Environmentally dependent and independent control of cell shape determination by Rho GTPase regulators in melanoma

In order to invade 3D tissues, cancer cells dynamically change cell morphology in response to geometric and mechanical cues in the environment. But how cells determine their shape in 3D versus 2D environments is poorly understood. Studying 2D versus 3D single cell shape determination has historically been technically difficult due to the lack of methodologies to directly compare the two environments. We developed an approach to study cell shape in 2D versus 3D by measuring cell shape at different depths in collagen using stage-scanning oblique plane microscopy (ssOPM). We find characteristic shape changes occur in melanoma cells depending on whether a cell is attached to a 2D surface or 3D environment, and that these changes can be modulated by Rho GTPase regulatory proteins. Our data suggest that regulation of cell protrusivity undergoes a ‘switch’ of control between different Rho GTPase regulators depending on the physical microenvironment.


Introduction
Taken together, these results reveal new context specific regulators of protrusivity and highlight 105 the ability of high-throughput plate based volumetric imaging to rapidly assay and identify proteins 106 in control of cell shape. instructions. On the second day after transfection, 10 5 cells/ml were re-suspended in 500 µl of 2.3 123 mgs/ml collagen rat tail (Gibco). A 100 µl volume of the collagen and cell mixture was dispensed 124 in quadruplicate wells onto poly-D-Lysine (0.1mg/ml) coated glass bottom view 96 well plates 125 (PerkinElmer). Plates were centrifuged @1200 rpm for 5 minutes at 4°C and incubated overnight 7 in a tissue culture incubator. After incubation, cells were fixed with 4% PFA methanol free for 30 127 mins at RT. Wells were stained with DRAQ5 at a concentration of 5 µM, to label nuclei. 128 Table 1 Oblique plane microscopy was performed using a setup reported previously [30] and a schematic 132 Volumetric imaging in 3 spectral channels (2 x fluorescence and 1 x scatter) was performed in 174 two stages. The first stage acquired the two fluorescence spectral channels (CAAX-GFP and 175 DRAQ5) for the entire plate. In the second stage, the image acquisition was repeated but now 176 with scattered light from collagen imaged on camera 1 with 488 nm excitation and in the absence 177 of an emission filter. Camera 2 was used to image the DRAQ5 channel for a second time. As 178 images of DRAQ5 were acquired in both stages, they could then be used to measure any drift 179 between image sets and thus enable the two sets to be co-registered. 180 181 Image reslicing and registration 182 The average background level was measured for each camera by taking the average pixel value 183 over a field of view acquired with no laser on. This was subtracted from the data prior to reslicing. 184 The 2D transform to co-register camera 1 and camera 2 was measured based on a 2-channel 185 fluorescence image acquisition of a sample of 100 nm four-colour fluorescent beads (TetraSpeck,186 Thermofisher) in 10% agarose. The x-y shift, magnification and rotation needed to co-register the 187 data acquired on camera 2 to that of camera 1 was measured manually using a custom script in 188 MATLAB (imtranslate, imrotate, imscale and fliplr: Image Processing Toolbox, MATLAB). This 189 transform was applied to all raw image data acquired on camera 2 prior to reslicing. 190 Raw ssOPM images are a set of image planes at 55 degrees to the optical axis of O1. The data 191 was transformed into conventional coordinates (z parallel to the optical axis, x&y&z 192 perpendicular). Reslicing was performed using a bi-linear resampling algorithm [35]. To increase 193 speed, a custom-written Java implementation of the algorithm was used. For all image 194 segmentation and cell shape analysis, raw camera images were binned by factor of 4 -prior to 195 reslicing -to reduce data volume and analysis times. Images presented in figures in this paper were resliced with factor 2 binning. After reslicing voxel sizes were 1x1x1 m 3 (for analysis) and 197 0.5x0.5x0.5 m 3 (for display). 198 The collagen channel was coregistered in 3D with the fluorescence channels using the imregtform 199 (Image Processing Toolbox, MATLAB) using the default optimizer parameters. The transform was 200 measured on the DRAQ5 channel, which was common to both acquisitions, and applied to the 201  Fig 3A). 211 To verify the robustness of the 3D segmentation used in this paper, two methods were tested. An 212 intensity-threshold-based approach using an Otsu threshold and an active contour method, which 213 uses energy minimisation ( Supplementary Fig 4A). Both methods generate a mask with minimal 214 user input so can be applied to large datasets. For both methods, cells and nuclei were segmented 215 in 3D. 216 Prior to segmentation, the tips of the parallelepiped-shaped volume imaged by the ssOPM image 217 acquisition were removed by cropping in the y direction. This removed any parts of the volume 218 which were not imaged over its full axial extent due to the light-sheet angle. 219 In the intensity-based method, thresholds were measured automatically for each field of view. The 220 nucleus threshold was selected using Otsu's method with a single level (multithresh, Image 221 Processing Toolbox, MATLAB). The cell body was masked using a similar method. In this case 222 there were 3 intensity levels in the image, background, brightly fluorescent cell membrane and 223 dim fluorescent protrusions. To include all parts of the cell in the final mask, the lowest threshold 224 found by a two-level Otsu method was used. 225 In the active contour method, an initial guess of the mask was generated using a threshold of 5 226 digital numbers (just above the background Imaging of 3 plates produced a segmented dataset of more than 3x10 4 cells. We aimed to remove 253 cells and nuclei that could not be segmented accurately due to low expression of the CAAX-GFP 254 transgene, as well as cells with surface areas that are conspicuously large due to under-255 segmentation. To remove cells that were improperly segmented due to low expression of the 256 CAAX-GFP construct, we removed cells that had both: (i) low CAAX-GFP intensity (an average 257 intensity of less than 1000 in camera digital numbers); and (ii) a nuclear to total cell volume ratio 258 of greater than 0.95. These cells were removed from subsequent analysis as they represent nuclei 259 with cells that cannot be appropriately measured due to low CAAX-GFP transgene expression. 260 Applying these criteria removed approximately 500 cells (~ 1.8 % of original total). 261

Coverslip localisation 263
The position of the coverslip was estimated using the fluorescence in the nucleus channel. To 264 account for spatial variations in the axial position of the coverslip over the field of view, the nucleus 265 channel was divided into 16 segments in the y direction and the average signal in each x-y plane 266 was found for each z position in each segment. For each segment, the coverslip location was 267 defined as the point when the signal first reaches 45% of its maximum value when moving in the 268 positive z direction. A full 2D map of coverslip height was then produced from the 16 269 measurements using bilinear interpolation, giving a smooth change of coverslip height across the 270 field of view. This map of coverslip height reflects the spatial variation across the field of view. 271 However, this is a relatively crude method and does not account for variations in nuclear intensity 272 between knockdowns and between plates, therefore there is a remaining global offset in coverslip 273 position for each well that is accounted for in the next step.

Feature reduction and feature normalisation 288
We originally computed more than 20 measurements of cell and nuclear shape features, and 289 subsequently reduced this set to a set of four features. We used clustering to remove the most 290 highly correlated features as follows. First, we used single cell data to calculate Pearson 291 correlation values between each feature using the 'cor()' function in R. Pearson correlation values 292 between features were hierarchically clustered using the hclust() function in R, and the 'complete' 293 linkage method (see Fig 2). The resulting cluster was partitioned into four groups, and a single 294 cell or nuclear feature was chosen as a representative feature for each group. To allow for 295 comparison between plates that were prepared and imaged on different days, we normalised 296 single cell measurements for cell and nuclear features within each plate. We performed 297 normalisation by dividing feature measurements for a single cell, by the plate median across all 298 conditions for cells in that feature. 299 300 Data analysis 301 As some of the cell and nuclear shape features we examined did not have a normal distribution, 302 we used non-parametric tests throughout this study. To test for differences in measures of central 303 tendency between two conditions we used a paired Wilcoxon test with BH adjustment. For tests 304 of difference in central tendency between more than two conditions we used a Kruskal-Wallis test, 305 followed by a Dunn's test to discern which conditions differed from control. Tests for differences 306 between groups were performed on summary statistics calculated for wells, with multiple wells 307 from three different plates in each test. Tests at the well level used at least 12 wells per treatment, 308 from three plates, with at least 50 cells in the dataset per condition. Statistical tests were proximal and distal cells appeared consistent between threshold and active contour based 326 segmentation. Overall this suggests that results are repeatable between segmentation methods. 327 As either segmentation method is viable for this dataset, the Otsu threshold approach was chosen 328 due to the shorter computation time. 329

Anisotropic PSF shape 331
In light-sheet microscopy, the point spread function (PSF) is usually not spherical. The FWHM 332 spatial resolution of this ssOPM system has been previously reported as 0.5 um in the plane of 333 the light sheet, with a light sheet thickness of 3.8 um at the waist. The light thickness increases 334 to 5.4 μm over a distance of 50 μm from the centre of the field of view [37]. We set out to establish 335 whether the anisotropic PSF might affect coverslip proximal cells -that are more likely to be flatter 336 -differently to coverslip distal cells. We therefore eroded the cell and nucleus masks from all 337 segmented data by an object approximating a worst-case PSF. This was performed using 338 MATLAB's imerode function with a morphological structuring element consisting of a 1x1x5 pixel 339 kernel angled at 45° to the coverslip plane (closest possible approximation to a 55° light sheet 340 angle) . The voxel size in the image data was 1 um 3 , so this corresponds to a 1x1x5 um 3 PSF. and nucleus volume were also tested for the test plate, but this metric was found to be affected 368 by the spatially varying light sheet thickness. This is expected for the thin flat cells used in the test 369 plate, which represent a deliberate worst case, as these cells are generally thinner than the light-370 sheet thickness. The cell volume metric was therefore not used in the analysis in this paper. We 371 concluded that the behaviours of 'cell surface area', 'angle between', 'protrusivity' and 'nucleus 372 minor axis' in the collagen plate are not explainable by the PSF shape alone and are dominated 373 by other factors. 374

376
An experimental paradigm to measure cell shape in distinct 377 physical environments 378 To study the differences in shape as cells transition between two mechanically and geometrically 379 distinct environments, we suspended cells in 2 mg/ml (initial concentration) type 1 collagen (rat 380 tail), seeded this mixture into glass-bottomed 96 well plates, and centrifuged them. We incubated 381 cells for 24 hours before fixation. Glass is a rigid or 'hard' substrate (on the order of 1 GPa), 382 whereas collagen at a concentration of 2 mg/ml is relatively elastic [15] (between 300 and 1600 383 Pascals) [38-42] (Fig 1A). We used this system to study control treated cells, and also cells 384 treated with siRNA targeting a variety of Rho-regulators that we had selected from preliminary 385 screening ( Fig 1B). To compare cell response between physical environments we used stage 386 scanning oblique plane microscopy (ssOPM) to image the geometry of cells with nuclei at different 387 distances from the glass coverslip (Fig 1C-E). In each single volume, approximately 200 cells 388 were imaged across a 144 μm z range. Using this imaging approach for all of the wells and 389 treatments in our study generated an initial dataset containing more than 30,000 individual cells. 390 A technical advantage of this system is that variables are internally controlled because 391 comparisons can be made between 2D and 3D microenvironments for a large number of cells in 392 the very same well of a 96 well plate. In the case of a single gene knockdown, protein depletion, 393 media conditions, collagen concentration and biological composition between 'proximal 394 embedded' and 'distal embedded' cells are shared between 2D and 3D cell microenvironments. 395 To quantify cell morphology, we imaged GFP signal in CAAX-GFP-expressing WM266.4 cells, 396 and visualized nuclei using DRAQ5 (Methods). Initially we generated over 20 measures of cell or 397 nuclear shape features ( Supplementary Fig 2A). We performed dimensionality reduction by using hierarchical clustering to group these measurements into four clusters of highly correlated shape 399 features (Fig 2A). A representative shape feature was selected from each cluster to be used 400 throughout our analysis. By considering the shape features in each cluster we were also able to 401 suggest an interpretation of the underlying biology that is tracked by each feature (Fig 2A). The 402 shape features chosen were 'Cell surface area', 'Angle between cell and nucleus', 'Nucleus minor 403 axis', and 'Cell protrusivity' (description of features in Supplementary Fig 2A). After feature 404 reduction, no features had an absolute correlation of greater than 0.5 (Fig 2A). Collectively we 405 refer to these shape features as global geometry features. shape we compared global geometry features of proximal and distal cells (Fig 2B). This revealed 416 significant stereotypic differences in cell morphology depending on whether cell nuclei were 417 proximal or distal to the coverslip. These changes included reduced cell protrusivity, and smaller 418 cell surface area in distal versus proximal cells (Fig 2B). In contrast to these cell shape features, 419 the nuclear geometry feature, length of the minor axis of the nucleus was not different between 420 proximal and distal cells (Fig 2B). Thus cells invading 3D collagen gels are typically less protrusive 421 and have a smaller surface area than in 2D environments. 422 Although our nuclear shape measures were not altered, we found changes in the relationship of 423 the cell to the nucleus. When cells were proximal to the coverslip the major axis of elongation of 424 the cell and nucleus was coordinated or 'coupled'. In contrast, for cells positioned away from the 425 glass coverslip we found an increase in the angle between the cell and nucleus. This suggests 426 the orientation of the nucleus was less constrained by cell geometry when positioned away from 427 the glass coverslip. 428 Previous studies have noted that WM266.4 cells exhibit extensive heterogeneity in morphology 429 when cultured either on stiff 2D plastic, or soft collagen matrices -adopting either amoeboid or 430 spindle forms [8,43,44]. But how the extent of this variability changes between 2D and 3D is poorly 431 understood. We tested whether variance in a range of shape features is changed between 432 proximal and distal settings. Due to the connection between cell protrusivity and metastatic 433 potential, we focused on variation in protrusivity (Fig 2E). To visualise these variations we grouped 434 the cells into three groups -below average, near average and above average protrusivity -and 435 selected 9 representative masks from the cell masks (Fig 2D). We also made a visual summary 436 of variation in protrusivity (without grouping the cells) by creating 'stacked maximum intensity 437 projections' (stacked-MIPs) (Fig 2F). The projected cells were the first 'n' cells in our dataset, 438 which matched the protrusivity criteria. These stacked-MIPs support the finding that protrusivity 439 is decreased when cells are positioned away from the coverslip. These observations are 440 consistent with those made by ourselves and others that WM266.4 cells alternate between round 441 and spindle forms in 3D gels, but are more homogenous on 2D stiff surfaces [7,8,44]. This 442 increased variance may reflect that there are more degrees of freedom away from the rigid 443 coverslip, which may make it less likely for cells to adopt stereotypic shapes. 444 Rho-regulators disrupt differences in shape between proximal and 446 distal cells 447 Having identified characteristic shape differences between cells that are proximal or distal to the 448 coverslip, we interrogated the molecular control of these shape changes. To do this we depleted 449 a range of Rho-regulatory proteins, and examined whether their function was affected by distance 450 from the coverslip. To select a set of Rho-regulators for study by ssOPM, we had conducted a 451 collection of preliminary screens on cells plated on collagen to look for genes that control cell 452 morphology and imaged by confocal microscopy (Fig 1B). We identified nine Rho-regulators that 453 have a potent influence on cell shape (Fig 1B). 454 We depleted these nine Rho-regulators and for eight of these we measured the effect on shape 455 transitions between 'proximal' and 'distal' cells ( Fig 1E and 3A and 3B). ECT2 depleted cells 456 frequently had a multinucleate phenotype consistent with failed cytokinesis (Fig 1E). This 457 phenotype indicated potent protein depletion from our treatments, however to focus on primary 458 effects of gene knockdown on shape we did not analyse these cells. To see how Rho-regulators 459 contribute to cell shape in proximal 2D and distal 3D cells, we projected our four global geometry 460 features ( Fig 3A) in principal component (PC) space ( Fig 3B). The two largest contributions to 461 PC1 were cell protrusivity and cell surface area, while the two largest contributions to PC2 were 462 the nucleus minor axis, and the angle between the cell and nucleus. 463 For control cells in PC space, coverslip proximal cells formed a cluster with lower variance than 464 coverslip distal cells. Coverslip distal cells explore shapes characterised by reduced protrusivity 465 and cell surface area (higher PC1) and lower nucleus minor axis and greater angle between cell 466 and nucleus (lower PC2). We found that depletion of some Rho-regulators generated overlap 467 between proximal and distal cells in PC space (Fig 3A). For example, this overlap was substantial 468 for cells depleted for FARP1. This was true but to a lesser extent for ARHGEF35 and SRGAP1-depleted cells. The shape convergence we saw in PC space was visually supported by generating 470 stacked-MIPs of 200 cells (Fig 3C), which indicated greater similarity between proximal and distal 471 cells for FARP1, compared to control. Taken together we identify FARP1, ARHGEF35 and 472 SRGAP1 as important for reducing differences in shape between proximal and distal contexts 473 (Fig 3A-C). Visualisation in PC space also suggested that Rho-regulators have stronger and more 484 stereotyped effects on cell shape in proximal compared to distal cells. In the proximal cells we 485 found that our four global shape features were sufficient to separate many Rho-regulator depleted 486 cells from control treated WM266.4 melanoma. This was especially the case for SRGAP1, 487 FARP1, TIAM2, RND3 and DOCK5 (Fig 3D), where depletion of these proteins in proximal cells 488 created combinations of shape features that were separable from control treated WM266.4 489 melanoma cells. This was in contrast to the distal context, where cell shape features overlapped 490 between control WM266.4 melanoma and Rho-regulator depletion (Fig 3D), and where cell This supported our finding that Rho-regulators have a potent and stereotypic effect on cell shape, 493 but this effect depends on distance of the cell from the glass coverslip. Our data suggests that for 494 distal cells the physical environment has an overarching control on shape and increases variability 495 in cell shape. 496 497 Control of cell protrusivity 498 We next sought to identify specific shape differences that are disrupted by Rho-regulator 499 depletion, and focused on cell protrusivity. It is critical to understand regulation of cell protrusivity 500 in WM266.4 melanoma because changes in protrusivity are linked to malignant cell migration. In 501 the past a comprehensive understanding of the genetic control of cell protrusivity has been 502 confounded by different control of protrusivity between 2D and 3D environments, as well as 503 between rigid and soft environments. We define protrusivity as 1-ratio of the mask volume to its 504 convex hull volume (Fig 4A & B). This is similar to the spreading metric used by Isogai et al [45]. 505 Increases in our protrusivity metric correlate with the number of cell protrusions, but also the 506 increased length of cell protrusions, and the angle between cell protrusions. 507 To understand how specific Rho-regulators control protrusivity when positioned near or far from 508 a rigid substrate, we compared median cell protrusivity of control and Rho-regulator depleted 509 WM266.4 melanoma cells. To account for the effect of cell microenvironment we made these 510 comparisons separately for proximal and distal cells (Fig 4C & D). In cells proximal to the 511 coverslip, depletion of FARP1, TIAM2, DOCK5 and RND3 each decreased protrusivity (Fig 4D). 512 In the distal context only, cells depleted for TIAM2 had reduced protrusvity compared to the control 513 WM266.4 melanoma cells (Fig 4D). TIAM2, the Rho-regulator that controlled protrusivity in distal 514 cells, also controlled protrusivity in proximal embedded cells, suggesting control of a particular 515 shape process near to the glass coverslip is associated with the ability to control the same process far from the glass coverslip. We noted that RND3 and DOCK5 depleted cells both appeared to be 517 associated with a reduction in cell protrusivity but also cell number (Fig 4C). Compared to control, 518 the average number of cells in wells depleted of DOCK5 and RND3 was reduced to approximately 519 65 and 77 percent, respectively. Therefore to focus on changes in protrusivity that were directly 520 related to shape control without complications introduced by cell survival, we continued our 521 analysis with FARP1 and TIAM2. 522 We visualised protrusivity in FARP1 and TIAM2 depleted cells using CAAX-GFP signal from 523 proximal and distal cells (Fig 4E) as well as through stacked-MIPs (Fig 4F). This confirmed that 524 FARP1 depleted cells are round when close to the coverslip, but have protrusivity similar to control 525 cells when positioned away from the coverslip (Fig 4E & F). In contrast, TIAM2 depleted cells 526 were round in both environmental contexts (Fig 4E & F). Taken together these results suggest 527 that in our collagen system control of protrusivity is environment-stiffness dependent for FARP1 528 but independent of environment for TIAM2. 529 530 Scale of regulation of protrusivity 531 Next we looked to find the distance over which protrusivity changes as cells are positioned away 532 from the coverslip in untreated cells. We also looked to characterise the different ranges or 533 distances over which FARP1 and TIAM2 control protrusivity. To resolve this we binned cells over 534 two micron intervals and plotted the mean cell protrusivity for distances up to 20 microns from the 535 coverslip ( Fig 5A). We found that compared to control, FARP1 and TIAM2 are each required for 536 cell protrusivity in cells with nuclei positioned up to 7-8 microns from the glass coverslip, but that 537 for distances beyond this only TIAM2 is required for protrusivity (Fig 5A & B).
To visualise the changes in protrusion that occur with distance from the coverslip, we plotted 539 stacked-MIPs in the XZ plane at a range of intervals for cells within the first 20 microns from the 540 coverslip ( Fig 5C). This confirmed that FARP1 depleted cells regain protrusivity beginning at 541 distances around 8-10 microns from the coverslip, and appear similar to control cells at distances 542 of 12 or more microns from the coverslip. Visual inspection also confirmed that TIAM2 treated 543 cells have a large reduction in protrusivity at all distances from the coverslip (Fig 5C). 544 Here our genetic perturbation data indicates that -for the gel used in this study -distances on the 545 order of 7 microns mark a threshold, beyond which the molecular control of protrusivity is 'handed 546 over' from FARP1 to other shape regulators including TIAM2. This suggests a model where the 547 control of protrusivity relies on both TIAM2 and FARP1 in the micro-environment close to the 548 coverslip, but that control of protrusivity 'switches' to rely on TIAM2 in the environment further 549 away from the coverslip (Fig 5D). These context specific roles may reflect different abilities of 550  (Fig 6B). This plot pooled both proximal and distal 568 cells and indicated a positive relationship between normalised cell and nuclear height at the single 569 cell level (Fig 6B). We also plotted frequency histograms for cell and nuclear height and noted 570 that FARP1 depleted cells had a shift towards increased cell and nuclear height (Fig 6B). 571 To visualise how cell and nuclear axial extent (height) change with distance from the coverslip we 572 binned cells by the position of their nucleus at one micron intervals, and calculated the average 573 cell and nucleus height (Fig 6C), as well as nuclear position. We used this information to generate 574 'glyphs' of cells that give an indication of how the relationship between cell and nucleus changes 575 with distance of the nucleus from the coverslip (Fig 6C). This plot suggested that FARP1 depleted 576 cells increase their height when nuclei are within 7 microns of the coverslip, but are similar to 577 control cells for distance beyond 7 microns. In contrast, we found that the height of TIAM2 578 depleted cells are similar to control cells within the first 7 microns of the coverslip, but are reduced 579 in height for distances beyond this. Statistical testing of the difference in height between proximal 580 and distal cells supported these observations (Fig 6D). 581 The increase in height for FARP1 depleted cells might be attributable to the concomitant increase 582 in nuclear height (Fig 6B). Changes in nuclear height on rigid surfaces have previously been 583 linked to maintenance of the perinuclear actin cap [46,47]. In contrast, the changes in cell height 584 in TIAM2 depleted cells (Fig 6C and 6D) are likely to be driven by reduced protrusivity in cells 585 away from the coverslip (Fig 4B-D), rather than directly by changes in nuclear geometry.
Considered together, these results show that FARP1 and TIAM2 are each required to regulate 587 cell height, but that they regulate height in different micro-environmental contexts. 588 589 Control of cell and nuclear coupling 590 Finally, we considered coupling of cell and nuclear orientation (Fig 6E and 6F). Coupling of cell 591 and nuclear orientation is frequently observed in mammalian cell systems, where it is important 592 for cell mechanotransduction and cell migration but this relationship breaks down in disease 593 contexts. To measure cell and nuclear coupling we calculated the angle between the orientation 594 of the major axis of the cell and the nucleus. We looked for changes in cell and nuclear coupling 595 by generating probability density plots for control WM266.4 melanoma, as well as cells depleted 596 for FARP1 and TIAM2. 597 Consistent with previous studies, we found that in control treated WM266.4 melanoma cells there 598 is a tight coupling of cell and nuclear orientation in proximal cells (Fig 6F). We found that this 599 coupling is reduced in cells with their nuclei positioned distal to the coverslip (Fig 6F). Given that 600 previous studies have also seen that changes in nuclear height are associated with breakdown 601 of cell and nuclear coupling [47], we examined this in FARP1 depleted cells and saw a tendency 602 for increases in the angle between the cell and the nucleus in proximal cells. 603 604 Comparison of cell shape in distinct environments reveals TIAM2 605 and FARP1 control a range of shape features but in different 606 physical environments.

607
Due to the importance of cell protrusivity in disease, we have focused on this metric of cell shape 608 and found a context-dependent control of protrusvity by a subset of Rho-regulators (Fig 4A-B). 609 However, we have found that shape differences between proximal and distal environments can 610 be characterised by the changes in three additional shape measures (Fig 3A). Therefore we 611 considered how each of the Rho-regulators in our study controlled these individual shape 612 features, and whether this control is physical context specific (Fig 6A-D). To test the effect of Rho-613 regulators across a range of shape features, we analysed well-median values for each of our 614 global geometry features and compared them between Rho-regulator depleted cells and control 615 cells. Comparisons were made using Kruskal-Wallis and Dunn's tests to compare controls to 616 treatment. 617 For cells proximal to the coverslip, we found that depletion of many Rho-regulators were able to 618 change multiple shape features. The broadest acting shape controllers in proximal cells were 619 FARP1, DOCK5, TIAM2 and RND3. The proteins that acted on fewer features were ARHGEF9 620 and ARHGEF35. PREX2 did not significantly change any shape features in this study (Fig 7A). In 621 contrast, for distal cells we found that cell geometry was relatively robust to depletion of the same 622 Rho-regulators (Fig 7B), and that fewer shape features were significantly changed by Rho-623 regulator depletion. 624 To summarise the breadth and context specificity of shape control by Rho-regulators in our study, 625 we generated a shape control matrix (Fig 7C). This highlights DOCK5 and TIAM2 as controlling 626 shape features in both proximal and distal contexts (Fig 7C). In contrast, FARP1 and RND3 stand out as broad-acting shape regulators that significantly changed each shape feature measured, 628 but were only effective in proximal or rigid physical regimes (Fig 7C). Here, we have used ssOPM to address these challenges by imaging thousands of cells in 641 collagen and in two distinct physical contexts. This approach creates an opportunity to understand 642 how cells respond to different geometrical and mechanical cues, and how a genetic perturbation 643 affects this response. We find that control treated WM266.4 melanoma cells have reduced 644 protrusivity and become more heterogeneous when positioned away from the coverslip. A 645 systematic depletion of Rho-Regulators revealed genes that modulate this transition. In particular 646 our data suggest that TIAM2 and FARP1 ordinarily function to promote protrusions in coverslip 647 proximal cells (Fig 4 and Fig 5). The context dependence of In LSFM the PSF is typically anisotropic due to the low excitation NA compared to detection. In 674 the ssOPM system used here the resolution was (0.5x0.5x3.8 μm 3 -0.5x0.5x5 μm 3 depending on 675 position in the light sheet). This may lead to cells which have a dimension < 5 μm to be extended 676 in the light sheet direction, depending on orientation. Deconvolution could be used to reduce this 677 effect [48] but 3D deconvolution of large datasets is slow and does not take into account the 678 spatially varying light sheet. As a simple test we eroded segmented masks with a structured 679 element object similar to the PSF. The same cell shape changes were found with both eroded 680 and uneroded masks. This suggests that, for this dataset, the PSF shape did not have a significant 681 effect compared to the biological effects. 682 A wide range of segmentation approaches can be used in fluorescence microscopy. We tested 683 intensity and active-contour-based segmentation. By visual inspection both methods produced 684 good masks of cells and nuclei. We further found that we would draw the same conclusions from 685 our data using either method. In the present study we use WM266.4 cells that have low levels of Rho GTP and produce both 696 hydrostatic blebs and actin based protrusion but are thought to be predominately mesenchymal 697 [7]. We also use a collagen concentration and polymerisation temperature associated with the 698 formation of highly reticular collagen networks, and nascent, unstable integrin-based adhesions 699 with low contractility [53]. 700 In this setting we identified FARP1, TIAM2, RND3 and DOCK5 as regulating protrusivity when 701 cell nuclei are proximal to the coverslip. In future it will be interesting to distinguish changes in the 702 amount of hydrostatic blebbing from actin-based pseudopodial protrusions by simultaneously 703 assessing plasma membrane and actin markers in the context of Rho-regulator depletion to 704 determine whether either type of protrusion is specifically controlled by these Rho-regulators. 705 Notably we also highlighted Rho-regulators where depletion increased cell protrusivity. In 706 particular, reduction of PREX2 and SRGAP1 tended to increase protrusivity indicating that these 707 Rho-regulars normally function to repress protrusivity, however these changes did not reach 708 statistical significance in this study. 709 710 Context specific protrusivity control 711 We found that most Rho-regulators were context dependent in the sense that they were more 712 potent in controlling protrusivity close to the coverslip. This may reflect major cell biological and 713 cytoskeletal changes that take place in response to rigidity sensing [54,55]. For example, FARP1, 714 RND3 and DOCK5 may modulate protrusivity through mechanisms that depend on the 715 abundance of integrin adhesions and filamentous actin organised into stress fibres, which are 716 associated with rigid substrates. In contrast TIAM2 was required for protrusion formation both 717 proximal to and far from the coverslip and may promote protrusivity independent from 718 environmental stiffness. 719 720 Changes in nuclear shape and alignment 721 For cells in close proximity to the coverslip, the reduced protrusivity in FARP1-depleted cells was 722 associated with an increase in nuclear axial extent ( Fig 6C) and reduced coordination in the angle 723 of elongation between cell and nucleus ( Fig 6F). Interestingly, increased nuclear height and 724 decreased coupling of cell-nuclear orientation were recently reported for loss of TIAM2 [47]. In 725 the case of TIAM2, increased nuclear height has been attributed to loss of nuclear capping actin, 726 and uncoordinated cell and nuclear orientation has been attributed to dysfunction of the 727 perinuclear actin cage [47]. In future it will be interesting to determine whether FARP1 regulates 728 perinuclear actin cap morphology, or is controlling nuclear axial extent by a separate mechanism. In this study we use the ability of ssOPM to image thousands of melanoma cells spanning 2D and 749 3D collagen environments. We find cells make characteristic changes between 2D and 3D and 750 that these changes can be modified by depletion of Rho-regulators. We find that cells in 3D 751 environments tend to reduce their protrusivity and their protrusivity also becomes more varied 752 and heterogeneous. Our data also suggest that cells respond to changes in environmental 753 parameters such as stiffness and geometry, over scales that are smaller than the diameter of the 754 nucleus. Furthermore, we identify TIAM2 and FARP1 as each controlling cell protrusivity but in 755 different physical contexts. Taken together our data indicate general reliance on TIAM2 for cell 756 protrusivity, and a context dependent switch from FARP1 dependent to FARP1 independent 757 control of protrusion between 2D and 3D settings. 758