ARF GTPases are central regulators of membrane trafficking that control local membrane identity and remodeling facilitating vesicle formation. Unraveling their function is complicated by the overlapping association of ARFs with guanine nucleotide exchange factors (GEFs), GTPase-activating proteins (GAPs), and numerous interactors. Through a functional genomic screen of three-dimensional (3D) prostate cancer cell behavior, we explore the contribution of ARF GTPases, GEFs, GAPs, and interactors to collective invasion. This revealed that ARF3 GTPase regulates the modality of invasion, acting as a switch between leader cell-led chains of invasion or collective sheet movement. Functionally, the ability of ARF3 to control invasion modality is dependent on association and subsequent control of turnover of N-cadherin. In vivo, ARF3 levels acted as a rheostat for metastasis from intraprostatic tumor transplants and ARF3/N-cadherin expression can be used to identify prostate cancer patients with metastatic, poor-outcome disease. Our analysis defines a unique function for the ARF3 GTPase in controlling how cells collectively organize during invasion and metastasis.

ARF GTPases are highly evolutionarily conserved regulators of membrane trafficking (Donaldson and Jackson, 2011; Sztul et al., 2019). ARF proteins co-ordinate membrane trafficking by regulating the local identity of the membrane to which they are recruited, such as through modulation of phospholipid composition via phosphatidylinositol kinases (Donaldson and Jackson, 2011; Nacke et al., 2021). This allows the recruitment of adaptor and coat proteins, facilitating membrane protein clustering and membrane deformation and ultimately leading to vesicle budding of encapsulated cargoes (Donaldson and Jackson, 2011). ARF GTPases are therefore central players in the localization of most membrane proteins and have emerged as key regulators of polarized cell behaviors underpinning cancer cell growth and metastasis (Casalou et al., 2020; Chen et al., 2022).

ARF GTPases cycle between GDP- or GTP-bound forms with the assistance of guanine nucleotide exchange factors (GEFs) and GTPase-activating proteins (GAPs; Adarska et al., 2021). Rather than consideration of ARF-GTP as active, and ARF-GDP as inactive, the full cycle of GTP loading of an ARF by a GEF to allow recruitment of effectors, followed by nucleotide hydrolysis by a GAP to return to GDP-ARF is required for ARF function (Sztul et al., 2019). Therein, an ARF GAP acts as both terminator and effector of the ARF GTPase cycle. In humans, five ARF GTPases are divided into three classes based on homology: Class I (ARF1, ARF3), Class II (ARF4, ARF5), and Class III (ARF6).

A complication in unraveling ARF GTPase function is their high degree of similarity in sequence and consequently their overlapping ability to associate with GEFs, GAPs, and interactors (Sztul et al., 2019). For instance, Class I ARFs (ARF1, ARF3; ARF2 was lost in humans during evolution) differ by seven amino acids in their N- and C-terminal regions, while their core ARF domain regions are identical. Moreover, of the 17 GEF and 23 GAP proteins, many of these share the ability to modulate nucleotide association on most ARFs in vitro. ARF GTPases can also act in amplifying loops, with a GEF acting as an ARF-GTP effector to activate further ARFs (Li and Guo, 2022; Padovani et al., 2014). This complexity makes it difficult to predict how ARFs and their regulators contribute to cellular behavior from individual single interactions and sets the stage for a systems-level analysis to identify how these potentially overlapping components functionally contribute to morphogenesis.

Here, we present a system-level characterization of ARF GTPase function in collective cellular behaviors using large-scale timelapse imaging of the morphogenesis of prostate cancer cells in 3D culture, machine learning to identify distinct resulting phenotypes, and molecular characterization of behaviors. This work identifies a key role for the poorly studied ARF3 GTPase in controlling how cells collectively organize into distinct phenotypes. ARF3 controls the modality of invasion, between leader cell-led chains of invasion versus collective sheet movement, by associating with and controlling turnover of the adhesion protein N-cadherin. ARF3 therefore acts as a rheostat for the modality of invasion, which regulates metastasis in vivo and can be used to identify prostate cancer patients with metastatic, poor-outcome disease. Our approach therefore allows elucidation of distinct functions of ARF GTPases in collective morphogenesis.

We interrogated the functional contribution of ARF GTPases, their GEFs, GAPs, and known interactors and effectors, which we term the “ARFome,” to cancer cell morphogenesis (Fig. 1 A). We engineered a lentiviral system that co-encodes an shRNA and membrane-targeted mVenus (mem:Venus) fluorescent protein to transduced cells (Fig. 1 B), and generated a highly validated library targeting all ARFs, GEFs, GAPs, and 72 known interactors (Fig. 1 C and Table S1). Examination of ARF GTPase expression across nine prostate cancer cell lines indicated that metastatic PC3 cancer cells showed high expression levels of all ARF GTPases compared to normal prostate cells (RWPE-1, PRECLH; Fig. S1, A–G), particularly in 3D compared to 2D culture (Fig. S1, H and I). PC3 cells also expressed almost all components of the ARFome (Fig. S1 J). When PC3 cells were plated on a thin coat of ECM as a suspension of single cells in low percentage ECM-containing medium, they formed heterogenous multicellular structures polarized around a central lumen, which we termed acini (Freckmann et al., 2022; Nacke et al., 2021). We used these PC3 acini to examine ARFome contribution to 3D morphogenesis as (i) they have high levels of all ARF GTPases, (ii) they, upon intraprostatic xenograft, provide a model for metastatic tumorigenesis, and (iii) we have shown that they can be used to identify ARF GTPase modules that regulate 3D invasion, in vivo metastasis, and predict patient survival (Nacke et al., 2021).

We developed a high-throughput, arrayed, live imaging-based screening approach to determine the 3D phenotype of ARFome component depletion on multi-day morphogenesis. Control (Scramble shRNA-expressing, Scr) 3D acini could be distinguished from non-shRNA-expressing acini by the presence (Fig. 1, D and E, white arrowheads) or absence (Fig. 1, D and E, red arrowheads) of mem:Venus fluorescence, respectively. 3D acini were imaged every hour for 96 h (Videos 1 and 2) and size, shape, and movement features were extracted for thousands of mem:Venus-positive acini per manipulation (Fig. 1, E and F; and Table S1). This live imaging approach revealed that multiple distinct 3D phenotypes occur in these cells in parallel, confirming our previous observations (Nacke et al., 2021). To detect these alternate phenotypes, we generated a machine learning classifier based on the Fast Gentle Boosting algorithm to define three morphogenesis classes with high fidelity to true user classification (91–97%): acini that are spherical (“Round”), acini that are elongated (“Spindle”), and those that are locally invading (“Spread”), which were then applied to classify and quantify the phenotype of all acini (Fig. 1 G). Application of these classes to timelapse sequences indicated that distinct phenotypes arose from single cells, and that cells could stay in the same cell state throughout observation or cycle between states to give rise to alternate phenotypes (Fig. 1 H).

To identify the phenotypes of individual ARFome component depletion, we compared the relative fold-change in frequency of Round, Spindle, and Spread phenotypes within each shRNA-expressing condition to control cells (Scr shRNA) over 96 h of observation (Fig. 2 A and Fig. S1 K). The resulting relative change in each phenotype across time allowed division of shRNAs against ARFome components into seven distinct Phenotype Groups based on clustering, including highly round (Group 3), increased local spreading (Group 1), or increased spindle-type behaviors (Group 2; Fig. 2, A and B; and Fig. S1 K). Some targets (22%, 27 out of 116) had different shRNAs mapping to different Phenotype Groups, likely reflecting differing knockdown (KD) efficiencies. Application of these groupings to network analysis of ARFome interactions from literature and publicly available databases indicated phenotypic clusters centered around different ARF GTPases (Fig. 2, C and D), which could not be easily appreciated based on connections between nodes alone due to the highly interconnected nature of the ARFome. ARF4 and ARF5 associated with Phenotype Groups 4 and 6 that are characterized by minimal change relative to control cells (Fig. 2, C and D). While ARF6 was associated solely with Phenotype Group 1, both ARF1 and ARF3 had one shRNA in each of Phenotype Groups 5 and 1, which displayed a modest but robust increase in Spindle and Spread behaviors, respectively. We therefore subsequently focused on exploring how these two highly similar Class I ARFs contributed to 3D multicellular morphogenesis.

We independently validated depletion of each Class I ARF GTPase using an orthogonal approach of lentiviral shRNA expression and stable antibiotic selection (Fig. S1, L and M). PC3 cells stably expressing shRNA (two per gene) to ARF1 or ARF3 were cultured in ECM and imaged every hour for 96 h, as described for the ARFome shRNA screen. Size, shape, and movement features were measured for each acinus and machine learning classifications applied to categorize and quantify Round, Spindle, and Spread phenotypes. Analysis of 3D phenotypes revealed that Area, as an indirect measure of growth, was unaffected by ARF1 or ARF3 depletion (Fig. 2, E–G). Either ARF1 or ARF3 depletion induced Spindle-type behaviors in acini at the expense of Round phenotype, but in the case of ARF3 the Spread phenotype was also induced (Fig. 2, E–G). This indicates that upon longer term selection for stable depletion these highly similar ARFs do not share identical phenotypes.

We examined Class I ARF contribution to a range of cellular behaviors. Depletion of neither ARF1 nor ARF3 affected cell proliferation in 2D or 3D culture (Fig. S2, A–E), corroborating a lack of effect on Area measurements in 3D culture (Fig. 2, F and G). The effect of Class I ARFs on individual cell shape in 2D culture was variably aligned with their respective collective 3D phenotypes upon ARF depletion (Fig. S2, E–H). ARF1 depletion increased the frequency of Spindle shape of single cells in 2D (Fig. S2, E–G), mirroring the collective Spindle phenotype induced in 3D upon ARF1 depletion (Fig. 2, E and F). In contrast, despite inducing both Spindle and Spread collective 3D behaviors, ARF3 depletion in 2D culture robustly induced Round single cell shape (Fig. S2, E, F, and H). Therefore, the effect of ARF3 on collective morphogenesis is specific to a 3D environment, not single cells. This emphasizes the requirement to examine ARF function in 3D systems that allow assessment of collective behaviors.

To determine the effect of Class I ARFs on collective behaviors, we examined the ability of wounded monolayers to invade, which can occur through wound repair via single-cell invasion, movement as a sheet, or as a leader cell-led chain of cells (Fig. 3 A; and Videos 3, 4, and 5). In the absence of exogenous ECM addition, this approach assays 2D migration. Plating of monolayers onto ECM and overlay of cells and wound with further ECM allows examination of collective invasion. Despite their differences in single cell shape effects, depletion of either ARF1 or ARF3 increased 2D migration ability, largely through the movement of single cells (Fig. 3, B and C). In 3D invasion contexts, either ARF1 or ARF3 depletion resulted in chain-type invasion mechanisms (Fig. 3, D and E; arrowheads), mirroring the induction of Spindle chains in 3D acinus culture in both conditions (Fig. 2, E–G). Co-depletion of ARF1 and ARF3 induced increased Spindle and Spread behaviors in both 3D and 2D culture and increased spindle-type invasion from wounded monolayers (Fig. 3, F–J). These data indicate individual and key roles for each of the Class I ARFs in suppressing invasion in cells and emphasize that the phenotype of ARF depletion is contextual on whether cells are assayed individually or collectively.

Given our observations that depletion of ARF1 or ARF3 altered shape and movement in both 2D and 3D (Fig. 2, E–G and Fig. 3, B–E), we examined whether over-expression of Class I ARFs would also affect these processes. Overexpression of mNeonGreen (mNG)-tagged ARF1 (ARF1-mNG) or ARF3 (ARF3-mNG), both of which localized to intracellular puncta compared to cytoplasmic mNG alone, did not affect cell growth in either 2D or 3D contexts (Fig. S2, I–M), similar to depletion of these ARFs. The shape of 2D single cells was modulated by ARF1-mNG or ARF3-mNG in the converse fashion to depletion of each ARF: ARF1-mNG overexpression increased the Round single-cell phenotype, while ARF3-mNG induced single cells to undergo spreading (compare Fig. S2, G and H to Fig. S2 N). This confirms distinct effects of ARF1 and ARF3 on 2D cell shape.

When examining the effects on cell movement, we observed that ARF1-mNG expression had no effect on 2D migration (Fig. S3, A and B) or 3D invasion (Fig. 4, A and B; white arrowheads demarcating chain-led invasion). In contrast, ARF3 drastically affected cell behaviors. ARF3-mNG overexpression increased both 2D wound closure and 3D invasion but did so by inducing sheet-like movement of the cell monolayer (Fig. 4, C and D; black arrowheads denoting sheet movement; Fig. S3, C and D). In 3D acinus culture, ARF1-mNG overexpression displayed largely no phenotypic alteration (Fig. 4 E and Fig. S2 I, bottom panel). In contrast, ARF3-mNG overexpression induced Spindle phenotypes at early time points that decreased over time, relative to control, while the Spread phenotype was robustly increased at all time points (Fig. 4 F and Fig. S2 I, bottom panel), mirroring the sheet like invasion of monolayers (Fig. 4 C; black arrowheads). This indicates a unique function of ARF3 as a rheostat that controls the modality of collective invasion; low ARF3 levels result in leader cell-led chain-type invasion, while elevated ARF3 levels switch cells to a collective sheet-movement invasive activity. It is important to note that these phenotypes manifest in 3D culture where collective activity is assayed for.

We mapped the unique ability of ARF3 to induce a collective sheet-type invasion phenotype by creating chimeras between ARF1 and ARF3, which only differ by 4 amino acids in their N-termini and 3 amino acids in their C-termini (Fig. 4 G). Examination of GTP-loading of ARF1, ARF3, and chimeras revealed that both ARF1-mNG and ARF3-mNG were GTP-loaded, with ARF1 displaying increased GTP levels compared to ARF3 (Fig. 4, G–I) and indicating that a lack of effect of ARF1 overexpression was not simply due to lack of GTP-loading of the tagged ARF1. ARF chimera with an ARF3 N-terminus and ARF1 C-terminus (3N/1C) displayed poor GTP loading despite robust expression, precluding full elucidation of the function of these alterations as this mutant may act as a dominant-negative GTPase. Conversely, ARF chimera with ARF1 N-terminus and ARF3 C-terminus (1N/3C) showed increased GTP-loading compared to ARF3 alone (Fig. 4, G–I).

We further characterized the functional effects on Class I ARF chimeras, examining their effect on localization, and 2D and 3D phenotypes (Fig. S3, E–K). Compared to the cytoplasmic and nuclear fluorescence of mNG alone, ARF3-mNG localized to intracellular puncta in 2D single cells and 3D acini (Fig. S3 E; white arrows; Fig. S3 H). The ARF 3N/1C-mNG chimera resulted in clustering of fluorescent puncta toward the cell periphery in 2D single cells (Fig. S3 E; black arrows; Fig. S3 F) and abrogated the ARF3-mNG induced increase in Spread phenotype observed in 2D (Fig. S3 G). A similar cell–cell contact-proximal localization was observed in 3D acini (Fig. S3 H). In contrast, the ARF 1N/3C-mNG chimera displayed enlarged puncta that were nonetheless reminiscent of the distribution of ARF3 in 2D single cells and 3D acini (Fig. S3, E and F). In acini, both ARF3-mNG and the ARF 1N/3C-mNG chimera puncta extensively co-localized with the Golgi marker GM130 and the recycling endosome marker RAB11 (Fig. S3 I), consistent with previous reports (Cavenagh et al., 1996; Kondo et al., 2012; Manolea et al., 2010). The ARF 3N/1C-mNG chimera, in contrast, maintained some colocalization with GM130 and RAB11, but the majority of localization occurred at cell–cell junctions. Phenotypically, the ARF3 N-terminus was dispensable, and C-terminus indispensable, to maintain sheet-type invasion (Fig. 4, J and K), and Spread-type acinus formation to levels reminiscent of ARF3 wild type (Fig. S3, J and K). These data indicate that ARF3 acts as a rheostat for the modality of invasion and that this function is dictated to the Class I ARFs by three unique residues in the ARF3 C-terminus (A174/K178/K180; Fig. 4 L).

We explored potential regulators and effectors of ARF3. In the morphogenesis ARFome screen, the ARFGEF PH and Sec7 Domain (PSD) displayed phenotypes similar to ARF3 (Fig. 2 C and Fig. S1 K). PSD is also known as Exchange Factor for ARF6 (EFA6A), due to its ability to strongly activate ARF6 GTP loading in solution; however, on membranes PSD is also a potent GEF for the Class I ARF, ARF1 (Padovani et al., 2014). Independent validation revealed that total levels of ARF3, but not ARF6, were increased upon PSD depletion (Fig. 5 A). Moreover, PSD KD resulted in a significant reduction of ARF3, but not ARF6, GTP loading (Fig. 5 B). PSD depletion mirrored the ARF3 depletion phenotype, resulting in increased Spindle and Spread behaviors in 3D and increased 3D chain-type invasion (Fig. 5, C and D; and Fig. 5 E, arrowhead; compare to Fig. 2, E–G and Fig. 3 E). Collectively, this suggests that in these cells PSD controls GTP loading of ARF3.

To identify a potential effector for ARF3, we examined the dual RAB11-GTP and ARF-GTP binding protein RAB11FIP4 (also known as Arfophilin-2), which displayed a similar phenotype to ARF3 in the morphogenesis screen (Fig. 2 C and Fig. S1 K). We confirmed that RAB11FIP4 associated with both endogenous and mNG-tagged ARF3 (Fig. 5, F and G) and colocalized in puncta with ARF3 (Fig. 5 H). Given the association of RAB11FIP4 with both ARF3 and RAB11, we examined whether ARF3 depletion affected RAB11FIP4 endosomal distribution. ARF3 depletion had no effect on the number or size of RAB11FIP4 puncta, nor of their distribution when segmenting the cells into periphery (“Periph”), juxtanuclear (“Juxta”), or the regions between (“Cyto”; Fig. 5, I and J). RAB11FIP4 depletion, however, mirrored ARF3 KD, inducing both Spindle and Spread phenotypes in 3D, and Chain-type ECM invasion (Fig. 5, K–N, arrowhead; compare to Fig. 2, E–G and Fig. 3 E). This allows us to propose a model wherein while the endosomal recruitment of Rab11FIP4, which is likely controlled by Rab11 (Hickson et al., 2003; Wallace et al., 2002a; Wallace et al., 2002b), is independent to ARF3 binding. PSD activation of ARF3-GTP loading facilitates ARF3 association with Rab11FIP4 on endosomes to suppress invasive 3D behaviors (Fig. 5 O).

As the PSD-ARF3-RAB11FIP4 module levels controlled the modality of collective movement in 3D, we examined whether ARF3 contributed to junctional organization between cells. Compared to mNG-expressing acini alone, ARF3-depleted acini displayed lowered overall F-actin intensity and a robust decrease of F-actin specifically at cell–cell, but not cell–ECM, junctions (Fig. 6 A, arrowheads; Fig. S4, A and B). In contrast, ARF3-mNG overexpression resulted in increased overall F-actin intensity, which was observed at the cell cortex (Fig. 6 A, arrows; Fig. S4, A and B). Analysis of the major cell–cell adhesion molecules E-cadherin and N-cadherin, which are co-expressed in PC3 cells, revealed that ARF3 levels associated with altered N-cadherin, but not E-cadherin, protein levels; ARF3 depletion decreased, while ARF3-mNG overexpression strongly increased, N-cadherin protein levels (Fig. 6, B–D). This decrease in N-cadherin protein levels upon ARF3 KD was mirrored by a decrease in N-cadherin mRNA levels (Fig. 6 E), initially suggesting that the ARF3 depletion phenotype could be explained by a transcriptional effect on N-cadherin. However, ARF3-mNG overexpressing cells also had decreased N-cadherin mRNA but displayed a robust elevation of N-cadherin protein. This suggests that ARF3 controls N-cadherin protein levels, with ARF3 levels uncoupling N-cadherin protein levels from mRNA levels.

N-cadherin could be recovered in ARF3 immunoprecipitants (Fig. 6 F) and co-localized with a subset of intracellular puncta positive for ARF3 (Fig. 6 G). Consistent with a decrease in total N-cadherin levels upon ARF3 depletion (Fig. 6, B and D), ARF3 KD did not affect the total cell area, but instead decreased both the number and the size of N-cadherin–positive puncta distributed throughout the cell (Fig. 6 H). Conversely, ARF3-mNG overexpression increased the total cell area, consistent with increased cell spreading (Fig. S2 N), and increased N-cadherin puncta number and size, particularly in non-peripheral regions (Fig. 6 I). This indicates that ARF3 controls the endosomal levels of N-cadherin.

We examined the identity of endosomes containing N-cadherin. Puncta positive for N-cadherin overlapped at a frequency of ∼20% with each of RAB4, RAB11, RAB11FIP4, and LAMP2 (Fig. 6, J and K). Depletion of ARF3 did not significantly affect localization of N-cadherin to RAB4 early recycling endosomes, LAMP2 late endosomes, or RAB11 recycling endosomes in general, but strongly increased localization of N-cadherin to RAB11FIP4-positive endosomes (Fig. 6 K). The percentage of RAB11 (P = 0.0111) and RAB11FIP4 (P = 0.0733) that co-localized with N-cadherin was decreased upon ARF3 depletion (Fig. 6 L). This suggests that ARF3 functions to couple N-cadherin specifically to a RAB11FIP4-positive subpopulation of recycling endosomes, to control total levels of N-cadherin protein.

To determine the consequence of ARF3 depletion-mediated uncoupling of N-cadherin trafficking at RAB11-RAB11FIP4 endosomes we examined N-cadherin surface distribution and turnover from the surface. While ARF3 reduction did not decrease surface N-cadherin, as determined by flow cytometry, ARF3-mNG overexpression robustly increased the steady-state N-cadherin surface levels (Fig. 6, M and N). Comparison of biotinylated N-cadherin at the cell surface (0 h, at 4°C without internalization) versus after 4 h of internalization (4 h, at 37°C), revealed that ARF3 depleted cells had accelerated, while ARF3-mNG overexpressing cells had delayed, turnover of N-cadherin from the surface (Fig. 6 O). ARF3 has been reported as part of the N-cadherin interactome (Li et al., 2019) and our data elucidate that ARF3 controls N-cadherin turnover from the cell surface by regulating association of internalized N-cadherin with RAB11-RAB11FIP4 recycling endosomes.

N-cadherin appeared to be a key cargo of ARF3 controlling morphogenesis. Unexpectedly, ARF3 and N-cadherin acted to mutually stabilize each other’s level; while N-cadherin levels decreased or increased upon ARF3 depletion or overexpression, respectively (Fig. 6, B and D), depletion of N-cadherin also decreased endogenous ARF3 levels (Fig. S4, C–F). In contrast, E-cadherin levels were not consistently changed upon alteration of either ARF3 or N-cadherin (Fig. 6, B and C; and Fig. S4 C). N-cadherin was essential for the switch between chain-type and sheet-type invasion, as depletion of N-cadherin increased 2D Spindle shape, chain-type invasion of 3D monolayers expressing mNG alone and completely reversed the sheet-type invasion of ARF3-mNG-expressing monolayers (Fig. S4 G; and Fig. 7, A and B). This was not simply due to a decrease in ARF3 levels upon N-cadherin depletion, as total levels of ARF3 were maintained upon ARF3-mNG expression in N-cadherin KD cells (Fig. S4, D and F). Moreover, N-cadherin depletion in control mNG-expressing cells phenocopied ARF3 depletion in 3D acini phenotypes (adoption of Spindle and Spread phenotypes; compare Fig. S4, H and I to Fig. 2, E and G), and completely reversed the spread-type phenotype of ARF3-mNG-expressing 3D acini (compare Fig. 4 F to Fig. 7, C and D). This effect on N-cadherin levels was also mirrored by depletion of the ARF3 effector RAB11FIP4 (Fig. S4 J). Taken together, these data indicate that ARF3, at Rab11-RAB11FIP4 endosomes, acts as a rheostat to control the turnover and total levels of N-cadherin to influence the modality of invasion (Fig. 7 E).

We examined the function of ARF3 in tumorigenesis in vivo through orthotopic intraprostatic xenograft of PC3 cells in control (mNG and Scr shRNA), ARF3-depleted (mNG and ARF3 shRNA), or ARF3-elevated (ARF3-mNG and Scr shRNA) conditions (Fig. 8 A). Mice were examined at timed endpoint of 8 wk, which allowed for examination of effects on both primary tumor formation and metastasis (Nacke et al., 2021). No difference in cell engraftment or prostate weights at timed endpoint were detected between any conditions (Fig. 8, B and C) suggesting no effect on primary tumor growth, similar to a lack of effect on 2D or 3D proliferation in vitro upon ARF3 manipulation (Fig. S2, A–D and Fig. S2, J–M). Mirroring the effect on switching collective movement modality in vitro, ARF3 depletion versus overexpression showed robust and alternate effects on metastasis in vivo. ARF3 depletion induced a fully penetrant metastatic incidence (the number of mice with a primary tumor that also possessed at least one metastasis) compared to a reduction in metastatic incidence in ARF3-overexpressing cells (100% in ARF3 KD, 67% in ARF3 overexpression, compared to 75% in control; Fig. 8 D, P = 0.0308). Moreover, ARF3 depletion also increased the number of organs with metastasis per mouse compared to ARF3 overexpression (Fig. 8 E, P = 0.01), as well as expanded the metastatic tropism to all organs examined, bar the stomach, while ARF3 overexpression resulted in metastasis to only very proximal organs (lumbar lymph nodes, mesentery, and spleen; Fig. 8 F).

We examined whether N-cadherin localization or levels were altered in ARF3-manipulated tumors, similar to 2D or 3D PC3 cells. Primary tumors from control mice displayed variable regions of both N-cadherin–positive and –negative labeling, as well as regions of N-cadherin with different intensity (Fig. 8 G, upper panels). While the average region of tumor positive for N-cadherin as well as weighted histoscore for N-cadherin intensity trended toward the corresponding effects observed in vitro for ARF3 depletion vs. overexpression, these did not reach statistical significance (Fig. 8, H and I). Rather, ARF3 manipulation affected the homogeneity of N-cadherin distribution across tumors. While the aforementioned patches of N-cadherin expression occurred in control tumors, ARF3-depleted tumors displayed larger patches of weak or no N-cadherin labeling, with some regions of often smaller N-cadherin positivity (Fig. 8 G, middle panels). In contrast, ARF3-overexpressing tumors displayed large areas of somewhat homogeneous N-cadherin expression with 40% of tumors exhibiting > 40% N-cadherin positivity in comparison to 25% of ARF3-depleted tumors (Fig. 8 G, bottom panels, H). Collectively, this suggests that rather than controlling set levels of high or low N-cadherin (from overexpression or KD, respectively) ARF3 overexpression effects an even distribution of N-cadherin expression across tumor cells.

These data suggest that ARF3 is an in vivo regulator of metastasis, not primary tumor formation, through control of N-cadherin levels. Moreover, this suggests that of the alternate modalities of collective movement that ARF3 can influence in vitro, while sheet-type movement conditions may allow local metastasis, only the spindle-type chain-based invasive modality is able to induce distant and widespread metastasis.

We examined whether levels of ARF3 and/or N-cadherin may identify patients with poor outcome or metastatic disease. We first compared normal and tumor tissue mRNA levels of ARF3 across tumor types from The Cancer Genome Atlas (TCGA) and the Gene Expression for Normal and Tumor database (which allows combination of multiple independent datasets; Park et al., 2019). This revealed that while ARF3 mRNA levels are widely altered in tumor versus normal tissue, the directionality of ARF3 mRNA alternation in tumors is dependent on tissue type and that ARF3 mRNA expression is not a consistent indicator of clinical characteristics (Fig. S5, A and B). This was corroborated in prostate cancer by a lack of consistent alteration in independent datasets comparing ARF3 mRNA levels across prostate normal tissue, primary tumor or metastasis (Fig. 9, A–E), or in the TCGA prostate adenocarcinoma dataset (PRAD Prostate; normal, n = 86; tumor, n = 323) for normal versus primary tumor or across Gleason Grades (Fig. S5, C and D).

Given the requirement of N-cadherin for ARF3 function in vitro, we examined whether the effect of ARF3 requires consideration of N-cadherin. Only four tumor types showed a consistent alteration in mRNA levels of N-cadherin (gene: CDH2) across the independent datasets: increased CDH2 levels in thyroid (THCA) and kidney (KIRP), decreased CDH2 mRNA levels in colon (COAD) and prostate (PRAD) adenocarcinoma (Fig. 9 F; and Fig. S5, E and F). Decreased CDH2 mRNA levels were also observed.

In our studies, the levels of ARF3 uncoupled N-cadherin protein from mRNA level (Fig. 6, B, D, and E). We therefore surveyed N-cadherin protein levels compared to mRNA from the TCGA prostate cohort (from Reverse Phase Protein Array; note ARF3 not profiled). In prostate cancer patients, CDH2 mRNA only modestly correlated with N-cadherin protein (Fig. S5 G). N-cadherin protein, but not mRNA levels, levels showed a modest decrease with progressive Gleason Grade (Fig. 9 H and Fig. S5 H). When divided into quartiles of expression, ascending N-cadherin protein expression, but not CDH2 mRNA levels, showed a significant, inverse decrease in the frequency of patients presenting a new neoplasm following initial therapy (protein, P = 0.0074; mRNA, not significant; Fig. 9 I), whether patients were with or without tumor (protein, P = 0.0228; mRNA, not significant; Fig. 9 J), and lymph node metastasis positivity (protein, P = 0.0213; mRNA, not significant; Fig. 9 K). This further confirms that N-cadherin protein is partially uncoupled from CDH2 mRNA levels, and that low N-cadherin levels identify patients with recurrent, metastatic tumors.

Analysis of progression-free survival of prostate cancer patients (TCGA) indicated that neither ARF3 nor CDH2 mRNA levels could stratify patient groups with altered survival (Fig. 9, L and M). In contrast, patients with lowest N-cadherin protein levels showed drastically reduced progression-free survival (compare lowest quartile [Q1] to all other patients, Q2–4; P = 0.0007; Fig. 9 N).

We examined whether combining N-cadherin protein expression with ARF3 mRNA levels would further stratify patient survival by comparing patient groups segregated by expression based on a median split (M1, low; M2, high; Fig. 9 O). Patients with low N-cadherin protein (M1) showed similar survival regardless of ARF3 mRNA levels (yellow and blue lines). In high N-cadherin expressing patients (M2) the mRNA levels of ARF3 divided survival patterns. Low ARF3 (M1) mRNA despite high N-cadherin expression (M2; green line) reduced survival to levels mirroring low N-cadherin protein. Conversely, having both high N-cadherin protein (M2) and ARF3 mRNA (M2; red line) identified patients with best progression-free survival. Similarly, high levels of CDH2 protein and ARF3 mRNA identified a patient group with lowest levels of new tumor formation after initial therapy, while all other groups showed similar rates (Fig. S5 I).

This clinical data is consistent with our in vitro data identifying a co-operation between N-cadherin and ARF3, wherein ARF3 and N-cadherin mutually control each other’s levels and function in tumorigenesis, with reduced N-cadherin protein associated with metastatic, recurrent prostate cancer. These data also indicate ARF3 as a contextual regulator of N-cadherin protein levels during tumorigenesis in prostate cancer patients.

The use of 3D culture allows the assessment of how individual genes or entire pathways contribute to collective cell behaviors. The application of a screening approach to 3D requires a number of adaptations not directly transferrable from the screening of 2D cell cultures. First, collective morphogenesis occurs over multiple days, requiring stable genetic manipulations to which siRNA transfections are poorly suited. Moreover, 3D phenotypes can be stereotyped but asynchronous or multiple phenotypes can occur in parallel in the same well. This requires live imaging of 3D morphogenesis to capture and quantify these considerations.

A bottleneck in the systematic screening of collective morphogenesis (arrayed, one manipulation per well) is the plating of multiple parallel manipulations into 3D culture at the same starting density. This is essential to ensure that phenotypes quantified do not simply represent morphogenesis from a different starting point, such as altered density. If the genetic manipulation of interest alters proliferation before plating into 3D culture, this is a technical challenge in assuring similar plating. We have overcome these obstacles using an arrayed shRNA library vector that co-encodes mem:Venus to allow plating at similar density in 3D culture. Through the use of phase contrast and fluorescent imaging of 96-well plates of 3D culture, tracking of morphogenesis of mem:Venus-positive acini, and machine learning–based classifications of distinct phenotypes, we were able to perform a functional genomic characterization of the ARFome contribution to collective morphogenesis.

It is important to note that despite using a library with high independent validation of target depletion from shRNAs, our approach is not an exhaustive analysis of every ARFome member. Rather, shRNAs are assigned into Phenotype Groups based on their relative change in Round, Spindle, or Spread acinus phenotype over time compared to a control (Scr) shRNA (Fig. S1 K). While this can detect phenotypes such as being highly Round (Group 3, e.g., IL6ST) or highly Spindle (Group 2, e.g., RAB11FIP3), phenotypes with modest change to control (Group 7, e.g., ARF4) were also identified. Modest changes can occur due to a bona fide lack of strong phenotype or could be due to inconsistent effect across the three independent instances we performed the screen. Therefore, lack of robust effect should be interpreted through the lens of such limitations of large-scale screens, rather than definitive demonstration of a lack of function of such ARFome members.

Of those ARFome members that exhibited notable phenotypes, our screen identified the Class I ARFs, the GEF PSD, and the effector RAB11FIP4 as repressors of Spindle- and Spread-type collective invasion. Particularly, we identify that loss of ARF3, not ARF1, in 3D culture phenocopies PSD and RAB11FIP4 loss of function. That depletion of Class I ARFs induced invasive activity was somewhat unexpected, as numerous studies, particularly in breast cancer cells, report a pro-invasive and pro-tumorigenic function of ARF1 (Boulay et al., 2008; Boulay et al., 2011; Haines et al., 2015; Lewis-Saravalli et al., 2013; Schlienger et al., 2014; Schlienger et al., 2016; Schlienger et al., 2015; Wang et al., 2020; Xie et al., 2016).

In our studies, we define that ARF3 has a function distinct from that of ARF1. Notably, these 3D phenotypes, which require multicellular collective function, were not recapitulated when looking at the shape of single cells in 2D, which may explain some of the differences to observations using 2D culture. Moreover, ARF3 expression was strongly induced in 3D culture, suggesting a requirement for collective function. That we identify such co-acting modules using a functional morphogenesis perturbation approach is notable as this would be difficult to predict from studies of single cells in the literature. Of ARFome expression in PC3 cells, PSD was the lowest expressed GEF, while RAB11FIP4 sits at approximately the mid-point of the ARFome interactors screened, which makes selection of these candidates non-obvious. PSD is also known as EFA6-A, which represents its prior consideration as a GEF for mostly ARF6. This is partially due to a low exchange activity of PSD from studies in solution (Hanai et al., 2016; Padovani et al., 2014). However, PSD is a potent GEF for ARF1 when both are present at membranes (Padovani et al., 2014). Indeed, in our studies, PSD was required for efficient GTP loading ARF3, but not ARF6. This underscores the power of our morphogenesis-based approach to identify co-acting molecules, rather than based solely on in vitro biochemical approaches.

We identified that RAB11FIP4 is a key ARF3 effector, and that N-cadherin is a key cargo protein of ARF3-RAB11FIP4 complex on RAB11 recycling endosomes. RAB11FIP4, also known as Arfophilin-2, is a dual Rab11- and ARF-binding protein that controls the organization of, and trafficking through, the Rab11 recycling endosome (Hickson et al., 2003; Wallace et al., 2002a; Wallace et al., 2002b). Indeed, ARF1 and ARF3 have been reported to control recycling endosome function independent to effects on the Golgi apparatus (Kondo et al., 2012). Notably, ARF3 was not required for RAB11FIP4 localization or distribution, concomitant with RAB11 being the major regulator of RAB11FIP4 endosomal recruitment (Hickson et al., 2003; Wallace et al., 2002a; Wallace et al., 2002b). Rather, ARF3 depletion increased N-cadherin retention in RAB11-RAB11-FIP4 endosomes. Similarly, KD of RAB11FIP3 in neuronal cells causes the intracellular retention of N-cadherin, which can only be rescued by expression of RAB11FIP3 with functional ARF-binding capacity (