6-----200322632932-----53.449.0-----2004236390 - - - ----43.7 - - - ---2004243678 - - - ----117.3 - - - ---200425536-------555.6------Mean6967716674 - - - -53.1-30.27.910.7 - - - 201026-8430------64.0-----201027-6846------33.2-----201028-9191------0.0-----201129-5052-------3.7-----201130-8189-------9.9-----201331-5195--48----87.1--6.0--201332-9179--25---13.6--72.2--201433-4358--156984--35.9--64.9-63.1-98.4201434-3050--49388--65.8---64-26.072.9201935-7365-65394970-11.4-11.646.232.64.4202036-4372--46----68.1---7.7--Mean - 646665375254--13.511.619.6-18.8-7.03 The broad sense heritability (H2b) and relative broad sense heritability reduction (RHR) in the primary breeding centers and simulated target environments. In the second scenario, the mean heritability estimates were 64% for Miesso (SE) and 66%, 65%, 37%, 52%, and, 54% for the TEs Kobo, Zema, Erer, Shewarobit, and Sheraro, respectively (Table 5). Out of the 36 experiments (2010–2020) in Miesso as SE and five TEs, nearly 19.4% showed low heritability, indicating a relatively stronger influence of environmental variation on the observed phenotypic performance in these trials. In comparison, when Melkassa served as the SE, heritability estimates were predominantly moderately high to very high, whereas medium to moderate high values were more commonly observed when Miesso was used as the SE. Notable, heritability estimates in Erer (TE) were relatively low, indicating reduced expression of genetic variance under the environmental conditions of this location. Relative Heritability Reduction (RHR) was used to quantify change in h2b when moving from the SE to the TEs. Enhanced expression of genetic variance was observed in TEs Miesso and Kobo relative to the SE Melkassa, as indicated by significant mean RHR values of -53.1% and − 30.2, respectively (Table 5). In contrast, the TEs Srinka and Zema exhibited reduced expression of genetic variance compared with Melkassa, with a positive mean RHR value of 7.9% and 10.7%, respectively. These results indicate that heritability increased in Miesso and Kobo relative to Melkassa, whereas it decreased in Srinka and Zema. However, changes inheritability alone does not determine selection efficiency, which also depends on the magnitude of genetic correlation between environments. When Miesso was used as SE, expression of genetic variance was relatively higher in the TEs Kobo, Shewarobit, and Sheraro as reflected by mean RHR values of -13.5%, -18.8%, and − 7.0, respectively. Conversely, under same SE, the TEs Zema and Erer displayed reduced expression of genetic variance, with positive mean RHR values of 11.6% and 19.6%, respectively. Direct response to selection (DR) and relative genetic gain reduction (RGGR) varied considerably across years, trials, and environments, reflecting strong genotype by environment interaction (Table 6). When Melkassa was used as the SE, mean DR values were comparable between Melkassa (578.2 kg ha–1) and Miesso (508.8 kg ha–1), while higher responses were observed in Srinka (844 kg ha–1), Kobo (654.9 kg ha–1), and Zema (610.2 kg ha–1). When Miesso served as SE, the highest mean DR was observed in Kobo (699.8 kg ha–1), followed by Shewarobit (562.5 kg ha–1), Zema (406 kg ha–1), Miesso (410.5 kg ha–1), Erer (304.7 kg ha–1), and Sheraro (227.9 kg ha–1). These results indicate that the expected response to selection differed substantially among environments, with relatively lower genetic gain predicted in Erer and Sheraro compared to other target environments. Importantly, variation in DR reflects differences in the expression of genetic variance and heritability across environments rather than differences in mean yield alone. Thus, environments with higher mean yield did not necessarily correspond to higher expected genetic gain. Table 6The genetic gain from direct response to selection (DR) and relative genetic gain reduction (RGGR) in the primary breeding centers and simulated target environments.YearExperimentDRRGGRMelkasaMiessoKoboSirinkaZemaErerShewarobitSheraroMiessoKoboSirinkaZemaErerShewarobitSheraro19941489.7-711.5-------45.3-----19942600.2-711.6-------18.6-----19953401.4-786.8-------96.0-----19964739.8-339.3------59.0-----19965666.4-277.0------58.4-----19966643.2-462.7------28.1-----19987157.0-1007.5-------541.8-----19998869.21058.5462.9------21.846.7-----199991035.5782.0537.3-----24.548.1-----199910538.7610.5604.3------13.3-12.2-----199911475.7807.6482.5------69.8-1.4-----199912406.81255.51505.4------208.6-270.1-----200013477.286.2-119.9333.5---81.9-74.930.1---200014954.3708.7-886.8---25.7-7.1---200015779.7465.3-799.1----40.3--2.5----200116467.6592.9-26.8200217316.4--976.1-------208.5----200218865.8--1731.8-------100.0----200319571.6468.9443.5-----18.022-----2003201316.3197.61007.2-----85.023-----200321275.8326.4884.0------18.3-221-----200322463.5115.2254.8-----75.145-----200423739.6670.1-----9.4 - - - ---200424180.8388.2------114.7 - - - ---20042522.2156.1------603.5------Mean578.2508.8654.9844.0610.2 - - - -37.47-54.76-52.5818.60 - - - 201026-572.1238.7------58.3-----201027-351.1342.6------2.-----201028-456.91138.2-------149.1-----201129-464.1579.2-------24.8-----201130-484.61086.8-------124.3-----201331-217.01067.6--486.6----392.0---124--201332-782.8761.8--168.4---2.7--78.5--201433-231.8438.2--114.1629.8268.7--89.1--50.8-171.7-15.9201434-161.5451.2--596.8455.725.6--179.5---270-182.284.2201435358.5950.7406121.7602389.4-165.21366-67.9-8.6202036-435.0642.7--340.3--47.8--21.8--Mean - 410.5699.8-406.0304.7562.5227.9 - -100.8--13.3-29.5-140.619.9 The genetic gain from direct response to selection (DR) and relative genetic gain reduction (RGGR) in the primary breeding centers and simulated target environments. The efficiency of the selection strategy was further evaluated using RGGR (Table 6). In the first scenario, where Melkassa served as the SE, most TEs, including Miesso, Kobo, and Srinka, exhibited negative mean RGGR values of -37.5%, -54.8%, and − 52.6%, respectively. These negative values indicate that direct selection within these TEs would result in greater genetic gain than indirect selection conducted in Melkassa. In contrast, Zema showed a positive RGGR value (18.6%), suggesting reduced relative genetic gain and lower efficiency of indirect selection from Melkassa for this environment. A similar trend was observed when Miesso was used as the SE. Indirect selection for the TEs Kobo, Zema, Erer, and Shewarobit resulted in negative mean RGGR values of -100.8%, -13.3%, -29.5%, and − 140.6%, respectively, again indicating that direct selection within these environments would be more effective. Conversely, Sheraro exhibited a positive RGGR value (19.9%), indicating reduced efficiency of indirect selection and a response pattern comparable to that observed for Zema. Overall, the observed variation in DR and RGGR confirms that genotype × environment interaction plays a major role in determining both the magnitude and stability of genetic gain across selection and target environments. The genetic correlation coefficients (rg) between the SEs and TEs ranged from weak to moderate level (Table 7), indicating significant genotype re-ranking across environments. When Melkassa was used as the SE, rg values with the TEs Miesso, Kobo, Srinka, and Zema ranged from 0.06 to 0.60, 0.04 to 0.62, 0.11 to 0.52, and 0.51 to 0.53, respectively, with corresponding mean rg values 0.28, 0.19, 0.26, and 0.52. When Miesso served as the SE, the mean rg values with the TEs Kobo, Zema, Erer, Shewarobit, and Sheraro were 0.14, 0.072, 0.14, 0.15, and 0.04, respectively, indicating generally weak genetic association between selection and target environments. The low to moderate genetic correlations values suggests limited predictability of genotype performance across environments, which may reduce the effectiveness of indirect selection. However, relatively strong genetic correlation (0.40 to 0.62) was observed in specific trials including Melkassa (SE) with Miesso (experiments 8, 9, 15, and 19), Kobo (experiments 12 and 20), Srinka (experiments 15 and 18), and Zema (experiments 13 and 14). These instances indicate that, under certain conditions, genotype performance was more consistent across environments, highlighting the context-dependent nature of genetic correlations. Table 7The genetic correlations (rg) and genetic gain from indirect selection (IR) in the primary breeding centers and simulated target environments.YearExperiment r g IRMiessoKoboSrinkaZemaErerShewarobitSheraroMiessoKoboSrinkaZemaErerShewarobitSheraro19941--0.15-------108.6-----19942-0.08------61.4-----19953-0.20------144.0-----19964-0.32------121.7-----19965-0.20------57.8-----19966-0.28------131.1-----19987--0.17-------82.9-----199980.560.08-----573.941.8-----199990.600.08-----464.363.9-----199910-0.060.35------37.8234.7-----1999110.270.17-----202.5116.5-----1999120.350.46-----378.6607.9-----200013-0.09-0.170.51----18.6-42.0189.0---2000140.15--0.53---102.6-476.7---2000150.54-0.50----256.1-404.4----200116-0.23-214.0200217---0.11-------92.3----200218--0.52------859.8----2003190.410.04-----181.622.1-----2003200.130.62-----38.1662.3-----2003210.250.22-----41.287.7-----2003220.240.27-----39.896.4-----2004230.37------209.3------2004240.34------90.1------2004250.13------8.0------Mean0.280.190.260.52 - - - 158.1141.1285.6332.9---201026-0.101------40.2-----201027-0.024------10.0-----201028-0.361------411.2-----201129--0.037-------20.9-----201130-0.046------47.4-----201331--0.02--0.09----18.0--44.5--201332-0.295---0.01---241.5---4.0--201433-0.135--0.0570.002-0.142-50.6--11.00.9-27.0201434-0.176--0.1160.1430.022-61.6--54.358.01.12019350.16-0.070.310.310.23-161.4-3150.8224.792202036-0.264--0.273---131.0--89.4--Mean - 0.140.070.140.150.04 - 101.5-31.041.094.522.0 The genetic correlations (rg) and genetic gain from indirect selection (IR) in the primary breeding centers and simulated target environments. In both SEs, the response to direct selection (Table 6) for grain yield consistently exceeded that of indirect selection (Table 7). When Melkassa was used as the SE, the mean genetic gains from direct selection were 578.2 kg ha–1 at Melkassa and 508.8, 654.9, 844, and 610.2 kg ha–1 in the TEs Miesso, Kobo, Srinka, and Zema, respectively (Table 6). In contrast, indirect selection practiced at Melkassa resulted in substantially lower genetic gains in these TEs, amounting to 158.1, 141.1, 285.6, and 332.9 kg ha–1, respectively (Table 7). Similarly, when Miesso served as the SEs, direct selection produced mean genetic gains of 410.5 kg ha–1 at Miesso and 699.8, 406, 304.7, 562.5, and 227.9 kg ha–1 in the TEs Kobo, Zema, Erer, Shewarobit, and Sheraro, respectively (Table 6). In contrast, indirect selection conducted at Miesso yielded markedly lower gains in these TEs, with mean responses of 101.5, 31, 41, 94.5, and 22 kg ha–1, respectively (Table 7). These results indicates that direct selection within target environments generally resulted in greater realized genetic gain than indirect selection from the SEs revealed superior efficiency of direct selection over indirect selection across both breeding scenarios. This pattern is consistent with the predominantly low to moderate rg observed between environments, which limit the effectiveness of indirect selection due to genotype re-ranking across environments. The relative efficiency of indirect selection was evaluated by comparing its effectiveness to that of direct selection across the TEs (Table 8). In all cases, the ratios of indirect to direct selection was below unity, indicating lower efficiency of indirect selection relative to direct selection. When Melkassa was used as the SE, the efficiency ratios for the TEs Miesso, Kobo, Srinka, and Zema were 0.23, 0.21, 0.33, and 0.56, respectively. Similarly, in case of Miesso served as the SE, the corresponding ratios for Kobo, Zema, Erer, Shewarobit, and Sheraro were 0.14, 0.08, 0.16, 0.17, and 0.06, respectively. The values indicate that indirect selection achieved only 6% to 56% of the genetic gains attainable through direct selection in the respective TEs. This consistent pattern suggests that selection conducted in the primary breeding environments does not fully translate into improved performance in the target environments. Table 8The relative efficiency of indirect selection to direct selection.YearExperimentIR/DRMiessoKoboSrinkaZemaErerShewarobitSheraro19941--0.15-----19942-0.09-----19953-0.18-----19964-0.36-----19965-0.21-----19966-0.28-----19987--0.08-----199980.540.09-----199990.590.12-----199910-0.060.39-----1999110.250.24-----1999120.300.40-----200013-0.22-0.350.57---2000140.14--0.54---2000150.55-0.51----200116-0.36200217---0.09----200218--0.50----2003190.390.05-----2003200.190.66-----2003210.130.10-----2003220.350.38-----2004230.31------2004240.23------2004250.05------Mean0.230.210.330.56 - - - 201026-0.17-----201027-0.034-----201028-0.36-----201129--0.04-----201130-0.04-----201331--0.01--0.09--201332-0.32---0.02--201433-0.11--0.100.00-0.10201434-0.14--0.090.130.04201935-0.18-0.080.420.370.24202036-0.2--0.26--Mean - 0.14-0.080.160.170.06 The relative efficiency of indirect selection to direct selection. Discussion The current breeding approach in Ethiopia is based on the development of varieties at a limited number of primary breeding centers, followed by the evaluation of promising genotypes in multi-environment trials for national release19,53. This approach assumes that genotypes selected at a few primary breeding environments (selection environments, SEs) will exhibit similar performance across diverse target environments (TEs). As a result of these efforts, 66 sorghum varieties were released for commercial production between 1970 and 202319. The objective of this study is not to call into question the efficiency of past breeding programs nor to undermine the substantial value of previously released varieties. Rather, our result indicate that the efficiency and impact of these breeding programs could be further enhanced through optimizing between selection and target environments, thereby increasing the consistency of genotype performance and the effectiveness of selection across diverse production conditions. Although the original experimental designs differed (e.g., lattice and RCBD layouts, plot sizes, and genetic materials), the use of within-experiment mixed-model analyses helps to reduce potential bias in variance component estimation. However, heritability estimates should be interpreted as entry-mean repeatability within individual trials rather than as generalized broad-sense heritability across environments. Consequently, estimates of genetic correlation and indirect selection efficiency are valid within experiments but should be interpreted cautiously when extrapolated across years or heterogeneous trial conditions. The experiments analyzed comprised diverse germplasm, including released varieties, standard checks, and local landraces, evaluated across different seasons, environments, and levels of experimental precision (e.g., plot size, number of replications, and design). The results showed that, in most cases, mean grain yield in the TEs (Kobo, Srinka, and Zema) exceeded that of observed in the SEs. The mean yield across experiments was comparable between Melkassa (SE) and Miesso (TE). Similarly, yield performance in the TEs Srinka, Erer, and Shewarobit generally exceeded that of Miesso (SE), whereas lower yields were observed in Zema and Sheraro relative to this SE. These differences in yield performance among environments likely reflect variation in climatic conditions, particularly precipitation and temperature (Fig. 1), as well as edaphic factors. This interpretation is supported by the strong genotype-by-environment interactions (G×E) effects detected (supplementary Tables S1, S2 and S3). Furthermore, yield variation within the same location across different experiments may be attributed to interannual climatic variability and differences in the genetic composition of the evaluated materials54,55. The analysis of relative yield reduction (RYR) revealed a consistent pattern with important implications for breeding strategy, particularly in evaluating the efficiency of a centralized selection approach. Relative yield reduction quantifies the difference in mean grain yield between selection and target environments. Positive RYR values indicate that mean yield in the selection environment exceeds that in the TEs, whereas negative values indicate superior performance in the TE (Table 3) as described by Saikumar56. When the RYR values are close to zero or not significantly different from zero, the two environments can be considered similar in terms of yield performance. These patterns demonstrate that genotypes selected in one environment often perform differently when evaluated in distinct TEs, indicating substantial environmental divergence. However, it is important to note that RYR reflects differences in environmental productivity rather than the efficiency of selection per SE. The observed RYR trend therefore provide evidence that selection and target environments are not interchangeable (Table 3 and supplementary Table S1), underscoring the importance of considering environment-specific or decentralized breeding strategies. The differential performance across environments, with yields varying substantially in both SEs and TEs (Table 3), highlights the agro-ecological diversity within the country38. This agro-ecological heterogeneity (Fig. 1) contributes to significant genotype-by-environment (G×E) interaction (supplementary Table S1) leading to differential yield responses32,57. The present analysis therefore supports the potential value of a more decentralized breeding approach, in which selection is conducted within or across a homogeneous target environment to identify genotypes with improved local adaptation and stability, thereby enhancing genetic gain and cultivar adoption across the regions57. Consistent with this view, De Sousa et al.58 reported that decentralized selection in the Ethiopia highlands can achieve 20–30% higher genetic gain in crop production. The result further indicates that, while breeding for wide adaptation may be economically attractive, achieving consistently superior performance across highly contrasting environments remains challenging due to differential genotypic responses and strong genotype-by-environment interaction. A genotype that performs best in one environment may be not necessarily rank similar in another reflecting the context-dependent nature of genetic expression. When varieties are developed primarily at centralized breeding sites but evaluated and released across diverse target environments, part of the genetic gain achieved at the selection environment may not fully translate to all target production environments. This can occur when locally adapted genotypes are not advanced during early selection stages due to performance under contrasting conditions. Consistent with this, Ceccarelli24 reported that centralized breeding may lead to the loss of a substantial proportion of locally adapted genotypes during early selection. Under such conditions, decentralized or environment-targeted breeding, defined as the development and selection of genotypes within specific or relatively homogeneous target environments, may provide an effective complementary strategy to better capture local adaptation and maximize realized genetic gain across diverse production systems59. Comparing heritability between the SE and TE provides an indicator of the potential response to selection60,76. When heritability is similar in both environments, it suggests comparable contributions of genetic and environmental effects to yield variation, which may support more consistent genotypes ranking across environments and improve the efficiency of indirect selection. In contrast, differences in heritability between SEs and TEs indicate environment-dependent expression of genetic variance, which may reduce the reliability of indirect selection and favor direct selection within target environments. According to Singh61, heritability values greater than 80% are considered very high, 60–79% moderately high, 40–59% medium, and below 40% low. In the present study, most environments exhibited high to very high heritability when Melkassa was used as the SE, whereas medium to high heritability values were observed when Miesso served as the SE (Table 5). This pattern reflects variation in the expression of genetic variance across environments and is consistent with the presence of genotype-by-environment interaction influencing trial expression. In general, negative RHR indicates higher heritability in the TE than in the SE, whereas positive values indicate the opposite (Table 5). The observed variation in heritability across environments provides additional evidence of genotype-by-environment (GxE) interactions (supplementary Tables S1), reflecting environmental-dependent expression of genetic variance. Collectively, these results suggests that the magnitude of genetic potential expressed by the evaluated populations is not uniform across sites. Consequently, selection conducted in a single primary environment, such as Melkassa or Miesso, may not consistently capture the full range of genotype performance expressed across diverse target environments, highlighting the importance of multi-environment or environment-targeted selection strategies. The superiority of direct selection over indirect selection was consistently demonstrated across environments, as reflected by response to selection (DR), indirect response (IR), and IR/DR ratios. Indirect selection efficiencies were consistently below unity, indicating that selection in primary breeding environments does not fully translate into genetic gain in target environments. This pattern is consistent with quantitative genetic theory, where selection response depends jointly on heritability and genetic correlation between selection and target environments. The efficiency of the selection strategy was further evaluated using relative genetic gain reduction (RGGR), which consistently confirmed the superiority of direct selection over indirect selection across environments (Table 6). These findings align with classical quantitative genetic theory, which suggests that selection is most effective when conducted in the target environment where improved genotypes are intended to be grown24. Ceccarelli15 reported that decentralized selection in target environments can achieve 1.5–2 times higher genetic gain compared to centralized station-based testing. Under such conditions, the expression of genes controlling target traits is better aligned with the production environment, thereby reducing the impact of G×E interaction and enhancing realized genetic gain62. Relative genetic gain reduction further confirmed that indirect selection in centralized environments is less efficient than direct selection in target environments. Overall, these findings indicate that substantial genotype-by-environment interaction reduces the effectiveness of centralized selection strategies. The genetic gain from direct response to selection (DR) varied significantly among environments (Table 7), highlighting differences in the expression of genetic variance across trials. This variation is consistent with the presence of genotype-by-environment (G×E) interaction (Supplementary Table S1), which influences both the magnitude and stability of selection responses58. Environments exhibiting stronger G×E interactions tended to show more variable and less predictable responses to selection, whereas those with relatively weaker G×E interaction displayed more stable and, in some cases, higher realized genetic gains (Table 7 and Supplementary Table S1). Similar patterns have been reported in sorghum under low-input conditions in West Africa63, faba bean under waterlogging stress37, maize under nitrogen deficiency50, and wheat under water stress64, where the advantage of direct selection increases with environmental divergence. The success of indirect selection largely depends on the genetic correlation between the SE and TE. In the present study, genetic correlation coefficients were generally low, indicating limited connectivity among environments and substantial environmental differentiation (Table 7). The variation in rg values suggests that yield expression is influenced by environmental-dependent genetic effect at multiple loci, consistent with previous findings that allelic effects may vary in magnitude and direction depending on the environmental conditions65,66. As environmental divergence increases, genetic correlations tend to decline and may approach zero or become negative, reducing the predictability of genotype performance across sites. Nevertheless, moderate to relatively high correlations (0.40–0.62) observed in specific cases indicate partial connectivity between certain environments, suggesting that some locations share similar genetic response patterns. Overall, the pattern of genotype re-ranking and variable genetic correlations suggests that the tested environments likely represent multiple mega-environments rather than a single homogeneous target population. Under such conditions, selection gains from one environment are not reliably transferable to others, limiting the effectiveness of centralized selection regardless of heritability levels26,58. While indirect selection can contribute to genetic progress and resources efficiency67,68, the combined evidence from DR, IR, IR/DR ratios, and RGGR clearly demonstrates the superiority of direct selection in TEs67. This reflects substantial G×E interaction, where different agro-ecologies generate distinct selection responses, limiting the effectiveness of centralized breeding environments69. This outcome is consistent with quantitative genetic theory, which states that selection response depends on both direct response in the target environment and genetic correlation between environments70. Together, these results strongly support decentralized, target-environment-specific breeding strategies that enhance genetic gain and improve cultivar adaptation65. The consistently low IR/DR ratios (< 1.0) further emphasize that indirect selection in primary breeding environments is less efficient than direct selection in target environments, confirming the importance of evaluating genotypes under the environments where they are intended to be grown50,71. The combined evidence from genetic correlations and selection efficiency indicates that sorghum breeding environments are not fully interchangeable and likely represent partially distinct adaptation environments. Consequently, breeding strategies relying on a limited number of centralized selection environments are structurally constrained under strong genotype-by-environment interaction72. These findings support decentralized or stratified breeding approaches, where selection is conducted within or across well-defined agro-ecological environments that better represent target production environments. Environment classification or reclassification based on genotype performance patterns could further improve selection efficiency by enhancing alignment between selection and target environments32. For hybrid breeding programs, testcross evaluation for parental selection should prioritize environments with high genetic representativeness of major target production environments73. In contrast, environments with low genetic correlation may be more suitable for identifying specifically adapted genotypes rather than broadly adapted parents74. Additionally, variation in management practices, input levels, and farmer preferences across regions may contribute to genotype re-ranking, emphasizing the importance of incorporating representative agronomic conditions into breeding trials75. Overall, integrating multi-environment genetic information with environmental stratification frameworks provides a more effective strategy for improving selection accuracy, optimizing resource allocation, and enhancing genetic gain. The results demonstrate that aligning selection environments with target production environments is critical for improving the efficiency and impact of sorghum breeding programs under heterogeneous agro-ecological conditions. Conclusion In Ethiopia, current breeding approaches for sorghum and other crops are largely centralized, operating from one or a few primary breeding centers. This approach assumes that genotypes performing well in these selection environments will also perform well across a wider range of TEs, thereby improving efficiency in the use of limited time and resources. The objective of this study was to evaluate whether optimizing the choice of selection and target environments, using historical multi-environment trial data, could enhance the efficiency of plant breeding programs. Our result show that grain yields, genetic variation, heritability and predicted genetic gains from direct selection were generally higher when selection was conducted within target environments than when selection was carried out indirectly in centralized selection environments. Comparisons of the relative efficiency of indirect selection in the selection environments versus direct selection in the target environments consistently demonstrated the superiority of direct selection. These findings suggest that the current selection environments (Melkassa and Miesso), which serve as primary breeding centers for sorghum, may not adequately represent the full range of production environments. In conclusion, the establishment of additional primary breeding centers at least in each region, strategically located and identified through systematic analysis such as the present study, may be needed to maximized genetic gain and improve varietal development. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We acknowledge the Melkassa Agricultural Research Center (MARC) of the Ethiopian Institute of Agricultural Research (EIAR), EIAR, Tigray Agricultural Research Institute (TARI), Oromia Agricultural Research Institute (OARI/IQQO) and Amhara Regional Agricultural Research Institutes (ARARI). We also extend our sincere appreciation to the entire national sorghum and millet improvement program coordination team at MARC. Author contributions TT, AS, and TL conceived the research and collected the data. TA and GG entered, analyzed, interpreted the data and drafted the first version of the manuscript. GK and GB revised and finalized the manuscript. All authors read and approved the manuscript. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data availability The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Teklay Abebe Teferi and Gebreslassie Gebru Berhe contributed equally to this work. Contributor Information Teklay Abebe Teferi, Email: teklayabebe6@gmail.com. Gebreslassie Gebru Berhe, Email: gebre192@gmail.com. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.