Multiple brace root phenotypes promote anchorage and limit root lodging in maize

Plant mechanical failure (lodging) causes global yield losses of 7-66% in cereal crops. We have previously shown that the above-ground nodal roots (brace roots) in maize are critical for anchorage. However, it is unknown how brace root phenotypes vary across genotypes and the functional consequence of this variation. This study quantifies the contribution of brace roots to anchorage, brace root traits, plant height, and root lodging susceptibility in 52 maize inbred lines. We show that the contribution of brace roots to anchorage and root lodging susceptibility varies among genotypes and this contribution can be explained by plant architectural variation. Additionally, supervised machine learning models were developed and show that multiple plant architectural phenotypes can predict the contribution of brace roots to anchorage and root lodging susceptibility. Together these data define the plant architectures that are important in lodging resistance and show that the contribution of brace roots to anchorage is a good proxy for root lodging susceptibility.


Introduction 51
A changing global climate is increasing the severity and prevalence of storm systems, which 52 threatens crop production. Crop failure from mechanical stress is called lodging and is defined as 53 the displacement of plants from vertical (Rajkumara, 2008). In cereal crops, lodging is reported 54 to cause between 7-66% yield loss (Carter and Hudelson, 1988 underlying factors that contribute to lodging susceptibility. However, defining these factors has 61 been limited by the reliance on unpredictable weather patterns to induce natural lodging events. 62 To overcome this limitation, measures of plant anchorage have been developed as a proxy for 63 root lodging (Erndwein et al., 2020). Using both natural lodging events and measures of plant 64 anchorage, several factors that impact lodging have been defined, and include field management 65 practices, plant stage at lodging, plant genotype, and plant architecture (e.g. root system 66 architecture and plant height) (Carter and Hudelson, 1988;Berry et al., 2004;Rajkumara, 2008). 67 68 In maize, the size and extent of the root system has been linked to root lodging resistance and 69 anchorage (Thompson, 1972 important phenotypes for root lodging resistance. In addition to root phenotypes, plant height is a 81 known factor in lodging susceptibility for many crops (Rajkumara, 2008). Yet in maize the link 82 between plant height and root lodging is unclear. One study showed no correlation between plant 83 height and root lodging in a natural root lodging event (Sharma and Carena, 2016), whereas 84 another study showed plant height as a secondary factor in root lodging models, which were 85 validated by natural root lodging events (Brune et al., 2018). The wide-spread interpretation of 86 previous results has been limited by a small number of genotypes per study, disparate 87 phenotypes analyzed in each study, and a primary reliance on univariate correlations. Therefore, 88 a detailed assessment of phenotypes at the population level is required to define how plant 89 architectures contribute to root lodging resistance and anchorage. 90 91 In this study, the brace root contribution to anchorage, brace root phenotypes, and plant height 92 were assessed in a population of 52 maize inbred lines. Additionally, root lodging susceptibility 93 was quantified after a natural storm event within the same population. Platform with a custom controller and side mounted FLIR 3.2 MP Color Blackfly camera (Figure 157 S1). Plots were marked with an orange stake and individual plants were marked with a yellow 158 popsicle stick for reference. Only the number of brace root whorls that entered the ground was 159 quantified from these images. 160 161 For the 2019 data analysis, we developed a semi-automated root tagging graphical user interface 162 (GUI) to optimize image processing ( Figure S2). This new workflow enables annotation and 163 analysis on the scale of 1-2 minutes per image. The height of the tag adhered to each plant was 164 used as a scale (pixels per inch) and phenotypes were measured for only brace root whorls that 165 entered the ground. These phenotypes are measured sequentially as 1) number of brace roots per 166 whorl (where whorl 1 is closest to the ground), 2) single outermost brace root width for the 167 highest whorl in the ground, 3) stalk width, and 4) a right triangle on each side of the plant from 168 the outermost root from highest whorl in the ground. The right triangle is used to extract the 169 height of the whorl on the stalk (a leg of a right triangle), the distance from the stalk-to-root 170 grounding (c leg of a right triangle), and the root angle (B angle of the right triangle). The brace 171 root spread width was calculated as the sum of the stalk width and the b leg from the triangle on 172 both sides of the plant. The root tagging workflow extracts pixel information and a post 173 processing script converts pixels to inches or degrees. Inches were converted into centimeters for 174 data analysis. Left and right (A and B) phenotyping outputs were compared with a Pearson 175 correlation analysis in R ver. 4.0.2 (R Core Team, 2013) with the corrplot package ver. 0.84 176 (Wei and Simko, 2017). A high correlation (r range 0.55-0.79) between A and B images 177 indicates that the root tagging workflow has high precision ( Figure S3), therefore measurements 178 were summed (root number per whorl) or averaged (all other phenotypes) to provide a per-plant 179 brace root phenotypic profile and reduce the effect of phenotyping error. all data sets were retained. When the genotype data clustered together, a taxon-partition was used 199 to assign both sets of sequence data to a terminal branch. When genotype data did not cluster, 200 both sets of sequence data were retained as individual branches. The phylogenetic relationships 201 can be visualized in Figure S4. A multiple regression model was built using architectural phenotypes to predict the None/All 241 ratio and root lodging susceptibility. Four models were tested: 1) the outcome (None/All ratio) 242 and predictors (architectural phenotypes) were assessed within the same year at the individual 243 plant resolution, 2) the outcome (None/All ratio) and predictors (architectural phenotypes) were 244 quantified within the same year at genotypic average resolution, 3) the outcome (None/All ratio) 245 was a multi-year genotypic average and predictors (architectural phenotypes) were individual 246 plant data, and 4) the outcome (lodging susceptibility) was genotypic average and predictors 247 (architectural phenotypes) were individual plant data. 248 249 Additionally, genotypes were grouped for categorical classification. For each of the outcomes 250 (the None/All ratio and root lodging susceptibility), data were first scaled and centered before 251 classified into one of three groups -low, average, or high. For each outcome, genotypes were 252 classified as low when they were more than 1 standard deviation below the mean, genotypes 253 were classified as high when they were more than 1 standard deviation above the mean, and 254 genotypes were classified as average when they were within +/-1 standard deviation of the 255 mean. Four models were tested: 1) the outcome (None/All ratio) and predictors (architectural 256 phenotypes) were assessed within the same year at the individual plant resolution, 2) the 257 outcome (None/All ratio) was a multi-year genotypic average and predictors (architectural 258 phenotypes) were individual plant data, 3) the outcome (lodging susceptibility) was genotypic 259 average and predictors (architectural phenotypes) were individual plant data, and 4) the outcome 260 (lodging susceptibility) was genotypic average and predictors (architectural phenotypes and 261 None/All ratio) were individual plant data. For each categorical classification model, a random 262 forest approach was used with 100 trees.  (Table S1). The None/All 278 ratio quantifies the reduction in anchorage upon removal of brace roots. Thus, a high None/All 279 ratio (close to 1) shows that brace roots have limited impact on anchorage, whereas a low 280 None/All ratio (close to 0) shows that anchorage is dependent on the brace roots in the ground. 281 The None/All ratio data was analyzed by an ANOVA and the effect of genotype (p<0.001), but 282 not year (p>0.05), was significant ( Figure 1, Table S3-S4). Thus, confirming our hypothesis that 283 the contribution of brace roots to anchorage varies among genotypes. 284 285 Figure 1. The brace root contribution to anchorage varies depending on genotype. A two-way analysis of variance (ANOVA) showed that the None/All ratio was impacted by genotype (p<0.05) but not year (p>0.05). A higher None/All ratio indicates that anchorage was not influenced by brace roots, whereas a lower None/All ratio indicates that anchorage is dependent on brace roots. Genotypes on the left have an overall higher None/All ratio, whereas genotypes on the right have a lower None/All ratio. The color of dot illustrates the year the data was collected. Outliers are outlined in black.

286
In support of the hypothesis that the contribution of brace roots to anchorage is genetically 287 controlled, the broad sense heritability (H 2 ) estimate was 0.30 (Table S5). However, when 288 considering population structure, the None/All ratio was identified as a homoplasious trait 289 ( Figure S4). For example, Hp301 ranked high (51) and SA24 ranked low (2) for the comparative 290 average None/All ratio across years (Figure 1), yet both genotypes are popcorn and cluster 291 phylogenetically ( Figure S4). Collectively, these findings establish that the contribution of brace 292 roots to anchorage is dependent on genotype but is not inherited from a common ancestor. 293 294 Plant phenotypes vary among genotypes 295 We hypothesized that the variation in the brace root contribution to anchorage (None/All ratio) 296 among genotypes is due to underlying variation in brace root phenotypes. captures RGB (Red Green Blue) images using a ground-based robot and extracts brace root 303 phenotypes through a semi-automated root tagging pipeline ( Figure S2). In addition to the 304 0.0 0.5

Ratio=None/All
Year 2018 2019 Low contribution

High contribution
All None Ratio = None/All phenotypes extracted with this pipeline, brace root whorl number in the ground, and plant height 305 were manually quantified for the 52 genotypes to provide an assessment of plant architecture for 306 one year (Figure 2A; Table S6). 307 308 All plant architectures that were measured varied significantly by genotype (Figures S5, Tables  309  S7-S8). A Principal Component Analysis (PCA) showed a dense, continuous distribution of 310 plants suggesting that a single phenotype was not driving distinct groups or clusters of plants 311 (Figure 2B). The first dimension of the PCA explains 30.40% of the variation and was primarily 312 driven by spread width and the stalk-to-root grounding (Tables S9-S10). The next 6 dimensions 313 explain an additional 64.10% of the variation (>90% total variation explained) and were each 314 driven by a single phenotype (Tables S9-S10). As expected, phenotypes that were highly 315 correlated ( Figure S6) also had eigenvectors that loaded in the same plane in the PCA ( Figure  316 2B). For example, the spread width and the stalk-to-root grounding eigenvectors load in the same 317 plane and direction ( Figure 2B) and were highly positively correlated (r = 0.99; Figure S6). 318 There is also a negative correlation between the root angle and root height on stalk (r = -0.61; 319 Figure S6) and eigenvectors load in the same plane but opposite directions ( Figure 2B). 320 Interestingly, this relationship shows that brace root whorls that attach higher on the stalk have a 321 more acute angle. Consistent with the genetic regulation of these phenotypes, H 2 for the majority 322 of phenotypes was greater than 0.25 with the highest H 2 at 0.78 for plant height (Table S5). 323 Thus, these data demonstrate that there is underlying variation in individual plant architectures 324 that may contribute to variation in the brace root contribution to anchorage.

23.327
Within-year Multi-year

Figure 2. Plant architecture varies among genotype and multiple architectures are important in the contribution of brace roots to anchorage. (A)
Brace root phenotypes were quantified with a semiautomated root tagging workflow and phenotypes were only extracted from the brace root whorls that entered the ground. Phenotypes 1, 4, 5, 6, and 7 were extracted for the top-most whorl that entered the ground. The number of roots on each whorl was quantified for all brace root whorls that entered the ground with W1 being the bottom-most whorl (closest to the ground) and W3 being the top-most whorl. The red triangle highlights the right triangle that is used to identify the height of the whorl (a leg of the right triangle), the stalk-to-root grounding (c leg of the right triangle), and the root angle (B angle of the right triangle). Brace root phenotypes and plant height were used as predictors in random forest models. highest correlation was between the number of brace root whorls that entered the ground and the 336 None/All ratio for within-year data (r = -0.41; Table 1). Interestingly, when phenotype data was 337 compared to multi-year None/All ratio data, the highest correlation was with the stalk-to-root 338 grounding (r = -0.36; Table 1). Overall, the correlations between plant architectures and the 339 None/All ratio were higher for within-year comparisons than for between year comparisons. 340 However, even for within-year comparisons, the correlations are moderate suggesting that 341 individual phenotypes cannot explain the contribution of brace roots in anchorage. 342 343 To further explore the relationship between plant architectures and the contribution of brace 344 roots to anchorage, machine learning models were developed. First, multiple regression models 345 were used to predict the None/All ratio from the plant architectures. For within-year models, 346 both individual plant and genotypic average data predicted the None/All ratio with an R 2 value of 347 0.23 (Table 2). However, when a multi-year genotypic average was used for the None/All ratio, 348 the R 2 value decreases to 0.19 (Table 2). Thus, within-year predictions are stronger than between 349 year predictions. 350 351 Next, a categorical classification approach was used to determine if plant architectures can 352 predict the relative classification (low, average, high) of the None/All ratio. Random forest 353 models were generated using the individual plant architectures as predictor variables and the 354 classification of within-year individual plant data or multi-year genotype average data as the 355 outcome. From these models, the None/All ratio category (low, average, or high) was predicted 356 with 66.00% accuracy (ROC AUC=0.76) for within-year data and 81.00% accuracy (ROC 357 AUC=0.87) for multi-year genotype average data ( Table 2). The increase in prediction accuracy 358 when considering genotypic averages is likely due to a reduction in noise that is inherent to each 359 measurement. These data further demonstrate that multi-year genotypic averages perform well 360 for the relative classification of the brace root contribution to anchorage as opposed to individual 361 plant data predicting specific outcomes in the multiple regression. 362 363 The predictors that are driving the random forest decision trees in each of the outcome models 364 were investigated to understand the plant architectures that are important for classifying the 365 brace root contribution to anchorage. For random forest models, the mean decrease in Gini value 366 can be used to assess the relative importance of each predictor in the decision tree. A higher Gini 367 value indicates a greater influence on the decision tree and thus a more important predictor. For 368 the random forest model that used within-year data, the top predictors included the single root 369 width, the number of brace root whorls in the ground, and plant height (Figure 2A). For the 370 model that used multi-year genotypic averages, the top predictors included plant height, the 371 stalk-to-root grounding, and the number of roots on whorl 1 (Figure 2A). The influence of 372 multiple phenotypes on prediction accuracy is consistent with the moderate univariate 373 correlations we reported (Table 1)  Genotype determines root lodging susceptibility 377 In 2020, a natural root lodging event occurred at our Newark, DE field site during the late-378 vegetative and early reproductive growth stages (65 DAP) after brace roots had entered the 379 ground. Tropical Storm Isaias brought heavy rains followed by sustained winds that averaged 380 23.8 miles per hour (mph) with gusts up to 59 mph (Table S11). Winds were from the West and 381 Northwest, but root lodging was not influenced by field position relative to the wind ( Figure S7). 382 This storm provided the unique opportunity to perform a detailed assessment of root lodging 383 within the same germplasm that we previously assessed for the brace root contribution to 384 anchorage and plant architectures. 385 386 Within the genotypes, there was a continuous distribution of root lodging susceptibility, with 826 387 plants remaining upright while 421 plants root-lodged ( Figure 3A). A genotypic-level 388 categorization identified 16 root lodging resistant genotypes, with no root lodging in either plot 389 replicate ( Figure 3B), and 36 genotypes that had variable root lodging in one or more plots 390 ( Figure 3C). As expected, the genotypes that were classified as lodging resistant had more brace 391 root whorls in the ground compared to those that were considered variable (p<0.05, Figure 3D). 392 The broad sense heritability (H 2 ) estimate of root lodging susceptibility was 0.58 (Table S5). 393 However, the phylogenetic relationship of genotypes showed that, like the brace root 394 contribution to anchorage, root lodging resistance is independent of population structure (Figure 395 S4). Together, these data are consistent with the genotype-specific, but not subpopulation 396 specific, susceptibility to root lodging. 397 398

400
The contribution of brace roots to anchorage is a good proxy for root lodging susceptibility 401 To determine if the None/All ratio is an appropriate proxy for identifying lodging resistant 402 genotypes, a pairwise Pearson correlation was calculated. The correlation between the multi-year 403 genotypic averages of the None/All ratio and root lodging susceptibility was surprisingly high (r 404 = 0.36; Table 1) given that the data were from different years and different growth stages. These 405 data provide additional support for the importance of brace roots for anchorage, with the positive 406 correlation showing that the greater the brace root contribution to anchorage (i.e., a lower ratio), 407 the less likely the plant is to root lodge. 408 409 To determine if the same phenotypes that are important for the None/All ratio are also important 410 in lodging susceptibility, pairwise Pearson correlations were run (Table 1)

Predictor importance -Mean decrease in Gini value
between phenotypes and the None/All ratio, there were overall moderate correlations between 412 individual phenotypes and root lodging susceptibility (Table 1). The two highest correlations  413 were with the number of whorls in the ground (r = -0.43) and plant height (r = 0.31). 414 Interestingly, these same phenotypes had high correlations with the None/All ratio between and 415 within-years. 416 417 When considering a multiple regression model using individual plant architectures to predict 418 genotypic averages of root lodging susceptibility, the R 2 value is 0.24 (Table 2). This is the same 419 predictive power as when using the plant architectures to predict the None/All ratio. Thus, these 420 data show that end-of-season plant architectural phenotyping data can be used to predict mid-421 season root lodging with surprising accuracy. 422 423 When the categorical classification approach was used, the relative genotype classification (low, 424 average, high) for root lodging susceptibility was classified with 72.00% accuracy (ROC 425 AUC=0.76; Table 2). With the inclusion of the None/All ratio, the prediction accuracy of root 426 lodging susceptibility increased slightly 78.00% (Table 2). Similar to the models that predicted 427 the None/All ratio (Figure 2A), the top predictor of lodging susceptibility was plant height. 428 However, unlike the None/All ratio, the rest of the plant architectures are approximately equal in 429 their contribution to the decision tree ( Figure 4). Overall, this data confirms the utility of multi-430 year data to provide insight into the plant architectures that promote anchorage and limit root 431 lodging. 432 433 434 The predictor importance for each of the random forest models was determined by the mean decrease in Gini value. A darker shade of blue indicates that the corresponding predictor had a higher mean decrease in Gini value and thus was more important in the decision tree, whereas a lighter shade of blue/white indicates that the corresponding predictor had a lower mean decrease in Gini value and thus was less important. Boxed percentages indicate the prediction accuracy for the corresponding model. Plant architecture predicted root lodging susceptibility with 72.00% accuracy, whereas the inclusion of the None/All ratio increased prediction accuracy by 7.00%. The top predictor for both models is plant height.

436
Plant height is positively correlated with brace root phenotypes but has opposing effects on root 437 lodging susceptibility. 438 Plant height has been historically associated with lodging, as evidenced by the development of 439 dwarf varieties of grain crops during the Green Revolution, which reduced the impact of lodging. 440 The moderate correlations between plant height and lodging susceptibility (r = 0.31) in this study 441 suggest that plant height alone cannot not explain root lodging in maize (Table 1). Interestingly,  442 there were positive correlations between brace root phenotypes and plant height ( Figure S6), but 443 they have opposing effects on lodging susceptibility (Table 1). In other words, taller plants are  444  more likely to root lodge, but taller plants also have larger brace root traits, which limits root  445 lodging. When genotypic averages were used, plant height was universally the most important 446 predictor (Figure 2A, Figure 4) in machine learning models. Together, these data show that the 447 contribution of multiple phenotypes is important for limiting root lodging and highlights the 448 importance of uncoupling plant height and root system traits for future plant improvement. 449

Discussion 450
This study was motivated by a limited understanding of how plant architectures contribute to 451 plant anchorage and root lodging resistance. Using a multi-year data acquisition strategy, we 452 assessed 1) the brace root contribution to root anchorage, 2) plant architectural phenotypes, and 453 3) root lodging susceptibility in a population of 52 maize inbred lines. Surprisingly, we found 454 that multi-year genotypic-level data were highly successful in predictive, categorical modeling. 455 456 In linking plant architectures to anchorage and root lodging resistance, we demonstrate that plant 457 height is important for both outcomes. Generally, genotypes classified as root lodging resistant 458 were shorter and with a greater brace root contribution to anchorage than genotypes that were 459 classified as root lodging susceptible. This is consistent with the prevailing idea from the Green 460 Revolution that short plants are more lodging-resistant. However, the correlation was moderate 461 (r = 0.31; Table 1), which indicates that plant height is not the only factor influencing root 462 lodging. Indeed, this is consistent with the variable reports on plant height and lodging in maize 463 (Brune et al., 2018;Sharma and Carena, 2016). In addition to plant height, we show brace root 464 phenotypes that collectively establish a wider base are associated with an increased contribution 465 to anchorage and root lodging resistance. Interestingly, these same phenotypes are associated 466 with increased plant height. Thus, a genetic decoupling of plant architectures is necessary to fully 467 optimize plants for root lodging-resistance. 468 469 Plant anchorage is an outcome of both above-and below-ground root systems, and their 470 interaction with the ground. Despite this complex relationship, we find that genotypes with a 471 higher brace root contribution to anchorage are more likely to resist lodging. This result 472 demonstrates that either brace roots are the primary contributor to anchorage, or the functional 473 and phenotypic assessment of the above ground root system is reflective of the function and 474 phenotype of the below ground. 475  Table 2. Prediction accuracy of supervised classification models. Supervised classification 630 models (linear regression and random forest models) were used to predict the None/All ratio and 631 lodging susceptibility from plant architectures. The average prediction accuracy and Receiver 632 Operating Characteristic (ROC) Area Under the Curve (AUC) values were extracted from 633 models. Models were run either on individual (within-year) data or genotypic averages for 634 outcomes. The predictors included in each model are indicated with an X. 635