Image projection and single-cell segmentation
The analysis was conducted on a total of 240 images from the wild type and mutants, with 18 z-stacks on average. These images were merged with max contrast projection, resulting in 41 images in the wild type, 42 images in the nai1-1 mutants, 40 images in the leb-1 bglu21-1 mutants, 40 images in the meb1-1 mutants, 38 images in the meb2-1 mutants, and 39 images in the meb1-1 meb2-1 mutants (Table 2). Subsequently, cell segmentation based on the red fluorescence of cell walls provided 12408 cell images in the wild type, 17205 cell images in the nai1-1 mutants, 9109 cell images in the leb-1 bglu21-1 mutants, 10664 cell images in the meb1-1 mutants, 6862 cell images in the meb2-1 mutants and 10357 cell images in the meb1-1 meb2-1 mutants (Table 2). Further, segregation of the 66605 cells from the 240 images resulted in 29629 cells that had ER body like features (Table 3).
Table 2
Summary of cell segmentation
Genotype
|
Setting
|
Objective
|
Days after Germination
|
Segmented Cells
|
Images
|
wild type (GFP-h)
|
1
|
20×
|
7
|
6181
|
19
|
2
|
5
|
830
|
5
|
7
|
1712
|
6
|
25×
|
5
|
1401
|
5
|
7
|
2284
|
6
|
nai1-1
|
1
|
20×
|
7
|
8901
|
20
|
2
|
5
|
1215
|
5
|
7
|
2712
|
6
|
25×
|
5
|
1514
|
5
|
7
|
2863
|
6
|
leb-1 bglu21-1
|
1
|
20×
|
7
|
5188
|
20
|
2
|
5
|
630
|
5
|
7
|
804
|
5
|
25×
|
5
|
1078
|
5
|
7
|
1409
|
5
|
meb1-1
|
1
|
20×
|
7
|
6723
|
20
|
2
|
5
|
292
|
5
|
7
|
1208
|
5
|
25×
|
5
|
695
|
5
|
7
|
1746
|
5
|
meb2-1
|
1
|
20×
|
7
|
2552
|
18
|
2
|
5
|
613
|
5
|
7
|
1428
|
5
|
25×
|
5
|
829
|
5
|
7
|
1440
|
5
|
meb1-1 meb2-1
|
1
|
20×
|
7
|
4893
|
19
|
2
|
5
|
1021
|
6
|
7
|
1797
|
5
|
25×
|
5
|
1117
|
5
|
7
|
1529
|
4
|
Table 3
Summary of ER body like features
Genotype
|
Objective
|
Days after germination
|
Cells with ER-body like features
|
No. of Images
|
wild type (GFPh)
|
20×
|
5
|
271
|
3
|
7
|
5949
|
21
|
25×
|
5
|
813
|
4
|
7
|
811
|
6
|
leb-1 bglu21-1
|
20×
|
5
|
251
|
3
|
7
|
4920
|
23
|
25×
|
5
|
504
|
5
|
7
|
732
|
4
|
meb1-1
|
20×
|
5
|
52
|
2
|
7
|
5234
|
22
|
25×
|
5
|
275
|
5
|
7
|
767
|
5
|
meb2-1
|
20×
|
5
|
71
|
2
|
7
|
1989
|
17
|
25×
|
5
|
529
|
4
|
7
|
494
|
5
|
meb1-1 meb2-1
|
20×
|
5
|
518
|
4
|
7
|
4509
|
19
|
25×
|
5
|
406
|
5
|
7
|
531
|
3
|
Image-wise and segmented cell-wise analysis
The z-scores of 40 features (6 spatial, 8 intensity and 26 Haralick features) were calculated from the merged micrograph images from the wild type and mutants based on GFP fluorescence of ER and ER bodies (Additional file 3A). Based on these features, we calculated the Pearson correlation coefficient (PCC) and conducted MDS analysis (Fig. 2 and Additional file 3B to H).
In the heatmap of the feature matrix, significant differences were found in the patterns between the plants that had ER bodies (e.g. wild type) and the plants that did not have ER bodies (nai1-1) in a specific experiment, which is the dataset with 7-day-old plants with PI staining setting 1 and 20× objective lens (Additional file 3A), indicating that the images can be separated into two groups depending on the presence or absence of ER bodies in the dataset. The MDS analysis showed that micrograph images can be separated on the scatter plot according to the morphology of ER bodies. The MDS1 and MDS2 axes explain 81.01 % and 24.15 % in the image-wise analysis, respectively (Fig. 2A). The separation between the images from the wild type and the nai1-1 mutant occurred along the MDS1 axis, showing that the axis indicates the presence or absence of ER bodies (Fig. 2A). The separation between the wild type plants and leb-1 bglu21-1 double mutants occurred along the MDS2 axis, suggesting that the axis explains the length of ER bodies since the leb-1 bglu21-1 double mutants have longer ER bodies compare to the wild type. We found that meb1-1 and meb1-1 meb2-1 mutants had a higher MDS2, suggesting that these mutants have shorter ER bodies (Fig. 2A). The micrograph images are even further separated in the scatter plot and explain 24.15% in the MDS2 axis and 19.79% in the MDS3 axis (Additional file 3B). We further conducted MDS analysis with two other image data sets (setting 2 with 20× objective, and setting 2 with 25× objective) from a different batch of experiments (Fig. 2B). In this data the variation in the image captured the difference of the objective lens and the age of the cotyledons in the MDS1 axis (77.41%). Further, we found the image variation with the presence or absence of ER bodies in the MDS2 axis (35.9%).
A similar trend was observed when the MDS analysis was done on segmented cell images of a dataset with 7-day-old plants, PI staining setting 1 and 20× objective lens. The cells that had ER bodies were clustered separately from the cells devoid of ER bodies in the MDS plots and showed that the maximum variation of MDS1, MDS2 and MDS3 were 70.93%, 28.74% and 21.36%, respectively (Fig. 2A and Additional file 3C). This suggests that the morphological parameters for the ER bodies are specific and discrete from that of the ER network. When we conducted the MDS analysis of segmented cells with the other image data sets (setting 2 with 20× objective, and setting 2 with 25× objective) for the respective groups of objective lens and age of the seedling, the separation between the cells with and without ER bodies were moderate. In the images with 20× objective, the variations explained were 53.69% and 58.59% in MDS1, 36.6% and 33.11% in MDS2, and 14.05% and 13.68% in MDS3 (Fig. 2C, Additional file 3E and H). The variations explained within the images from the 25× objective were 57.34% and 53.53% in MDS1, 32.5% and 34.91% in MDS2, and 15.17% and 16.81% in MDS3 (Fig. 2C, Additional file 3F and G). Therefore, although the estimation is robust according to image taking methodology, the same experimental setting is desirable to predict the MDS analysis precisely.
The feature data of cell-wise images were subjected to k-means clustering (k = 60) within each group (Fig. 3) to determine the group of cells having distinct ER body phenotypes. The k-means clusters segregated the cells that were devoid of ER bodies and were similar to the cells of the nai1-1 mutant, and the remaining cells can be attributed to their genotype in the features. Clusters that showed ER body like features were considered for further analysis to evaluate the overall effect of the genotype in explaining the morphological diversity of ER bodies. The features from the clusters of segmented cells across different experimental settings were integrated. Further, we investigated individual images of the clusters including ER body images (Fig. 3). The images of the clusters showed similar ER body morphology within each cluster, but apparent variations between the clusters. This indicates that the k-means cluster analysis grouped the cells having similar phenotypic variants throughout the genotypes. In clusters 10, 12, 19, 31 and 54, we observed cells mostly belonging to plants without leb-1 bglu21-1. In clusters 2, 18, 28, and 44, we observed cells mostly belonging to mutants. Clusters 2, 7, 50, 51 and 54 revealed morphologically distinct ER bodies. Cluster 16 was identified as an autofluorescence like feature, presumably noise images. With this approach we excluded the cell images from the nai1-1 mutant and from stomata cells with no ER bodies as well as auto-fluorescence. Consequently, 29629 cells were classified from 66605 cell-wise images as having ER bodies after k-means clustering analysis. At this resolution the differences in the ER body morphology across the mutants and the wild type could be compared. Further, anomalies in the texture features within ER bodies were detected from the clustered cells.
Feature analysis of cells having ER bodies
We re-examined the morphological variations of ER bodies with pooled cell-wise images that show ER bodies. However, we found a technical variation between 20× and 25× objective lens on integrating the z-score values of intensity, Haralick and spatial features (Fig. 4A). Therefore, we performed constrained ordination on the ER body phenotype within these cells by using PCC distances to find the variances among the features (Fig. 4B). We set genotype as a fixed factor and the others (objective lens, plant age, staining method) as random factors in the analysis. The variation explained in CCA1 and CCA2 was 67.49% and 30.7%, respectively. After conditioning the random factors we observed a significant difference in the feature diversity depend on the genotypes. The significance was determined by conducting a permutational multivariate analysis of variance (PERMANOVA) test on constrained ordination for Pearson correlations within each cell and feature (p-value < 0.05, 1000 iterations) (Additional file 4). Despite the moderate morphological variation, constrained ordination revealed that genotype difference could be explained at 1.37%. We found that images could be grouped depending on the ER morphology, such as long, rounded and aggregated ER bodies (Fig. 4B). Thus, we grouped the images according to the k-mean clustering (Fig. 3) within genotypes and performed constrained ordination on the mean of features from the grouped images. In the pooled dataset, we found that the features of the mutants could be distinguished from each other according to the difference in their ER body morphology (Fig. 4C). The variation explained by genotype was 7.56% (p-value < 0.01) and the variation shown in CCA1 and CCA2 was 60.14% and 31.77%, respectively (Fig. 4C). The scatter plot shows the wild type and the meb1-1 mutant placed in the centre, while the leb-1 bglu21-1 with long ER bodies shifted to the upper-left, and the meb2-1 and meb1-1 meb2-1 mutants with rounded and aggregated ER bodies shifted to the right (Fig. 4C and Additional file 3I).
The FDA and MDA provided the proportion of cell images predicted to be of a certain genotype. In MDA analyses we observed that a small proportion (< 30%) of cells from mutants were predicted to be from the wild type, suggesting that the mutant plants have cells that show similar ER body features to those of the wild type (Fig. 4D). However, substantial levels (> 60%) of mutant cell features were predicted to be in their respective genotypes, suggesting that each mutant had ER bodies with specific morphological features (Fig. 4D). Also, a proportion of cells from the mutants were still predicted as belonging to different genotypes, indicating the tendency of similarity in features across the genotypes (< 10%). The FDA analysis showed the discrete proportion of ER bodies were predicted to be discrete with 100% identity to their respective genotype. This suggests that the variation within the genotype is non-linear. The variation in the features estimated within the genotype is represented in the box plots (Additional file 5), which shows that the mutants had a distinct ER body morphology from the wild type.
Dynamics of the cellular feature
We used time-lapse image analysis to examine the difference in ER body movement between the wild type and the mutants. We observed ER body movement across time in both wild-type and mutant plants, but noticed that there was a reduction in the ER body movement in the meb2-1 and meb1-1 meb2-1 mutants (Fig. 5A and movies in additional file 6 to 10). We calculated the average of the ER body displacement from their initial position and found that the overall ER body movement was highest in the wild type across time (Fig. 5B). A similar trend was observed in the leb-1 bglu21-1 and meb1-1 mutants, but not in the meb2-1 and meb1-1 meb2-1 mutants. Statistical analysis revealed that movement was reduced in the meb2-1 and meb1-1 meb2-1 mutants compared to the wild type, leb-1 bglu21-1 and meb1-1 mutants (FDR ≤ 0.01) (Fig. 5C). These findings suggest that the MEB2 protein is involved in ER body movement.