Introduction

Cancer remains a major health problem and one of the leading causes of mortality worldwide1. This is mostly due to the development of metastatic disease and the lack of powerful tools to detect invasive cancer cells at an early stage. Remarkably, during cancer progression, cells undergo a multitude of morphological alterations in which the cytoskeleton is known to play a key role. However, the identification of these subtle cellular adaptations remains a critical limitation, contributing to late diagnosis and poor disease outcomes.

It is well established that the cytoskeleton is a complex and dynamic three-dimensional (3D) network able to modulate cellular architecture and, therefore, to interfere and dictate many cellular processes, namely cell migration, invasion and metastasis2. The cytoskeleton is a scaffold of interlinked microtubules, microfilaments, and intermediate filaments that altogether form the major mechanical structure of the cell, pivotal in coordinating cell shape and intracellular signalling. These fibers differ in their stiffness, polarity and assembly, sensing the microenvironment to respond accordingly3. In particular, microtubules are composed of α-/β-tubulin heterodimers that associate into long filaments and are responsible for chromosome segregation during cell division, as well as for directing proteins to the leading edge of migrating cells2,4. In contrast, actin filaments switch between monomeric and polymeric structures forming membrane protrusions such as lamellipodia, filopodia and pseudopodia, which generate the driving force for cell movement5,6. Intermediate filaments, including vimentin and keratin, connect the nucleus and cytoplasmic compartment with the cell surface, thus maintaining cellular integrity and mediating the interplay with the extracellular matrix (ECM)6.

To date, several studies have demonstrated that alterations in cytoskeletal components or interacting proteins can impact cancer cell behaviour. For instance, abnormal actin expression has been reported in many cancer types, affecting cell proliferation, migration and resistance to therapy5. Mutations in tubulin genes have also been found to be implicated in drug resistance7,8, whereas actin and tubulin distribution patterns have been associated with unique Epithelial to Mesenchymal Transition (EMT) signatures during carcinogenesis9,10. Likewise, abnormal activity of cytoskeleton binding partners, such as Filamin A, is correlated with metastatic potential in hepatocellular carcinoma and prostate cancer11,12,13.

Despite the vast number of studies, most of the research on cytoskeletal reorganization has been focused on the expression levels of cytoskeletal components or interactors and mostly based on fluorescence intensity or distribution14,15,16. Therefore, studies addressing fiber organization and topological patterns of this complex scaffold could provide an innovative perspective that will be determinant to understand how specific conformations translate into functional properties.

In this work, we propose a new pipeline to evaluate the cytoskeletal architecture of invasive cells, encompassing an algorithm for automated extraction of microtubule structural patterns. Immunofluorescence images were used to devise a framework of analysis, which include fiber morphology, orientation, compactness, and radiality. For validation, cells expressing mutant E-cadherin, leading to loss of cell-cell adhesion and an invasive phenotype, were investigated and compared with wild-type cells, while taking into consideration cell-ECM interaction. Overall, our data revealed that the proposed framework is able to distinguish unique microtubule signatures making it an efficient strategy to recognise cells with disseminating properties. We envision that our approach could provide a valuable tool for research purposes to study cytoskeleton rearrangements in a broad range of biological processes.

Results

A new image-based pipeline is able to dissect cytoskeletal organization features

The cytoskeleton undergoes dramatic reorganization during cancer progression but the identification of such modifications remains challenging. In this work, we have developed a new bioimaging pipeline to characterize and quantify cytoskeletal fiber organization from immunofluorescence images of cells stained for α-tubulin. For this purpose, preprocessing and processing methodologies were sequentially applied in order to properly segment nuclei and cytoskeletal fibers, and to extract accurate features. Images of cells stained for the nucleus and the cytoskeletal component α-tubulin were first subject to deconvolution, a technique to remove noise and blur, improving contrast and resolution. Notably, multiple images were acquired for each channel along the Z axis and projection of the Z-stacks through maximum intensity projection (MIP) in deconvoluted 2D images was applied. These images were then processed with a Gaussian filter to smooth the fluorescence signal of cytoskeletal components, a Sato filter to highlight curvilinear structures, and a Hessian filter to generate binary images (Fig. 1a). Ultimately, binary images were skeletonized to enable the calculation of specific cytoskeletal parameters. In this context, line segment rearrangements and graph networks were used as processing strategies of the cytoskeletal structures, allowing automatic extraction of Line Segment Features (LSFs) and Cytoskeleton Network Features (CNFs), denoted by line segments and nodes, respectively (Fig. 1b; Table 1). Importantly, automatic nuclei segmentation was also performed to define nuclei centroids and area, required for the extraction of nuclei-cytoskeletal ratios.

Fig. 1
Fig. 1The alternative text for this image may have been generated using AI.
Full size image

Preprocessing and processing framework of cytoskeletal analysis. (a) Results of the preprocessing pipeline are shown for immunofluorescence images of CHO cells with distinct morphologies. α-tubulin is stained in red and DAPI (blue) marked nuclei. Line segment arrangements and graph representations are shown for each preprocessed cell, with black dots as nodes and red pixelated paths as graph edges. (b) Schematic representation of Line Segment Features (LSFs) and Cytoskeleton Network Features (CNFs) analysed for each cell. LSFs, considering individual or close-by segments, were calculated using nuclei centroids as reference points. For each line \(\:i\), \(\:\left({x}_{i}^{1},{y}_{i}^{1}\right)\) and \(\:\left({x}_{i}^{2},{y}_{i}^{2}\right)\) represent the start and end points; \(\:{d}_{i}\) represents the distance to the i’th neighbour line; \(\:{\theta}_{i}\) is the angle with the horizontal axis; \(\:{L}_{i}\) is the length and \(\:{D}_{i}\) is the distance to the nucleus’ centroid, \(\:\left({x}_{N},{y}_{N}\right)\); \(\:{\alpha}_{i}\) represents the smallest angle between the line and the vector from a pixel \(\:\left({x}_{N},{y}_{N}\right)\) to line i. CNFs were extracted from the graph representation of the skeleton. Each square represents a pixel, whereas each edge \(\:i\) is a set of \(\:{n}_{j}\) adjacent pixels and \(\:{L}_{i}^{E}\) represents its total length.

These data suggest that the proposed computational framework is able to scrutinize cytoskeletal organization features, which will be crucial to distinguish unique geometrical patterns associated with invasive cells.

Table 1 Features extracted from deconvoluted cytoskeletons (DCF), line segment rearrangements (LSF) and graphs (CNF).
Full size table

Cells with distinct morphology exhibit evident differences in cytoskeletal architecture

To further investigate the architecture of the cytoskeleton, a total of ten cytoskeletal features were recovered by exploring LSFs and CNFs, allowing a thorough characterization of this complex fiber network. These comprise orientation, morphology, quantity, compactness, radiality, bundling, parallelism, connectivity, complexity, and cytoskeleton-nucleus interconnection. Remarkably, we verified that cells with distinct morphologies display different cytoskeletal structures, which can be characterized and quantified by our computational pipeline. For comparative analysis of qualitative and quantitative outputs, features were assessed in groups of three cells. Interestingly, line segments and graph edges were able to accurately reproduce fiber distribution and organization across cells phenotypically different, as well as to generate a corresponding quantitative profile at single cell level (Fig. 2a). In particular, we have used angular distribution (θi) of the cell cytoskeleton to determine the Orientational Order Parameter (OOP) and evaluate fiber alignment and orientation. A lower angular distribution corresponds to well aligned fibers and thus to a higher OOP value. Accordingly, we observed that the cell with lower angular distributions, as a result of well aligned fibers (cell (c), Fig. 2a), is the cell with the higher OOP (0.475). In contrast, cells presenting a wider range of angles and thus disorganized fibers display lower OOP values (cell (a), 0.019; cell (b), 0.139). Fiber length and its intercellular variability were determined through LiE, demonstrating that cells with higher fiber lengths also presented a higher length variability. Another important aspect of cytoskeletal organization is the quantity of polymerized fibers in a cell, as determined by the number of lines (Nl). We have identified 50, 105 and 292 fibers in cells (g), (h) and (i) respectively. The number of fibers in a specific cell is highly variable and is probably related to the cell state and behaviour. Nonetheless, cytoskeletal fibers may be found in similar numbers in many cells but these might be positioned in a more confined region or sparsely distributed. To better understand this phenomenon, we have evaluated fiber compactness by measuring the number of fibers per cell area. While in cell (j), cytoskeletal fibers are more dispersed in the cytoplasm and clearly distinguishable (Nl/Ac 0.421 μm− 2), in the other cells, cytoskeletal fibers are more compactly distributed (cell (k), Nl/Ac 1.539 μm− 2; cell (l), Nl/Ac 2.039 μm− 2).

Fig. 2
Fig. 2The alternative text for this image may have been generated using AI.
Full size image

Cytoskeletal architecture of cells with different morphologies. (a) For each cytoskeletal feature, illustrative examples of cells are displayed following α-tubulin staining. Graphs show the quantification for the respective cells. Line segments and graph edges are coloured according to the 2D feature, except for radiality, in which the result represents the probability of lines showing \(\alpha \: \in \:\left[ {0,\:15} \right]^{\b{o}}\) for all pixels of the segmented cytoskeleton. Yellow dots represent points of higher probability. Standard deviations are shown for morphology and cytoskeleton-nucleus interconnection. For features yielding only a score per cell, no deviations are presented. (b) Comparative analysis of cytoskeletal features in cells with distinct cytoskeletal organizations. OOP, Orientational order parameter; \(\:{D}_{CN}^{W}\), Weighted nucleus-cytoskeleton centroid distance; RS, Highest radial score; FD, Fractal dimension; \(\:{N}_{E}\), Number of edges (branches); \(\:{R}_{\left(e,e\right)}\), Ratio of endpoint-to-endpoint branches; \(\:{N}_{l}/{A}_{C}\), Number of lines-cytoskeletal area ratio.

Since the geometry of fiber assembly is known to dictate cell behaviour17, we have expanded our analysis to radiality relative to the respective nucleus centroid. No evident radial distribution was detected in cell (m), a small and round cell, as confirmed by the lower radial score (RS, 0.266). Intermediate values were observed for cell (n) (0.302) with fibers nucleating from the centre of the cell, whereas a more prominent radial pattern is detected in cell (o) (0.564), in which fibers nucleate from the centroid of the nucleus.

Taking advantage of our pipeline, we were also able to study the interconnection between the cytoskeleton and the nucleus, based on the principle that cell migration depends on the tight coordination of both cellular structures. For this purpose, distances between fiber and nucleus centroids were measured, revealing that cytoskeletal fibers present variable distances to nuclei centroids. In general, small cells, with fibers close to the nucleus, show shorter fiber-nucleus average distance (cell (p), Di 3.14 μm), when compared to cells with extended cytoplasm and, as expected, longer distances (cell (q), Di 7.73 μm; cell (r), Di 15.94 μm).

In addition to the above mentioned, we have examined fiber bundling, parallelism, connectivity and complexity, although such features were less effective in discriminating cells with different phenotypes (Supplementary Fig. 1). Figure 2b highlights the performance of selected parameters in discriminating similarities and differences between cells. For instance, similar fractal dimension (FD) values were observed representing comparable complexity among cells, while orientation is unique for each cell, with distinctive OOP values. Overall, these results demonstrate that our pipeline is able to attain a multitude of metrics, which is required for a comprehensive perspective of the cytoskeletal architecture.

Cells with disrupted E-cadherin and invasive potential have distinct cytoskeletal architecture

Following the analysis of single cells, we have further validated the applicability of our framework by performing a global analysis in non-invasive and invasive cells. In particular, we have explored a well-established model of cells expressing E-cadherin – a critical invasion suppressor – or a deleterious variant of the protein, leading to loss of cell-cell adhesion and an invasive phenotype. The selected variant leads to the amino acid deletion p.L13_L15del, impacting both E-cadherin expression and function, and was identified in a family with hereditary diffuse gastric cancer18,19. Importantly, cells were grown in laminin, which provides a supportive environment for cell growth.

As shown in Fig. 3a, cells expressing wild-type or mutant E-cadherin have distinct cytoskeletal organization. Mutant cells exhibited significantly lower OOP values, as a result of the broader fiber angular distribution and of a more disorganized fiber pattern (Fig. 3b, p < 0.001). This is in contrast to the higher OOP values observed in wild-type cells, indicative of well aligned fibers. Consistent with this, evaluation of Circular Variance, as an alternative orientation parameter, also shows a significant increase suggestive of highly disperse orientations.

Fig. 3
Fig. 3The alternative text for this image may have been generated using AI.
Full size image

Microtubule organization of invasive and non-invasive cells. (a) Representative images of cells transfected with wild-type and a mutant form of E-cadherin. (b) Features evaluating fiber organization include orientational order parameter and circular variance. (c) Morphology was assessed through estimation of line segment and branch length, as well as branch intercellular dispersion. (d) Compactness, (e) radiality and (f) quantity of fibers are displayed for mutant and wild-type contexts. (g) Attributes such as the distance between nuclei and fiber centroids, and ratio nuclear/cytoskeletal area are shown for invasive and non-invasive cells. For each condition, at least 150 cells were analysed. Graphs represent median with inter-quartile range 25-75%. ∗∗, p ≤ 0.01 and ∗∗∗∗, p ≤ 0.0001. WT, non-invasive wild-type cells; M, invasive mutant cells.

To assess cytoskeletal morphology, fiber size was investigated through several metrics based either on LSFs and CNFs, considering fibers as line segments or graphs, respectively. As determined by line segment and branch lengths, a reduction was detected in fiber length in mutant cells, when compared with wild-type cells (Fig. 3c, p < 0.0001). Interestingly, the variability in fiber length was significantly lower in mutant cells, suggesting a more uniform fiber size than that of wild-type cells (branch intracellular dispersion, p < 0.0001).

Concerning fiber abundance, its mean number per cell was found to be significantly lower in mutant cells (p < 0.0001). In particular, the number of line segments and the number of branches were respectively 166.9 and 187.3 for wild-type cells, and 104.4 and 115.8 for mutants (Fig. 3f). A similar tendency was observed for fiber intensity. A better portrait of fiber organization was achieved by considering spatial distribution. We have found that cytoskeletal fibers of mutant cells are more densely packed and distributed in a less radial manner than those of wild-type cells. Remarkably, a lower distance between nuclei centroids and fibers centroids, as well as a marked increase in the ratio nuclear-cytoskeletal area was observed in mutant cells, which could partially explain higher fiber density (Fig. 3d-g).

Taken together, these results demonstrate that invasive cells with E-cadherin dysfunction have compact but highly disorganized fiber orientations, which might be advantageous for the plasticity and fast spreading of cancer cells through the ECM.

Discussion

The cytoskeleton is a highly dynamic structure known to orchestrate a myriad of cellular processes. However, how the cytoskeleton is modulated conferring cells an increased adaptability to the microenvironment is far from understood and, despite technological advances, evaluation of cytoskeletal remodelling remains challenging. It is therefore urgent to establish novel methodologies to better understand the organization of this complex fiber network and its impact on cancer cell behaviour.

In this study, we developed a novel computational pipeline to identify biologically relevant architectural patterns of the cytoskeleton and investigate fine-tuned alterations associated with cancer progression. We proposed a model, which encompasses image preprocessing and feature extraction steps followed by an accurate quantitative analysis of cytoskeletal features, ultimately predicting cellular fate (Fig. 4). Specifically, our results have shown that the proposed approach can successfully recognize microtubule organization through a panel of metrics obtained from the corresponding fiber line segments and graph networks. Using this framework, we have identified an architectural pattern in cells with invasive abilities for which microtubules were shown to be shorter, to have disperse orientations and to be more compactly distributed.

Fig. 4
Fig. 4The alternative text for this image may have been generated using AI.
Full size image

Proposed pipeline for the identification of cytoskeleton architectural features. (a) The model involves image processing techniques, namely deconvolution, Gaussian filtering and skeletonization, for precise extraction of line segments and graph networks. (b) Quantitative analysis of cytoskeletal architecture includes ten metrics that are able to capture morphology, abundance, and complexity of fibers at single cell level or in cell populations. Comparative analysis of data can ultimately provide a reliable prediction of cellular response.

Notably, the developed pipeline constitutes a powerful tool to investigate the dynamics of the cytoskeleton in homeostatic and in diverse disease contexts. Aside from cancer, cytoskeletal abnormalities have been reported to contribute to neurodegenerative conditions, schizophrenia, myopathies, or aging20,21,22,23.

Herein, the computation of a panel of ten metrics in a straightforward manner contrasts with most studies on cytoskeletal organization, which are often limited to a few features, such as total signal quantification14,15,24. In addition, most studies take advantage of state-of-the-art software, although these methods are operator-dependent, requiring manual curation of filament edges and lines25,26. These limitations account for the lack of well-established methodologies to reliably characterize the cytoskeletal organization and determine the functional effect of subtle alterations in the cell. To the best of our knowledge, this is the first open access pipeline that can be used for the simultaneous analysis of ten metrics, based on 2D immunofluorescent images of cells stained with cytoskeletal markers. Besides compiling classical organizational features, our algorithm includes parallelism, radiality and cytoskeleton-nucleus interconnection, which represent major added-value metrics. Importantly, although we have focused on α-tubulin fibers, this method can also be applied to other cytoskeletal proteins displaying similar structures, as actin and intermediate filaments.

Over the past decades, fluorescent markers for cytoskeletal proteins have been used to monitor changes in the cytoskeleton, and metrics built according to the research purpose. For instance, in order to evaluate the effect of the microtubule-disrupting drug Nocodazole, microtubule and actin filaments were quantified by Lichtenstein et al., but the recovered information was restricted to total fiber-associated fluorescence, and fiber length and orientation27. More recently, other studies have emerged focusing on alternative parameters related to the geometry of the cytoskeleton. Orientation, parallelness, bundling and density were described in plant cells, though based on Image J software, that still requires intensive operator manipulation28. Using human lung cancer cells, Basu et al. investigated changes during EMT through assessment of the OOP, focusing on the angular distribution of cytoskeletal fibers. This method enabled a paradigm shift by identifying an intermediate state between pure “epithelial” cells and those with a mesenchymal phenotype9. The orientation of vimentin fibers was also shown to be critical for mesenchymal migration in human fibroblasts29.

In addition, numerous reports employed skeletonization following binary segmentation to extract filament networks, breaking them at intersections and using disconnected component algorithms to discriminate individual filaments30. However, this process often generates small fragments, distorting the original shape and leading to the overcounting of filaments due to the creation of extra branches. To mitigate such issues, post-processing strategies have been described. For instance, Alioscha-Perez et al. developed a filament merging strategy based on iteratively connecting short and on fixed-length line segments, according to overlap and alignment criteria31. Zhang et al. merged network paths after skeletonization by applying geometric constraints, concerning similarity, proximity, and continuity30. More recently, Liu et al. employed a modified ResNet to detect filament junctions and endpoints in order to cope with skeletonization artifacts, including junction points transformed into multiple junctions that yield numerous false short filaments15. SOAX has also emerged as a relevant post-processing algorithm for quantification of network structures, and demonstrated to provide accurate data regarding filaments or bundles of filaments32. Nevertheless, these work best for spatially sparse or sufficiently diluted structures, as opposed to our dense cytoskeletal structures. Our strategy addresses this drawback by exploring intensity distribution as a complementary source of information, which is independent of the skeletonization process. A major advantage of the present approach is its applicability to 2D images, which are used in most fluorescent microscopy techniques to trace specific molecules and intracellular mechanisms. To obtain a more detailed portrait of the structures, 3D images would be of great interest, although they increase the level of complexity and are computationally more demanding.

The potential of our pipeline lies on the possibility to integrate additional cytoskeletal parameters and to perform comparative analysis of either single cells or cell populations. Overall, in this study, we were able to comprehensively analyse the cytoskeletal architecture at the individual cell level and at the global level. It is plausible to foresee that multiple features, rather than a single one, will allow a faithful characterization of cytoskeletal patterns, while generating a phenotypic signature associated with particular cellular mechanisms. As determined for this dataset, the above mentioned organizational features can be used in future investigations to identify potentially invasive cancer cells or molecular targets, with impact in cancer research and clinical applications.

Materials and methods

Cell culture

Chinese Hamster Ovary (CHO) cells were obtained from the American Type Culture Collection (ATCC). Briefly, cells were grown in a humidified incubator at 37 °C, 5% CO2 in α-MEM (+) medium (Gibco, Invitrogen) supplemented with 10% fetal bovine serum (Hyclone) and 1% penicillin/streptomycin (Gibco, Invitrogen). Cells were cultured on 6-well plates and transiently transfected with vectors encoding either the wild-type E-cadherin or the c.38_46del variant (p.L13_L15del), as previously described19. Cells were then cultured on Corning™ BioCoat™ Poly-D-Lysine/Laminin 8-well culture slides (Corning) and grown until 70–80% confluency was reached.

Immunofluorescence and image acquisition

Cells cultured on culture slides were fixed with 4% paraformaldehyde for 20 min. Following a 10 min wash in phosphate buffered saline (PBS), cells were permeabilized with 0.1% Triton X-100 in PBS for 15 min at room temperature. Cells were blocked with 3% bovine serum albumin (BSA) in PBS and stained overnight at 4 °C with α-tubulin rabbit primary antibody (Invitrogen, #PA5-16891). Subsequently, Alexa Fluor 594 goat anti-rabbit (Invitrogen) was applied for 1 h in the dark. Nuclei were stained with DAPI (Sigma-Aldrich, 0.1 µg/ml in PBS) for 15 min and coverslips were mounted on slides using Vectashield medium (Vector Laboratories). Images were acquired on a Carl Zeiss Apotome Axiovert 200 M Fluorescence Microscope (Carl Zeiss, Jena, Germany) with a 40x objective (Plan-Apochromat 40x/1.3 Oil DIC (UV) VIS-IR M27), using an Axiocam HRm camera and the Zeiss Axion Vision 4.8 software. For DAPI and α-tubulin channels, multiple images were acquired along the Z axis (10 Z-stacks). Images were saved with a resolution of 16-bit and pixel width equal to 0.16 μm.

Image preprocessing

RGB immunofluorescence images were first deconvoluted using Deconvolution Express from the Huygens Software (Scientific Volume Imaging) and Maximum Intensity Projection (MIP) was applied to obtain single 2D grayscale images of cytoskeletal structures and nuclei. Preprocessing of the nuclei channel included a binarization step based on inspection-tuned Otsu threshold. This threshold was slightly overestimated to improve distinction of adjacent nuclei and each resulting mask was dilated with a 5 × 5 disk structuring element. Finally, the intensity weighted centroids (\(\:{C}_{N}\) and \(\:{C}_{N}^{W}\)), area (\(\:{A}_{N}\)), and delimiting contours of masks were obtained. Centroids/contours associated with masks with an area below 600 pixels were considered noise and thus discarded.

Deconvoluted cytoskeletal images were subjected to the preprocessing filters Gaussian, Sato and Hessian, followed by skeletonization of the binary mask. Image preprocessing was performed using Python 3.6, and the parameters used for tuning are displayed in Supplementary Table 1. Of note, overlapped cells, cells with unclear structures, or with missegmented nuclei were manually segmented. For this purpose, the Python module ROIPoly (https://github.com/jdoepfert/roipoly.py) was used to define polygonal ROI, producing individual binary masks for single cells.

Processing and deconvoluted cell features (DCF)

After segmentation, the deconvoluted cytoskeletal structure was used to extract DCF, while its skeletonized version was used to obtain LSF and CNF employing different processing strategies. For feature extraction from a 2D deconvoluted image, the area \(\:{A}_{C}\) relative to the cytoskeleton was estimated with hysteresis thresholding33. The resulting area \(\:{A}_{C}\) of the threshold mask was subsequently used for computation of \(\:{N}_{L}/{A}_{C}\) and \(\:{A}_{N}/{A}_{C}\). Within this segmentation mask, intensity distribution-based features were obtained (Table 1 and Supplementary Materials). For each condition, at least 150 cells were analysed.

Line segment features (LSF)

To extract LSFs, the Python library Skan designed for skeleton images34 was used to individualize cytoskeletal filaments, allocating them into sets of adjacent pixel coordinates. Line segments were detected automatically by fitting each filament to line segments through an Adapted Linear Regression (ALR) method. For a given cell with \(\:{\text{N}}_{\text{L}}\) lines detected, this method outputs a set \(\:L=\left\{\left(\left({x}_{i}^{1},{y}_{i}^{1}\right),\left({x}_{i}^{2},{y}_{i}^{2}\right)\right):i\in\:\{1,\dots\:,{N}_{L}\}\right\}\) where each element \(\:i\) corresponds to the two extremities (1 and 2) of the detected line. The information provided by the 2D spatial distribution of line segments, enabled the quantification of several properties, such as local and global distribution of orientations, alignments, lengths, and positions. Some features required the nucleus centroid \(\:\left({x}_{N},{y}_{N}\right)\) as a reference point within the cell. For radiality, all the \(\:{\upalpha\:}\) angles were considered, \(\:{\upalpha\:}=\{{{\upalpha\:}}_{i}:i\in\:\{1,\dots\:,{N}_{l}\left\}\right\}\) and quantification was achieved through \(\:R{S}_{x,y}\), defined as the probability of finding a line segment with \(\:{{\upalpha\:}}_{i}\in\:{\left[\text{0,15}\right]}^{\circ\:}\). In this case, \(\:RS\) is defined at the pixel \(\:\left({x}^{*},{y}^{*}\right)\) that maximizes \(\:R{S}_{x,y}\), \(\:RS\).

Cytoskeleton network features (CNF)

To evaluate CNFs, nodes and edges were identified using the 4-connected Von Neumann neighbourhood method, in which skeleton pixels that touch another pixel’s edges are considered connected and part of the same skeleton object35. This algorithm, included in Skan, generates a set of vertices connected by 1-pixel wide edges. Each undirected network is an ordered tuple, \(\:G=\left(V,E,\:I\right)\), where V and E are finite sets, termed vertices and edges of G, respectively. Skan outputs each edge of \(\:G\), \(\:{e}_{i}\), which is a sequence of pixel coordinates encoded in a vector as defined in Supplementary data.

Statistical analysis

Two-tailed unpaired \(\:t\)-test was performed to determine statistically significant differences. In all statistical tests, results are shown as: ns, not significant; *, p \(\:\le\:\) 0.05; **, p \(\:\le\:\) 0.01; ***, p \(\:\le\:\) 0.001 and ****, p \(\:\le\:\) 0.0001.