The Robo4 cytoplasmic domain is dispensable for vascular permeability and neovascularization

Vascular permeability and neovascularization are implicated in many diseases including retinopathies and diabetic wound healing. Robo4 is an endothelial-specific transmembrane receptor that stabilizes the vasculature, as shown in Robo4−/− mice that develop hyperpermeability, but how Robo4 signals remained unclear. Here we show that Robo4 deletion enhances permeability and revascularization in oxygen-induced retinopathy (OIR) and accelerates cutaneous wound healing. To determine Robo4 signalling pathways, we generated transgenic mice expressing a truncated Robo4 lacking the cytoplasmic domain (Robo4ΔCD). Robo4ΔCD expression is sufficient to prevent permeability, and inhibits OIR revascularization and wound healing in Robo4−/− mice. Mechanistically, Robo4 does not affect Slit2 signalling, but Robo4 and Robo4ΔCD counteract Vegfr2-Y949 (Y951 in human VEGFR2) phosphorylation by signalling through the endothelial UNC5B receptor. We conclude that Robo4 inhibits angiogenesis and vessel permeability independently of its cytoplasmic domain, while activating VEGFR2-Y951 via ROBO4 inhibition might accelerate tissue revascularization in retinopathy of prematurity and in diabetic patients.

R OBO4 was identified 15 years ago in a bioinformatics search for endothelial-specific transcripts and initially named Magic Roundabout, to denote its homology to Robo receptors and its selective expression on endothelial cells in actively growing tumour vessels 1 . Robo4 encodes a 140 kDa protein that comprises a 60 kDa extracellular domain (ECD) with two immunoglobulin (Ig)-like domains and two fibronectin-repeat regions, a transmembrane domain and an intracellular domain (ICD) devoid of known catalytic sequences 2 . Robo4 is upregulated in endothelium of embryonic blood vessels and in micro-vessels of patients with tumours and vascular injury, but is also seen in quiescent endothelium [1][2][3][4][5][6][7] . Its endothelial-specific expression is regulated by activator protein-1 (AP-1) binding to a 3 kb fragment of the promoter [7][8][9] . Besides endothelial cells, Robo4 is also expressed in hematopoietic stem cells and regulates their trafficking [10][11][12][13] . The highly endothelial-specific Robo4 expression has prompted considerable interest in this molecule to selectively target or image tumour vascularization 9,14-16 .
Robo4 À / À mice are viable and fertile, demonstrating that Robo4 function is dispensable for developmental angiogenesis 17,18 . Adult Robo4 À / À mice exhibit enhanced corneal neovascularization induced by VEGF implantation in the eyes, and Robo4 knockout breast tissue implanted into pregnant mice develops more vessels when compared with wildtype controls 18,19 , suggesting that Robo4 has anti-angiogenic effects in certain tissue contexts. In addition, Robo4 also maintains vascular barrier function, as shown in Robo4 À / À mice, which have increased VEGF-induced dermal permeability 17,18 . Furthermore Robo4 À / À mice show increased angiogenesis in oxygen-induced retinopathy (OIR) 17 . OIR is widely used to model ocular neovascular disease 20,21 . Neonatal mice are subjected to 75% oxygen exposure for 5 days (P7-P12), resulting in vascular obliteration. After being placed back to room air, OIR retinas enter into a proliferative phase, leading to regrowth of normal vessels (revascularization) and formation of pathological pre-retinal neovascular tufts driven by angiogenic factors. Thus, OIR models can be used to examine vessel loss, physiological revascularization and pathological neovascular tuft formation 20,21 .
How Robo4 signals to maintain vascular barrier function is currently unclear. Initial studies reported that Robo4 acts as a receptor for Slit family proteins (Slit1-3) that regulate crossover of axons at the mid-line of the developing central nervous system 2,17,[26][27][28][29] and also mediate cardiovascular development and angiogenesis 15,[30][31][32][33] . Slit2 binding to Robo4 was proposed to counteract VEGF-driven angiogenesis and vascular permeability by signalling through the Robo4 ICD 17 that can bind to paxillin and other cytoskeletal modifying proteins 2,34,35 . However, Slit2 does not bind to Robo4 in Biacore binding assays, and Slit2 binding to vasculature is unaltered in Robo4 À / À mice 14,18 . In fact, crystal structure analysis showed that the critical Slit-binding residues in the Robo1 and 2 extracellular domains are not conserved in mammalian . Alternatively, Robo4 could affect angiogenesis by modulating Slit2 signalling through Robo1 and 2. A recent study showed that tamoxifen-inducible deletion of Slit2 in postnatal mice leads to severe retinal angiogenesis defects, demonstrating potent proangiogenic functions for Slit2 (ref. 32). Tamoxifen-inducible deletion of Robo2 on a Robo1-null background phenocopies the Slit2 mutant defects, demonstrating that Slit2 provides proangiogenic signals via Robo1 and 2 (ref. 32). These in vivo studies support pro-angiogenic Slit effects previously observed in vitro [40][41][42][43][44][45] . ROBO1 and 4 can be co-immunoprecipitated from HUVEC lysates, suggesting that Robo4 could restrain angiogenesis by acting as a dominant negative regulator of Slit2 signalling through Robo1 and 2 (ref. 35).
A third model of Robo4 action is that it signals through UNC5B, another endothelial-specific guidance receptor [46][47][48][49] . High-throughput screening of 41,500 secreted proteins and ECD domains with a Robo4-ECD-domain fused to the Fc portion of human Immunoglobulin (Robo4-ECD-Fc) identified the UNC5B-ECD-Fc as the only Robo4-binding protein 18 . The Robo4 ECD binds to the UNC5B ECD with a Kd of 12 nM (ref. 18). Robo4-ECD-Fc binds to cells transfected with UNC5B, and Robo4 binding to blood vessels is lost in Unc5B À / À mice, together providing strong evidence for a direct interaction between both receptors. The Robo4 ECD activates signalling through Unc5B, which inhibits Src activation downstream of Vegfr2, thereby attenuating VEGF-mediated sprouting angiogenesis 18 . These data suggest that the Robo4 ECD might be sufficient to mediate Robo4 actions in the vasculature.
To distinguish between the different possible modes of Robo4 action, we generated mice lacking the Robo4 cytoplasmic signalling domain. If this domain was endowed with signalling capacity, mice should develop a phenotype resembling Robo4 knockouts. However, we find that the Robo4 cytoplasmic domain is dispensable for its effect on angiogenesis. We show here that Robo4 À / À mice have increased ocular permeability and revascularization when subjected to OIR and exhibit accelerated healing of cutaneous wounds. Robo4DCD expression inhibits OIR revascularization, vessel permeability and wound healing in Robo4 knockout mice. Signalling studies show that the ROBO4 ECD signals via the UNC5B ICD to prevent activation of VEGFR2 Y951. The data suggest that promoting Y951 activation via ROBO4 blockade might represent an opportunity to enhance ocular revascularization in retinopathy of prematurity and wound healing in diabetic patients.

Results
Generation of mice lacking the Robo4 cytoplasmic domain. We used a mouse Robo4DCD cDNA construct (aa 1-522) that expressed the ECD and TM domain but lacked 495% of ICD sequence and was fused to GFP ( Supplementary Fig. 1). The construct was placed under the control of a tetracyclineresponsive element (TET-off system) and used to generate transgenic mice. These mice were crossed with a CDH5promoter-driven tetracycline-transactivator (CDH5-tTA) transgenic mouse line 50 . The double transgenic mice (hereafter Robo4 þ / þ ;Robo4DCD mice) were expected to express CDH5driven Robo4DCD in endothelial cells and additionally expressed endogenous Robo4. Mice carrying either of the single transgenes expressed endogenous Robo4 but not Robo4DCD and were used as controls (Robo4 þ / þ ;Stg mice) (Fig. 1a). We also generated mice expressing Robo4DCD in the absence of endogenous Robo4 (hereafter Robo4 À / À ;Robo4DCD mice), by intercrossing the double transgenic mice with Robo4 À / À mice 18 .
Robo4 þ / þ ;Robo4DCD and Robo4 À / À ;Robo4DCD mice were born at the expected Mendelian frequency in the absence of doxycycline treatment (Fig. 1b), indicating that neither lack of the full-length protein, nor lack of its cytoplasmic domain affected embryonic vascular development. Western blot analysis with an anti-Robo4 antibody on mouse lung endothelial cell (MLEC) protein extracts confirmed the presence of a single 140 kDa Robo4 band in wild-type mice, lack of expression in Robo4 À / À and Robo4 À / À ;Stg mice and expression of a shorter B100 kDa Robo4DCD form corresponding to the ECD (60 kDa), TM domain and GFP (together approximately 40 kDa) in cells isolated from Robo4 À / À ;Robo4DCD mice ( Fig. 1c). Quantification of western blots showed that expression levels of Robo4DCD were similar to endogenous Robo4, indicating that transgenic mice expressed physiological levels of the mutant protein (Fig. 1d). Because it was GFP-tagged, the 100 kDa Robo4DCD protein could also be detected by western blot with an anti-GFP antibody in mouse lung lysates from Robo4 þ / þ ;Robo4DCD mice but not from Robo4 þ / þ ;Stg mice ( Fig. 1e). Anti-GFP staining of Robo4 þ / þ ;Robo4DCD postnatal retina vasculature, adult aorta endothelium and primary MLECs showed specific transgene expression in endothelial cells ( Fig. 1f-h). The Robo4DCD-GFP fusion protein was primarily localized on the cell surface, with enrichment in the lateral junctional areas (Fig. 1g,h). Expression of mCherry-tagged Robo4DCD construct in HUVECs confirmed cell surface localization of the mutant protein ( Fig. 1i; Supplementary Fig. 2). We also detected a soluble Robo4 (sRobo4) form of 60 kDa using immunoprecipitation with an anti-Robo4 antibody in human serum and in supernatant collected from human umbilical artery ECs (HUAECs, Fig. 1j). Based on its molecular weight, sRobo4 is likely to encode most of the ECD. Expression of Robo4-full length (FL) and DCD constructs in porcine aortic endothelial cells (PAECs) led to sRobo4 secretion into the supernatant, whereas untransfected PAECs did not express sRobo4 (Fig. 1j). sRobo4 was also detected in mouse serum from wild-type and Robo4 À / À ;Robo4DCD mice, but was absent in Robo4 À / À mice (Fig. 1k). ELISA analysis of serum confirmed the presence of sRobo4 in Robo4 þ / þ ;Robo4DCD mice but not in Robo4 À / À mice (Fig. 1l). The serum level of sRobo4 is B7 ng ml À 1 in Robo4 þ / þ mice, B25 ng ml À 1 in Robo4 þ / þ ;Robo4DCD and B20 ng ml À 1 in Robo4 À / À ;Robo4DCD mice, demonstrating moderate overexpression of sRobo4 in transgenic mice (Fig. 1l).
Robo4DCD inhibits OIR revascularization. To assess effects of Robo4DCD on developmental angiogenesis, we analysed postnatal retinas of Stg and Robo4DCD expressing mice on both wild-type and Robo4 knockout background. We did not see any differences in vessel outgrowth and vessel branching between Robo4DCD expressing mice and their control littermates at P7 ( Supplementary Fig. 3a).
After return to room air, hypoxia in the avascular area triggered re-growth of normal vessel sprouts from centrally located veins and the remaining capillaries in the periphery (Fig. 2a,b), and pre-retinal neovascular tufts (Fig. 2e, see arrowheads for tufts). Compared with Robo4 þ / þ , Robo4 À / À mice showed increased revascularization, characterized by a significant decrease of the retina avascular area, and increased sprouting from veins ( Fig. 2a-d). Expression of Robo4DCD in Robo4 À / À mice prevented the increased revascularization ( Fig. 2a-d), indicating that Robo4DCD rescued Robo4 function during ocular revascularization. Intraperitoneal injection of recombinant sRobo4 protein encoding His-tagged ECD 18 in Robo4 À / À mice inhibited OIR revascularization (Fig. 2a-d), demonstrating that sRobo4 is sufficient to activate signalling.
Robo4DCD inhibits vascular leak and wound healing. As Robo4 À / À mice exhibit enhanced VEGF-driven dermal vessel permeability 17,18 , we tested effects of Robo4DCD in a Miles assay. To this end, Evans blue was injected intravenously followed by an intradermal injection of saline or VEGF into Robo4 þ / þ , Robo4 À / À ;Stg and Robo4 À / À ;Robo4DCD mice. After 30 min, animals were sacrificed, perfused with PBS and extravasated dye was imaged and quantified. With saline injection, Robo4 À / À ;Stg mice exhibited slightly increased leakage of the Evans blue dye into the skin when compared with Robo4 þ / þ and Robo4 À / À ;Robo4DCD mice ( Fig. 4a,b). Injection of VEGF into the control wild-type mice led to a robust leakage of Evans Blue (Fig. 4a,b). Robo4 À / À ;Stg exhibited significantly increased leakage of the Evans blue dye into the skin (Fig. 4a,b). In contrast, VEGF injection caused much less dye leakage in Robo4 À / À ;Robo4DCD mice ( Fig. 4a,b). To test whether this effect was specific for the VEGF pathway, we injected another potent inducer of vessel permeability, histamine. Histamineinduced permeability was similar in Robo4 þ / þ ;Robo4DCD and in Robo4 þ / þ ;Stg mice ( Supplementary Fig. 4), suggesting that Robo4DCD specifically targets VEGF-induced permeability signalling and is sufficient to prevent this process.
To determine the functional outcome of enhanced revascularization and vessel permeability in Robo4 À / À mice, we next studied cutaneous wound healing. Robo4 À / À mice and Robo4 À / À ;Stg mice exhibited significantly faster wound closure than wild-type littermates at 7 days after wounding ( Fig. 4c-e). Expression of Robo4DCD in Robo4-null background was sufficient to prevent accelerated wound healing ( Fig. 4c-e).
To determine the mechanism responsible for the enhanced VEGF response, we treated control and ROBO4 siRNA knockdown cells with VEGF for 5 and 15 min and examined phosphorylation of the VEGFR2 intracellular tyrosine residues Y951, Y1175 and Y1214. VEGF induced a significant increase of phosphorylation at all sites in control siRNA treated cells ( Fig. 5a; Supplementary Fig. 6a). ROBO4 knockdown further enhanced phosphorylation on Y951, but not on Y1175 or Y1214 (Fig. 5a-c; Supplementary Fig. 6a). Phosphorylated VEGFR2 Y951 activates c-Src 23,24 , and c-Src phosphorylation levels were also increased in ROBO4 knockdown cells (Fig. 5a,d) Robo4 Robo4 -/- Robo4 -/-% of tuft area 20 15 and Src phosphorylation in cells isolated from Robo4 À / À mice compared with Robo4 þ / þ mice, while Y1173 activation was similar between genotypes ( Fig. 5e-h). Treatment of ROBO4 knockdown HUVECs with recombinant sRobo4 was sufficient to reduce VEGF-induced Y951 activation, suggesting that the ROBO4 cytoplasmic domain was dispensable for effects on VEGF signalling ( Supplementary Fig. 6b). To test this further, we re-expressed siRNA-resistant full-length mouse Robo4 in ROBO4 knockdown HUVECs, which prevented the increase in VEGF-induced Y951 phosphorylation (Fig. 5i,j; Supplementary Table 1). Next, we expressed siRNA-resistant mouse Robo4DCD in ROBO4 knockdown HUVECs, which was sufficient to prevent the excessive pY951 and p-Src in response to VEGF (Fig. 5k,l; Supplementary Table 1). Consistently, both constructs also inhibited increased HUVEC monolayer permeability, wound closure and fibrin gel sprouting angiogenesis seen with VEGF treated, ROBO4-silenced cells ( Supplementary  Fig. 5). In primary MLECs, we observed that VEGF-induced Y949 and Src phosphorylation are reduced in Robo4 À / À ; Robo4DCD mice compared with Robo4 À / À ;Stg littermates, while Y1173 activation is similar between genotypes (Fig. 5m,n;  Supplementary Fig. 6c). Thus, ROBO4 selectively targets the VEGFR2 Y951/949-Src signalling cascade independently of its cytoplasmic domain (Fig. 5o).
ROBO4 is dispensable for Slit2 signalling. To determine if enhanced angiogenesis in the absence of Robo4 could be due to dominant negative effects on Slit2 signalling through Robo1, we tested if ROBO4 affected Slit2 signalling in HUVECs in vitro (Fig. 6a). ROBO4 siRNA knockdown in HUVECs did not affect mRNA expression levels of ROBO1 and 2 ( Supplementary Fig. 7). Stimulation with recombinant Slit2 and western blot analysis showed that Slit2 induced AKT phosphorylation, and that ROBO4 silencing did not affect the level of Slit2-induced p-AKT (Fig. 6b). Expression of Robo4DCD in ROBO4 siRNA knockdown cells did not affect Slit2-induced AKT activation (Fig. 6c). We next tested effects of ROBO4 on Slit2-induced angiogenic sprouting in 3D fibrin gels. Again, neither ROBO4 siRNA, nor Robo4DCD affected sprouting induced by Slit2 (Fig. 6d). In contrast to ROBO4, silencing of ROBO1 and 2 abolished Slit2-induced AKT activation (Fig. 6e). Reconstitution of ROBO1 and 2 knockdown cells with adenovirus encoding full-length siRNA-resistant rat Robo1 rescued Slit2-induced AKT activation ( Fig. 6e; Supplementary Table 1). AKT activation of ROBO1 and 2 knockdown cells was not rescued by expression of Robo1DCD (Fig. 6e), indicating that the Robo1 ICD is required for Slit2 signalling while ROBO4 does not interfere with Slit2 mediated AKT activation and sprouting.
To understand which domains of UNC5B were required for VEGF signalling, we generated various Unc5B cytoplasmic domain deletions (Fig. 7d). Expression of a rat Unc5B construct deleted in its entire cytoplasmic domain (rUnc5BDCD) in UNC5B knockdown cells failed to rescue VEGF-induced VEGFR2 pY951 phosphorylation (Fig. 7g,h; Supplementary Fig. 7c), suggesting that the Unc5B signalling domain was required to modulate VEGFR2 activation. The Unc5B cytoplasmic domain contains a death domain (DD) implicated in cell survival, a UPA domain (abbreviated for conserved in Unc5B, Pidd and Ankyrin) and a ZU5 domain 51 (Fig. 7d). Expression of rUnc5BDDD lacking the death domain in UNC5B knockdown cells rescued VEGF-induced VEGFR2 pY951 phosphorylation (Fig. 7g,h;  Supplementary Fig. 7c), indicating that the DD is dispensable for Unc5B effects on VEGFR2 signalling. In contrast, expression of rUnc5BDUPA in UNC5B knockdown cells failed to rescue VEGF-induced VEGFR2 pY951 phosphorylation (Fig. 7g,h;  Supplementary Fig. 7c), indicating that the UPA domain contains residues required for modulation of VEGFR2 signalling. To test if the UPA domain was sufficient to target VEGFR2 Y951, we generated a construct containing only the cytoplasmic rUnc5B UPA domain linked to Unc5B transmembrane and extracellular domain (Fig. 7d). Expression of this construct in UNC5B knockdown cells was sufficient to rescue VEGFR2 Y951 activation (Fig. 7g,h; Supplementary  Fig. 7c). None of the Unc5B constructs affected VEGFR2 Y1175 phosphorylation, demonstrating a highly specific effect of the activated UPA domain on Y951 (Fig. 7g,h; Supplementary  Fig. 7c). Finally, expression of rUnc5BFL in cells silenced for both UNC5B and ROBO4 failed to rescue Y951 activation (Fig. 7i,j), demonstrating that UNC5B cannot modulate VEGFR2 in the absence of ROBO4. For uncropped versions of all immunoblots shown in this study, see Supplementary Fig. 8. These data support a model in which ROBO4 decreases VEGFR2 Y951 activation by binding to UNC5B and activating signalling through the UNC5B UPA domain (Fig. 7a).

Discussion
We show here that Robo4 deletion enhances revascularization in OIR models and during cutaneous wound healing. Transgenic expression of near-physiological levels of Robo4DCD suppresses VEGF-induced vessel permeability, excessive angiogenesis and would healing in Robo4 À / À mutants. These results demonstrate that Robo4 restrains angiogenesis independently of its cytoplasmic domain. The OIR retina undergoes two types of angiogenesis: revascularization and neovascular tuft formation 20,21 . Revascularization is considered as physiological 'healing' of the injured retina and leads to formation of normal vasculature that eliminates ischemia, while pathological neovascularization results in formation of fragile balloon-like vessels prone to bleeding, hence aggravating ischemic tissue injury 20   The data are inconsistent with a model where Robo4 signals through its ICD to suppress angiogenesis. Thus, even though the Robo4 ICD can interact with various cytoplasmic signalling molecules 2,26,34,35 , this association appears dispensable for the in vivo Robo4 effects described in this study. Immunolocalization studies in HUVECs had previously shown that full-length  ROBO4 shuttles between the plasma membrane and intracellular vesicles, while Robo4DCD is readily detectable at the plasma membrane 18,35 . We speculate that the ROBO4 ICD may control its trafficking between intracellular vesicles and plasma membrane, perhaps via association with some of the previously identified binding partners, many of which are implicated in endocytosis and vesicle trafficking [52][53][54][55] .The molecular details of this remain to be elucidated, but appear of minor relevance for ROBO4 biological function.
Interestingly, we find that a soluble 60 kDa form of Robo4 is detectable in supernatant from Robo4 expressing cells and in mouse and human serum. The size of sRobo4 suggests that it is a proteolytic fragment that contains most of the ECD, including the N-terminal Ig-like domains that bind UNC5B (see below) 18 .
Thus, sRobo4 appears to be endowed with biological activity and its release into the circulation could activate Robo4 signalling to inhibit VEGF-induced permeability and angiogenesis. Increased plasma Robo4 levels in patients with tumours or vascular injury 3,56 may therefore reflect a physiological strategy to protect vasculature from hyperpermeability and excessive angiogenesis. Alternatively, sRobo4 could interfere with the interaction between membrane-bound Robo4 and UNC5B and inhibit UNC5B signalling.
Having established that the ICD is dispensable for ROBO4 function, we tested if ROBO4 restrained angiogenesis by heterodimerization with ROBO1. MLECs and HUVECs express 10-fold higher levels of ROBO4 than of ROBO1 (ref. 32), and both receptors can be co-immunoprecipitated from endothelial cell lysates 35 , suggesting that ROBO4 could affect Slit2 signalling through ROBO1. Since ROBO4 lacks Slit2 binding residues 36-39 , its heterodimerization with ROBO1 could lead to formation of complexes unable to bind and transduce Slit2 signals. Inducible loss of Slit2 or combined deletion of Robo1 and 2 function leads to severe angiogenic sprouting defects in OIR models 32 , while Robo4 deletion leads to enhanced sprouting angiogenesis (Fig. 2), suggesting that Robo4 could restrain ocular vascularization by blocking Slit2 signalling through Robo1 and 2. However, in contrast to Slit2 mutant mice, neither Robo4 À / À nor Robo1 À / À ; Robo4 À / À showed any developmental angiogenesis defects 32 . Moreover, our signalling experiments in HUVECs failed to reveal any effect of ROBO4 knockdown or Robo4DCD expression on Slit2 signalling. Instead, we find that Slit2 activation of downstream AKT can be completely blocked by ROBO1 and 2 knockdowns or ROBO1DCD expression. These data are fully consistent with genetic evidence from studies with Slit2 and Robo1 and 2 receptor loss-of-function mutants 32 and reveal opposing actions of Robo1, 2 and 4 in ocular vascularization in vivo: Slit2 signals through Robo1 and 2 to promote OIR angiogenesis, while Robo4 restrains this process independently of its cytoplasmic domain. Thus, heterodimerization of ROBO1 and 4 does not appear to play a biological role in the context of angiogenesis in vivo.
Although ROBO4 failed to show effects on Slit2 signalling, it readily affected VEGF signalling by targeting a specific VEGFR2 intracellular tyrosine, Y951, which controls VEGF-induced c-Src activation, vascular permeability and pathological angiogenesis [22][23][24][25] . Since ROBO4 knockdown enhanced VEGFinduced VEGFR2 Y951 and downstream Src phosphorylation, ROBO4 appears to function as a selective inhibitor of this site's activation. By contrast, ROBO4 fails to affect other intracellular VEGFR2 phosphosites including Y1175, which is critical for PLC-g and downstream ERK activation 57 . Exchange of the Y1173 residue in mouse VEGFR2 (Y1175 in human VEGFR2) for phenylalanine results in arrested endothelial cell development and embryonic death, similar to the global Vegfr2 À / À phenotype 58 . Selective targeting of Y951 but not Y1175 by ROBO4 is consistent with the phenotype of Robo4 À / À mice and may explain why these mice are deficient in angiogenesis under pathological condition, but not angiogenesis during development.
Expression of Robo4DCD was sufficient to rescue VEGFinduced Y951/Y949 activation and Src phosphorylation in vitro and in vivo, demonstrating that the ROBO4 ICD is dispensable for its effects on VEGF signalling. We had previously shown that the ROBO4 ECD binds to the ECD of another endothelial guidance receptor UNC5B 18 , prompting us to investigate if ROBO4 could signal through UNC5B to counteract VEGFR2 Y951 activation. Indeed, we find that like ROBO4 knockdown, UNC5B knockdown enhances Y951 but not Y1175 phosphorylation in response to VEGF and that re-expression of Unc5B-FL rescues Y951 activation in UNC5B but not in ROBO4 knockdown cells, supporting a function of UNC5B downstream of ROBO4. Structure-function analysis showed that effects on Y951 depend on the UPA domain in the UNC5B ICD, together supporting a model where ROBO4 affects Y951 by binding and signalling through UNC5B UPA.
How UPA signals remains to be clarified. The domain organization pattern of the UNC5 ICD (that is, ZU5-UPA-DD) is also found in ankyrins, scaffolding proteins regulating the assembly of specialized membrane microdomains 59 , and PIDD, scaffolding proteins controlling programmed cell death [60][61][62] . Crystal structure analysis has shown that when engaged with ligand (such as endothelial ROBO4), the UNC5B ZU5 and DD domains adopt a closed conformation, leaving the UPA domain exposed and thus able to interact with cytoplasmic signalling molecules 51 . UPA contains several tyrosines that might constitute docking sites for kinases or phosphatases regulating VEGFR2 Y951, but the molecular details remain to be established. Of note, our demonstration that UPA is sufficient to mediate ROBO4-UNCB5 effects on Y951 indicate a critical role for this domain and allow focusing future experiments on this domain. In addition to Robo4, Unc5B binds to various other secreted ligands and transmembrane molecules including Netrins, BMP family members, Neogenin and the fibronectin and leucine-rich transmembrane protein Flrt3 (refs 63-67). Of these, Flrt3 is also expressed in endothelial cells, and conditional Flrt3 knockout mice develop developmental retinal hypervascularization similar to that seen in Unc5B-mutant mice 18,68 . Thus, in contrast to Robo4 mutants, which show normal developmental angiogenesis, both Flrt3 and Unc5B mutants show impaired developmental angiogenesis, suggesting that additional signalling pathways relevant to vascular development are targeted by this interaction. While full elucidation of Unc5B signalling is beyond the scope of the current work, the data shown here demonstrate that endothelial ROBO4 targets VEGFR2 Y951 via the UNC5B UPA domain, thereby counteracting Y951-mediated vessel permeability and angiogenesis. Selectively enhancing VEGFR2 Y951 activation by blocking ROBO4 function targets revascularization without affecting neovascular tuft formation. The reason for this differential behaviour might be because tuft formation involves proliferation, which is not regulated by Y951 but via the VEGFR2 1175 residue. Taken together, we propose that Robo4 blockade might represent an opportunity to accelerate revascularization in retinopathy of prematurity and wound healing in diabetic patients.

Methods
Mice. The Yale University Institutional Animal Care and Use Committee (IACUC) reviewed and approved all mouse experiments.
Antibodies and reagents. The detailed information of all antibodies used in this study is listed in Supplementary Table 3 Cell culture and cell isolation. Primary MLECs were prepared as previously described 69 . Briefly, mouse lungs were dissected, minced and digested in 1 mg ml À 1 type I collagenase (Worthington Biochem. Corp., #LS004196) for 1 h. Endothelial cells were purified using CD31 Ab-coated Dyna-beads and cultured in DMEM supplemented with 10% fetal bovine serum (FBS), 100 units/ml penicillin, 100 mg ml À 1 streptomycin, 100 mg ml À 1 endothelial cell mitogen (Biomedical Technologies, Inc, #BT-203), and 10 units per ml heparin. After 5-7 days, endothelial cells were further enriched by ICAM-2 Ab-coated Dyna-beads to more than 90% purity.
HUVECs were obtained from Yale Vascular Biology & Therapeutics (VBT) program and maintained in medium supplemented with 2% FBS, VEGF and other necessary growth factors and cytokines (EGM-2 BulletKit, Lonza Inc., Rockland, ME). The cells were verified by CD31 and VE-cadherin staining for endothelial cell identify. Wi-38 fibroblasts were purchased from ATCC and maintained in DMEM containing 10% FBS, 100 units per ml penicillin, 100 mg ml À 1 streptomycin, 2 mM L-glutamine and 1 mM sodium pyruvate. All cells were free of mycoplasma contamination.
Recombinant mouse sRobo4 (aa 1-478) construct was generated by PCR from mouse Robo4 cDNA and cloned into the eukaryotic expression vector pRK5 as fusion to a C-terminal 6x Histidine tag (sRobo4-His). The construct was transiently transfected into CHO cells to produce the protein. The recombinant protein was purified to 495% purity by affinity chromatography using NiNTA affinity purification (NiNTA Superflow, Qiagen) and reconstituted in PBS/0.2 M NaCl.
ELISA. ELISA was performed using human anti Robo4-1 mAb 18 and a biotinylated anti-Robo4 Ab (R&D Systems, #BAF2366) as coating and detection Ab, respectively. In brief, 96-well plates (Nunc) were precoated with 1 mg ml À 1 anti-Robo4-1 mAb overnight, washed with PBST (0.05% Tween-20 in PBS) and blocked with PBS/2% goat serum/1% BSA. Thereafter, each well was incubated with 100 ml mouse serum sample for 2 h, 100 ml of 0.5 mg ml À 1 biotinylated anti-Robo4 Ab (diluted in PBS/1% BSA) for 2 h and streptavidin-HRP (diluted in PBS/1% BSA) for 30 min at room temperature, with appropriate PBST washes after each incubation. The wells were developed with ELISA development substrate mix (R&D Systems) for 10-20 min, and the reaction was stopped with stop solution followed by absorbance reading at 450 nm. Concentrations of sRobo4 in mouse serum were calculated using a standard curve generated from mouse sRobo4-HIS.
Immunoprecipitation. Cell culture supernatants, and human or mouse serum were precleared twice with protein A/G magnetic beads (Thermo Scientific, #88803) and incubated with anti-Robo4-1 mAb in the presence of appropriate protease inhibitors overnight at 4°C, followed by 1 h incubation with protein A/G beads. Precipitates were washed three times in PBS, resolved in 2XLaemmli's sample buffer and sRobo4 was detected by Western blot using R&D anti-Robo4 Ab.
Whole-mount staining of retinas. The eyes of mouse pups were fixed in 4% PFA for 20 min. The retinas were dissected and incubated with fluorescent labelled-IsoB4 in Pblec buffer (1 mM MgCl 2 , 1 mM CaCl 2 , 0.1 mM MnCl 2 , 1% Triton X-100 in PBS) overnight. After 3 Â 15 min washing with PBS, the retinas were mounted in fluorescent mounting medium (DAKO Inc.).
Oxygen-induced retinopathy. OIR was performed as described 20,21 . Briefly, the breeding mother and P7 neonatal pups of both genders were exposed to 75% O 2 until P12. The pups were then exposed to room air for an additional 5 days until P17. In some experiments, P12 OIR pups were subjected to intraperitoneal (I.P.) injection of 300 mg g À 1 recombinant mouse sRobo4 from P12 to P16. Eyes were collected at P17 and the retinas were stained with IsoB4. The avascular, sprouting, tuft and total retina areas were measured using ImageJ.
To test OIR retinal vessel leakage, a mixture of 70 kDa fixable rhodaminedextran (200 mg kg À 1 body weight) and fluorescent Alexa 647-conjugated IsoB4 (5 mg kg À 1 body weight) was injected into P17 OIR mice (5-6 g) retro-orbitally. Mice were sacrificed after 5 min or as indicated. The eyes were fixed in 4% PFA for 20 min. The retinas were dissected and incubated with fluorescent Alexa 488conjugated IsoB4 in Pblec buffer for 10 h. Stained retinas were flat-mounted and imaged with Alexa 488, rhodamine and Alexa 647 channels. To quantify dextran retained in vessels, mean fluorescent intensity of rhodamine in Alexa 647-labelled vessel area was measured using image J, followed by normalization with Alexa 647 intensity.
Miles vascular permeability assay. Six-to-eight week old mice of both genders were injected intravenously with 100 ml PBS containing 1% Evans blue. 15 min afterwards, 50 ng VEGF (in 20 ml saline) or 50 ng histamine (in 20 ml PBS) was injected intradermally into the shaved back skin. After 30 min, the animals were perfused with 20 ml PBS, and back skin was dissected and imaged. To extract the dye, skin biopsies were collected, dried, weighed and incubated in formamide solution at 56°C overnight. The absorbance of the extracts was measured with a spectrophotometer at 620 nm. Evans blue content was calculated using a standard curve of Evans blue in formamide and normalized to saline/PBS injected control wild-type mice.
In vivo wound healing assay. Mouse in vivo wound healing assay was described previously 70 . Eight-to-ten week old male mice were shaved and depilated on the back. Wounds were created by a 6-mm biopsy punch on the back skin. Wound images were acquired with a Leica M125 microscope equipped with a digital camera on days 0 and 7. Wound areas were measured by ImageJ and presented as percentage of the initial wound area at day 0.
In vitro sprouting assay. In vitro endothelial cell sprouting assays were performed as described 70 . Briefly, a layer of fibrin was made on the bottom of 24-well plates by mixing solubilized fibrinogen (10 mg ml À 1 in EBM-2) with thrombin (final concentration 1 U). SiRNA-transfected and adenovirus infected HUVECs were suspended in fibrinogen solution and plated on top of the fibrin gel. After the upper layer of gel was solidified, Wi-38 fibroblasts were plated on the top. Cells were replenished with EBM-2 supplemented with 2% FBS in the presence of VEGF or Slit2 for 5 days. After trypsinizing fibroblasts on the top, sprouts were labeled with 4 mg ml À 1 Calcein (Life Technologies) for 1 h, photographed with a fluorescence microscope and the area covered by endothelial sprouts measured by ImageJ.
In vitro endothelial permeability assay. HUVEC permeability assay was performed in 24-well plates containing transwell inserts (6.5-mm diameter, 0.4-mm pore size polycarbonate filters, Corning Costar Corporation) according to manufacturer's instruction. In brief, HUVECs were transfected with siRNA on day 0 and infected with GFP or Robo4 adenovirus on day 1. On day 2, cells were plated onto fibronectin (10 mg ml À 1 )-coated transwell inserts at a density of 150,000 cells per cm 2 and cultured in complete media for 3 days to allow formation of cell monolayer. To test transwell permeability, FITC-dextran (70 kDa, Molecular Probes) with final concentration of 1 mg ml À 1 and 5 nM VEGF was added into the upper compartment of the inserts. After 30 min, FITC-dextran intensity in the medium of the lower compartment was measured using a fluorescence spectrophotometer.
In vitro wound healing assay. HUVECs were transfected and infected with the indicated siRNA and adenovirus. After 48 h, cells were plated into 6-well plates, cultured until confluence and starved overnight in 0.2% FBS/EBM-2. Scratch wounds were created with a 200-ml pipette tip. The wounded cell monolayer was cultured in 0.2% FBS/EBM-2 supplemented with 6 nM VEGF-A. Cell-free areas were photographed at 0 h and 20 h post wounding under an inverted light microscope connected with a digital camera.
Quantitative real-time PCR. Total RNAs were purified from HUVECs using the RNeasy Plus Mini Kit (Qiagen, #74134) and reverse transcribed to cDNAs using IScript Reverse Transcriptase III (Bio-Rad, #170-8891) according to manufacturers' instructions. QRT-PCR was performed using the resulting cDNAs and the corresponding primers (Supplementary Table 2). The data were first normalized to GAPDH, and the relative expression levels of different genes were calculated.
Signalling studies. HUVECs were transfected with 20 pmol siRNA per well in 6-well plates using RNAiMax transfection reagent (Invitrogen, #13778) according to the manufacturer's instruction. Infection with GFP, Robo4 or Unc5B adenovirus was done one day after siRNA transfection. After 48 h, cells were starved 10 h to overnight in EBM-2 supplemented with 0.2% FBS and treated with 1 mg ml À 1 (6 nM) recombinant Slit2 or 25 ng ml À 1 (3 nM) VEGF-A. For MLEC signalling experiments, primary MLECs were starved 10 h to overnight in DMEM containing 2% FBS and stimulated with 5 nM VEGF-A. Cell lysates were collected at the indicated time points and subjected to 4-15% SDS-PAGE followed by immunoblotting using appropriate primary and second Abs. See Supplementary  Fig. 8 for all uncropped immunoblots.
Statistics. No statistical methods were used to determine sample size before experiments. Animals were selected for animal experiments based on their genotypes and proper age and gender as described in 'Methods' section, which was a pre-established criterion before the experiment. No randomization and blinding were used. Samples with equal variances were tested using Mann-Whitney U test or two-tailed Student's t-test between groups. When the variances were unequal, samples were tested using Welch's t-test. P value o0.05 was considered to be statistically significant.
Data availability. All data that support the findings of this study are available within the article and its Supplementary Information File and from the corresponding author upon reasonable request.