GraFT facilitates automated FS tracing and tracking
To develop a robust framework for FS tracing and tracking that can be readily applied to analyse the cytoskeleton in cells, we relied on a network representation of imaging data6,19–21. Network representation of filamentous networks usually place nodes on the intersection of multiple, potential FSs. Since AFSs are highly heterogeneous, bent, and dynamic, we opted to provide a finer network representation by allowing the inclusion of more nodes, other than those where potential FSs intersect. This was achieved by using the Visvalingam-Whyatt algorithm22 (see Material and Methods). After standard image preprocessing steps to enhance filament-like structures and removal of artifacts (see Material and Methods), we obtained undirected weighted network representations of the analysed images (Fig. 1A, B).
FS tracing was performed for each of the connected components (i.e. pieces) of the resulting network (Fig. 1C). To this end, the nodes in each connected component were classified based on their connectivity, given by the number of immediate node neighbours (Fig. 1D). Nodes that are of connectivity different than two (i.e. one or larger than two) served as start or end nodes of putative FSs. To identify FSs, we relied on a constrained depth first search (cDFS): Starting from designated node(s) of degree one (Fig. 1D), cDFS visits a direct neighbour if the angle between the consecutive edges traversed by the cDFS was above a given threshold. GraFT allows that the threshold is user-defined or calculated based on the average angle of adjacent edges in a connected component (Materials and Methods). Whenever cDFS visits a node of degree different from two, it stores the traversed path for further inspection. cDFS is performed until there are no more nodes to be visited, resulting in a list of traversed paths (Fig. 1E). The longest among these paths is then selected as an FS (Fig. 1F). This is particularly relevant for the tracing of AFSs, since long and persistent actin bundles are found in different cells23,24. The process of identifying and removing the longest among the remaining identified paths is in turn repeated until there is no edge included in at least one FS (Fig. 1F). As a result, GraFT first identifies the longest FS followed by finer, shorter FSs until all edges are covered by paths traversed by the cDFS.
If a time-series of images is provided as input, GraFT applies the approach described above to identify FSs in each image. To track FSs in the time series, we cast the tracking as an assignment problem. This involves creating a cost matrix based on the pairs of end-nodes of each FS in two consecutive images (described in Materials and Methods). To generate the cost matrix, distances were calculated between potential FSs in consecutive frames using the spatial position of the two end nodes of each FS in an image. Additionally, unmatched FSs were retained for a certain number of time steps to account for the possibility of FSs moving in and out of focus, with the exact number of time steps calculated from the data.
The output of GraFT includes several properties of the traced and tracked individual FSs, such as: FS length, rod length, intensity (per length), bendiness ratio and angle to the major cell axis as well as movement of tracked FSs (see Materials and Methods). FS length is defined by the number of pixels forming the FS and the image resolution, while rod length is the Euclidean distance between the end nodes of the FS (see Figure S1 A). Bendiness is the ratio between these two measurements, with values of bendiness closer to one indicating less bent FS. Intensity is determined by the sum of the underlying pixels in the original image found within the mask covering the defined FS, and allows the determination of intensity per length. Angles of the FS are determined relative to the major cell axis and are calculated using end nodes (Figure S1 A). We also quantified the movement of tracked FSs over time, based on the average distance of individual nodes and their projections on the tracked FSs between consecutive time points (Figure S1 B).
GraFT outperforms state-of-the-art approaches for FS tracing and tracking
To evaluate the performance of GraFT, we compared it against three contending state-of-the-art approaches, including: TSOAX13, based on SOACs25,26, that allows FS tracing and tracking on image data alone, DeFiNe21, that uses network-based representation (here provided by GraFT) for FS tracing, and FilamentSensor 2.017, which pre-processes the image-data to segment, trace, and track bent FSs. To measure and compare performance of these approaches, we determined the precision of segmentation, tracing, and tracking of in silico generated filamentous networks (Fig. 2A, see Materials and Methods). These data sets included various combinations of noise and blur, to investigate the robustness of the FS tracing, as well as different rates of (dis)appearance of FSs between consecutive images, to evaluate the performance of FS tracking. The simulated noise and blur are representative of typical confocal microscopy data (Fig. 2C,D, left and right sub-panels, see Materials and Methods).
To evaluate the precision of traced FSs, we calculated the ratio of correctly traced FSs and all identified FSs. The segmentation precision was assessed using the Jaccard Index (JI), which measures the similarity between the true segmentation and the predicted segmentation. We did not measure the accuracy of segmentation, as its (rather low) value is affected by the empty space in the image.
Blur and noise are intrinsic to light images, and are expected to affect performance of FS tracing and tracking. Therefore, it is important to investigate the robustness of computational approaches for FS tracing to blur and noise. Our results demonstrated that GraFT and TSOAX exhibited excellent performance in the absence of noise, while GraFT outperformed the other approaches with all other noise levels (Fig. 2C). Further, TSOAX showed high JI for medium noise levels, but exhibited reduced tracing precision. As expected, TSOAX showed poor performance for intermediate to high noise levels since this method lacks pre-processing steps to address noise16. In addition, both DeFiNe and FilamentSensor 2.0 displayed poor FS tracing (Fig. 2C). Further, GraFT and TSOAX both demonstrated high JI for all blur levels, indicating good performance on segmentation (Fig. 2D). However, the performance of TSOAX on FS tracing, was, similar to that of DeFiNe and FilamentSensor 2.0, poor across varying blur levels (Fig. 2D). FilamentSensor 2.0 was unable to trace the closed ring structure, regardless of the noise or blur level, and instead divided it into many FSs. DeFiNe failed to identify all FSs in the images, indicating that the mathematical formulation of DeFiNe is not optimal for tracing FSs that are bent. Therefore, we concluded that GraFT outperformed state-of-the-art tracing approaches of FSs in image data.
To simulate FS movement, we used the same FSs and created a time-series data set allowing (dis)appearance to evaluate the performance of the compared approaches. We found that TSOAX was not successful in FS tracking (Figure S2). More specifically, TSOAX correctly tracked only one FS over two consecutive frames. By contrast, FilamentSensor 2.0 exhibited reliable tracking of FSs as long as they appeared in consecutive frames, with the exception of the ring structure. Nevertheless, FilamentSensor 2.0 was unable to track FSs that disappeared and reappeared among frames. However, GraFT tracked all FSs (Figure S2), and thus outperformed both FilamentSensor 2.0 and TSOAX in terms of tracking capabilities.
To demonstrate the robustness of tracking by GraFT, we created sparse in silico data containing four FSs, with different degrees of curvature. We implemented different FS movement across frames as well as other types of deformations over time. To create a coarse version of this in silico data set, we removed every second image from the original time-series data. The change in convex to concave FS was captured well, as shown by the rising values, plateauing, and then further increasing values (Figure S3 A). Other aspects of FS deformations and movements in fine and coarse data were similarly well addressed by GraFT (see Figure S3 B to D). These in silico analyses and comparisons demonstrated that GraFT captures static and temporal FS properties well, making it suitable for real-world image analyses of filamentous networks.
GraFT captures actin cytoskeleton dynamics across hypocotyl cell development
To investigate changes of the actin cytoskeleton across cell growth and development, we imaged cells of five-day-old etiolated Arabidopsis thaliana (Arabidopsis) seedling stems (hypocotyls), expressing the actin marker mNeonGreen-FABD27. Here, we imaged cells using spinning disk confocal microscopy close to the apical hook at the top of the hypocotyl (before rapid cell expansion; called young), at the middle of the hypocotyl (rapidly expanding cells; called expanding) and at the bottom close to the junction between hypocotyl and root (mature cells; called mature; Fig. 3A-D)28.
We employed GraFT to firstly study network properties of the AFSs in the cell types. Here, we aimed to identify different network properties between the cell types using network assortativity, which is a measure of how heterogenous a network is with respect to degrees of neighboring nodes. A negative assortativity indicates that the network is heterogenous, i.e. consisting of a mix of high degree nodes that link to low degree nodes. By contrast, assortativity close to 0, is a non-assortative or random network, displaying no association between the degrees of nodes and those of their neighbours. We found a shift in assortativity (from random to heterogenous actin network) when moving from young to mature cells of the hypocotyl, where expanding cells had the most heterogenous network (Fig. 3F). We also found that the actin cytoskeleton density was lowest for the expanding cells (Fig. 3G).
We next conducted a detailed spatiotemporal analysis of the individual AFSs. We first divided AFSs into four groups based on FS length, namely, smaller than 3.26 µm (very short), between 3.26 to 6.51 µm (short), 6.51 to 21.7 µm (long), and over 21.7 µm (very long) (see Table S1 for numbers of AFSs in each group). The distribution of AFSs in each group varied between cell types, the young cells had the highest number of FSs in the very short group, whereas mature cells had most FSs in the short group. The mature cells had the lowest number of both very short and very long FSs, indicating that these cells contained mainly FSs of short to intermediate length (Table S1). We found that mean bendiness ratio was significantly reduced for all FS length groups in young cells compared to mature cells. Additionally, the very long group for expanding cells contained straighter FSs than the other two cell types (Fig. 3H). We also found that all groups, except for the one comprising the very long Fs in young cells in comparison to the other cells showed higher mean filament intensity per length, as a measure for bundling (Fig. 3I).
In general, when we compared mean bendiness ratio with mean intensity per length, we found that the very short to short FSs appeared straight and faint, while the group of the longest AFSs consisted of bundled AFSs. Here, we observed that very long bundled AFSs were mainly present in young and expanding cells but largely absent in mature cells (mean length of the group of very long AFS: Upper: 64.6, Middle: 74.3, Bottom: 43.5 µm). Notably, the very long AFSs of the expanding cells appeared most bundled, and most straight when compared to the other cell types. We also observed that very short to long AFSs exhibited higher mean movement than the very long AFSs. This was most apparent for the very long AFSs in expanding cells that appeared nearly static (Fig. 3J). These very long AFSs (in expanding cells) aligned better with the major cell axis than that of very long AFSs of the other cell types (Fig. 3E). We therefore concluded that network properties and dynamic behaviour of FSs can be distinguished between cells at certain developmental stages of seedling stems.
In summary, our analysis showed that expanding cells had substantial differences in AFS properties, with the sparsest network containing long and static AFSs, when compared to the two other cell categories of the hypocotyl. We also found that all AFSs in the very short, short and long groups exhibited considerably more movement than the ones in the very long group, consistent with the observation that short actin filaments are dynamic, while long AFSs are more bundled and rigid. These results were obtained in a fully automated way using GraFT.
GraFT captures relevant actin cytoskeleton reorganization with Latrunculin B treatment
Next, we employed GraFT to investigate whether we can provide detailed analyses of how the actin polymerization inhibitor Latrunculin B (LatB)29 impacts the AFSs. Here, we used time-series from TIRF microscopy (increased axial resolution) of the actin cytoskeleton in Arabidopsis hypocotyl cells containing Lifeact-Venus labelled actin. This set-up allowed us to also showcase the performance of GraFT across microscopy methods, fluorescence intensity and resolution. The experimental procedures followed those described by Ma et al.30. We examined five-day-old Arabidopsis hypocotyl cells (etiolated and light-grown) treated with LatB (0.4 µM or 1 µM for 0–10 min, 10–20 min and 20–30 min) as well as untreated control seedlings (Fig. 4A-H). We used lower LatB concentration for the dark-grown seedlings as we noted that they were more sensitive to the drug.
We used GraFT to explore the network density of the actin cytoskeleton and found that it decreased with increasing LatB treatment time (Fig. 4O), corroborating that LatB inhibits AFS polymerization. Similar to the analysis above, we next divided the AFSs into four groups based on FS length, i.e. smaller than 3.25 µm (very short), between 3.25 to 6.5 µm (short), 6.5 to 19.5 µm (long), and over 19.5 µm (very long). However, we used only three groups for 20–30 min LatB treatment in dark conditions due to insufficient data for statistical analyses. With increasing LatB treatment times, we found a noticeable reduction in the number of filaments in each length category (Table S2), which corresponded well with the reduced network density with treatment times.
We next assessed the AFS dynamics in the different treatments. In light-grown seedlings, the mean AFS movement was notably affected by LatB treatment (Fig. 4I). Every treatment group displayed significant differences, typically reduced mean movement, from the control group across all four length groups. A similar trend, although to a lesser extent, was observed in dark-grown seedlings (Fig. 4L). Notably, the longest AFSs in both light and dark-grown cells exhibited negligible movement as the treatment duration extended. Comparing mean movement of AFSs in dark and light-grown cells, we found that the AFSs generally were more dynamic in dark-grown cells than those in light-grown cells (Control light: 1.4 µm/s, dark: 1.5 µm/s; 0–10 mins light: 0.9 µm/s, dark: 1.2 µm/s; 10–20 mins light: 0.9 µm/s, dark: 1.2 µm/s; 20–30 mins light: 1.2 µm/s, dark: 1.4 µm/s). It is possible that the increased AFS dynamics in dark-grown cells could be attributed to rapid cell elongation or perhaps be a consequence of the differences in LatB concentration.
Not surprisingly, the intensity per length (bundling) increased with longer AFSs across control as well as treatments in light grown cells. However, extended treatment time led to further increase in intensity per length (Fig. 4K), at least in light-grown cells. This indicates that a decrease in density for prolonged LatB treatment resulted in depolymerisation of thinner FSs, indicating that bundled AFS are less affected by the LatB treatment. Interestingly, we did not observe the same behaviour in the AFS intensity of dark-grown seedling cells (Fig. 4N). Here, control cells exhibited the highest intensity per length in the very short and short groups, but this trend was reversed for the two longest groups. The mean bendiness ratio increased with extended treatment times in both light and dark-grown cells, except for the group treated the longest with LatB in light-grown cells, which did not exhibit any significant differences from the control group (Fig. 4J, M). We observed differences between light and dark grown cells for the mean circular angle of the AFSs; however, the treatments did not appear to significantly affect the angles (Fig. 4P). The circular variance for light-grown cells indicated greater variation between cells with treatment, while dark-grown cells did not exhibit any noticeable effect on variation (Fig. 4Q).
To sum up, as the LatB treatment time increased, we observed severe alterations to the actin cytoskeleton, which appeared sparser. Moreover, the AFSs exhibited reduced movement, indicating that AF depolymerization affected both AFS dynamics and bundling, and reduced the number of very long AFSs under prolonged treatment. These effects were more pronounced in light-grown cells, perhaps as a consequence of differences in growth between dark and light-grown cells and in LatB concentrations.
GraFT captures important actin cytoskeleton dynamics and characteristics under virulence factor treatment
The actin cytoskeleton has been implicated in plant pathogen interactions30,31. For example, plant-type-I formin acts as a molecular sensor responsible for actin remodeling in response to the two virulence factors (VFs): Xanthomonas campestris pv. campestris diffusible signal factor (DSF) and pathogen-associated molecular pattern (PAMP) flagellin (flg22)30. Here, we employed GraFT to investigate and quantify the spatiotemporal behaviour of the actin cytoskeleton in response to the two VFs (and only DMSO as control). DSF treatments reduced the actin network density, whereas flg22 had similar actin density, when compared to the control (Fig. 5G), at least in response to the concentrations used here.
To investigate individual AFSs, we again grouped them into the four length categories as per above (refer to Table S3 for group counts). The distribution of AFS lengths varied slightly across different treatments, with a consistent trend of a higher proportion of very short AFSs and a lower proportion of very long AFSs in the treated cells. Notably, over half of the AFSs in DSF treatment were very short, while flg22 treatment mirrored the distribution pattern in the control more closely (see Table S3). The mean movement was significantly reduced in DSF treated cells for short to very long AFS groups compared to control; however, DSF treatment had little impact on mean movement on very short AFSs. By contrast, flg22 treatment had a slightly higher mean movement, with the exception of very long filaments (Fig. 5D), indicating an opposite effect from DSF treatment. When inspecting mean movement of all AFSs, we see that DSF treatments resulted in a general significant slowdown, whereas flg22 had a significant higher movement from control (Control: 0.6, DSF: 0.5, flg22: 0.7 µm/s). The treatments did not affect mean bendiness ratio, except for very short AFS treated with flg22 (Fig. 5E), indicating that the VF molecules did not affect this property of the AFSs much. Both DSF and flg22 treatments altered filament intensity (bundling) (Fig. 5F). Treated cells displayed significantly dimmer AFSs across all length groups when compared to untreated cells, suggesting that both VFs affect bundling and density of FSs. Additionally, the circular mean, portraying angular direction, of AFS in treated cells deviated from control cells, exhibiting increased variance (Fig. 5H, I), indicating that both treatments disrupted network directionality. Previous studies found that treatment with flg22 resulted in a denser actin network compared to control30,32. We did not see the same behaviour; Instead, the cell density was similar to control cells, due to the different image preprocessing and quantification of properties, which may neglect very fine FSs. We stress that in previous studies it was not possible to measure AFS intensities, mean length nor mean movement. Here, we show that the mean movement is affected by flg22 treatment, giving a more active actin cytoskeleton, as well as a drop in AFS bundling. We also see that the mean length of all AFS’s is not the same, flg22 treated cells have on average the longest AFSs, whereas DSF treated cells have the shortest (Control: 5.9 µm, DSF: 5.3 µm, flg22: 6.5 µm), showcasing that flg22 interfere with actin polymerization by causing mean longer FSs.
We conclude that DSF and flg22 treatments had opposite effects on AFS movement, where DSF caused more static FSs, and decreased actin network density, while flg22 caused more active FSs with similar actin network density to that of control. Both treatments interfered with directionality, causing AFSs to orient less well with the major cell axis. It is plausible that this is a direct consequence of how the VFs act, but could also be due to differences in the efficiency of the molecules to access the cells and in how they bind to their respective receptors to impact the actin cytoskeleton.