Twenty-five genotypes from the International Institute of Tropical Agriculture (IITA) yam breeding program were used in this study. Details of the genotypes and testing locations are presented in Supplementary Tables 1, 2 . The experiments were conducted at Ikenne, Abuja, and Ibadan, Nigeria, at the following coordinates: Ikenne (l6.85°N, latitude 3.69°E), Abuja (9.13°N, 7. 23°E), and Ibadan (7. 26°N, 3.54°E). The genotypes were planted using a 5 × 5 alpha lattice design, with three plants per plot at a spacing of 1 m × 1 m, and the experiment was replicated twice during the 2022 and 2023 planting seasons.

The yam tubers were harvested eight months after planting and a single representative tuber per plot from the five environments (a combination of three locations and two seasons) was used for sample preparation. Each sampled tuber was divided longitudinally from the proximal to the distal end. The tuber was peeled, washed, and towel-dried. From the middle section, 100 g was used for crude starch extraction, while the remaining portions were diced into 1 m × 1 m cubes to ensure proper representation of the entire tuber for sensory and instrument-based assays. A 500 g portion of each yam sample was weighed for quality assessment, focusing on sensory attributes of both pounded and boiled yam genotypes. Additionally, an instrumental assay was conducted exclusively for the pounded yam product. Cooking and pounding were performed using the QASA machine (QYP-6000, capacity 3.6 L capacity, Qasa, Cheerfengly Industrial Co., Ltd., Taipei City, Taiwan). A total volume of 380 ml of water was measured using a measuring cup and poured into each machine pounder. The cooking time was set to 10 min for each sample, after which 80 g was collected, placed in a ziplock bag, and stored in a warmer to retain temperature. The remaining cooked yam were then pounded for 3 min in the machine. Each sample was wrapped in cling film and placed in a ziplock bag. All samples were properly labeled with a marker to avoid mix-ups and to facilitate easy identification.

Ten panelists, all experienced yam eaters, participated in the assessment of the boiled and pounded yam samples. Each panelist was given 5 g of coded pounded yam and boiled yam samples, one at a time, for assessment. Descriptors, as indicated in Table 1 , were used for the evaluation.

Textural profile analysis was conducted using the TA.XT Plus Stable Microsystems Texture Analyzer (Serial No: 2-P6-Z10447-01-V0038D577; Stable Microsystems, Godalming, UK), coupled with a compression TA-40 platen cylindrical probe. The pounded yam samples from each genotype were molded for uniformity using an open-ended, custom-made cutter plastic with dimensions of 3.6 cm × 2.2 cm. The molding process was replicated three times to ensure accuracy and reliability. Pre-test speed, test speed, and post-test speed were set at 1 mm s⁻¹, 1.75 mm s⁻¹, and 5 mm s⁻¹, respectively. A trigger force of 10 g and a strain of 40% were used for the double compression test. The temperature of each cut sample was measured using an infrared thermometer (AD-5612A), ranging from 40°C to 55°C before placement on the TPA machine. The weight of the samples ranged approximately 25 g to 30 g. This machine simulates human chewing behavior.

A whole yam tuber was sampled and peeled, and 100 g from the middle portion was weighed and grated using a manual kitchen shredder. After shredding, the samples were transferred to a blender (IFM-C20G crush miller; Iwatani, Osaka, Japan) and fine-milled. Two hundred milliliters of water were added to facilitate blending for 2 min per sample. The samples were then poured into a 180 μm by 200 mm sieve (model no. 019-214 775-01; TokyoScreen Co. Limited, Tokyo, Japan), which was placed on a mesh for stability with a larger bowl underneath. Three liters of water was added to facilitate the washing of crude starch through the sieve, and the mixture was left for 1 h to complete the process, leaving the residue behind. The supernatant was removed, and the crude starch was air-dried and weighed after reaching a constant weight was achieved. The result was expressed as a percentage using the following formula:

Data analyses were performed using various packages in the ‘R’ environment for statistical computing (R Core Team, 2022, version 4.2.2). The sensory data for boiled and pounded yam—including variables such as texture, mealiness, appearance, color, aroma, taste, stretchability, moldability, and overall acceptability—were analyzed using multinomial and ordinal logistic regression models to evaluate the effects of genotype (Geno), environment (Env), and their interactions, using the “nnet” and “ordinal” packages in R.

Multinomial logistic regression models were fitted to assess the relationship between sensory attributes and explanatory variables.

Where:

p(yi = 1) is the probability that the response variable yi equals 1

1 − p(yi = 1) is the probability that the response variable yi equals 0

β0 is the intercept term, which represents the log-odds of the outcome yi = 1 when there are no predictors in the model.

β1 is the coefficient for the Environment (Env) predictor, which quantifies the change in the log-odds of the outcome associated with a one-unit change in the environment variable.

β2 is the coefficient for the Genotype (Geno) predictor, which quantifies the change in the log-odds of the outcome associated with a one-unit change in the genotype variable.

Envi represents the value of the environment predictor for observation.

Genoi represents the value of the genotype predictor for observation i.

For variables such as appearance, color, aroma, taste, overall acceptability, stretchability, and moldability, each coefficient (log odds) indicates the likelihood of being at or below the reference category (Meccakusa) and above it, relative to a baseline category. A positive coefficient indicates that the genotype is associated with higher log odds of receiving a lower variable rating, suggesting the genotype is more likely to be rated in the lower categories. A negative coefficient indicates that the genotype is associated with lower log odds of receiving a lower variable rating, suggesting the genotype is more likely to be rated in the higher categories. However, for texture and mealiness, the values represent regression coefficient (log odds) for each genotype relative to the reference category (Meccakusa). A positive coefficient indicates that the genotype is more likely to fall into the outcome category compared to the reference category (Meccakusa). A negative coefficient suggests the opposite.

Pearson correlation analysis was performed using the ‘metan’ package in R software version 4.1.1 (R Core Team, 2022) ‘corr_coef’ function from the package was used to plot the graph.

The multi-trait genotype ideotype distance index (MGIDI) was used to rank the genotypes based on multiple trait values, as suggested by Olivoto et al. (2019). A fixed effect model, Patterson and Williams (1976) was carried out on sensory data generated using the model:

For the first stage of the MGIDI, each trait was rescaled using this model

where the symbols indicate the following for trait j and genotype i: rXij is the rescaled two-way table; η_nj is the new maximum value after rescaling; φnj is the new minimum value after rescaling; η₀j is the original maximum value; φoj is the original maximum value; and θ_ij is the original value for the ith genotype. Each column ranged from 0 to 100, considered the desired sense of selection (increase or decrease), and sustained the correlation structure of the original set of variables. The values obtained after rescaling for ηnj and ϕnj in a state of positive gains (ϕnj = 0 and ηnj = 100) and in a state of positive gains (φnj = 100 and η_nj= 0) were used Olivoto et al. (2019). The second stage was to compute an exploratory factor analysis (FA) through rXij to group correlated traits into factors and then estimate the factorial scores for each row/genotype/treatment. The scores were then obtained from the data collected for the dimensionality reduction of traits and relationship structure using the following model:

where the letters indicate the following: F is the g × f matrix with the factorial score, Z is the g × p matrix with the rescaled means, A is the p × f matrix of canonical loading, R is the p × p correlation matrix between the traits, g is the number of genotypes, f is the factors retained, and p is the measured traits. The third stage was to compute an ideal genotype. For this, a [1 × p] vector was considered to be the ideotype matrix using the Euclidean distance between the scores of the genotypes, and the ideal genotypes were determined by the MGIDI index, as shown

where MGIDIi is the multi-trait genotype-ideotype distance index for the ith genotype, Fij is the score of the ith genotype in the jth factor (i = 1, 2, ..., g; j = 1, 2, ..., f), being g and f the number of genotypes and factors, respectively, and Fj is the jth score of the ideotype. The genotype with the lowest MGIDI is then closer to the ideotype and therefore presents desired values for all the analyzed traits.

The Pearson correlation plot ( Figure 2 ) shows the relationships between crude starch content and various traits related to pounded and boiled yam among 25 white Guinea yam genotypes. The texture of boiled yam was positively correlated with the texture of pounded yam (r = 0.60, p <0.01). The taste of pounded yam showed a strong positive correlation with its overall acceptability (r = 0.91, P <0.001). The color of pounded yam was also strongly positively correlated with its overall acceptability (r = 0.92, p <0.001). The taste of boiled yam showed a strong positive correlation with its overall acceptability (r = 0.93, p <0.001). Crude starch was positively correlated with the appearance, color, aroma, and taste of boiled yam (r = 0.57, 0.62, 0.59, and 0.51; p <0.01) respectively. Crude starch was also positively correlated with mealiness, the overall acceptability of boiled yam, and the color of pounded yam (r = 0.43, 0.50, and 0.40; p <0.05, respectively). Crude starch was also negatively correlated with the texture of pounded yam (r = −0.50, p <0.05). Other traits were positively correlated with each other (p <0.001, p <0.05), as shown in the correlation plot, while some correlations were non-significant ( Figure 2 ).

The Pearson correlation coefficient between instrumental textural profile analysis, sensory attributes on pounded yam, and crude starch traits for 25 white yam genotypes across five environments ( Figure 3 ) showed significance at various levels. There was a strong positive (P <0.001) correlation between the overall acceptability of pounded yam with aroma (r = 0.93), taste (r = 0.91), color (r = 0.92), and appearance (r = 0.94). There was a positive correlation between appearance and color (r = 0.94; p <0.001). Instrumental springiness was negatively correlated with hardness (p <0.05, r = −0.44) but positively correlated with chewiness (p <0.05, r = 0.45) and cohesiveness (p <0.001, r = 0.95). There was a negative correlation between stretchability and hardness (r = −0.57; p <0.01). Crude starch was positively correlated with gumminess, chewiness, and resilience (r = 0.75, 0.75, and 0.69; p <0.001, respectively). There was negative correlation between sensory texture with gumminess (r = −0.72) and hardness (r = −0.69; p <0.001). A negative correlation was observed between sensory texture and chewiness (r = −0.61; p <0.01). There was also a negative correlation between sensory texture and crude starch and adhesiveness (r = −0.50, −0.48; P <0.05, respectively). Gumminess was also positively correlated with crude starch, hardness, and adhesiveness (r = 0.75, 0.83, 0.65; (p <0.001, respectively). Some of the traits were also not significantly correlated with each other ( Figure 3 ).

The Log odds ratio (LOR) results ( Table 2 ) present the ordinal and multinomial logistic regression outputs for all variables assessed in the boiled yam genotype performance, with Meccakusa used as the reference genotype (LOR = 0). For the traits appearance, color, aroma, taste, and overall acceptability, each coefficient (log odds) indicates whether a genotype is at, below, or above the reference category (Meccakusa) relative to a baseline category. Both negative and positive coefficients were observed across genotypes at various significant levels, while some were not statistically significant. However, for texture and mealiness, the values represent regression coefficients (log odds) for each genotype relative to the reference category (Meccakusa). All genotypes showed positive coefficients for texture log odds, whereas for mealiness, both positive and negative log odds were observed.

For appearance, genotypes such as TDr1741073 (LOR = 0.904) and TDr1738003 (LOR = 0.231) had positive log odds, while genotypes like TDr1719029 (LOR = −2.795***) and TDr1719069 (LOR = −2.378***) showed significantly negative values.

For taste, TDr1741073 (1.417***), TDr1739023 (1.086**), and TDr1741044 (0.927**) were top performers, whereas TDr1719029 (−1.744***) and TDr1738023 (−0.956**) recorded low LORs. In terms of texture, most genotypes performed well, with TDr1739046 (2.321*), TDr1738074 (2.266***), and TDr1741042 (2.106***) exhibiting high and significant LORs.

For the pounded yam, the log odds ratio ( Table 3 ) present the ordinal and multinomial logistic regression results for all the variable assessed on the pounded yam genotypes. For variables such as appearance, color, aroma, taste, moldability, stretchability, and overall acceptability, each coefficient (log odds) indicates the likelihood of being at or below the reference category (Meccakusa), or above it, relative to a baseline category. Both positive and negative coefficients were observed across the variables. However, for texture and mealiness, the values represent regression coefficients (log odds) for each genotype relative to the reference category (Meccakusa). Both positive and negative coefficients were observed at various levels of statistical significance.

For appearance, genotypes such as TDr1741073 (LOR = −0.019) and TDr1719020 (−0.181) had comparatively higher (less negative) log odds, whereas TDr1738023 (−2.76*) and TDr1719069 (−2.825***) recorded significantly negative values. For color, TDr1741073 (1.172***) was the only genotype with a significant positive LOR, while most others, such as TDr1738023 and TDr1716069, had strongly negative coefficients. Texture was a strong differentiator, with top-performing genotypes including TDr1716069 (4.048***), TDr1739058 (3.424***), and TDr1741044 (3.195***), while TDr1719055 (−8.865) showed a large negative value. For mealiness, most genotypes had negative or extreme values; TDr1719020 (−13.035***), TDr1739046 (−15.026***), and TDr1739049 (−15.681***) recorded particularly poor performance. In contrast, TDr1719069 (1.338) and TDr1719029 (1.292) showed the highest positive LORs for mealiness.

Significance levels for boiled and pounded yam traits (p <0.1, p <0.05, p <0.01) are included to indicate the strength of evidence for differences in trait ratings relative to the reference genotype.

Figure 4 illustrates the variation in sensory quality among yam genotypes evaluated for pounded yam. The x-axis represents the different genotypes, with Meccakusa as the baseline reference, while the y-axis displays the log-odds ratio, reflecting the likelihood of a sensory attribute being perceived in each genotype relative to the baseline. The different colors correspond to the sensory attributes measured for pounded yam quality. The log-odds ratios varied across genotypes, highlighting differences in sensory qualities. Among the genotypes, TDr1741073 was perceived as having the best sensory quality for pounded yam compared to the baseline, Meccakusa.

Similarly, Figure 5 presents the sensory quality of white Guinea yam genotypes for boiled yam. The different colors represent the sensory attributes measured for boiled yam across the genotypes. The log-odds ratios varied, with genotypes showing values above zero being preferred for sensory attributes compared to the baseline, Meccakusa. The results indicate that TDr1741073, TDr1738003, TDr1741042, and TDr1739023 exhibited superior sensory quality for boiled yam relative to other genotypes, using Meccakusa as the reference.

The MGIDI index identified one factor based on boiled yam traits evaluated. Factor analysis one was associated with texture and mealiness. The average communality and uniqueness accounted for 76% and 24% of the total genetic variability in the data, respectively ( Table 4 ). Of the boiled traits evaluated, six showed desired selection gain using the MGIDI index. A total genetic gain of 22.11% was observed for the assessed multi-traits where increases were desired, and 2.85% for those where a decrease was desired ( Table 4 ).

Based on the MGIDI index for crude starch traits, four genotypes were also identified (TDr1741073, TDr1741042, TDr1719055, and TDr1738003) as high-performing genotypes for multiple traits assessed across 25 white yam genotypes ( Figure 6b , Supplementary Table 4 ).

The MGIDI index identified two factors based on the pounded yam traits evaluated. Factor Analysis one was associated with all pounded yam traits except stretchability and texture. FA2 was associated with texture and stretchability. The average communality and uniqueness accounted for 85% and 15% of the total genetic variability in the data, respectively ( Table 4 ). Pounded yam traits showed desired selection gain using the MGIDI index. Increases are desired for the assessed multi-traits, with a total genetic gain of 21.65% and −1.04% for traits where a decrease was desired.

Based on the MGIDI index, four genotypes were also identified (TDr1741042, TDr1741073, TDr1738074, and TDr1739067) as high-performing genotypes for multiple traits assessed across 25 white yam genotypes ( Figure 6c ; Supplementary Table 5 ).

Combining the characteristics of sensory and instrument-based traits, the MGIDI identified four genotypes—TDr1741073, TDr1739058, Meccakusa, and TDr1739046—as best performing genotypes. Hence, the TDr1741073 genotype was the top performer across multiple traits among the 25 genotypes tested in six different environments ( Figure 6c ).