A total of 265 new inbred lines derived from CIMMYT’s east Africa mid-altitude medium maturity breeding program (EA-PP1) were used in the present study. The selected 265 lines were developed from a set of elite drought donor lines in an elite-by-elite line combination. Specific drought donors like LapostaSequia-C7-F71, LapostaSequia-C7-F78, LapostaSequia-C7-F103, LapostaSequia-C7-F180, CML386, and DTPWC9 were used to develop these lines. In addition to drought tolerance, lines with good GCA and per se performance for high yield potential as well as resistance to several foliar diseases were also included in this study. An incomplete line-by-tester design (Kempthorne, 1957) was used to generate 795 testcrosses by crossing 265 lines to six single-cross testers. A total of 93 lines belong to CIMMYT heterotic group A while the remaining 172 lines belong to heterotic group B. All lines were crossed to three testers from opposite heterotic groups ( Supplementary Table S1 ). Pollen from each line was used to manually pollinate the single cross testers. The testcrosses were formed during the short rainy season (October–March) in the Kiboko maize research station in Kenya. The inbred lines were selected through rigorous phenotypic evaluations in the breeding nurseries and screened for foliar diseases. The testers used in the study are good combiners and are used as a female parent to adapt the three-way hybrid to mid-altitude environments. Detailed information on the pedigree of the inbred line is presented in Supplementary Table S1 .

A total of 795 testcrosses were evaluated in a multilocation trial under optimum and managed drought conditions. The experiments were interconnected by six commercial hybrid checks (DK8031, H513, PH3253, Pioneer 3253, DH04, and WH505). An Alpha (0,1) lattice experimental design with two replications was used. Each genotype was planted in two-row plots of 5 m in length, spaced at 0.75 m between rows and 0.25 m between hills. Each hill was planted with two seeds and later thinned to one plant per hill at 3 weeks after seedling emergence to adjust the final plant population to 53,333 plants/ha. Basal fertilizer application was performed at planting using di-ammonium phosphate (DAP) fertilizer at the rate of 60 kg N and 60 kg P₂O₅ per hectare. Six weeks after emergence, all experiments were top dressed with nitrogen fertilizer at the rate of 60 kg N/ha. All the experiments were kept weed-free by manual weeding and herbicide control.

The testcrosses were evaluated in Kenya at four locations under optimum management (Kakamega, Kiboko, Kirinyaga, and Shikutza) and at three locations (Kiboko, Homabay, and Mtwapa) under drought stress conditions ( Table 1 ). Drought stress experiments were conducted in the off season where it does not rain during cropping period. The drought stress fields are irrigated with a drip irrigation system. Irrigation was applied twice weekly for 3 to 5 h depending on the potential evapotranspiration (Trachsel et al., 2016). Irrigation of the trials was stopped at ~750 GDD after planting (2 weeks before the anticipated date of flowering) to induce drought stress (Trachsel et al., 2016; Ertiro et al., 2017). The well-watered or optimum trials were conducted in the main season under rainfed conditions. The trials were also supplemented with irrigation throughout the growth cycle to avoid drought stress. The management of drought stress experiment and phenotyping were according to the procedures outlined in the drought phenotyping protocols by CIMMYT (Zaman-Allah et al., 2016; Zaidi, 2019).

Data were collected per plot for each experiment under managed drought and optimum conditions for days to 50% anthesis (AD, as number of days after planting when 50% of the plants per plot shed pollen), days to 50% silking (SD, as number of days after planting when 50% of the plants per plot show silks), anthesis–silking interval (ASI) (as the difference between AD and SD), plant height (PH, measured as the length in centimeters from the base of a plant to the insertion of the first tassel branch of the same plant for 10 representative plants per plot), and ear height (EH, measured as the length in centimeters from the base of a plant to the internode of the top ear of the same plant for 10 representative plants per plot). At harvest, field weight [weight of dehusked ears (cobs) in kilograms per plot] and grain moisture (MOI, measured using a moisture meter on grain sampled from the center of five representative ears per plot) were recorded. GY was calculated using the field weight of ears per plot and a shelling percentage of 80, and adjusted to a moisture content of 12.5%. For drought trials, all trait measurements were performed according to the procedures outlined in the drought phenotyping protocols by CIMMYT (Zaman-Allah et al., 2016; Zaidi, 2019).

The restricted maximum likelihood method was utilized to conduct analysis of variance for each trial as well as a combined analysis across environment years using the META-R statistical package (Alvarado et al., 2020). The following linear mixed model was used:

where Y i j k b is the observed trait of interest; μ is the overall mean; E j is the effect of j t h environment; R ( E ) k j is the effect of k t h replication within j t h environment; B ( R E ) b k j is the effect of b t h block within environment; R ( E ) k j is the effect of k t h replication within j t h environment; B ( R E ) b k j is the effect of b t h block within k t h replication in j t h environment; G i is the effect of the i t h genotype; GE i j is the i t h genotype by j t h environment interaction and ϵ i j k b is effect of the residual error. Residuals are assumed as independent and identically distributed, ϵ ∼   i i d N ( 0 , ϵ i j k b 2 ). Heritability in the broad sense (Hallauer et al., 2010) was computed as:

where σ g 2 , σ g e 2 and σ ϵ 2 is the genotype variance, genotype-by-environment interaction variance, and residual variance, respectively. The number of replications and environments is denoted by r and e, respectively. The correlation between traits under managed drought and optimum conditions was calculated using the “cor” function in R (R Core Team, 2023). A correlation heatmap and the distribution of phenotypic values of the traits were generated using the ggplot2 R package.

Line-by-tester analysis was conducted by excluding the checks from the analysis. The total variance of the testcrosses values was partitioned into the GCA variance due to the line and tester as well as the specific combining ability variance (SCA) due to line × tester (Trachsel et al., 2016; Liu et al., 2021). Data were analyzed using the AGD-R version 4 software (Rodríguez et al., 2015). Analyses were performed for within (Equation 1) and across locations (Equation 2) using the joint linear mixed model as:

where Y i j k m is the observed value, μ is the overall mean; R k is the effect of k t h replication; B ( R ) k m is the effect of m t h incomplete block within k t h replication; L i is the GCA effect of the i t h line; T j is the GCA effect of the j t h tester; L T i j is the SCA effect of the i t h line by j t h tester, and ϵ i j k m is the residual effect assumed to be independent and identically distributed as ϵ ∼   i i d N ( 0 , ϵ i j k m 2 ).

where Y i j d k m is the response value, μ is the overall mean; E d is the effect of d t h environment; R ( E ) k d is the effect of k t h replication within d t h environment;   B ( R E ) m k d is the effect of m t h incomplete block within k t h replication in d t h environment; L i is the GCA effect of the i t h line; T j is the GCA effect of the j t h tester; L T i j is the SCA effect of the i t h line by j t h tester; L E i d is the GCA effect of i t h line interacting with d t h environment; T E j d is the GCA effect of j t h tester interacting with d t h environment; L T E i j d is the SCA effect of i t h line by j t h tester interacting with d t h environment and ϵ i j d k m is the residual effect assumed to be independent and identically distributed, ϵ ∼   i i d N ( 0 , ϵ i j d k m 2 ).

The effects of environment and replication within the environment were considered fixed, while line, tester, line × tester, and their corresponding interaction with environment and block within environment within replication were considered random. The restricted maximum likelihood procedure was used to estimate all variance components (Rodríguez et al., 2015). Tests of the significance of the variance component estimates and model comparison were performed using the likelihood ratio test (Stram and Lee, 1994; Giampaoli and Singer, 2009).

The estimates of the GCA effects for line, tester, and SCA effects of the line-by-tester interaction were estimated according to the procedure outlined by Singh and Chaudhary (1985). To test the significance of GCA and SCA at the 5% level of significance, the values of GCA for each line and tester and the SCA values of each hybrid were divided by their corresponding standard error, to compute their respective t-values. The calculated t-values were then compared with the tabular t-values at their corresponding error degree of freedom (Singh and Chaudhary, 1985). To assess the relative significance of GCA and SCA effects, Baker’s ratio was estimated using the variance of GCA of lines and testers and SCA based on the following formula:

where σ 2 G C A is the GCA variance of lines or tester and σ 2 S C A is the SCA line × tester variance. The closer the ratio to unity, the greater the predictability of a hybrid performance based on GCA effects alone (Baker, 1978).

The proportional contribution of total genetic variance due to lines, tester, and the interaction between lines and tester was computed as suggested by Singh and Chaudhary (1985). Pearson’s correlation of testcross hybrid performance (TCP) with the sum of GCA effects of both the parents r(GCA, TCP) using a leave-one-hybrid-out cross-validation was explained in detail by Schrag et al. (2009). All analyses were performed using the ASReml-R software version 3.0 (Butler et al., 2009).

Individual location analyses for GY under managed drought and optimum conditions revealed significant (p < 0.05) genotypic variance and the heritability estimates ranged from 0.17 to 0.73 ( Supplementary Table S2 ). Results of combined analysis across locations under optimum and managed drought conditions are shown in Table 2 . The analysis showed highly significant (p < 0.01) genotypic variance among the testcrosses for all the traits indicating that genetic variation among testcross hybrids was abundant. The ratio of genotype to genotype-by-environment interaction (GEI) variance for GY was higher in the optimum condition compared to managed drought stress condition, suggesting that GEI interaction was greater under drought conditions. The GEI revealed a highly significant (p < 0.01) variance for GY, AD, SD, ASI, PH, and EH under optimum and drought conditions. The coefficients of variation for all the traits under optimum and drought conditions were below 25% except for ASI. The estimates of broad-sense heritability for all traits were relatively higher under optimum conditions (0.53 to 0.86) than under drought stress conditions (0.25 to 0.85). The low experimental coefficient of variation and high broad-sense heritability for most of the traits reflected good experimental accuracy. A wide range of phenotypic variation was observed among the testcrosses for all the traits under drought and optimum conditions ( Figure 1 ). Drought stress considerably reduced GY, AD, and SD compared to optimum. Furthermore, drought stress prolonged the interval between AD and SD in the testcrosses. Plant height was reduced by drought stress compared to the optimum condition.

The mean performance of GY and other agronomic traits for the 20 selected best testcross hybrids and six commercial checks under managed drought and optimum conditions is presented in Tables 3 and 4 , respectively. Under drought stress, GY averaged 2.15 t/ha, with a range of 0.41 to 4.89 t/ha. In contrast, under optimum conditions, the average GY was significantly higher at 7.39 t/ha, ranging from 3.95 to 9.77 t/ha. Under drought stress, it took 58 to 79 days (average, 67 days) to reach 50% anthesis (AD). Under optimum conditions, AD averaged 68 days, ranging from 62 to 74 days. For silking (SD), it took 60 to 79 days (average, 70 days) under drought stress and 60 to 72 days (average, 66 days) under optimum conditions. The ASI averaged 2.8 days (range, −2 to 10 days) under drought stress and 1 day (range, −1.4 to 3.1 days) under optimum conditions ( Tables 3 , 4 ). However, under optimum conditions ( Table 3 ), PH varied from 172 to 265 cm with a mean of 226.4 cm under optimum conditions. Under managed drought conditions, it ranged from 153.2 to 263.1 cm with an average of 214.7 cm. Generally, plant and ear height were taller under optimum conditions than under drought stress conditions ( Tables 3 , 4 ).

Under managed drought conditions, testcrosses S13_5 × T2 (4.9 t/ha), S13_5 × T5 (4.6 t/ha), S7_9 × T2 (4.50 t/ha), S13_3 × T2 (4.4 t/ha), and S13_3 × T3 (4.3 t/ha) are among the top-performing hybrids, with GY outperforming the best commercial check (WH505; Table 3 ). The yield performance of the best testcross hybrid under managed drought conditions was almost double that of the best commercial check ( Table 3 ). Under optimum conditions, the best testcross hybrid (S2_8 × T1) outyielded the best commercial hybrid by 2.7 tons ( Table 4 ). Overall, drought stress reduced the GY by 20%. Under the optimum condition ( Table 4 ), testcross S2_8 × T1 (9.8 t/ha), S4_10 × T1 (9.46 t/ha), S6_4 × T1 (9.39 t/ha), S6_4 × T4 (9.35 t/ha), S2_8 × T6 (9.34 t/ha), and S2_14 × T4 (9.3 t/ha) are the best-performing hybrids. All the 20 best hybrids outyielded the best commercial check (WH505).

Insights into the relationships among traits is crucial for breeders when selecting suitable traits to incorporate in selection criteria. Significant (p < 0.001) coefficients of phenotypic correlation were observed among most traits across drought and optimum conditions ( Figures 2A–C ). Significant (p < 0.001) and positive correlations were identified between GY and PH and EH. PH and EH implied that selection improvement of these traits under optimum conditions could lead to an increase in GY ( Figure 2B ). Furthermore, significant (p < 0.001) and negative correlations observed between GY and AD, SD and ASI, and ASI with PH, EH, and AD under optimum conditions implied that that these traits are reliable selection indices for yield improvement in optimum conditions. A significant (p < 0.001) and negative correlation was observed between GY and AD_DS, SD_DS, and ASI_DS under drought stress. This implied that selection of GY alone selection may not be effective.

The genotypes and GEI variance components were significant at the 0.05% probability level for all traits under OPT and DS conditions ( Table 5 ). The GCA variance for lines and testers were highly significant (p < 0.001) for majority of the traits under both DS and OPT conditions, indicating high genetic variation among these genotypes. In contrast, GCA variance for testers was not significant for GY under DS conditions. The SCA variance for line × tester interactions was not significant for all the traits under DS and OPT conditions ( Table 5 ). Variances for line × environment and tester × environment interactions were significant (p < 0.001) for most of the traits under DS and OPT conditions. However, tester-by-environment interaction variances were not significant for GY, PH, and EH under DS conditions. SCA variances were not significant for most of the traits under DS and OPT conditions ( Table 5 ). Baker’s ratio ( Tables 5 , 6 ) was closer to or equal to 1 for all the traits under DS and OPT conditions.

The proportional contribution due to GCA variance of lines to the total genotypic variance varied from 81% to 100% under DS and OPT conditions ( Figure 3 ). In contrast, the contribution of GCA variance of tester to total genotypic variance ranged from 0% to 4% under DS and from 0% to 18% under OPT conditions ( Figure 3 ). Similarly, the SCA variance varied from 0% to 1% under DS compared to 0% to 3% under OPT conditions. Overall, proportional contribution due to GCA variance of lines was by far larger than the contribution due to tester and the interaction between line and tester variance.

The summary statistics of GCA effects of lines and testers for GY and other traits under DS and OPT conditions is shown in Figure 4 and in Tables 6 and 7 . The values of GCA estimates varied among the lines and testers for all the traits. Under DS conditions ( Table 5 ), the GCA effect of GY varied from −0.65 to 1.00 with 53% and 47% of lines found to be positive and negative effects, respectively. Lines S13_5, S13_6, S13_12, S6_11, S7_2, S2_8, S8_7, S2_7, S8_8, S2_15, and S8_19 were positive with high GCA effects for GY ( Figure 3 ). Tester T5 had the highest GCA effect, whereas tester T4 had the lowest GCA effects for GY. The GCA effects for other traits are also shown in Tables 6 and 7 .

The GCA effects of ASI ranged from −0.92 (S13_5) to 1.46 (S10_8), while tester GCA effects varied from −0.08 to 0.09 ( Table 7 ). More than half of the lines and testers showed a negative GCA effect for ASI. Under OPT conditions ( Figure 3 ; Table 6 ), GCA effects varied considerably among the lines and testers. The GCA effects for GY ranged from −0.01 to 0.72. Out of the total lines, 55% of the lines showed positive GCA effects for GY, whereas 45% of the lines displayed negative GCA effects for GY. Line S2_8 recorded the highest GCA effects for GY followed by lines S10_1, S6_4, S10_14, S2_14, S10_15, S8_7, S2_3, S8_19, and S6_9. Testers T1 and T5 showed a positive and high GCA effect for GY. Across DS and OPT management conditions ( Figure 4 ; Tables 6 , 8 ), lines S2_8, S10_1, S6_4, S10_14, S2_14, S10_15, S8_7, S2_3, S8_15, and S13_5 consistently exhibited a high and positive GCA effect for GY and a high and negative GCA effect for AD, SD, ASI, PH, and EH.

The summary statistics of SCA effects on selected testcross hybrids for all the traits evaluated under DS and OPT conditions is presented in Tables 7 and 9 . Across all the management conditions, some of the selected testcross hybrids showed consistently high and positive SCA effects for GY and a high negative SCA effect for other traits. This indicates that these testcross hybrids are good specific combiners for GY and possess desirable plant architectural characteristics including shorter number of days to anthesis and silking, reduced period between anthesis and silking, shorter plant stature, and lower ear placement. Prediction correlation between the observed testcross hybrid performance and the GCA-predicted hybrid performance was high for GY with r = 0.95 under OPT and r = 0.90 under DS conditions ( Figure 5 ). Prediction correlations were also high for AD and PH under both OPT and managed DS conditions.