Deep learning reveals a damage signalling hierarchy that coordinates different cell behaviours driving wound re-epithelialisation

Tissue damage triggers a complex series of overlapping cell and tissue movements that recapitulate embryonic morphogenetic episodes, which together will stave off infection and eventually repair the wound back to something approaching its pre-wounded state (Gurtner et al., 2008; Eming et al., 2014; Peña and Martin, 2024). One of the key tissue movements contributing to wound healing is re-epithelialisation, whereby the epidermal cells at the cut wound edge and the sheet of epithelium behind them advance to seal the wound gap. Individual and concerted cell contractions and shape changes, as well as cell movements and cell divisions, all contribute to the closure of the epithelial gap (Park et al., 2017; Aragona et al., 2017; Razzell et al., 2014; Tetley et al., 2019; Turley et al., 2024a) but how much each of these cell behaviours contributes to the ultimate goal is not known; neither do we know precisely which signals of those known to be activated soon after tissue damage (e.g. the wound calcium wave, JNK signalling and signals released by wound inflammatory cells) might regulate each cell behaviour, nor how any of these cell behaviours might be compensated for if one or more of the others fail.

In recent years, there have been several elegant live imaging studies of wound re-epithelialisation in mouse models using sophisticated state-of-the-art imaging strategies to enable the capture of cell movements and cell divisions within the advancing epidermis. One of these studies targeted wounds in the murine tail, where the epidermis is devoid of distracting hairs (Aragona et al., 2017). This study provided definitive evidence that the front rows of wound edge cells migrate while those further back proliferate (Aragona et al., 2017). Another study, this time using the thin hairless epidermis of the ear, showed the first high-resolution imaging of an advancing wound epidermis. This too indicated zones of migration at the leading edge and proliferation further back, and was of sufficiently high resolution to enable tracking of fluorescently labelled epidermal nuclei, and to quantify both migratory speed and rates of proliferation (Park et al., 2017).

Because tissues in mouse and human are opaque, any live-imaging approaches are, by definition, technically extremely challenging. Before the studies described above, it had not been feasible to directly observe mammalian re-epithelialisation, and even these pioneering new studies provide rather limited volumes of data and offer less than optimal spatial and temporal resolution. As a complement to murine approaches, several recent studies have used the translucency and genetic tractability of Drosophila embryos, larvae and pupae to investigate wound re-epithelialisation (Razzell et al., 2014; Tetley et al., 2019; Tsai et al., 2018; Turley et al., 2024a; Wu et al., 2009). These studies have been useful in revealing the contributions not only of leading edge epithelial cells, where a contractile actomyosin purse-string helps draw these cells forward, but also of cells behind the leading edge, where tissue fluidity – driven by cells jostling within the sheet – is pivotal for releasing tissue tension within the ‘following’ epithelial sheet, and thus enabling the front row cells to advance forward (Razzell et al., 2014; Tetley et al., 2019).

Collecting large movie datasets of healing wounds is more achievable in translucent models such as Drosophila and zebrafish, but requires new methods for analysis of the data (Turley et al., 2022). Indeed, recently, we and others have used deep learning algorithms (U-NetCellDivision) to precisely quantify when and where cells are dividing within unwounded tissues and in the advancing wound epidermis within the Drosophila pupa (Turley et al., 2024a; Villars et al., 2023). In our testing dataset, we counted the number of correctly detected divisions (true positives) and the errors (false positives and false negatives) to compute a metric to measure the accuracy of the model, defined as the F1 score, which ranges from 0 to 1. Our model achieved a high score of 0.964 (Turley et al., 2024a). We used another deep-learning algorithm (U-NetBoundary) to segment the cell boundaries, enabling us to reveal how daughter cells shuffle post-division to align with the global tension in the tissue (Turley et al., 2024a). The F1 score for the cell boundary model is again high at 0.931. These models allow us to quickly and accurately quantify cell behaviours on a large scale, as required for the current study.

Here, we have used deep learning models to integrate this cell division data with the two other cell behaviours contributing to wound re-epithelialisation: cell shape changes and cell motility. We achieved this by using additional deep learning algorithms to detect cell divisions and the shapes of cells in the wound epithelium; in parallel, we also quantified cell migration by tracking the nuclei of cells in time-lapse movies using a single-particle tracking algorithm (Turley et al., 2024a).

Ultimately, we would like to determine the contributions of each cell behaviour to the wound closure effort and how each of these cell behaviours is regulated by conserved wound-induced signals. To that end, we have analysed large datasets of the three cell behaviours extracted from live time-lapse confocal microscopy of wounded Drosophila pupae. We first characterised the typical re-epithelialisation of wounds in wild-type tissue, before genetically manipulating several well-documented wound healing signals – the calcium wave [the first damage signal released after wounding (Razzell et al., 2013; Sipka et al., 2021; Tu et al., 2019)], JNK signalling (Hammouda et al., 2020; Weavers et al., 2019) and, finally, the innate immune inflammatory response (Thuma et al., 2018; Weavers et al., 2018; Weavers and Martin, 2020) – in order to determine which of these sources of signals might mediate the regulation of each of these cell behaviours.

The wing of the Drosophila pupa is an ideal model for investigating the cell biology of wound healing because it is immobile and optically translucent, enabling long-term, high-resolution imaging, and also because of its considerable genetic tractability (Etournay et al., 2015; Heemskerk and Streichan, 2015; Popović et al., 2017; Weavers and Wood, 2016). Pupae at 18 h after pre-puparium formation (APF) were wounded with a laser after removal of the outer pupal case (Fig. 1A). We generated these lesions in a consistent region of the wing (Fig. 1B) because mechanical stresses inevitably will vary across the developing wing and could influence wound re-epithelialisation (Butler et al., 2009; Etournay et al., 2015; Hufnagel et al., 2007; Osterfield et al., 2013; Paci and Mao, 2021; Popović et al., 2017; Turley et al., 2024a). The wing at this developmental stage consists of two flattened epithelial sheets with intervening haemolymph, in which reside innate immune cells (called haemocytes in Drosophila) and fat body cells (Franz et al., 2018; Fig. 1C). Wounds were generated in pupae expressing E-cadherin-GFP and Histone2Av-RFP (to label epithelial cell boundaries and nuclei, respectively), with a micropoint laser, and subsequently imaged by confocal microscopy with 3D stacks taken every 2 min for 3 h.

We tracked and quantified the three cell behaviours that contribute to wound re-epithelialisation – cell migration, cell shape changes and cell divisions – by means of automated algorithms. We used the plug-in TrackMate to track the nuclei of cells over time in 3D using the Histone2Av-RFP channel data (Tinevez et al., 2017). Cell migration was measured by averaging the radial component of the velocity of the nucleus of the cell at given times after wounding and at different distances from the leading edge (Fig. 1D). Next, we applied the first of our deep learning models, U-NetBoundary (Turley et al., 2024a), to detect cell boundaries (from the E-cadherin-GFP channel), and thus identify individual cells (Fig. 1E). Cell elongation relative to the wound was calculated by computing individual q-tensors for each cell from their cell boundaries (Olenik et al., 2023). These q-tensors are traceless matrices that comprise a component for cell elongation and another for cell orientation (see Materials and Methods for further details). A heatmap of cell elongation relative to the wound indicates which cells are elongated perpendicular to (pointing towards) the wound edge (in red) and those aligned along the wound margin (in blue) (Fig. 1B′). To quantify the shape of cells adjacent to the wound, we averaged the q-tensors of the cells, binning them into groups of cells as a function of distance from the wound, and the time after wounding.

Similarly, we detected cell divisions within the wounded epithelium using a deep learning model called U-NetCellDivision (Turley et al., 2024a), which uses dynamic cell boundary and nuclear information from the GFP and RFP channels (Fig. 1F). We binned these dividing cells into radial bands (annuli) extending out from the wound edge, and for various time points post-wounding, to determine division density for each tissue annulus back from the wound edge.

To determine how wound size impacts the cell biology of repair, we generated wounds of two different sizes; large wounds had an initial area between 700 and 1100 µm², and small wounds were between 200 and 400 µm² (Turley et al., 2024a). The large wounds took, on average, 50 min to close to 20% of their initial size, and small wounds healed the same amount in 25 min (Fig. 2A). At this late stage of wound closure, both wound sizes have leading edge cells that make contact with one another to seal the wound closed (Wood et al., 2002). This last part of wound closure is visually very noisy, and the leading edge becomes unclear. Hence, we define a wound as closed at this point, as the majority of re-epithelialisation has occurred (see Fig. 2B-B″ for time course of a large wound healing and Movies 1 and 2 for repair of small and large wounds, respectively).

We next quantified the spatiotemporal distribution of each of the three contributing cell behaviours during wound closure. Our heatmaps indicate the weighted mean velocity of cell nuclei towards the wound as a function of distance from the wound edge and time after wounding, with the weight corresponding to the visible area of tissue in our videos. Red highlights cells migrating towards the wound and blue highlights cells moving away from it (Fig. 2C,F). For small wounds, cell rows extending up to 50 µm (or 10 or more rows of cells) behind the leading edge, migrate toward the wound; this migration persists throughout the 25 min of healing time. For larger wounds, cells migrate faster and for a longer time, which is as expected due to the longer distance needed to be travelled for effective closure of these wounds (Fig. 2C,F). Interestingly, two distinct phases of migration are observed during the closure of large wounds: for the first ∼30 min post-wounding, cells migrate rapidly but, subsequently, migration slows over the next ∼15 min as the wound closes (Fig. 2C,F). Additionally, we observe a gradient of cell velocity, with cells migrating faster the further back they are from the wound. This would suggest an increase in cell density in this region of the tissue, which we also measured in Fig. S1D,E, showing the predicted density rise.

Next, we examined the cell shape changes during wound closure. In the cell shape heatmaps we calculated the mean q-tensor of the cells for each distance from wound and time after wounding. This was used to determine cell elongation relative to the wound, with cells elongating towards the wound centre indicated by red in the heatmap, those elongating along the wound edge highlighted in blue, and those in white having the same shape as the tissue average (Fig. 2D,G, and see Materials and Methods for details). The q-tensor is a dimensionless quantity, and we have normalised its value using the maximum absolute value of cell elongation around large wounds. For small wounds, cells at the leading edge, and one row back from this edge, rapidly elongated perpendicular to the wound margin (Fig. 2D); these cells subsequently became rounder and were no longer elongated after 10 min; this shape reversion contributes to wound closure. In large wounds too, cell elongation occurred in leading edge cells after wounding, but here, cell elongation propagated back much further into the neighbouring tissue, with cells up to 30 µm (approx. 7-8 rows) back from the leading edge affected (Fig. 2G). Similar to small wounds, cells rapidly reverted back to a rounder shape such that by 25 mins post wounding the mean cell shape had returned to the tissue average (Fig. 2G). This initial period of cell shape change aligns with the early ‘fast phase’ of cell migration (Fig. 2F).

The third cell behaviour contributing to wound re-epithelialisation is cell division. We quantified the density of proliferation around wounds as a function of space and time, and these data are presented as heatmaps in (Fig. 2E,H and Turley et al., 2024a). During the initial period, when cells are rapidly migrating and changing their shapes, we see a clear suppression of cell division in the wound edge epithelium, reaching a minimum at 60 min. This suggests that cell proliferation and the other contributing wound cell behaviours might be mutually exclusive, which is also suggested in studies of mammalian wounds (Aragona et al., 2017; Park et al., 2017). For small wounds, the zone of suppressed cell divisions extends 30 µm back from the wound edge and lasts until 80 min after wounding (Turley et al., 2024a). In larger wounds, this suppression of cell divisions has a similar time course, but extends over a much larger area, stretching back over 100 µm (>25 cell rows) back from the wound, beyond the field of view. After these periods of reduced cell division, we observed a synchronised burst of proliferation beginning 90 min after wounding and lasting for 1 h (Turley et al., 2024a). This occurs in an annulus extending between 20 to 70 µm back from the closed small wound sites and 20 to >100 µm for large wounds (Fig. 2E,H). This surge of proliferation could be a consequence of synchronised wound signals inducing cells to divide to replenish cells lost during wounding and/or cells that would have divided, if not restricted by wounding, becoming synchronised as suppression ends.

To quantify the significance of these changes in cell behaviours, we measured them relative to an unwounded wing. Fig. S1A-C shows heatmaps of cell behaviours around a ‘virtual wound’ in unwounded tissue. This ‘virtual wound’ is a point in the wing around which we quantify the cell behaviours as if it were the centre of an actual wound (see Materials and Methods for details). These heatmaps show uniformity in space and in cell behaviours, with only small fluctuations that are much smaller than effects seen due to wounding. We also quantify the signal-to-noise ratio of the cell behaviours to determine whether these responses are robust across different pupae. Fig. S2 shows heatmaps of the signal-to-noise ratio for each cell behaviour, and for small and large wounds. In all cases, wound healing cell behaviours exhibit a high signal-to-noise ratio well above one, indicating consistency across pupae.

After characterising re-epithelialisation of wounds in wild-type pupal wing tissue, we generated transgenic flies in which key wound healing signals had been perturbed. We began by blocking the well-characterised calcium wave, which is the first damage signal released after tissue injury (Razzell et al., 2013; Weavers et al., 2019). This was achieved by expressing trpm^RNAi to knock down the Trpm calcium channels through which this signal propagates (Razzell et al., 2013); this effectively inhibits the spread of the wound Ca²⁺ wave (Fig. 3D,D′). As ubiquitous expression of trpm^RNAi is partially lethal, we temporally controlled its expression using the Gal4-UAS system with a temperature-sensitive Gal80^ts (McGuire et al., 2004). We kept embryos and larvae at 18°C during early development (to prohibit RNAi expression before pupal stages) and raised the temperature to 29°C once at the pre-pupal stage to activate trpm^RNAi expression.

As flies develop faster at higher temperatures, pupae were raised for 14.75 h at 29°C (instead of 18 h at 25°C) to achieve an equivalent level of development. To confirm the consistency of developmental stage, we used cell elongation as a proxy for wing development (Athilingam et al., 2021; Etournay et al., 2015) and found no significant difference between the temperature shift paradigms (Fig. S3A). Changes in wound healing behaviour could be a consequences of inherent failures in the capacity of the cells to migrate or to change their shape in knockdown flies, and so we first quantified the global mean migration (Fig. S3B,E,H) and q-tenser (Fig. S3C,F,I) in unwounded tissues and found minimal differences between wild-type tissues and all the genetically perturbed tissues we studied. For cell division density, we found some deviation during development between wild-type and knockdown tissues (Fig. S3D,G,J) and here we corrected for these underlying differences in our comparisons of wild-type versus knockdown wounds (see Materials and Methods for details).

We first quantified the area of wounds over time by comparing control (which had been raised with the same temperature shifts) and trpm^RNAi fly lines, and observed a slight reduction in the healing rate (Fig. 3E; Movie 3). We went on to investigate whether this gross defect in healing might be linked to any particular changes in the contributing cell behaviour(s). The control (14.75 h APF) pupae exhibited a typical wound response with regards to cell migration and shape changes, and a clear (delayed) burst of divisions back from the leading edge (Fig. 3A-C). However, all three cell behaviours were altered in trpm^RNAi expressing pupae (Fig. 3F-H). Epithelial cell migration at the wound edge was initially slower for 30 min after wounding, but migration subsequently persisted for longer than in control wounds, perhaps to compensate for the initially slow start (Fig. 3I). Cell shape changes were also altered, with cells at the leading edge remaining highly elongated for a more extended period (Fig. 3G,J). These two changes in cell behaviour could explain the slower healing of wounds lacking a fully functional calcium wave (Fig. 3E). We also observed that the synchronised burst of proliferation present in wild-type wounds is considerably reduced to approximately half the number of dividing cells in wounded trpm^RNAi pupae (Fig. 3H,K). For all these genetic perturbation experiments, we measured the effect on large wounds, as they generally displayed a more extensive cell behavioural response (see earlier) and we reasoned that any differences after genetic manipulation might be clearer.

Next, we disrupted JNK signalling, which is known to be key in developmental morphogenetic episodes such as dorsal closure (Tafesh-Edwards and Eleftherianos, 2020), and in wound healing, both in the formation of actomyosin cables and for associated cell shape changes (Kwon et al., 2010; Lee et al., 2019).

JNK signalling was blocked by expressing a dominant-negative version of Drosophila JNK (termed basket, Bsk^DN) in the pupal epithelium, as previously used for inhibiting JNK signalling (Lee et al., 2019; Tafesh-Edwards and Eleftherianos, 2020; Weavers et al., 2019). We showed that this knockdown was effective by live imaging JNK activity via TRE-DsRed, a transgenic reporter that indicates active JNK signalling, around control and Bsk^DN laser wounds (Fig. 4A,A′). JNK signalling is known to be downstream of the Ca²⁺ wave (Weavers et al., 2019) and so, unsurprisingly, we see a similar, but not identical, effect on aspects of wound closure. Again, there was a reduction in the overall rate of wound closure, with wounds taking 80 min to close to 20%, which is 30 min more than for wild-type pupal wounds (Fig. 4B; Movie 4). However, although our automated tracking data revealed that rates of cell migration in Bsk^DN pupae were similar to wild-type tissues (Fig. 4C,F), cells in the wound epithelium appeared unable to effectively change their shape (Fig. 4D,G). Cell elongation around the wound margin failed to propagate back through the tissue. Cells at the leading edge did elongate initially, similar to wild-type wounds, but this elongation then took 80 min to resolve, rather than only 25 min in control pupae. This failure of cells to change shape, and thus not contribute to wound closure, could be a major reason for the slower healing in pupae with blocked JNK signalling. We also observed a loss of post-wound suppression of cell divisions after JNK knockdown from 40-100 min, as previously described for larval cell death lesions (Cosolo et al., 2019). Our data indicate that although the timing of the wound-induced cell proliferative burst is similar to controls, the number of dividing cells is reduced (Fig. 4E,H), although to a lesser degree than for the trpm^RNAi pupae.

Last, we ablated the innate immune cells of the flies (termed haemocytes) to block the wound inflammatory response. The inflammatory response is believed, through studies in various model organisms, to both clear away wound debris and also release signals that aid in orchestrating various aspects of wound closure, including wound re-epithelialisation (Antsiferova and Werner, 2012; Weavers and Martin, 2020). Haemocytes were ablated in pupae by expressing the pro-apoptotic gene reaper (rpr) specifically in haemocytes using srp-Gal4 in combination with a temperature-sensitive Gal80^ts, similar to that performed previously (Weavers et al., 2016); for this genetic manipulation, the shift to 29°C for 14.75 h (to reach an equivalent developmental stage to 18 h APF at 25°C) was lethal, so instead we maintained pupae at 18°C for 21 h then shifted them to 29°C for 5 h. Once again, we confirmed there were no statistical differences in developmental stage of control versus ‘immune-ablated’ wings (Fig. S3A); Fig. 5A-C shows the cell behaviours of control flies raised using the same temperature regime.

We confirmed that this genetic perturbation successfully killed the pupal haemocytes and so prevented the standard immune response to wounds (Fig. 5D,D′). We found that wounds in these immune-ablated pupae are also slower to heal (Fig. 5E; Movie 5), but the contributing disruptions in cell behaviour are subtly different from those where calcium signalling or JNK signalling are blocked. This time, after immune cell ablation, cell shape changes remained unaffected, whereas cell migration is severely hindered. We observed almost no cell migration for the first 40 min and a slow mean nuclei velocity even after 45 min, by which time control wounds have ceased their cell migration (Fig. 5F,I). In contrast to our studies of repair in JNK knockdown wounds, we observed that cell shape changes were relatively normal, with little difference between control and immune-ablation flies (Fig. 5G,J). Wound-induced cell proliferation in these pupae showed the same spatial temporal pattern to that of control wounds but with a clear overall reduction in divisions (Fig. 5H,K). These results indicate that signals from haemocytes appear to be needed for both epithelial cell migration and for stimulating cell division.

In this study, we have developed automated tools to extract information from high-resolution videos of epithelial wound closure in the Drosophila pupal wing. Our deep learning model takes advantage of both dynamic videos and focused 3D stacks to maximise information input to models for the detection of cell boundaries and nuclei, enabling us to quantify cell shape changes, cell migration and cell divisions that together contribute to wound re-epithelialisation. Using these AI methods allows us to generate large datasets and to rapidly analyse them with minimal human input. Our code and data are publicly available at https://github.com/turleyjm/woundHealing. The current cell boundaries algorithm does not provide perfect segmentation, but has still enabled us to quantify meaningful cell behaviour data from the epithelial tissue surrounding a healing wound.

We quantified the three main epithelial cell behaviours known to contribute to wound healing – cell migration, shape changes and divisions (Aragona et al., 2017; Park et al., 2017) – to determine their contributions to the healing of small and larger wounds. We began by characterising wild-type wound healing. Cell migration and shape changes dominate during the closure of the wounds, while divisions were suppressed initially. Towards the end of healing and after closure, a synchronised burst of divisions occurs, perhaps to regenerate the cells lost during wounding (Turley et al., 2024a). Cells migrated quicker and for longer in large wounds compared with small wounds. At the leading wound edge, cells became highly elongated; this effect is both stronger and propagates further back into the tissue for larger wounds. These cells subsequently changed their shapes, expanding into the wound site to aid in the closure of the wound gap. Only later does cell division resume, with a synchronised burst of proliferation beginning immediately behind the leading edge and extending further from the wound in larger wounds (Turley et al., 2024a).

What damage-induced signals might be regulating these cell behaviours? We have begun to address this systematically in our model, which is amenable to precise quantification of changes in each of these cell behaviours after genetic knockdown of various signalling pathways. We have individually knocked down the first damage signal, i.e. the wound calcium wave (Razzell et al., 2013; Weavers et al., 2019), the immediate early JNK response signal [which transcriptionally activates many signals downstream of tissue damage (Tafesh-Edwards and Eleftherianos, 2020)] and the wound inflammatory response [which is believed to orchestrate many of the tissue responses of wound repair (Antsiferova and Werner, 2012; Weavers and Martin, 2020)]. As JNK signalling and macrophage recruitment to wounds are both known to be, at least partially, downstream of calcium signalling, we expect to see some degree of overlap in the cell behaviour effects of these knockdown studies.

For the first time, we have the tools to test how each of these classes of signals might impact the various contributing wound cell behaviours and indeed compensate for one another. We find that cell migration is slowed when the calcium wave is blocked and is almost stopped in tissues denied a wound inflammatory response, but remains largely unchanged after blocking JNK signalling (Fig. 4K). Cell shape changes were stalled in wounds where the calcium wave was blocked and were almost completely halted in JNK knockdown tissues, but were normal in tissues lacking an inflammatory response (Fig. 6). The synchronised burst of cell divisions in wound epithelium was spatially normal but 50% reduced in tissues denied a calcium wave. JNK knockdown tissues appeared somewhat desensitised to wounds with regard to cell division, with both a smaller initial suppression and then a reduced subsequent burst in divisions; in tissues without an inflammatory response there was also a general dampening in proliferation. Overall, we found that pupal tissues inhibited from experiencing a normal calcium wave displayed a mix of phenotypes reflecting both JNK and inflammation-defective flies with regard to their cell migration and shape changes (Fig. 6). This is consistent with the calcium wave, at least in part, being upstream of both of these signals (Razzell et al., 2013; Weavers et al., 2019).

Potential next steps using this system will be to examine additional sources of signals and to dissect further with regard to the signals delivered by inflammatory cells, by knocking out individual growth factor and/or trophic signals from haemocytes, rather than simply ablating the lineage as we have done here. This approach might allow investigations of which pathways the immune cells are using to influence both migration and divisions of epithelial cells. We also could attempt to target each of the cell behaviours one by one, to not only determine their contribution to wound healing but to also study how the other two behaviours are impacted. Can they compensate for the loss of one behaviour or will they also be negatively impacted? For example, using flies with perturbed cell cycles, we could probe the consequences for wound repair of tissues that cannot suppress cell divisions during wound closure or that fail to trigger a burst of synchronised proliferation.

Other cell behaviours, such as intercalations, which have also been linked to wound healing and development (Etournay et al., 2015; Guirao et al., 2015; Osterfield et al., 2013; Razzell et al., 2014; Tetley et al., 2019), are currently inaccessible from our data due to non-perfect segmentation. Nevertheless, this might be overcome by using dynamic deep learning algorithms similar to those used in our cell division detection model. Changes in cell area and/or density might also play a key role in wound healing, but this is currently difficult to study in the Drosophila pupal wing because of its highly heterogeneous cell density. Future work on tissues such as this will need to measure cell density before wounding and locally track how this evolves during repair. Such studies could ultimately automate measurements of the individual contributions of shear across tissues that might impact a repairing wound.

Applying AI and mathematical tools to biological systems can reveal new results and increase our understanding of underlying mechanisms (Turley et al., 2022). Quantifying properties and behaviours of cells in a tissue on a large scale via automated tools such as deep learning has many applications extending from developmental biology through to cancer. Here, we offer a first insight into what signals regulate each of the cell behaviours that contribute to the evolutionarily conserved process of wound re-epithelialisation. In future, it will be possible to dissect out these signals and how they regulate wound closure in more nuanced ways, and to do similar for the significantly more-complex repair of mammalian skin wounds.

Drosophila stocks were raised and maintained on Iberian food according to standard protocols (Greenspan, 1997). All crosses were performed at 25°C. The following Drosophila stocks were used: E-cadherin-GFP (B#60584), Histone2Av-RFP (B#23651), Act-Gal4 (B#4414), Tub-Gal80^ts (B#7108), UAS-Trpm^RNAi (B#31672), TRE-DsRed (B#59011), UAS-bsk^DN (B#9311), UAS-GCamp7f (a gift from John Gillespie, University of Bristol, UK), Srp-Gal4 (Brückner et al., 2004), UAS-rpr (B#5824) and UAS-nuclear-red-stinger (a gift from Brian Stramer, King's College London, UK; Barolo et al., 2004). Unless otherwise stated, Drosophila mutants and transgenic lines were obtained from the Bloomington Stock Centre.

Drosophila pupae were aged to 18 h APF at 25°C unless stated otherwise. Dissection, imaging and wounding were all performed as previously described (Weavers et al., 2018). The time-lapse movies were generated using a SP8 Leica confocal microscope. Each z-stack slice consisted of a 123.26×123.26 µm image (512×512 pixels) with a slice taken every 0.75 µm. The flies were wounded on a wide-field microscope, which has a nitrogen-pumped micropoint ablation laser (tuned to 435 nm, Andor) attached, before being quickly transferred to the confocal microscope (Turley et al., 2024a).

For the genetic perturbation experiments, we use the Gal4-UAS system to explore the role of these genes in wound healing (McGuire et al., 2004). As many of these genetic manipulations are key to development and are lethal or partially lethal, we used the temperature-sensitive Tub-Gal80^ts to control expression of the UAS-driven gene (McGuire et al., 2004). This requires us to maintain the genetically perturbed flies at 18°C during their development before shifting them to 29°C before imaging. For fly lines where the calcium wave or JNK signalling were to be blocked, we shifted flies to 29°C once they had developed to pre-pupae for 14.75 h APF, such that they had developed to the same stage as pupae aged at 18 h APF at 25°C. For immune-ablation fly lines, this was still lethal. Therefore, they were maintained at 18°C for 21 h then shifted to 29°C for 5 h, which ablated the immune cells, and the pupae developed to the same stage used in previous experiments. We compare the results of these genetically perturbed pupae with those of control flies that had undergone the same temperature shift schedule.

We confirmed the effectiveness of our genetic perturbations in various ways: to observe the calcium wave, we imaged a single plane in the epithelium of flies expressing UASGCamp7f (Weavers et al., 2019) with a rapid frame rate of 0.651 ms/frame to catch this flash and spreading wave for up to 1 min post wounding. Loss of JNK signalling in UAS-bsk^DN flies was demonstrated using the downstream reporter TRE-DsRed. A z-stack of the wound was taken 5 h post wounding for both control and knockdown tissues. The effectiveness of our immune cell ablation was seen using flies expressing nuclear-red-stinger in haemocytes, with a z-stack taken 45 min after wounding.

Image analysis was performed using ImageJ and python. The code used to quantify cell behaviours and make figures for paper can be found in our GitHub repository (under databases.py, paper_BiologyWound.py). Videos and datasets used are available at Zenodo (Turley et al., 2024b) (https://doi.org/10.1101/2023.03.20.533343) and https://github.com/turleyjm/woundHealing. Further details on the deep learning algorithms used to extract cell behaviours can be found in our previous publication (Turley et al., 2024a). A description of the methods used is given below.

We developed our own pipeline to analyse the three key cell behaviours involved in wound closure. Many of the main aspects of this data processing have been discussed in our previous publications (Olenik et al., 2023; Turley et al., 2024a). This workflow comprises a series of automated deep learning (and other) algorithms to extract this information with minimal human input. We quantify four properties from the movies: wound area, cell velocity, cell shape and cell divisions. First, a deep learning model called U-NetTissue segments the tissue into tissue and non-tissue binary masks. Non-tissue could be a wound or parts of the tissue above or below our frame of reference. Sometimes minor manual corrections are needed to make the masks more accurate, mostly when wounds have almost closed and images become very noisy due to wound debris and immune cells. These masks of wound closing can be used not only to measure wound area over time but also as our frame of reference to measure cell behaviours contributing to healing.

We measured the cell migration via a single particle tracking algorithm called TrackMate. This is deployed on the 3D stack of cell nuclei (via the Histone2-RFP channel). By tracking these nuclei, we can quantify cell migration. The mean velocity is calculated and subtracted from the individual velocities for each nucleus to remove the general migration of the tissue and only quantify the deviations from this flow. Using the wound mask discussed above, we can find the centre point of the wound and then calculate the radial component of the deviation velocity of cells in the tissue relative to the central point. This is carried out for every cell at every timepoint.

Next, we calculated the elongation and orientation of each cell. This is caried out by focusing the 3D stack and converting it into a 2D image, which will be inputted to a deep learning model called U-NetBoundary, which detects cell boundaries. To maximise information input to our network, we use our own focusing algorithm, which is modified from similar stack focusers (Turley et al., 2024a). Our method returns an RGB image with the most in-focus pixels in green, then pixels above and below in red and blue, respectively. Once the boundaries have been detected, a watershed algorithm is applied via Tissue Analyzer (Etournay et al., 2016; Olenik et al., 2023; Turley et al., 2024a). From the cell boundaries, we calculate the q-tensor for each cell. This dimensionless and traceless second-rank tensor contains information about the elongation and orientation of each cell (Olenik et al., 2023). We compute the q-tensor relative to the centre of the wound and measure the value of δQ⁽¹⁾, which is defined as the mean elongation of cells towards the wound. Below is the derivation of this property:

Last, we detect cell divisions using our final deep learning algorithm U-NetDivision, further details of which were recently published (Turley et al., 2024a). The network can accurately find cell divisions from dynamic 2-channel (Ecadherin-GFP and Histone2-RFP) videos consisting of five frames. The AI algorithm finds both the position and time in which the divisions occur.

Now that we have extracted information about the wound and three cell behaviours from the videos, we can quantify their properties in space and time relative to the wound. First, to calculate division density relative the wound, we calculate the distance from the edge of a wound to the divisions using a distance transform (Fabbri et al., 2008). We then find all the divisions in a band of a given radius and width. To quantify the density of divisions, we divide the number of divisions by the area of the band. Using both the distance transform and our tissue mask, we can calculate the area of the tissue in each band. Once the wound has closed, we can no longer perform a distance transform using the wound edge, so we instead take the centre of the last timepoint before the wound closes as the ‘wound site’ from which we take our distance transform. As the tissue is still developing and moving, we track this point over time using the tracking software TrackMate (Tinevez et al., 2017). By calculating the average velocity of the cells around the wound site, we can track this point and use this as our frame of reference to measure the distance from the wound site.

For quantifying cell migration and shape changes, we use the distance transform to find all the cells in a band of a given radius and width. We then take the mean of either their radial component of the deviation in velocity of cells or cell elongation relative to the wound (δQ⁽¹⁾). As the q-tensor is a dimensionless quantity, for ease of comparison we normalised to the 5th percentile of the absolute value of δQ⁽¹⁾ for large wounds.

We calculated the signal-to-noise ratio for each time and distance from the wound by determining the mean velocity in each of our videos and then dividing this by the standard deviation of the velocity. This was used to make the signal-to-noise ratio heatmaps. The same method was used for cell elongation relative to wound and division density to generate their signal-to-noise ratio heatmaps.

We found that there was a change in the density of cell divisions for unwounded tissues between controls and knockdown flies. Therefore, when comparing wounded controls and knockdown flies, we adjusted for this difference in baseline division density. First, we measured the division density over time in unwounded tissues (Fig. S3D,G,J). This was smoothened using a moving average over three data points to reduce the noise. Taking the density of division heatmaps for the wounded knockdown (e.g. Fig. 3H) and control (e.g. Fig. 3C), we subtracted the moving average of the corresponding unwounded tissue. Subsequently, we compared the differences in these heatmaps to find changes in the wound healing response (e.g. Fig. 3K).

To compare wounded and unwounded data, we generated a ‘virtual wound’ in unwounded tissue by defining a ‘wound site’ in the middle of the tissue in the first frame of the video. This acts in a similar way to a closed wound moving with the mean velocity of the tissue and is a reference point around which we can measure cell behaviours.

We thank members of the Weavers, Martin, Chenchiah and Liverpool groups for helpful discussion. We also thank the Wolfson Bioimaging Facility (Bristol, UK), particularly Stephen Cross, for help setting up pyimagej, for helpful conversations and for sharing useful resources. We thank Adam Grieve for collecting pupae when F.R. was heavily pregnant. We are grateful to the Drosophila research community, FlyBase and the Bloomington Stock Centre (Indiana, USA), for various fly lines and reagents. J.T., I.V.C. and T.B.L. thank the Isaac Newton Institute for Mathematical Sciences, Cambridge, for support and hospitality during the programme ‘New Statistical Physics in Living Matter: Non Equilibrium States Under Adaptive Control’, where some of the work on this paper was undertaken.

The peer review history is available online at https://journals.biologists.com/dev/lookup/doi/10.1242/dev.202943.reviewer-comments.pdf