Cell mixing induced by myc is required for competitive tissue invasion and destruction

Live imaging of myc-driven competition in healthy Drosophila tissues shows that cells expressing higher levels of myc actively mix with the neighbouring cells, which increases the probability of eliminating neighbouring cells. Cells are known to compete with each other: fast-proliferating cells (winners) can cause the demise of slower-proliferating (loser) cells. This process can be triggered by the cancer-associated protein Myc. To investigate interaction between the winners and losers in living organisms, Eduardo Moreno and performed live imaging of myc-driven competition in healthy Drosophila tissues. They find that that cells expressing higher levels of myc actively mix with the neighbouring cells, increasing the probability that they will eliminate the competition. Specifically, cell–cell intercalation is driven by differences in tension at the interface between various cell combinations: winner–winner, winner–loser and loser–loser. At a molecular level, variations in the levels of F-actin at cell junctions due to differential levels of the membrane lipid PIP3 result in tension differences. The outcome is tissue destruction and invasion. Cell–cell intercalation is a well-established phenomenon during development, but these findings suggest that it can also occur in the context of disease. Cell–cell intercalation is used in several developmental processes to shape the normal body plan1. There is no clear evidence that intercalation is involved in pathologies. Here we use the proto-oncogene myc to study a process analogous to early phase of tumour expansion: myc-induced cell competition2,3,4,5,6,7. Cell competition is a conserved mechanism5,6,8,9 driving the elimination of slow-proliferating cells (so-called ‘losers’) by faster-proliferating neighbours (so-called ‘winners’) through apoptosis10 and is important in preventing developmental malformations and maintain tissue fitness11. Here we show, using long-term live imaging of myc-driven competition in the Drosophila pupal notum and in the wing imaginal disc, that the probability of elimination of loser cells correlates with the surface of contact shared with winners. As such, modifying loser–winner interface morphology can modulate the strength of competition. We further show that elimination of loser clones requires winner–loser cell mixing through cell–cell intercalation. Cell mixing is driven by differential growth and the high tension at winner–winner interfaces relative to winner–loser and loser–loser interfaces, which leads to a preferential stabilization of winner–loser contacts and reduction of clone compactness over time. Differences in tension are generated by a relative difference in F-actin levels between loser and winner junctions, induced by differential levels of the membrane lipid phosphatidylinositol (3,4,5)-trisphosphate. Our results establish the first link between cell–cell intercalation induced by a proto-oncogene and how it promotes invasiveness and destruction of healthy tissues.

Cell-cell intercalation is used in several developmental processes to shape the normal body plan 1 . There is no clear evidence that intercalation is involved in pathologies. Here we use the proto-oncogene myc to study a process analogous to early phase of tumour expansion: myc-induced cell competition [2][3][4][5][6][7] . Cell competition is a conserved mechanism 5,6,8,9 driving the elimination of slow-proliferating cells (so-called 'losers') by faster-proliferating neighbours (so-called 'winners') through apoptosis 10 and is important in preventing developmental malformations and maintain tissue fitness 11 . Here we show, using long-term live imaging of mycdriven competition in the Drosophila pupal notum and in the wing imaginal disc, that the probability of elimination of loser cells correlates with the surface of contact shared with winners. As such, modifying loser-winner interface morphology can modulate the strength of competition. We further show that elimination of loser clones requires winner-loser cell mixing through cell-cell intercalation. Cell mixing is driven by differential growth and the high tension at winner-winner interfaces relative to winner-loser and loser-loser interfaces, which leads to a preferential stabilization of winner-loser contacts and reduction of clone compactness over time. Differences in tension are generated by a relative difference in F-actin levels between loser and winner junctions, induced by differential levels of the membrane lipid phosphatidylinositol (3,4,5)-trisphosphate. Our results establish the first link between cell-cell intercalation induced by a proto-oncogene and how it promotes invasiveness and destruction of healthy tissues.
To analyse quantitatively loser cell elimination, we performed longterm live imaging of clones showing a relative decrease of the protooncogene myc in the Drosophila pupal notum (Fig. 1a, b and Supplementary Video 1), a condition known to induce cell competition in the wing disc 3,4 . Every loser cell delamination was counted over 10 h and we calculated the probability of cell elimination for a given surface of contact shared with winner cells (Fig. 1c, d and Methods). We observed a significant increase of the proportion of delamination with winner-loser shared contact, whereas this proportion remained constant for control clones (Supplementary Video 2 and Fig. 1d). The same correlation was observed in ex vivo culture of larval wing disc (Extended Data Fig. 1

and Supplementary Video 3). Cell delamination in the notum was apoptosis dependent (Supplementary Video 4;
UAS-diap1, 1 out of 922 cells delaminated, 4 nota) and expression of flower lose (fwe lose ), a competition-specific marker for loser fate 12 , was necessary and sufficient to drive contact-dependent delamination ( Fig. 1d and Supplementary Videos 5 and 6). Moreover we confirmed 11,12 that contact-dependent death is based on the computation of relative differences of fwe lose between loser cells and their neighbours (Extended Data Fig. 2a-e). Thus, cell delamination in the notum recapitulates features of cell competition 4,12 .
This suggested that winner-loser interface morphology could modulate the probability of eliminating loser clones. Using the wing imaginal disc, we reduced winner-loser contact by inducing adhesionor tension-dependent cell sorting 13 (Fig. 2d) and observed a significant reduction of loser clone elimination (Extended Data Figs 2f, g and 3a-c). This rescue was not driven by a cell-autonomous effect of E-cadherin (E-cad) or active myosin II regulatory light chain (MRLC) on growth, death or cell fitness (Extended Data Fig. 3d-f) but rather by a general diminution of winner-loser contact. Competition is ineffective across the antero-posterior compartment boundary 14 , a frontier that prevents cell mixing through high line tension 15 . Accordingly, there was no increase in death at the antero-posterior boundary in wing discs overexpressing fwe loseA in the anterior compartment ( Fig. 1g, h). However, reducing tension by reducing levels of myosin II heavy chains was sufficient to increase the shared surface of contact between cells of the anterior and posterior compartments (Fig. 1e, f, P , 10 25 ), and induced fwe lose death at the boundary (Fig. 1g, h). Altogether, we concluded that the reduction in surface contact between winners and losers is sufficient to block competition, which explains how compartment boundaries prevent competition.
Loser clones have been reported to fragment more often than controls 14,16 , whereas winner clones show convoluted morphology 14,17 , suggesting that winner-loser mixing is increased during competition. This could affect the outcome of cell competition by increasing the surface shared between losers and winners. We used clone splitting as a readout for loser-winner mixing. Two non-exclusive mechanisms can drive clone splitting: cell death followed by junction rearrangement (Fig. 2a), or junction remodelling and cell-cell intercalation independent of death (Fig. 2b). To assess the contribution of each phenomenon, we systematically counted the proportion of clones fragmented 48 h after clone induction (ACI) ( Fig. 2c; see also Methods and Extended Data Fig. 4a, b). We observed a twofold increase in the frequency of split clones in losers (wild type (WT) in tub-dmyc) versus WT in WT controls ( Fig. 2d and Extended Data Fig. 4c). Overexpressing E-cad or active myosin II was sufficient to prevent loser clone splitting, whereas blocking apoptosis or blocking loser fate by silencing fwe lose did not reduce splitting ( Fig. 2d and Extended Data Fig. 4c). Finally, the proportion of split clones was also increased for winner clones either during myc-driven competition (UAS-myc, UAS-p35; Fig. 2d and Extended Data Fig. 4c) or during Minute-dependent competition 2 (WT clones in M 2/1 background; Fig. 2d and Extended Data Fig.  4c). Altogether, this suggested that winner-loser mixing is increased independently of loser cell death or clone size (Extended Data Fig. 4d) by a factor upstream of fwe, and could be driven by cell-cell intercalation. Accordingly, junction remodelling events leading to disappearance of a loser-loser junction were three times more frequent at loser clone boundaries than control clone boundaries in the pupal notum (Fig. 2e, f and Supplementary Video 7, P , 10 24 ). The rate of junction remodelling was higher in loser-loser junctions and in winner-winner junctions than in winner-loser junctions (Extended Data Fig. 5a, b). The preferential stabilization of winner-loser interfaces should increase the surface of contact between winner and loser cells over time. Accordingly, loser clone compactness in the notum decreased over time whereas it remains constant on average for WT clones in WT background (Fig. 2g, h; P , 10 24 ). Similarly, the compactness of clones in the notum also decreased over time for conditions showing high frequency of clone splitting in the wing disc, whereas clone compactness remained constant for conditions rescuing clone splitting (compare Fig. 2d with Extended Data Fig. 5d, e). Altogether, we concluded that both Minute-and myc-dependent competition increase loser-winner mixing through cell-cell intercalation.
We then asked what could modulate the rate of junction remodelling during competition. The rate of junction remodelling can be cell-autonomously increased by myc (Extended Data Fig. 5c and Supplementary Video 8). Interestingly, downregulation of the tumour suppressor PTEN is also sufficient to increase the rate of junction remodelling 18 through the upregulation of phosphatidylinositol (3,4,5)-trisphosphate (PIP3). We reasoned that differences in PIP3 levels could also modulate junction remodelling during competition. Using a live reporter of PIP3 that could detect modulations of PIP3 in the notum (Extended Data Fig. 6a, b), we observed a significant increase of PIP3 in the apico-lateral membrane of tub-dmyc-tub-dmyc interfaces compared with WT-WT and WT-tub-dmyc interfaces (Fig. 3a, b). Moreover, increasing/reducing Myc levels in a full compartment of the wing disc was sufficient to increase/decrease the levels of phospho-Akt (a downstream target of PIP3 (ref. 19), Fig. 3c), whereas fwe loseA overexpression had no effect (Extended Data Fig. 6c). Similarly, levels of phospho-Akt were relatively higher in WT clones than in the surrounding M 2/1 cells (Extended Data Fig. 6d). Thus differences in PIP3 levels might be responsible for winner-loser mixing. Accordingly, reducing PIP3 levels by overexpressing a PI3 kinase dominant negative (PI3K-DN) or increasing PIP3 levels by knocking down PTEN (UAS-pten RNAi) were both sufficient to induce a high proportion of fragmented clones (Fig. 3d, e) and to reduce clone compactness over time in the notum (Extended Data Fig. 5d, e), whereas increasing PIP3 in loser clones was sufficient to prevent cell mixing (Fig. 3d, e). Moreover, abolishing winner-loser PIP3 differences through larval starvation 20 (Extended Data Fig. 6e-g) prevented loser clone fragmentation (Extended Data Fig. 6h, i), the reduction of clone compactness over time in the notum (Extended Data Fig. 5d, e) and could rescue WT clone elimination in tub-dmyc background (Extended Data Fig. 6j, k). We therefore concluded that differences in PIP3 levels are necessary and sufficient for loser-winner mixing and required for loser cell elimination.
We then asked which downstream effectors of PIP3 could affect junction stability. A relative growth decrease can generate mechanical stress 21,22 that can be released by cell-cell intercalation. Accordingly, growth reduction through Akt downregulation is sufficient to increase clone splitting (Extended Data Fig. 7a, b) and could contribute to loser

LETTER RESEARCH
G2015 Macmillan Publishers Limited. All rights reserved clone splitting. However, Akt is not sufficient to explain winner-loser mixing because, unlike PIP3, increasing Akt had no effect on clone splitting (Extended Data Fig. 7a, b). PIP3 could also modulate junction remodelling through its effect on cytoskeleton 23 and the modulation of intercellular adhesion or tension 1 . We could not detect obvious modifications of E-cad, MRLC or Dachs (another regulator of tension 24 ) in loser cells (Extended Data Fig. 7c-f). However, we observed a significant reduction of F-actin levels and a reduction of actin turnover/ polymerization rate in loser-loser and loser-winner junctions in the notum (Fig. 4a, b, P , 10 25 ; see also Extended Data Fig. 8 and Supplementary Video 9). Similarly, modifying Myc levels in a full wing

RESEARCH LETTER
disc compartment was sufficient to modify actin levels ( Fig. 4c), and F-actin levels were higher in WT clones than M 2/1 cells (Extended Data Fig. 9a). This prompted us to test the role of actin organization in winner-loser mixing. Downregulating the formin Diaphanous (Dia, a filamentous actin polymerization factor 13 ) by RNA interference (RNAi) or by using a hypomorphic mutant was sufficient to obtain a high proportion of fragmented clones (  Fig. 5d, e). This effect was specific to Dia as modulating Arp2/3 complex (a regulator of dendritic actin network 13 ) had no effect on clone splitting ( Fig. 4d and Extended Data Fig. 9b). Thus, impaired filamentous actin organization was necessary and sufficient to drive loser-winner mixing. These actin defects were driven by the differences in PIP3 levels between losers and winners (Extended Data Fig. 10). Thus Dia could be an important regulator of competition through its effect on cell mixing. Overexpression of Dia was indeed sufficient to reduce loser clone elimination significantly (Extended Data Fig. 9d) without affecting Hippo/YAP-TAZ pathway 25 (Extended Data Fig. 9e). Filamentous actin has been associated with tension regulation 13 . We therefore asked whether junction tension was modified in winner and loser junctions. The maximum speed of relaxation of junction after laser nanoablation (which is proportional to tension 26 ) was significantly reduced in loser-loser and winner-loser junctions compared with winner-winner junctions (Fig. 4e, f and Supplementary Video 10). This distribution of tension has been proposed to promote cell mixing 27 . Accordingly, decreasing PIP3 in clones reduced tension both in low-PIP3-low-PIP3 and low-PIP3-normal-PIP3 junctions, whereas overexpressing Dia in loser clones or starvation were both sufficient to abolish differences in tension ( Fig. 4f and Supplementary Video 10), in agreement with their effect on winner-loser mixing and the distribution of F-actin. Thus the lower tension at winner-loser and loser-loser junctions is responsible for winner-loser mixing. Altogether, we concluded that the relative PIP3 decrease in losers increases winner-loser mixing through Akt-dependent differential growth and the modulation of tension through F-actin downregulation in winner-loser and loser-loser junctions (Fig. 4g).
Several modes of tissue invasion by cancer cells have been described 28 , most of them relying on the departure of the tumour cells from the epithelial layer 29 . This study suggests that some oncogenes may also drive tissue destruction and invasion by inducing ectopic cell intercalation between cancerous and healthy cells, and subsequent healthy cell elimination. myc-dependent invasion could be enhanced by other mutations further promoting intercalation (such as PTEN). Stiffness is increased in many tumours 30 , suggesting that healthy cellcancer cell mixing by intercalation might be a general process. Activation of progesterone-sensitive gal4 (gal4 switch) was done by using fly food mixed with RU486 at 1 mg ml 21 or 50 mg ml 21 . Full development occurred in the hormone-containing food (egg laying and larval development). Act,cd2,Gal4 clones were induced with an 8 min heat shock. Generation of Flower loseA ::mcherry knock-in. The Flower knock-in fly was made by genomic engineering 41 . The genomic engineering in ref. 41 is a two-step process consisting of ends-out gene targeting followed by phage integrase phi31mediated DNA integration. A founder knockout line was established with a genomic deletion of the flower locus at position 3L: 15816737-15810028. The knock-in construct was composed of the deleted flower locus with a mCherry fusion after exon 5 (see Extended Data Fig. 2a, specific for loseA isoform). The knock-in construct was done by site-directed mutagenesis to remove the stop codon and add a restriction site, used to insert mCherry after exon 5. The knockout of flower and the knock-in Flower loseA ::mcherry were tested by PCR and sequencing.
The vectors used for generating the FlowerloseA::mcherry were as follows: pGX-attP, knockout vector; pGEM-T, used for the site-directed mutagenesis; pGEattB GMR , knock-in vector.
Dissected wing discs were mounted in Vectashield (Vectorlab) and imaged on a Leica confocal SP2 using a 363 water immersion objective numerical aperture 1.3 or a Leica confocal SP8 using a 363 oil objective with numerical aperture 1.4. Images shown are maximal z-projections containing the adherens junction plane.
Adult eye pictures were taken on a Leica MZFLIII dissecting scope with a Nikon DXM1200 colour camera at the same magnification. Eye area was measured in Fiji 45 . Pupal notum and wing disc live imaging, death probability calculation. Pupae were collected 48 or 72 h ACI and dissected 18-20 h after pupae formation. Pupae were prepared as indicated in ref. 43 and imaged on a confocal spinning disc microscope (Till Photonics) with a 340 oil objective (numerical aperture 1.35) or a point scanning confocal microscope Leica SP8 with a 363 objective (numerical aperture 1.4). Large-field views of the tissue were obtained by tile imaging (6-12 tiled positions). The z-stacks (1 mm per slice) were recorded every 5 min using autofocus at every cycle and every tiled position using E-cad::GFP signal. Videos were recorded in the nota close to the scutellum region in the vicinity of the aDT and pDT macrochaetes. Videos shown are cropped from larger fields of view after maximum projections, correction for bleach (using Fiji) and correction for tissue drift (Stack reg plugin, Fiji).
Every cell delamination event in the clones was tracked over 10 h. We excluded cells in the midline where spontaneous delaminations were occurring 46 . The proportion of apical perimeter shared with winners (sum of the winner-loser junction lengths over the total apical perimeter) was measured 1 h before delamination at the junction plane (E-cad::GFP signal or actin belt visualized with utABD::GFP) using imageJ. We then measured the proportion of apical perimeter shared with winners for all the cell in the clones at time 0 in every video (excluding cells in the midline) manually or by using CellPacking analyser for skeletonization 47 and a home-made macro (Igor Pro Software, Wavemetrics). Probabilities of death were then obtained by dividing the number of delaminating cells by the total number of loser cells in each shared perimeter category. Note that similar correlations were obtained by using the number of junctions shared with winners divided by total number of junctions (not shown).
Ex vivo culture of wing-discs used clone8 media, as indicated in ref. 21, using a point scanning confocal microscope Leica SP8 with a 363 objective (numerical aperture 1.4), 36 h ACI (WT in tub-dmyc). Discs showing rapid drop-down of cell division were excluded from the analysis. Removal of the signal from the peripodal cells and selection of the signal from the junction plane were performed by using a home-made Matlab macro for selective plane projection (inspired by ref. 21). For every x-y pixel, the z-plane with the maximal E-cad::GFP intensity (calculated by summing pixel intensity on a 50 pixel 3 50 pixel square at every plane) was kept and used to retrieve RFP signal in the same plane. The measurement of probability of death was performed as in the notum. Imaging of fwe loseA ::mcherry KI was also performed in ex vivo wing discs using the same projection procedure.
Junction remodelling was manually counted at the interface of the clones. Each event leading to the disappearance of a junction between two RFP-positive cells was counted. Remodelling events were only counted if the new topology was maintained until the end of the video (10 h; Fig. 2f and Extended Data Fig. 5c). The total number of remodelling events was then divided by the total number of junctions analysed. For Extended Data Fig. 5a, we counted every remodelling event occurring over 10 h for single junctions, and calculated the proportion of junctions undergoing a single remodelling event and the probability of undergoing an additional remodelling event (after a first remodelling event). All winnerwinner junctions tracked shared one vertex with a loser cell, whereas loser-loser junctions tracked also shared one vertex with a winner cell. All winner-loser junctions were tracked. Image quantification. Clone fragmentation, clone size and clone compactness. Clones were counted as fragmented when GFP-positive cells were separated by a single GFP-negative cell in the apical area (using E-cad or phalloidin staining) 48 h LETTER RESEARCH ACI in the wing pouch. Discs showing too high a clone density were excluded (.20 clones per wing pouch). This technique accurately evaluated the number of clones that had recently split. The probability of obtaining false positive results was expected to be low because the probability of observing two independent clones separated by a single cell required the initial recombination to occur in two cells separated by a single cell, as well as the absence of division of the cell in-between for the subsequent 48 h (on average, cells should divide more than three times 48 ). Accordingly, we did not observe lower fragmentation when it was evaluated using twin clones (Extended Data Fig. 4a, b) where 23b-gal-positive clones were counted as fragmented when cells were separated by single negative cell, and when the sister clone (23GFP) formed a single continuous group. However, we slightly underestimated the total number of fragmented clones because the proportion of all the split 23b-gal clones was close to 12% (including groups separated by more than one cell) (Extended Data Fig. 4a, b).
Clone size was measured on 9-mm-wide maximum projection containing the apical junction plane in the wing pouch. GFP-positive patches were automatically recognized and segmented with a homemade Fiji macro, and used to measure the average clone area, the density of clones ((number of clones/area of the wing pouch) 3 10,000 5 number of clones per 10,000 mm 2 ) and the total surface of the wing pouch covered by GFP-positive cells (total clone area/area of the wing pouch).
Clone compactness is defined as C 5 4p 3 (area/perimeter 2 ). Clone compactness was measured for every clone in the notum by drawing their contour using Fiji at t 0 and 10 h later. The fold change was calculated as (C 10 h 2 C 0 h )/C 0 h .
Death induction at compartment boundary. Measurement of death induction at the antero-posterior compartment boundary was done by counting the number of TUNEL-positive cells in the wing pouch in the posterior compartment (control compartment) and in an ROI in the anterior compartment (a band having a width of three cells along the antero-posterior compartment boundary). The compartment boundary was detected with Patched staining. The z-projections of the three different genotypes (dpp-gal4/UASfweloseA::HA; zip 2 /zip ebr ; and zip 2 /zip ebr ; dpp-gal4/UASfweloseA::HA) were randomized by assigning random file names, and TUNEL cell counting was performed blindly. The expected random distribution of dead cells in the ROI was calculated by taking the average ratio of surface of the ROI over the total surface of the wing pouch for ten wing discs. The zip 2 /zip ebr larvae were sorted by using a fluorescent balancer chromosome (Cyo, act-GFP, Bloomington).
Intensity measurements. The intensity measurements for fwe loseA ::mcherry were done in ex vivo cultured wing disc on selective plane projections (as described previously). For each cell in the loser clones, we measured the membrane intensity using Fiji (with a width line of six pixels) divided by the average membrane intensity of all the cells measured in the same disc, and measured the proportion of perimeter shared with winner cells.
PIP3 intensity measurements were performed on maximum projections of cell apical area (4 mm) in living notum expressing ubi-tGPH::GFP. Cytoplasmic signal was removed using two subsequent background substractions on Fiji (Rolling Ball radius 200 pixels and 5 pixels). Junction signal was the mean intensity on a line of width 6 pixels (1 pixel 5 0.148 mm). Measurements of utABD::GFP junction intensity were performed similarly on a line of width 6 pixels after a single background subtraction (rolling ball radius 200 pixels).
Measurements of line intensity profile in the wing discs (Extended Data Fig. 10) were performed by using a line of width 100 pixels in the dorsal compartment parallel to the dorso-ventral compartment boundary (1 pixel 5 0.267 mm) on maximal z-projections using Fiji. Intensity profiles were measured for Patched (Ptc) and utABD::GFP or Dia. Position 0 was determined by detecting the minimum of the derivative of Ptc intensity profile (calculated on Igor Pro Software after smoothing of the profile, 300 pixels averaging window), which corresponded to the boundary between the anterior and the posterior compartments. The distance from the most anterior side of the profile to the antero-posterior boundary was normalized to 1. Each intensity profile was divided by its mean. The average intensity profile was then calculated by averaging the normalized intensity values obtained for each disc at the same relative distance to the antero-posterior boundary.
FRAP. The utABD::GFP has been previously used to assess actin dynamics 39,49 . FRAP experiments with utABD::GFP were performed in ex vivo cultured wing disc in dimethylsulfoxide (DMSO; 0.2%, control) and after treatment with Jasplakinolide (2 mM, Life Technologies) using a confocal spinning disc microscope (Till Photonics) and a 360 oil objective (numerical aperture 1.35). Bleaching was done using 100% of the power of a 488 solid-state laser in diffraction-limited ROI after acquiring three time points (one frame per 0.5 s). Recovery was recorded on 30 s. Intensity recoveries were obtained by measuring the mean intensity in an ROI of 15 pixels 3 15 pixels in Fiji (1 pixel 5 0.10 mm) containing the bleached region. Each curve was normalized by the intensity profile of a neighbouring control region (20 pixels 3 20 pixels) to correct for bleaching due to imaging. For each normalized curve, we subtracted the intensity at t 0 (post-bleach) and then divided by the average intensity of the three first points (pre-bleach). The averages of all the recovery curves were then shown. FRAP in the notum was performed similarly 48 h ACI (WT losers in tub-dmyc). Characteristic times of recovery were calculated by fitting the normalized curves with Igor Pro software using the equation I(t) 5 A -B exp(2t/t), where t is the time (in seconds), t is the characteristic time of recovery, A the mobile fraction and A 2 B the initial intensity after bleaching. For comparison between winner-winner and winner-loser junctions in the same cell, we performed simultaneous bleaching and recovery recording in two identical ROIs (one in a winner-winner junction, one in a winner-loser junction of the same cell). Each recovery curve was normalized and fitted as mentioned above. We then calculated the fold change of characteristic time of recovery for each cell: (t w-w 2 t w-l )/t w-w .
The endo-Ecad::GFP FRAP experiments were performed similarly using a Leica SP8 point scanning confocal microscope (363 oil immersion objective, numerical aperture 1.4), using an argon laser for imaging and bleaching (488 nm). Recovery was assessed over 200 s, one frame per second. The average recovery curves were obtained in the same way as for the utABD::GFP FRAP experiments. Laser ablation. Junction laser ablations were performed in the notum using a twophoton infrared laser on a two-photon microscope (two-photon FluoView 1000 Olympus, Center for Microscopy and Image Analysis, University of Zurich) using a 325 water objective (numerical aperture 1.05). Imaging and ablation were performed with 950 nm wavelength, 1.3% power for imaging, 30% power and 300 ms exposition for the ablation scanning along a line perpendicular to the junction. Relaxation of vertices (visualized with E-cad::GFP) was tracked for at least 10 s (one frame per 0.594 s, one frame per 0.891 s for Fig. 4f 'starved'). Positions of vertices were then tracked using CellTrack software 50 and the evolution of vertex distance over time was fitted with an exponential function (A 0 1 A 1 exp(A 2 t)) on the first ten time points after ablation using Igor software. The values of V max were the derivatives of the exponential fit at t 0 . Statistics. All the error bars shown in the figures are s.e.m. or 95% confidence intervals for probabilities of death. Statistical tests were Fisher's exact tests (twosided) for comparison of proportions or Mann-Whitney non-parametric tests for all other experiments, except in Extended Data Fig. 8f where we used a one-sample t-test with 0 as reference value (normality tested with a Shapiro-Wilk test). Difference in variance was always below 2. No statistical methods were used to set sample size. Experiments were not randomized, and we did not analyse experiments blindly, except for the result of Fig. 1h. Every experiment was done at least three independent times. The same control was used for all the quantifications of clone splitting because it was based on three independent experiments and gave reproducible values (for example compare Fig. 2d, WT in WT, and Extended Data Fig. 4a, b).

RESEARCH LETTER
Extended Data Figure 2 | Contact-dependent death is triggered downstream of flower. The transmembrane protein Flower is a central regulator of competition 12 . The fwe lose isoforms (loseA and loseB) are induced downstream of several competition contexts and their expression is necessary for loser elimination and sufficient to drive cell elimination when contacting WT cells 12 . The contact-dependent communication could occur upstream of fwe (for instance by modifying the levels of induction of fwe lose ) or downstream of fwe induction. Several pieces of evidence indicate that it occurs downstream of fwe induction. First, cell death also correlated with shared apical perimeter in clones homogenously expressing fwe loseA (Fig. 1c, red curve). Second, using a knock-in fusion fwe loseA ::mcherry (Extended Data Fig. 2a), we could show that fwe lose induction did not correlate with the surface of contact shared with winners (Extended Data Fig. 2b, c) as previously suggested by in situ experiments for fwe lose (ref. 12). Finally, the probability of elimination of clones overexpressing fwe lose is proportional to the relative differences in fwe lose levels inside and outside the clones (Extended Data Fig. 2d, e). Altogether, this suggested a model where cells can compute the relative differences of fwe lose levels with all their neighbours through an unknown molecular mechanism. a, Schematic of the modified fwe locus (left) and the resulting messenger RNA of the three isoforms (right). Orange rectangles are exons. The 59 and 39 untranslated regions are shown in purple. Exon 5 is specific to each isoform. The red box shows the localization of the mCherry tag at the end of the exon 5 of fwe loseA . Note that the vector backbone was conserved in the knock-in line (white, AmpR). b, Two examples of selective plane z-projection of ex vivo cultured wing discs expressing fwe loseA ::mcherry KI in WT clone in tub-dmyc background (purple) 36 h ACI representative of 12 discs. The clone contour is shown in purple (right). Scale bars, 10 mm. The intensity profile of the white dotted line is shown below. Bottom right, a lateral view of fwe loseA ::mcherry and its accumulation in the apico-lateral region. c, Scatter plot of fwe loseA ::mcherry membrane intensity in loser cells in wing disc (WT in tub-myc, y axis) against the proportion of perimeter shared with winner cells (x axis). One dot represents one cell. Left: adult eye of a WT fly (oregonR), and flies with abnormal eye morphology due to induction of JNK-dependent death in the eyes (eye-specific gal4, GMR-gal4 and UAS-eiger, the fly orthologue of TNF 35,51 ) expressing b-gal (control), diap1 (apoptosis inhibition), E-cad or sqhE20E21 (active MRLC), representative of 30, 34, 37, 29 and 29 adults, respectively. Right: averaged eye surface in pixels (a.u., arbitrary units); n 5 number of flies; error bars, s.e.m. P values, Mann-Whitney tests. f, E-cad and active MRLC do not modify the probability of loser death for a given surface of contact with winners. Probability of loser cell elimination in the pupal notum for a given surface of contact shared with winners in myc-dependent competition (purple, from Fig. 1d), in WT cells in WT background (control, dotted green, from Fig. 1d), or in losers overexpressing E-cad (dotted red) or active MRLC (sqh-E20E21, dotted pink). Statistical tests are Fisher's exact tests performed with mycdependent competition (purple) (NS, P . 0.05). Error bars, 95% confidence intervals.

RESEARCH LETTER
Extended Data Figure 4 | Clone fragmentation does not correlate with clone size. a, Twin clones 48 h ACI marked with two copies of GFP (green) and absence of b-gal or two copies of b-gal (red) and absence of GFP (FRT40A ubi-nlsGFP/FRT40A bcat-bgal). Left: non-fragmented clones. Middle and right: fragmented clones (the GFP sibling clone is used as a reference) with clone cells separated by a single cell (middle) or more than one cell (right). b, Proportion of fragmented clones 48 h ACI in WT GFP clones in WT background (blue, from Fig. 2d) quantified with the one cell distance criteria. Same quantification in FRT40A ubi-nlsGFP/FRT40A bcat-bgal where 23bgal clones were counted as split when clone cells were separated by a single cell and were associated with a continuous group of sibling 23GFP cells. This quantification showed no differences with the WT GFP clones in WT background, demonstrating that our method does not produce false positive results. However, it slightly underestimates the total number of fragmented clones (compare with 'all', where every split 23bgal clone is counted); n 5 number of clones. Statistical tests are Fisher's exact tests performed with WT GFP clones in WT background (blue). c, Wing discs 48 h ACI in control (WT in WT) and in supercompetition assay with loser cells expressing b-gal, UAS-ecad, UAS-sqhE20E21, UAS-diap1 (an endogenous apoptosis inhibitor) or UAS-p35 (a bacterial caspase 3 inhibitor) and fwe lose RNAi; or after induction of winner clones (UAS-p35, UAS-myc in WT, p35 is necessary to block the cell autonomous death induced by high myc overexpression 52 , and WT in M 2/1 where WT clones have no GFP). White arrowheads show fragmented clones. Insets show close-up view of representative clones (see Fig. 2d for number of clones analysed). Scale bars, 100 mm. d, Scatter plot showing the proportion of fragmented clone (y axis) against the average size of clone (x axis) 48 h ACI for all the different genotypes used in this study (see legend). One dot represents one fragmentation assay. There is no correlation between clone size and clone splitting. Pearson correlation coefficient 5 0.14. Note also that without the outlier point (UAS-pten RNAi in WT) the correlation is close to 0 (correlation coefficient 5 20.036). LETTER RESEARCH