Background

Grain production is critical to food security and depends on many factors attributed to gene-environment interactions [19]. Some yield-related traits have higher heritability and better stability under different growth conditions [20] and have been used to improve maize breeding in many studies [21]. An important focus in maize yield studies is exploring the phenotypic plasticity of maize ears, with a particular focus on kernel number per ear—a key trait in high-yield maize breeding [22].

Efficient and accurate characterization of ear phenotypes in different environments is fundamental to plasticity research. Although modern harvesting equipment can automatically and efficiently measure traits such as ear number, bulk density, and kernel water content in the field, the characterization of other traits such as kernel number per ear, which are significantly regulated by genetics, must be done manually [22,23,24,25]. Manual measurement is inefficient, subjective, prone to error, and able to generate only limited indices. Although some previous studies have tried to explore the effectiveness of automated measurement methods, they are limited by efficiency or lack of several key ear phenotypes. One such method treats images on a homogenous background with traditional image processing technology, capturing each kernel as an object and obtaining the length and width of the kernel [26]. It can be difficult to use this method to obtain maize kernel objects on an ear, however, due the nonhomologous background (i.e., other kernels on the ear). Miller et al. [27] have solved this problem; however, they ignored key ear traits such as the number of kernels per ear and the number of rows per ear. Warman et al. [28] proposed novel algorithms to focus on these ear traits, but the approach is limited by the simplicity of the imaging platform (a non-product concept device) and low analysis efficiency (one ear per analysis). Therefore, there is a need to develop a high-throughput automated phenoty** platform that integrates the ear phenotypes concerned by Miller et al. [27] and Warman et al. [28], as well as many other critical phenotypic characteristics.

To this end, we developed a maize ear phenoty** platform called MAIZTRO that combines all necessary hardware and software and allows efficient, accurate, and high-throughput field measurement of multiple ear traits. Using MAIZTRO, we evaluated the phenotypic plasticity of ear phenotypes in 3819 transgenic maize inbred lines targeting 717 genes and identified adaptive genes in plants grown in a variety of environments. Our aim was to use MAIZTRO to investigate the plasticity of ear traits using a transgenic maize inbred population and identify potential regulatory genes that perform well under different field conditions.

Results

High-throughput platform for phenoty** maize ear traits

To address the challenges of measuring maize ear traits in the field, we developed a phenoty** platform called MAIZTRO that combines all necessary software and hardware (Fig. 1). The software integrates image-processing algorithms and two trained models, including a semantic segmentation model based on the full convolution network (FCN) algorithm for dividing different kernel regions and a kernel counting model based on the random forest (RF) algorithm for counting different kernel types. Two manually labeled maize ear image datasets were used to train the FCN and RF models. First, a maize ear image dataset was generated comprising 30,000 scanned images (representing 30,000 ears) that covered inbred lines or hybrids of tropical, subtropical, and temperate maize varieties. Next, two copies of the datasets were created. In one dataset, ears were divided using boundary boxes to define areas with different kernel types, such as bare areas without kernels, diseased kernel areas, and normal kernel areas. This dataset was used to generate paired images for FCN training. In the other dataset, different colored points were used to annotate kernel types. This dataset was used to generate paired images for RF training. Cross-validation ensures the model’s generalization capabilities by allocating training and validation data at a ratio of 8:2. Finally, we used the parameters of the trained models as the basis for analysis via MAIZTRO software (Fig. 1C), which was used to operate the camera (Fig. 1F) to obtain images of up to 18 ears at a time. Traditional contour detection was used to segment a single maize ear and obtain linear measurements such as length, width, and perimeter. The trained models were then incorporated into the analysis to determine kernel type area and kernel number (Fig. 1C). The results were saved to the database, from which data could be downloaded as needed (Fig. 1D). In order to obtain complete information, photos were taken of both the front and back of each ear.

Fig. 1
figure 1

Overview of MAIZTRO software and hardware. A Model training datasets. B Algorithm training. C Software graphical user interface (GUI). D Database management. E Hardware exterior. F Hardware interior. G Information concerning the field plots is input by scanning a barcode. H Up to 18 ears can be analyzed at a time. I Operation requires two people: one to operate the machine and one to sort the ears. Full convolution network (FCN) and random forest (RF) models were trained using different datasets. Scanned ear images were analyzed based on the trained model, and results were stored in the database

Full size image

MAIZTRO uses integrated software and hardware and is highly automated with modular integration (Additional file 1: Figs. S13). The device is mobile and equipped with a display, camera, light source, PC, ear measuring disk (Fig. 1F), and handheld scanner (Fig. 1G). The scanner is used to read a barcode corresponding to the sample plot information for each ear. MAIZTRO can scan 18 ears at once and determine various kernel regions. It is designed for simple operation and easy training. With two users operating the instrument, 18 ears can be scanned approximately 700 times a day, allowing complete phenotypic characterization of approximately 12,600 ears per day (Fig. 1H, I). To evaluate the accuracy of measurements obtained by the platform, we used a number of static characteristics described by Kalantar-Zadeh [29] (see the “Methods” section for details). Evaluation of these static characteristics was based on a new data set and considered the impact of different operators, equipment, and durations on the performance prediction (Table 1). The average accuracy for number of rows per ear, number of kernels per row, ear length, ear width, and normal kernel number was greater than 95% for most static parameters (range 90.7% to 99.9%). The average accuracy for bald tip length measurements was lower, but still greater than 74% (range 74.9% to 99.9%). The number of abnormal kernels was not assessed.

Table 1 Static characteristics used to assess the accuracy of MAIZTRO
Full size table

Transgenic maize inbred population

An efficient transgene-development platform with high genetic transformation efficiency was established in the past 10 years by the Center for Crop Functional Genomics and Molecular Breeding of China Agricultural University to create maize inbred lines with specifically targeted genes using transgene technology [30, 31]. Using this platform, 3819 transgenic lines targeting 717 genes were created over a period of 2 years. These lines represent specific gene overexpression or gene knockout generated by CRISPR/Cas9 technology in the same genetic background, namely the wild-type inbred line B73-329 [32,33,34]. The targeted genes were selected from a set of functional maize genes defined by scientists from China Agricultural University or from cDNA libraries.

In 2018 and 2019, the 3819 lines and line B73-329 were planted at several sites in the main maize production regions of China. In 2018, transgenic lines targeting 453 genes (413 overexpressed genes, 40 knockouts) were planted at Gongzhuling, Anyang, and Zhuozhou (Additional file 2: Table S2). In 2019, transgenic lines targeting 354 genes (294 overexpressed genes, 60 knockouts) were planted at Gongzhuling, Anyang, and Shangzhuang (Additional file 2: Table S3). Of the lines planted in 2018, plants harboring 90 individual transgenes were replanted and verified in 2019, for a total of 717 independent genes analyzed (Fig. 2A). Using Gene Ontology enrichment analysis to annotate and classify the functions of these genes, we identified 138 biological processes for 510 genes covering a wide range of physiological and biochemical processes (Additional file 2: Tables S4 and S5). The other 207 genes lacked Gene Ontology annotation.

Fig. 2
figure 2

Characteristics of 717 genes involved in maize ear phenotype and phenotypic plasticity. A Numbers of genes studied in 2018 and 2019. B Correlation matrix plot of ear phenotypes of transgenic inbred lines planted in 2018 (data for 2019 are in Additional file 1: Fig. S5). C Distribution of genetic values and linear plasticity of phenotypes in 2018 (data for 2019 are in Additional file 1: Fig. S8). Dotted line indicates the wild-type phenotype. D Quartile coefficients of dispersion for linear and nonlinear plasticity of phenotypes in 2018 and 2019

Full size image

Variation and correlation among ear phenotypes

The trait values for 15 important ear phenotypes were measured for transgenic and wild-type lines using MAIZTRO. Most phenotypes conformed to a normal distribution, whereas negative-effect indices had obviously skewed distributions (Additional file 1: Figs. S1 and S2). The trait values for most ears were close to 0 for negative-effect indices, including bald tip length, proportion of bald tip area, and proportion of empty area (Additional file 1: Figs. S1 and S2).

Variations in ear phenotypes among lines were calculated across different environments using the coefficient of variation (CV), and CV distributions were plotted (Additional file 1: Figs. S3 and S4). For most phenotypes, CV values for wild-type were in the middle of the distribution (dotted lines in Additional file 1: Figs. S3 and S4), and those for transgenic lines were distributed around the values for wild-type (red shading). The CV values for kernel-related characteristics such as bald tip length, proportion of bald tip area, and proportion of empty area varied more than the other traits in both years. Ear length also varied substantially.

We next analyzed correlations among the various phenotypic characteristics (Fig. 2B; Additional file 1: Figs. S5) and found that relationships were similar in 2018 and 2019, although they were stronger in 2019. There was a strong positive correlation between normal kernel number and ear length, ear width, and number of kernels per row. The relationships between ear shape and other phenotypes were weak.

Variation and correlation of phenotypic plasticity in ear traits

The trait values for 15 ear phenotypes for 717 genes were measured in multiple environments and analyzed using a Bayesian Finlay-Wilkinson regression (FWR) model for each phenotype [35]. This model can estimate the main effect and slope for each maize line (transgenic or wild-type), from which the residual of each observation can be calculated before estimating residual variance (i.e., nonlinear plasticity). The slope from the FWR model reflects the linear response of a genotype to the environment in the tested population (i.e., linear plasticity). A slope equal to one indicates that the transgenic line had an overall average response to the environment, whereas a slope equal to zero indicates that the line did not respond to the environment. The residual variance of each line was used as a measure of model fitting; a large residual variance indicates a lack of genetic basis for the environmental response and thus a poor linear model fit [35]. In our study, the correlations among the mean phenotype value (also called genetic value, g), linear plasticity (b), and nonlinear plasticity were analyzed (Additional file 1: Figs. S6 and S7). The g and nonlinear plasticity values correlated positively with bald tip length, proportion of bald tip area, and proportion of empty area, but correlated negatively with proportion of normal kernel area. For most phenotypes, the correlation between g and b values was stronger than those between g and nonlinear plasticity.

The dispersion of all phenotypes was evaluated using the quartile coefficient of dispersion (Fig. 2D). Most phenotypes had variable b values (range of quartile coefficient of dispersion, 0.1–0.3, Fig. 2D), indicating diverse plasticity mechanisms among the different genotypes. Dispersion values were similar in 2018 and 2019. Dispersion of b values was lower for bald tip length, ear length, and proportion of bald tip, but higher for ear diameter and kernel width (Fig. 2D). Nonlinear plasticity had a higher degree of dispersion than b values in 2018 but a similar degree of dispersion in 2019 (Fig. 2D). The g and b values (Fig. 2C, Additional file 1: Figs. S8) were nearly normally distributed in both years, with wild-type almost always appearing in the middle of the distribution (Fig. 2C, Additional file 1: Figs. S8).

Screening of target genes

The phenotypic plasticity of plants using the FWR method is represented by the g and b values [7, 28, 39]. The platform is simple to operate without complex knowledge after simple training and does not require complex configuration as with the system of Miller et al. [27]. Compared to Warman et al. [28], which can measurement only one ear at a time and focuses on kernel phenotype, MAIZTRO offers high-efficiency and comprehensive measurement of 15 ear traits from 18 ears at a time. Therefore, using this platform, we can measure all ears in a plot to eliminate sampling errors. MAIZTRO is especially suitable for use in the field because the closed setting prevents interference from environmental factors, which are more difficult to eliminate using the simple devices of Makanza et al. [39] and Warman et al. [28]. Therefore, MAIZTRO is suitable for multi environment assessment of ear phenotype of a large number of materials in the field.

Some plant phenotypes are more plastic than others [7, 39]. In our study, most phenotypic values and their variation showed a normal distribution. Wild-type values were always in the middle of the distribution, indicating that the targeted genes affected phenotypic variation in different ways and to varying degrees. In our analysis using the FWR model, the b value variation (linear plasticity, quartile coefficient of dispersion) was mostly in the range of 0.1 to 0.3 and was similar in 2018 and 2019; nonlinear plasticity showed a similar trend, although it differed between the 2 years (Fig. 2D). These results suggest that the impact of phenotypic plasticity on maize ears was essentially uniform across the transgenes analyzed. The distributions of g and b values were similar to those of phenotypic value and variation (a normal distribution with wild-type in the middle), indicating that the genes studied also have different effects on phenotypic plasticity. Our data support the idea that phenotypic plasticity is genetically regulated [7, 10,

Methods

Sample material and planting

The maize lines used in this study included 3819 transgenic inbred lines targeting 717 genes and one wild-type inbred line (B73-329) as a reference (Additional file 2: Table S2 and S3). Each transgenic line facilitated single-gene overexpression or knockout relative to the wild-type line. The recipient for all transgenic lines was B73-329, and targeted genes were overexpressed or knocked out using CRISPR/Cas9 technology [32, 34].

Transgenic kernels were planted in different environments in two different years, including 2217 transgenic lines (453 genes) in 2018 and 1602 transgenic lines (354 genes) in 2019. Lines planted in 2018 included 40 gene knockouts and 413 overexpressed genes (Additional file 2: Table S2). Lines planted in 2019 included 57 gene knockouts and 297 overexpressed genes (Additional file 2: Table S3). The planting sites in 2018 were Gongzhuling (N 43.5°, E 124.8°) in Jilin Province; Anyang (N 36.0°, E 114.1°) in Henan Province; and Zhuozhou (N 39.5°, E 115.8°) near Bei**g. The planting sites in 2019 were Gongzhuling, Anyang, and Shangzhuang (N 40.1°, E 116.2°) in Bei**g. In each planting year, the same lines were planted at each of the three sites. Kernels were planted according to the same block design at each site, and microenvironments were divided by block (Additional file 2: Table S2 and S3). Additional file 2: Table S1 summarizes the meteorological characteristics of the sites.

Analysis of ear phenotypes

In each year, all three sites were harvested simultaneously when 95% of ears reached full maturity. MAIZTRO was then used to analyze the following 15 ear phenotypes for all ears (Fig. 1): bald tip length, ear circumference, ear diameter, ear length, ear shape, ear width, kernel thickness, kernel width, normal kernel number, shrunken kernel number, number of kernels per row, number of rows per ear, proportion of bald tip area, proportion of empty area, and proportion of normal seed area. Ear shape was calculated as the ratio of ear width at one-third the length of the ear from the tip (first third) to ear width at one third the length of the ear from the base (last third). For the last three traits, proportion was calculated as the ratio of total area ranging from 0 to 1. For all traits, we filtered out outliers exceeding 1.5 times the interquartile range in the population prior to analysis.

Model training and verification of MAIZTRO

MAIZTRO calculates parameters primarily related to ear contour and kernel type using the FCN and RF algorithms, which require a large amount of data for training to obtain ideal performance (Fig. 1). We collected images of 30,000 ears from inbred lines and hybrids of tropical, subtropical, and temperate maize varieties covering a rich ear diversity. Based on these images, we manually created two labeled image sets (Fig. 1). The first set consisted of pairs of RGB (red, green, blue) images and manually labeled segmentation images and was used to train the FCN model. The second set consisted of a series of paired RGB images and manually labeled images in which points of different colors represented different kernel types (for example, blue for normal kernels) and was used to train the kernel-counting RF model.

Six indicators were used to evaluate the measurement performance of the platform for major phenotypes including number of rows per ear, number of kernels per row, ear length, ear width, bald tip length, and normal kernel number [29].

Accuracy was defined as the agreement between a single measured value and the real value. A sample of 200 ears was measured using MAIZTRO to give the measured value (MV) and manually to give the real value (RV). The accuracy for each ear was calculated as:

$${\text{Accuracy}} = 100 - \left( {\frac{{{\text{MV}} - {\text{RV}}}}{{{\text{RV}}}} \times 100} \right)$$

The accuracy for the platform was then calculated as the average accuracy from 200 ears.

Trueness was defined as the agreement between an average measured value from multiple measurements and the real value. The same 200 ears were measured one time manually and three times using MAIZTRO. The same person measured each ear all three times using MAIZTRO to ensure consistency of the process (i.e., measuring the ear, removing it, and measuring it again). Trueness was calculated as the accuracy of the average of the three measurements.

Precision was defined as the repeatability of multiple measurements under the same user-platform conditions without regard to real value. The same 200 ears were measured three times by the same person using MAIZTRO with the same process. The CV was calculated for each ear and then the average CV of all 200 ears was calculated.

Reproducibility was defined as the repeatability of multiple measurements under different user-platform conditions. For this, 100 ears were measured by five different people operating five different platforms. Each ear was measured five times total. The CV of each ear was calculated and then the average CV of all 100 ears was calculated.

Technical repeatability was defined as the ability of a single instrument to produce the same results across several measurements. Using the same platform, 18 ears were measured 1000 times by the same person with fixed placement (i.e., without being removed). The CV was calculated for each ear and then the average CV was calculated.

Robustness was defined as the ability of a sensing system to produce the same output value when measuring the same object over a period of time. The same 18 ears were measured once a day for 10 days. The CV was calculated for measurements made on different days for each ear, and then the average CV was calculated.

Stability analysis

Phenotype plasticity of each transgenic inbred line and the wild-type line was estimated across environments by applying the FWR model with the FW package in R software [35]. For this model, the individual phenotype observed in an environment could be expressed as:

$${y_{ij}} = {\upmu } + {g_i} + \left( {1 + {b_i}} \right){h_j} + {{\upvarepsilon }_{{\text{ij}}}}$$

where yij is the phenotype of line i (either transgenic or wild-type) collected in the jth environment, μ is the population mean, gi is the main effect of line i and reflects the estimated value of the mean phenotype, (1 + bi) is the linear plasticity of line i over the environments, (1 + bi) hj indicates the phenotypic change of line i in a given environment, and ɛij is an error term; variance was recorded as a measure of the nonlinearity of the response to the environment [7, 11].