Dissecting the phenotypic components of crop plant growth and drought responses based on high-throughput image analysis

Abstract

Significantly improved crop varieties are urgently needed to feed the rapidly growing human population under changing climates. While genome sequence information and excellent genomic tools are in place for major crop species, the systematic quantification of phenotypic traits or components thereof in a high-throughput fashion remains an enormous challenge. In order to help bridge the genotype to phenotype gap, we developed a comprehensive framework for high-throughput phenotype data analysis in plants, which enables the extraction of an extensive list of phenotypic traits from nondestructive plant imaging over time. As a proof of concept, we investigated the phenotypic components of the drought responses of 18 different barley (Hordeum vulgare) cultivars during vegetative growth. We analyzed dynamic properties of trait expression over growth time based on 54 representative phenotypic features. The data are highly valuable to understand plant development and to further quantify growth and crop performance features. We tested various growth models to predict plant biomass accumulation and identified several relevant parameters that support biological interpretation of plant growth and stress tolerance. These image-based traits and model-derived parameters are promising for subsequent genetic mapping to uncover the genetic basis of complex agronomic traits. Taken together, we anticipate that the analytical framework and analysis results presented here will be useful to advance our views of phenotypic trait components underlying plant development and their responses to environmental cues.

Figure 1.

Experimental Design. (A) The growth stages of spring barley. (B) High-throughput phenotyping of barley plants in a LemnaTec system (http://www.lemnatec.com/). (C) Plants were monitored in a noninvasive way under control and drought stress conditions. Drought stress (in dash box) was treated at the stage of “stem extension” as indicated in (A).

Figure 2.

Pipeline for Analysis of High-Throughput Phenotyping Data in Barley. (A) The workflow used for barley phenotyping data analysis. High-throughput imaging data from the LemnaTec system were imported and processed using the barley analysis pipeline in the IAP system. The extracted phenotypic traits were further processed and evaluated (see Methods). (B) Input (left) and result (right) images in the analysis pipeline. Shown are images from 44-d-old plants (the last day of stress phase) captured by VIS, FLUO, and NIR cameras from the side view. (C) Classification of phenotypic traits. Traits are classified into four categories: color-related, NIR-related, FLUO-related, and geometric features, based on images obtained from three types of cameras and two views. (D) Phenotypic traits revealing the stress symptom. Left: An example shows a NIR-related trait over time. Right: heat map shows NIR intensity difference, measured by the ratio value between control and stress plants. Blue indicates low difference, whereas red indicates high difference. Note that plants from different genotypes show different patterns, indicating their different stress tolerance.

Figure 3.

Phenotypic Similarity Revealed by Genotype Similarity. (A) and (B) Clustering analysis of phenomic profiling data. HCA (A) and a six-by-six self-organizing map (SOM) (B) were used to reveal the phenotypic similarity of all the investigated barley plants based on the highly reproducible traits. In (A), colored bars along the top of the heat map reflect the sampled agronomic group assignment (groups 1 to 3 and DH) as labeled. Colored bars along the left indicated the corresponding genotypes of individuals as listed in the key. The lower panel shows the median correlation values among individual plants from the same agronomic groups and different groups. In (B), plants with similar genotypes or treatments tend to be at nearby map locations. Control and stress plants are colored and indicated in blank and filled points, respectively. The numbers in the key show the number of plants from the same genotypes belonging to the control or stress group. (C) Phenotypic similarity trees showing the phenotypic relationship of plants from agronomic groups 1 to 3 under control (left; blank shapes) and stress (right; filled shapes) conditions. The trees were constructed from overall phenotypic distance matrices (see Methods). (D) Scatterplot indicating the degree of correlation of phenotypic distance between genotypes under both control (x axis) and stress conditions (y axis). Mantel test was performed to examine whether the phenotypic distances in the two conditions correlate with each other. P value was calculated with Monte-Carlo simulation (with 10,000 permutations). Genotype pairs that are far away from the regressed line (red) are labeled and colored (orange, small distances in control and large distances in stress; blue, otherwise).

Figure 4.

Phenotypic Profile Reflects Global Population Structures in the Temporal Scale. (A) Projections of top six PCs based on PCA of phenotypic variance over time. The percentage of total explained variance is shown. The stress period is indicated by the dashed box. (B) Scatterplots showing the PCA results on DAS 44 (explained the largest variance). The first six PCs display 83.3% of the total phenotypic variance. The component scores (shown in points) are colored and shaped according to the agronomic groups (as legend listed in the box). The component loading vectors (represented in lines) of each variable (traits as colored according to their categories) were superimposed proportionally to their contribution. See also Supplemental Figures 6 and 7.

Figure 5.

Dissection of the Sources of Phenotypic Variance. (A) Dissecting the phenotypic variance over time by linear mixed models. For phenotypic data before stress treatment, is confounded with . Filled circles represent average variance of each component computed over all traits, and solid lines represent a smoothing spline fit to the supplied data. Error bars represent the se with 95% confidence intervals. The numbers of traits with significance at P < 0.001 are indicated above the bars. The stress period is indicated in dashed box. (B) The total experimental CV (colored in gray) and genetic CV across lines (green for control, orange for stressed, and blue for the whole set of plants) over time. Data points denote the average CV value over all geometric traits. Solid lines denote the loess smoothing curves and shadow represents the estimated se. (C) Statistical significance of genotype effect (left), treatment effect (middle), and their interaction effect (right), as detected by linear mixed models. The shading plot indicates the significance level (Bonferroni corrected P values) in terms of LOD scores (-log probability or log of the odds score). Traits are sorted according to their overall effect patterns. Trait identifiers are listed on the right, which are given according to Figure 6A. G, genotype; E, environment (treatment).

Figure 6.

Trait Heritability and Trait-Trait Genetic and Phenotypic Correlations. (A) Heat map showing broad-sense heritability (H²) of the investigated phenotypic traits over time (left), as exemplified by the digital volume (bottom right). Box plot (top right) shows the average heritability of phenotypic traits from the four categories (right). Error bars, se with 95% confidence intervals. (B) Network visualizing significant phenotypic (r_p; left) and genetic (r_g; right) correlations among the 54 image-derived traits and three manual measurements (brown nodes). For visualization purpose, only significant correlations are shown (P < 0.01 for r_g and r_p, and r_p > 0.5). Trait identifiers are given as in (A) and colored according to their classification as indicated. Positive correlations are shown by solid lines in red, and negative correlations are shown by dashed lines in blue. (C) Pearson’s correlation of r_g and r_p over time. The test of relationship between matrices of r_g and r_p was performed using Mantel’s test, as exemplifying on the right panel.

Figure 7.

Modeling of Plant Growth Based on Digital Biomass. (A) Plant growth prediction based on fitting of the digital volume using five different mechanistic models. The quality of fit (R²) of each model is given. The best-fitted model-logistic model can be considered as the growth curve of barley plants. Several logistic-model derived parameters such as the “inflection point” (IP; a time point with the maximum growth rate) and “maximum biomass” (the maximum growth capacity) are indicated. Dots represent data points derived from images and curves represent the least-squares fit to the observed data. Shown is the result of fitting for a Victoriana plant. See also Supplemental Data Set 2. (B) Pairwise comparison of model-derived parameters, image-derived data, and manually determined FW or DW for control plants. Each point in the dot plots (bottom-left quadrants) represents one plant from a specific genotype as colored and labeled at the bottom. Pearson’s correlation coefficients are indicated in top-right quadrants. (C) Curve fitting of digital volume in drought stress conditions. Plant growth before rewatering is modeled by one quadratic function and three different bell-shaped functions. Growth in recovery phase is modeled by a linear function. Three vertical lines from left to right: the first inflection point, the time of maximum biomass, and the second inflection point estimated from the best-fitted model (bell-shaped model 3). See also Supplemental Data Set 3. (D) Pairwise comparison of model-derived parameters, image-derived data, and manual measurements for stressed plants. (E) Comparison of plant growth between control and stress conditions. R_IP represents the growth rate (px³/day) at the inflection point of control plants. R_rec denotes the recovered growth rate (px³/day) in recovery phase of stress plants. ϵ_stress, referred to “stress elasticity” calculated as the ratio of R_rec and R_IP. Two drought tolerance indexes, yield stability index (YSI) (Bouslama and Schapaugh, 1984) and stress susceptibility index (SSI) (Fischer and Maurer, 1978), are provided for comparison.