In vivo interaction screening reveals liver-derived constraints to metastasis

It is estimated that only 0.02% of disseminated tumour cells are able to seed overt metastases1. While this suggests the presence of environmental constraints to metastatic seeding, the landscape of host factors controlling this process remains largely unclear. Here, combining transposon technology2 and fluorescence niche labelling3, we developed an in vivo CRISPR activation screen to systematically investigate the interactions between hepatocytes and metastatic cells. We identify plexin B2 as a critical host-derived regulator of liver colonization in colorectal and pancreatic cancer and melanoma syngeneic mouse models. We dissect a mechanism through which plexin B2 interacts with class IV semaphorins on tumour cells, leading to KLF4 upregulation and thereby promoting the acquisition of epithelial traits. Our results highlight the essential role of signals from the liver parenchyma for the seeding of disseminated tumour cells before the establishment of a growth-promoting niche. Our findings further suggest that epithelialization is required for the adaptation of CRC metastases to their new tissue environment. Blocking the plexin-B2–semaphorin axis abolishes metastatic colonization of the liver and therefore represents a therapeutic strategy for the prevention of hepatic metastases. Finally, our screening approach, which evaluates host-derived extrinsic signals rather than tumour-intrinsic factors for their ability to promote metastatic seeding, is broadly applicable and lays a framework for the screening of environmental constraints to metastasis in other organs and cancer types.

It is estimated that only 0.02% of disseminated tumour cells are able to seed overt metastases 1 .While this suggests the presence of environmental constraints to metastatic seeding, the landscape of host factors controlling this process remains largely unclear.Here, combining transposon technology 2 and fluorescence niche labelling 3 , we developed an in vivo CRISPR activation screen to systematically investigate the interactions between hepatocytes and metastatic cells.We identify plexin B2 as a critical host-derived regulator of liver colonization in colorectal and pancreatic cancer and melanoma syngeneic mouse models.We dissect a mechanism through which plexin B2 interacts with class IV semaphorins on tumour cells, leading to KLF4 upregulation and thereby promoting the acquisition of epithelial traits.Our results highlight the essential role of signals from the liver parenchyma for the seeding of disseminated tumour cells before the establishment of a growth-promoting niche.Our findings further suggest that epithelialization is required for the adaptation of CRC metastases to their new tissue environment.Blocking the plexin-B2-semaphorin axis abolishes metastatic colonization of the liver and therefore represents a therapeutic strategy for the prevention of hepatic metastases.Finally, our screening approach, which evaluates host-derived extrinsic signals rather than tumour-intrinsic factors for their ability to promote metastatic seeding, is broadly applicable and lays a framework for the screening of environmental constraints to metastasis in other organs and cancer types.
The importance of microenvironmental conditions of the host organ for metastatic seeding has long been recognized.The seminal 'seed and soil' hypothesis postulates that metastatic cells will seed and colonize only favourable environments 4 .However, we still lack a comprehensive overview of the signals from the metastasis-accepting organs that promote or suppress the establishment of secondary tumours.Indeed, retrospective analysis may identify factors that allow metastases to thrive long term, but does not capture early events of metastatic seeding, and cannot distinguish between cellular interactions that are cause or consequence of metastatic outgrowth.Here we devised a screening approach for functional testing of cell-cell interactions in vivo, aimed at the identification of host-derived factors that determine the fate of disseminated tumour cells (DTCs) at the time of seeding.We used it to interrogate which hepatocyte-derived signals promote or suppress seeding of colorectal cancer (CRC) liver metastases and identify plexin B2 as a crucial regulator of liver colonization.

Screening interactions in a mosaic liver
Hepatocytes constitute 60% of the liver by cell number and 80% by mass 5 .We therefore hypothesized that early interactions with these cells might influence the ability of extravasated DTCs to seed metastases.To test this, we developed an experimental strategy for pooled perturbation of hepatocyte-tumour interactions during seeding (Fig. 1a).First, hundreds of genes are stably overexpressed in hepatocytes using CRISPR-mediated transcriptional activation (CRISPR-a), resulting in a 'mosaic liver' containing multiple perturbed environments; then, tumour cells are delivered to the liver by intrasplenic injection.We assumed that, after cancer inoculation, DTCs interacting with hepatocytes overexpressing a seeding-promoting factor would seed and grow, while DTCs interacting with hepatocytes overexpressing a suppressing factor would fail to form metastases.The effect of a perturbation on seeding can therefore be inferred by its enrichment in metastatic or non-metastatic areas, indicating a Article seeding-promoting or seeding-suppressing effect, respectively (Fig. 1a).Thus, even if perturbed seeding events are not directly observed, their outcome can be retrospectively assessed by the presence of a metastasis.
Molecular interactions between surface proteins on tumour cells and hepatocytes are probably among the first events to occur after DTC extravasation into the space of Disse.We therefore designed a library of sgRNAs targeting ligands and receptors (LRs) expressed by quiescent and regenerating mouse hepatocytes 8 .Moreover, we included components of the Hippo signalling pathway, the activation of which suppresses melanoma metastases 9 ; claudins recently implicated in CRC dissemination 10 ; orphan class C G-protein-coupled receptors (GPCRs) with an unknown role in metastatic seeding; and 100 safe-targeting sgRNAs as negative controls 11 (Extended Data Fig. 1c).
The library (3 sgRNAs per gene, 997 sgRNAs in total; Supplementary Table 1) was cloned into transposon vectors containing a CMV-GFP reporter, and co-injected with SB100X into AlbCre;dCas9-SPH mice at a concentration resulting in 0.5% GFP + hepatocytes (Extended Data Fig. 1d,e).One week after injection, we isolated CD31 − CD45 − GFP + hepatocytes using fluorescence-activated cell sorting (FACS) and performed targeted sgRNA amplification from genomic DNA followed by sequencing.Correlation analysis of sgRNA abundances in the pre-and post-injection library revealed stable sgRNA distribution and high library retention, indicating no loss of perturbation diversity (Extended Data Fig. 1f).Notably, the introduction of sgRNAs into non-proliferative hepatocytes, rather than into tumour cells, ensures unaltered library distribution throughout the experiment, avoids bottleneck effects arising from poor grafting of tumour cells in vivo and prevents library skewing towards perturbations that confer a proliferative advantage.Moreover, the elevated number of hepatocytes in the adult mouse liver (150 million) 12 enables screening of 997 sgRNAs library at high coverage (750×), while also ensuring a multiplicity of infection (MOI) lower than 1 (Extended Data Fig. 1g).
Our approach depends on the assumption that a DTC interacting with a hepatocyte with a seeding-promoting perturbation would have increased chances of survival, and the corresponding sgRNA would therefore be enriched in metastatic areas.To record the proximity of perturbed hepatocytes to metastases, we introduced the sLP-mCherry niche-labelling system 3 in Villin-creER T2 ;APC fl/fl ;Trp53 fl/fl ;Kras G12D ; Smad4 KO (AKPS) organoids (AKPS sLP-mCherry ) and observed efficient perimetastatic hepatocyte mCherry labelling in vivo after intrasplenic injection (Extended Data Fig. 1h-j).Thus, hepatocytes with successful transposon insertion (GFP + ) can be separated by fluorescence-activated cell sorting (FACS) as either proximal to metastasis (metastasis-proximal, mCherry + GFP + ) or distant from metastases (metastasis-distal, mCherry − GFP + ) (Extended Data Fig. 1k).
We conducted three screening experiments with independently amplified sgRNA library batches (Extended Data Fig. 2a).In total, 7 Alb-cre;dCas9-SPH mice and 5 non-Cre littermate controls were injected with sgRNA library and intrasplenically injected with AKPS sLP-mCherry organoids.Metastases were allowed to grow for 2 weeks, after which metastasis-proximal and metastasis-distal hepatocytes were isolated using FACS.The amount of sorted cells across all experiments and mice resulted in cumulative coverage of 1,000× for Alb-cre; dCas9-SPH mice and 500× for the littermate controls (Extended Data Fig. 2b,c).We scored genes based on the enrichment of their inferring sgRNAs in metastasis-proximal versus metastasis-distal hepatocytes (Fig. 1b).The top-scoring differentially enriched perturbations were consistent across individual mice and library batches (Fig. 1c), indicating high robustness of our screening strategy, and were not differentially enriched in non-Cre littermates (Extended Data Fig. 2d).sgRNAs strongly enriched in metastasis-distal hepatocytes induced overexpression (OE) of the tumour necrosis factor family cytokine lymphotoxin-β (Ltb), as well as several genes involved in acute phase response such as serum amyloid 1 (Saa1), amyloid precursor protein (App), ceruloplasmin (Cp) and α2 macroglobulin (A2m) (Fig. 1d).The depletion of sgRNAs targeting these genes in metastatic areas suggests that their upregulation prevents seeding of DTCs, possibly by inducing local immune activation.Indeed, Saa1 was suggested to attract macrophages to the tumour invasive front 13 , whereas amyloid protein (APP) deposition recruits neutrophils in several cancers 14 .Conversely, sgRNAs enriched in the proximity of metastases induced OE of epithelial growth factor (Egf), a known driver of metastatic CRC 15 , as well as other regulators of morphogenesis (Gpc3 and Psen1), and several genes that are involved in axon guidance such as Plxnb2, Nrp2, Sema3b, Ncam1 and Nenf (Fig. 1b-d).Our screen therefore implicates neurotrophic factors as promoters of metastatic seeding in the liver.This is consistent with reports of DTCs hijacking axonal morphogenesis pathways to interact with endothelial cells 16,17 ; however, their role in tumour-hepatocyte interactions has not been explored.
We next sought to cross-validate the results of our screen in transcriptional and mutational data of human liver metastases.We first tested whether any of the identified seeding-regulating factors (SRFs; 62 genes, top and bottom decile in GFP + mCherry + hepatocytes) were predicted to engage in tumour-hepatocyte interaction at the metastatic edge.We therefore generated spatial transcriptomics data of a human CRC liver metastasis with an extensive tumour-liver border, and performed cell type deconvolution using published single-cell RNA-sequencing (scRNA-seq) datasets 18,19 (Extended Data Fig. 2e).We then predicted LR pairs between metastasis-proximal hepatocytes and the tumour edge that potentially regulate expression changes between the tumour edge and core (Methods and Extended Data Fig. 2f,g).Among the 109 active LR pairs, 22 involved SRFs, including the chemoattractants App, Saa1, Cp and Ltb and the axon guidance molecules Plxnb2, Nenf and Nectin2 (Fig. 1e).Next, we extracted LRs from a published dataset of genomic alterations enriched in liver metastases compared with in matched primary tumours 20 and identified their interaction partners expressed by hepatocytes (Methods and Extended Data Fig. 3a).We found that 21 LRs mutated in liver metastases potentially interact with SRFs, with deleted LRs mainly predicted to interact with suppressing factors, and amplified LRs with promoting factors, possibly suggesting a selection of these interactions (Fig. 1f).Finally, SRFs are also predicted to interact with LR-encoding differentially expressed genes in liver metastases compared with in matched primary CRC in two independent scRNA-seq datasets 21,22 (Extended Data Fig. 2h).Together with our screening results, these analyses demonstrate the ability of our screening platform to capture disease-relevant interactions, and implicate hepatocyte-derived chemoattractants and axon guidance cues as regulators of metastatic seeding in the liver.
To test the direct effect of SRFs on cancer growth, we devised a small interaction screen based on co-culture of hepatocytes and cancer cells.sgRNAs targeting SRFs were transfected in an arrayed manner in primary hepatocytes isolated from Alb-cre;dCas9-SPH mice, or in immortalized mouse hepatocytes (AML12) stably expressing dCas9-SPH (Extended Data Fig. 3b).AKPS sLP-mCherry organoids dissociated into single cells were then sparsely seeded onto the hepatocyte monolayer and allowed to grow for 5 days before colony counting (colony forming units, CFU) (Fig. 1g).OE of App and Saa1 did not result in significantly altered CFU values with respect to the non-targeting sgRNA (sgNT) or untransfected controls, suggesting that their effect on seeding in vivo might be mediated by local recruitment of a third cell type (Fig. 1h and Extended Data Fig. 3c).Conversely, we observed increased CFU values after Plnxb2, Psen1 and Gprc5b OE, indicating their direct involvement in the interactions between hepatocytes and tumour cells (Fig. 1h).In particular, both CRISPR-a-mediated and lentiviral Plxnb2 OE in AML12 cultures had a very potent effect on AKPS seeding (Fig. 1i and Extended Data Fig. 3d-g).Moreover, addition of recombinant mouse plexin B2 ectodomain (rmPlexin B2) on AKPS single cells significantly increased CFUs both in the presence and absence of hepatocytes, suggesting that plexin B2 directly binds to tumour cells (Fig. 1j and Extended Data Fig. 3h).Notably, these results could be recapitulated by adding recombinant human plexin B2 (rhPlexin B2) on patient-derived CRC organoids (PDOs) with or without immortalized human hepatocytes (PTA-5565) (Fig. 1k and Extended Data Fig. 3i,j).Together with our transcriptomic and mutational analysis, these results implicate plexin B2 in direct interactions between hepatocytes and metastatic cells, which we next sought to investigate in vivo.

Plexin B2 is required for liver seeding
Plexin B2 is widely expressed in epithelial cells of most mouse tissues, where it localizes to the basolateral membrane 23 .Although its functions are mostly characterized in neural development 24 , the phylogenetic emergence of the plexin family predates the appearance of the nervous system 25 , and recent studies have unravelled roles of plexin B2 in several tissue contexts 26,27 .We used adeno-associated virus 8 (AAV8) to broadly deliver sgRNAs targeting Plxnb2 to hepatocytes of Alb-cre;dCas9-SPH mice in vivo (Extended Data Fig. 4a-e).Consistent with the results of our screen, Plxnb2 OE induced a threefold increase in metastatic foci after intrasplenic injection of AKPS organoids (Fig. 2a,b).Notably, Plxnb2 OE also promoted grafting of syngeneic pancreatic ductal adenocarcinoma (Ptf1a-cre;Kras G12D/+ ;Trp53 flox/+ , KPC) and melanoma cells (Tyr-creER;BRaf CA ;Pten lox/lox , D4M-3A), suggesting that its seeding-promoting effect also applies to other cancers that frequently metastasize to the liver (Fig. 2c,d).To test the requirement of hepatocyte-derived plexin B2 for metastatic seeding, we performed hepatocyte-specific ablation of plexin B2 through AAV8-mediated delivery of Plxnb2-targeting sgRNAs in Alb-cre;Cas9 mice or Alb-cre to Plxnb2 flox/flox mice (Plxnb2 KO; Extended Data Fig. 4d,f,g).In both of the experimental models, loss of plexin B2 almost completely prevented metastatic outgrowth of AKPS liver metastases, as revealed by histological analysis and in vivo bioluminescence imaging (BLI; Fig. 2e,f and Extended Data Fig. 4h).We further assessed the influence of hepatocyte-derived plexin B2 on seeding of spontaneous liver metastases from colorectal tumours generated by colonoscopy-guided submucosal injection of organoids (Extended Data Fig. 5a).Notably, 86% of the Plxnb2-OE mice developed spontaneous liver metastases 8 weeks after orthotopic tumour inoculation (AKPS organoids), while only 29% of control littermates did (Fig. 2g and Extended Data Fig. 5b,c).Conversely, Plxnb2 deletion significantly decreased the incidence and numbers of spontaneous liver metastases of Apc flox/flox ; Trp53 flox/flox ;Tgfbr2 flox/flox ;Kras G12D ;Akt1 myristoilated (APTAK) organoids, which exhibit a high metastatic potential 28 (Fig. 2h and Extended Data Fig. 5d-f).At steady state, plexin B2 is widely expressed by hepatocytes-with higher expression in portal areas (Extended Data Fig. 5g,h).Notably, although plexin B2 immunoreactivity is higher in peritumoral hepatocytes (Extended Data Fig. 5i,j), Plxnb2 expression is unaltered in the livers of mice bearing AKPS colon tumours, suggesting that Plxnb2 is not upregulated by primary-tumour-secreted factors nor systemic effects but, rather, due to a local response to the presence of liver metastases (Extended Data Fig. 5k).Notably, ablation of Plxnb2 5 days after intrasplenic AKPS organoid injection delayed but did not prevent metastasis formation, indicating that the presence of plexin B2 on peritumoral hepatocytes is required for metastatic seeding, but not to sustain growth (Extended Data Fig. 5l,m).
Histological analysis revealed that hepatic plexin B2 levels substantially change the morphology and microenvironment of AKPS metastases.The few remaining lesions in Plxnb2-KO livers exhibited cellular disarray and often lacked gland structures, suggesting that the absence of plexin B2 impairs epithelial morphogenesis in liver metastases (Fig. 2i).Instead, lesions in Plxnb2-OE livers consisted mostly of epithelial cells, contained fewer fibroblasts positive for α-smooth muscle actin (α-SMA) and periostin (POSTN) that instead surround AKPS liver metastases in wild-type (WT) livers, and exhibited an extensive CD146 + vascular network (Fig. 2i,j and Extended Data Fig. 6a,b).Importantly, α-SMA and CD146 immunoreactivity, as well as transcriptionally predicted cellular composition, were unaltered before tumour inoculation, indicating that increased metastatic seeding was not due to alterations in the liver environment induced by Plxnb2 OE but, rather, to direct tumour-hepatocyte interactions (Extended Data Fig. 6c-f).

Plexin B2 induces epithelialization
We next sought to investigate the mechanism by which plexin B2 controls liver colonization.scRNA-seq revealed that, in AKPS organoids, 2 h treatment with rmPlexin B2 was sufficient to induce a shift towards a more proliferative cell population, with increased expression gene sets related to MAPK and JNK signalling, cell junction assembly and morphogenesis (Fig. 3a).Consistent with an induction of proliferation, rmPlexin B2 increased frequencies of EdU + cells in AKPS organoids and PDOs (Extended Data Fig. 7a,b).Lesions in Plxnb2-OE livers further contained a higher density of Ki-67 + epithelial cells, indicating a proliferative advantage in vivo, as well as lower levels of cleaved caspase 3 (Fig. 3b and Extended Data Fig. 7c).We profiled AKPS liver metastases in Plxnb2-OE and control livers using single-nucleus assay for transposase-accessible chromatin with sequencing (snATAC-seq) and snRNA-seq (Extended Data Fig. 7d,e).Gene set enrichment analysis (GSEA) in tumour cells confirmed upregulation of terms related to cell division, morphogenesis and assembly of cell projections (Fig. 3c,d and Extended Data Fig. 7f).Consistent with cytoskeletal remodelling, phalloidin staining indicated notable apical F-actin accumulation in epithelial gland structures of AKPS liver metastases in Plxnb2-OE livers (Fig. 3e).Treatment of KPC cells and CRC PDOs with rmPlexin B2 similarly induced F-actin focal aggregation in vitro (Extended Data Fig. 7g,h).Notably, lesions in Plxnb2-OE livers also exhibited strong downregulation of genes involved in immune recognition, such as Cd74, B2m and H2-D1, coinciding with diminished CD4 + T cell infiltration (Fig. 3d and Extended Data Fig. 7f,i).
To reveal the transcription factors mediating the effects of plexin B2 in metastases, we used the chromatin profile of tumour cells to identify differentially accessible peaks and their associated motifs (Fig. 3f).With respect to the controls, metastases growing in Plxnb2-OE livers exhibited increased activity of several members of the SP/KLF family of zinc-finger transcription factors (Fig. 3g).Consistent with these results, we detected increased nuclear levels of KLF4 in AKPS liver metastases growing in Plxnb2-OE livers, while it was absent from lesions in Plxnb2-KO livers (Fig. 3h).Moreover, expression of Klf4 as well as its predicted target genes was increased in tumour cells in Plxnb2-OE livers, as well as in AKPS organoids treated with rmPlexin B2 (Extended Data Fig. 7j).
Klf4 is expressed by differentiated cells of the colonic epithelium, but its expression is lost in CRC, in which it acts as tumour suppressor by inhibiting epithelial-mesenchymal transition (EMT) [29][30][31] .In AKPS metastases colonizing Plxnb2-OE livers, increased KLF4 coincided with elevated EPCAM immunoreactivity (Extended Data Fig. 7k).We therefore hypothesized that reactivation of KLF4 at secondary sites might promote epithelialization of tumour cells through reversion of EMT, which is thought to be essential for successful metastatic outgrowth of several carcinomas [32][33][34] .Treatment of two-dimensional (2D) AKPS cultures with the KLF4 inhibitor WX2-43 (ref.35) induced a mesenchymal-like phenotype, altering colony morphology, size and actin cytoskeleton, and reduced the frequency of Ki-67 + cells (Extended Data Fig. 7l-n), indicating that KLF4 suppresses mesenchymal traits and promotes proliferation in AKPS organoids.KLF4 inhibition further reduced seeding of AKPS organoids in vitro, which could be rescued by co-treatment with rmPlexin B2 (Extended Data Fig. 7o).
AKPS organoids in culture show a hybrid EMT state with non-overlapping transcriptional signatures of EMT and mesenchymalepithelial transition (MET), and a mix of E-cadherin (ECAD) high ZEB1 low and ECAD low ZEB1 high cells (Extended Data Fig. 8a,b).Notably, nuclear ZEB1 levels are decreased after treatment with rmPlexin B2 and conversely increased by KLF4 inhibition (Extended Data Fig. 8c).In vivo, ZEB1 is absent from metastases growing in Plxnb2-OE and control livers, while AKPS lesions in Plxnb2-KO livers retain epithelial ZEB1 expression (Fig. 3i and Extended Data Fig. 8c).In the absence of plexin B2, AKPS metastases also lack expression of ELF3 and GRHL2, two transcription factors that preserve epithelial identity by suppressing EMT 36,37 (Extended Data Fig. 8d).These data implicate hepatocyte-derived plexin B2 as an inducer of epithelialization of AKPS liver metastases, and suggest that reversion of EMT is required for DTC seeding and adaptation to the liver environment.Consistent with a liver-induced epithelialization of metastases, AKPS organoid lines derived from metastases exhibit morphological and transcriptomic evidence of epithelial morphogenesis, and have an impaired ability to establish colonies when seeded in vitro as single cells, indicating increased susceptibility to anoikis (Extended Data Fig. 8d-f).

Plexin B2 interacts with class IV semaphorins
Interactions between plexin B2 and its canonical ligands, class IV semaphorins, have been implicated in promoting tumour invasion and metastasis by means of cytoskeletal remodelling and activation of RAC1 signalling [38][39][40][41] .The seeding-promoting effect of rhPlexin B2 on PDOs in vitro was indeed prevented by antibody-mediated blockade of the semaphorin-binding domain of plexin B2 (Fig. 4a).Moreover, treatment of AKPS organoids with rmPlexin B2 increased in vitro and in vivo seeding in a RAC1-dependent manner (Extended Data Fig. 9a,b), suggesting that class IV semaphorins mediate the seeding-promoting effects of hepatocyte-derived plexin B2 on DTCs.
In AKPS liver metastases, SEMA4A + tumour cells contact hepatocytes at the metastatic leading edge, and class IV semaphorins (Sema4a, Sema4c, Sema4d and Sema4g) are widely expressed (Extended Data Fig. 9c,d).SEMA4A, SEMA4C, SEMA4D and SEMA4G are also detected in human colon and primary CRC 42 (Extended Data Fig. 9e).Notably, in two scRNA-seq datasets of human CRC 18 , expression of the semaphorin genes and KLF4-target genes is significantly upregulated in high-relapse cells (HRCs) 43 and in intrinsic consensus molecular subtype 3 (iCMS3) cells 44 , but not in Lgr5 high cells (Fig. 4b-d and Extended Data Fig. 9f,g).Semaphorin expression also coincides with a MET signature and expression of KLF4-target genes (Extended Data Fig. 9h).Conversely, the core epithelial HRC signature is significantly upregulated in metastatic cells grown in Plxnb2-OE livers, as well as in AKPS organoids after treatment with rmPlexin B2 (Fig. 4e and Extended Data Fig. 9i).Of note, semaphorin levels are unaltered in liver metastases compared with in matched primary tumours 21,22 (Extended Data Fig. 9j).These analyses indicate that high semaphorin expression marks a subpopulation in the primary tumour with elevated liver metastatic potential.Indeed, in a large cohort of patients with CRC, increased expression of SEMA4A, SEMA4C and SEMA4D, but not SEMA4G, is associated with reduced recurrence-free survival (Extended Data Fig. 9k).Moreover, copy-number variation (CNV) analysis in the COAD dataset 45 indicates that SEMA4A, SEMA4C and SEMA4D are commonly found amplified, while SEMA4G is often deleted in patients with CRC (Extended Data Fig. 9k).Cumulatively, these data support the role of plexin B2-semaphorin-KLF4 signalling in promoting liver seeding, and might explain the differential ability of distinct tumour cell subpopulations to successfully form hepatic metastases.
To test the requirement of semaphorin genes for liver metastases, we performed simultaneous partial knockdown of Sema4a, Sema4c, Sema4d and Sema4g in AKPS organoids (AKPS Sema4KD ), and observed downregulation of gene sets involved in epithelial morphogenesis, cell adhesion and RAC1 GTPase activity (Extended Data Fig. 10a-c).Notably, when co-injected with control organoids in competitive seeding assays, AKPS Sema4KD organoids exhibited reduced grafting ability in vivo (Fig. 4f and Extended Data Fig. 10d).To achieve complete deletion of semaphorins, we generated quadruple KO organoids (AKPS Sema4KO ; Extended Data Fig. 10e and Supplementary Table 2).While AKPS Sema4KO organoids show unaltered proliferation in vitro, they exhibit significant grafting impairment when inoculated orthotopically (Extended Data Fig. 10f-h).To assay liver colonization, we injected AKPS Sema4KO organoids intrasplenically, and found significantly decreased

Article
metastatic burden compared with control organoids (Fig. 4g).Notably, KLF4 immunoreactivity is lost in AKPS Sema4KO metastases, which further exhibit diminished EPCAM and ECAD expression and high ZEB1 levels (Fig. 4h,i).Loss of semaphorins in AKPS liver metastases therefore phenocopies loss of plexin B2 on hepatocytes, supporting the notion that plexin-semaphorin interactions are required for metastatic seeding.

Discussion
Historically viewed as a late-stage event in cancer progression, spreading of DTCs has recently gained recognition as an early phenomenon in tumorigenesis 46,47 .Recent studies have elucidated mechanisms that promote DTC survival in circulation 48 , regulate DTC dormancy [49][50][51] or promote recurrence 43 .Yet, owing to the technical challenges of tracking single extravasated tumour cells, the environmental determinants of DTC adaptation and survival in a foreign organ environment remain largely unclear.Identification of these factors might reveal a therapeutic window to target metastasis at its most vulnerable point: before the establishment of a growth-promoting metastatic niche.The results presented here identify the interaction between hepatocyte-derived plexin B2 and class IV semaphorins on tumour cells as a necessary inducer of KLF4-mediated epithelialization of liver metastases and lay a methodological framework to deepen our understanding of metastatic seeding.

Online content
Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-024-07715-3.

Animal experiments
Hydrodynamic tail vein injection.Hydrodynamic injection, an efficient method to deliver nucleic acids to the liver, involves the rapid injection (6-8 s) of a large volume (8-12% body weight) of saline (0.9% sodium chloride) into the tail 52 .Mice were placed in a restraining device, and the tail was warmed with a red light lamp to induce lateral tail-vein dilatation.The injection site was cleaned and disinfected using an alcohol swab.Transposon plasmid (PT4-CMV-GFP (Addgene, 11704) or PT4-U6-sgRNA-CMV-GFP) and SB100X transposase OE plasmid pCMV(CAT)T7-SB100X (Addgene, 34879) were equilibrated at room temperature, transferred to a 3 ml syringe mounted with a 27 G needle, and then injected into the tail vein in a continuous motion.
The mouse was removed from the restrainer and signs of recovery were monitored.

Intrasplenic injection of tumour cells followed by splenectomy.
Intrasplenic injection of tumour cells followed by splenectomy was performed as previously described 53 .In brief, the mice were placed into a container connected to an oxygen-isoflurane inhalation device (4% for induction and 2-3% for maintenance).After 5 min, anaesthetized mice were placed onto a thermal pad (37 °C), isoflurane gas was continuously supplied by a nose cone and sterile eye ointment was applied to avoid corneal dehydration.Anaesthesia depth was monitored regularly by testing toe and tail pinch reflexes and by observing the rate, depth and pattern of respiration.All of the surgical instruments were sterilized before use and the surgical procedure was performed under aseptic conditions.The skin over the surgical site was shaved and disinfected with betadine.Using aseptic technique, an incision was made in the skin and peritoneal wall to expose the spleen.A sterile gauze was placed under the spleen.For each mouse, four 50 μl domes of AKPS organoids were collected, washed of Matrigel in ice-cold PBS and mechanically dissociated using a 20 G needle on a 20 ml syringe.Dissociated organoids were pelleted at 290g for 3 min, and resuspended in 0.04 ml sterile PBS and injected under the splenic capsule with an insulin syringe (BD, MicroFine, 0.3 ml, 30 G).Alternatively, 100,000 KPC cells in 0.04 ml sterile PBS were injected.After 10 min, the splenic artery and vein were closed by ligation.Immediately after, the spleen was resected.Subsequently, the wound was washed three times with sterile PBS.The peritoneal wall was closed with an absorbable polyglactin suture (Vicryl 4-0 or 5-0 coated) and the skin was closed with wound clips.The mice were monitored for weight loss and the experiment was terminated maximally 3 weeks after tumour cell injection.
Colonoscopy-guided submucosal injection of CRC organoids.The procedure for the colonoscopy-guided orthotopic injection of mouse colonic organoids was adapted from a previously published protocol 54 .At 36 h after passaging, AKPS or APTAK organoids were mechanically dissociated, resuspended in OptiMEM (Gibco) (one 50 μl Matrigel dome in 50 μl of OptiMEM per mouse for AKPS, three 50 μl Matrigel domes in 50 μl of OptiMEM per mouse for APTAK).Mice were anaesthetized by isoflurane inhalation and placed onto their back on a heating pad (37 °C).The colons were evacuated of stool with prewarmed PBS (37 °C) (Gibco) using the plastic tubing from an intravascular catheter (BD) mounted onto a 50 ml syringe (B.Braun).The organoid solution was injected with custom injection needles (33 gauge, 400 mm length, point style 4, 45° angle, Hamilton), a syringe (Hamilton) and a colonoscope with integrated working channel (Storz).The needle was brought into contact with colonic mucosa and 50 μl of organoid solution was quickly delivered to form a submucosal injection bubble.The mice were then monitored until the experimental or humane end point was reached.

Tail vein injection of AAV8.
A 100 μl solution of 1 × 10 11 -2 × 10 11 AAV viral genomes in sterile saline was loaded into a 1 ml syringe mounted with a 27 G needle.Mice were placed in a restraining device, and the tail was warmed with a red-light lamp to induce lateral tail-vein dilatation.The injection site was cleaned and disinfected using an alcohol swab, and then the AAV-containing solution was injected into the tail vein.The injections were performed 1 week before tumour inoculation for Plxnb2 overexpression, 2 weeks before or 5 days after tumour inoculation for Plxnb2 knockout, or 2 weeks after tumour inoculation for primary tumour experiments.
In vivo BLI.At 5 min before imaging, mice were injected intraperitoneally with luciferin substrate (d-luciferin, 150 mg per kg body weight in 0.15 ml PBS).During imaging on the Lumina S5 In Vivo Imaging System (IVIS, PerkinElmer), mice were anaesthetized with isoflurane and kept warm on a heated stage.
In vivo CRISPR-a screen Library cloning.sgRNA sequences for dCas9-SPH-mediated OE were retrieved from the Caprano library 55 and obtained as oPool from Integrated DNA Technologies with Gecko flanking sequences 56 .Three sgRNAs were selected per target from the Caprano Set A, for a total of 897 sgRNAs targeting 299 genes.One hundred safe-targeting sgRNAs 11 were also included.The library (Supplementary Table 1) was cloned into the transposon vector PT4-CMV-GFP (Addgene, 117046) and amplified as described previously 56 .In brief, oPools were resuspended to 1 μg μl −1 in water and incubated for 2 h at 37 °C, then amplified by PCR (a list of the primer sequences is provided in Supplementary Table 3): 1 μl library (1 ng μl −1 ) was mixed with 12.5 μl NebNext MasterMix, 1.25 μl oligo reverse primer (10 μM), 1.25 μl oligo forward primer (10 μM) and 9 μl water and incubated in a thermocycler (98 °C for 30 s; 20 cycles of 98 °C for 10 s, 63 °C for 10 s, 72 °C for 12 s; then 72 °C for 2 min).The PCR products were purified using the Qiagen QIAquick PCR Purification Kit, eluted in 30 μl buffer EB (Qiagen) and separated on a 2% agarose gel in Tris-borate EDTA (TBE) buffer with SybrSafe Dye.The transposon vector was digested overnight with the Bsmb-v2 enzyme and run on a 2% agarose gel in TBE buffer with SybrSafe Dye.The 150 bp sgRNA amplicon and the digested vector band (missing 1,000 bp filler) were excised and gel-extracted using the QIAquick Gel Extraction Kit.Both the vector and insert were processed for isopropanol purification by incubating 50 μl eluted DNA with 50 μl isopropanol, 0.5 μl GlycoBlue Coprecipitant and 0.4 μl 5 M NaCl.The reactions were vortexed and incubated at room temperature for 15 min, before centrifugation at 16,000g for 15 min at room temperature.The precipitate was washed twice with 1 ml ice-cold 80% ethanol and air-dried for 1 min before resuspension in 10 μl water.The DNA concentration was measured using the NanoDrop system.The Gibson assembly reaction mix containing 10 μl MasterMix, 330 ng vector, 50 ng insert and water to 20 μl was incubated at 50 °C for 1 h.After a second isopropanol precipitation, the cloned transposon libraries were resuspended in 5 μl Tris-EDTA (TE) buffer and incubated at 55 °C for 10 min.The DNA concentration was measured using the NanoDrop system.
Library amplification.Escherichia coli electrocompetent bacteria (E.cloni; Biosearch Technologies) were thawed on ice for 20 min before addition of 1 μl of Gibson reaction and electroporation (1,800 V) in a MicroPulser Electroporator (Bio-Rad).A total of 975 μl of prewarmed recovery medium was added and the bacteria were incubated 37 °C at 250 rpm for 1 h.Bacteria were plated onto prewarmed 30 cm 2 square agar plates with ampicillin and incubated overnight at 37 °C.The plates were scraped with 30-50 ml Luria-Bertani broth and plasmids were purified using the endotoxin-free MaxiPrep kit (Macherey-Nagel).Precipitated DNA was resuspended in 1 ml water and the concentration was measured using the NanoDrop system.pCMV(CAT)T7-SB100 (Addgene, 34879) was amplified in competent E. coli under chloramphenicol selection and purified using the Endotoxin-Free MaxiPrep kit (Macherey-Nagel).

Isolation of metastasis-distal and metastasis-proximal hepatocytes.
Mice were euthanized by raising the CO 2 concentrations, then the abdomen was opened and a G22 cannula was inserted into the inferior vena cava.The liver was perfused with 20 ml Hanks buffer (0.5 mM EDTA and 25 mM HEPES in HBSS) followed by 15 ml digestion buffer (15 mM HEPES and 32 μg ml −1 liberase (Roche) in low-glucose DMEM).After initial swelling of the liver, the portal vein was cut to allow outflow.After perfusion, the gallbladder was removed and the liver was transferred to a Petri dish with 10 ml digestion buffer and squished with a cell scraper to release the hepatocytes.Liberase was inactivated by adding 10 ml isolation buffer (10% fetal bovine serum (FBS) in low-glucose DMEM).The cell suspension was filtered through a 100 μm cell strainer and centrifuged at 50g for 2 min.The supernatant was removed and the pellet was washed again twice with 20 ml isolation buffer.Cells were resuspended in 2 ml FACS buffer (2 mM EDTA, 0.5% BSA in PBS) and stained with Zombie Violet (1:500) (BioLegend, 423113), TruStain FcX (anti-mouse CD16/32) antibody (BioLegend, 101320, 1:50), PE/Cy7 anti-mouse CD31 (BioLegend, 102418, 1:300) and BV570 anti-mouse CD45 (BioLegend, 103135, 1:300) for 25 min at 4 °C.Cells were washed and filtered through a 70 μm strainer.CD45 − CD31 − hepatocytes that contained a sgRNA (GFP + ), were divided into metastasis-proximal (mCherry + ) and metastasis-distal (mCherry − ) by drawing different sorting gates on an AriaIII sorter (BD Biosciences) with 100 μm nozzle (the gating strategy is shown in the Supplementary Information).Cells were collected in PBS, centrifuged at 800g for 5 min and the pellet was stored at −20 °C.
Genomic DNA extraction and targeted guide amplification and sequencing.Genomic DNA extraction was performed as described in the 'Isolation of genomic DNA with NucleoSpin Blood Kits and PCR pre-check' protocol of the Broad Institute's Genetic Perturbation Platform.In brief, the pellet was equilibrated at room temperature, resuspended in 200 μl PBS and incubated with proteinase K and lysis buffer mixture at 70 °C overnight.Then, 1 μl RNase A was added for 5 min at room temperature, followed by column purification using the NucleoSpin Blood Mini Kit (Macherey Nagel).DNA was eluted in 25 μl elution buffer (prewarmed at 70 °C), incubating on column for 5 min before centrifugation, then the concentration was measured using the NanoDrop system.sgRNAs were target amplified from both post-injection libraries and the plasmid library using an equimolar mixture of staggered P5 primers and P7 primers with sample-specific barcode (the sequences are provided in Supplementary Table 3) under the following conditions: 10 μl titanium buffer, 8 μl dNTPs, 5 μl DMSO, 0.5 μl P5 primer mix (100 μM), 40 μl water, 1.5 μl Titanium Taq polymerase, 25 μl DNA (maximum, 10 μg), 10 μl P7 primer (5 μM).The reactions were incubated in a thermocycler with the following programme: 95 °C for 5 min; 28 cycles of 95 °C for 30 s, 53 °C for 30 s and 72 °C for 20 s; and 72 °C for 10 min.PCR products from different samples were pooled according to the number of reads required to ensure 200-1,000 reads per sorted cell.After 1× SPRIselect bead (Beckman Coulter, B23319) purification, PCR products were eluted in 50 μl TE buffer.The quality and quantity of libraries were assessed using the dsDNA high-sensitivity kit (Life Technologies, Q32854) on the Qubit 4 fluorometer (Thermo Fisher Scientific) and using the high-sensitivity D1000 reagents and tapes (Agilent, 5067-5585, 5067-5584) on the TapeStation 4200 or Bioanalyzer (Agilent Technologies) system.Libraries were sequenced using the NextSeq kit (Illumina) with 75 bp single-end read chemistry and 9 bp index read, adding 10% PhiX spike-in (Illumina).
Replicates and coverage.The procedure outlined above (from cloning to sequencing) was repeated independently three times (three batches).In summary, the sgRNA coverage (that is, the number of sorted cells/997 sgRNAs) added up to a total of 1,000× for Alb-cre;dCas9-SPH mice (n = 7) and 500× for non-Cre littermate controls (n = 5).
Metastasis-proximal and metastasis-distal libraries from the same mouse were considered as paired samples.As sgRNAs in paired samples are considered to be independent sgRNAs (3 sgRNAs in 7 mice are thus considered to be 21 independent sgRNAs), paired testing yields consistent effects between paired replicates.This analysis was repeated for the individual library batches.The paired function was not used to compare perturbation enrichment in Alb-cre;dCas9-SPH versus non-Cre littermates, as an equal number of samples is required.Instead, sgRNA counts were added up from all Alb-cre;dCas9-SPH and non-Cre mice and then the standard MAGeCK RRA test function was applied.Results were visualized using MAGeCKFlute 60 and ggplot2 62 .

Spatial transcriptomics of human CRC liver metastases
Visium library preparation.A sample of human CRC hepatic metastasis (CB522586, 44 year old, male) with clear tumour-liver borders was selected from a commercial biobank (Origene).Non-consecutive sections were cut with a thickness of 10 μm and placed onto two capture areas of the 10x Visium Spatial Gene expression slide using the cryostat (Leica, CM3050S).The tissue optimization kit was used to determine the permeabilization times (24 min), and cDNA libraries were generated according to the manufacturer's instructions (10x Genomics).
The quality and quantity of libraries were assessed using the dsDNA high-sensitivity kit (Life Technologies, Q32854) on the Qubit 4 fluorometer (Thermo Fisher Scientific) and the high-sensitivity D1000 reagents and tapes (Agilent, 5067-5585, 5067-5584) on the TapeStation 4200 (Agilent Technologies) system.Paired-end sequencing was performed for all of the libraries (read 1:28 bp; index read: 10 bp; read 2: 82 bp) on the NovaSeq 6000 (Illumina) system using NovaSeq SP Reagent Kits (100 cycles).
Data analysis.Binary base call (BCL) files were demultiplexed using Bcl2fastq v.2.20.0.422 (Illumina) and preprocessed using Space Ranger (v.1.1.0or v.1.2.0; 10x Genomics).Spot transcriptomes were deconvoluted with Spotlight 63 using published scRNA-seq data as reference.Specifically, two datasets of primary CRC 18 and liver tumour microenvironment 19 were integrated using the Seurat integration method 64 .Edge spot selection was performed using the CellSelector function in Seurat, and the FindMarkers function was used for differential gene expression analysis of metastasis proximal versus distal and metastasis core versus centre.NicheNet 65 was used to predict LR interactions at the metastatic leading edge (503 possible interactions).The ligand activity analysis from NicheNet was used to estimate the potential of these interactions to regulate differentially expressed genes (DEGs) between metastasis edge and core, yielding 109 LR pairs with regulatory potential.These were then intersected with SRFs (top and bottom decile of the screen, 62 factors).

Cross-validation with human and transcriptional mutational data
Interactors of liver-metastasis-specific mutations.The Genomic Features of Organotropism dataset 20 was used to extract genes with mutations more frequently occurring in liver metastases as compared to primary tumours.This set of genes was then parsed with CellPhoneDB (v.3) 66 , CellTalkDB 67 and NicheNet 65 to filter for ligand and receptors and compile a list of their known interactors.The obtained interactors were then filtered for expression by hepatocytes according to the Human Protein Atlas (v.22.0;https://www.proteinatlas.org/) 68.Specifically, we filtered out genes with transcript per million (TPM) ≤ 0.5 in the RNA GTEx tissue gene data, normalized TPM ≤ 0. Interactors of liver-metastasis-specific DEGs.Two published datasets of primary CRC tumours and matched liver metastases 21,22 were downloaded and imported into Seurat.Tumour cells were subsetted on the basis of EPCAM expression, then DEGs were calculated between liver metastases and primary CRC using the FindMarker function in Seurat.DEGs were parsed using CellPhoneDB (v.3) 66 , CellTalkDB 67 and NicheNet 65 to filter for ligand and receptors and compile a list of their known interactors, which were intersected with SRFs (top and bottom decile, 62 factors).

Integration of the luciferase reporter for BLI.
To generate the pLVX-fireflyLuc-IRES-zsGreen1 vector, the protein coding sequence of firefly luciferase was amplified from an in-house plasmid, and cloned into EcoRI/BamHI-linearized pLVX-IRES-zsGreen1 (Takara) by InFusion (InF-fireflyLuc-F: TATTTCCGGTGAATTCCACCATGGAAGACG CCAAAAAC and -R: GAGAGGGGCGGGATCCTTACACGGCGATCTTT CCGCC).Sanger (Microsynth) and whole-plasmid (PlasmidSaurus) sequencing confirmed the identity of the construct and the absence of unwanted mutations.Lentiviral preparation and transduction of organoids was conducted as described above.Successfully transduced organoids were selected by FACS on the basis of GFP fluorescence and gating for live cells.
shRNA-mediated semaphorin KD.The sequences of shRNAs targeting Sema4a, Sema4c, Sema4d and Sema4g were obtained from The RNAi Consortium shRNA Library (Broad Institute) and cloned in an arrayed manner in a lentiviral vector expressing GFP as a selection marker based on a published plasmid backbone 72 .The EV control expressed a puromycin-resistance cassette for selection.Sanger (Microsynth) and whole-plasmid (PlasmidSaurus) sequencing confirmed the identity of each construct and the absence of unwanted mutations.Lentiviral preparation and transduction of organoids was conducted as described above.Organoids transduced with shRNAs targeting semaphorins were dissociated into single cells, incubated with Zombie Violet (1:500, BioLegend, 423113) and selected by FACS based on GFP fluorescence and gating for live cells.AKPS organoids transduced with the EV were also dissociated into single cells and subjected to sorting (only live-cell gate), and then selected by puromycin as described above.Semaphorin knockdown was assessed using RT-qPCR.For competitive seeding assays, AKPS Sema4KD and AKPS EV organoids were grown separately and then mixed at a 1:1 ratio and mechanically dissociated for intrasplenic injection.A small fraction of the injection mix was seeded in three domes to estimate the injection ratios.
Patient-derived CRC organoids.Human CRC organoids were obtained from the Visceral Surgery Research Laboratory at the University of Basel.Tissues from primary and liver metastases of patients with CRC were obtained from the University Hospital Basel after patient consent and ethical approval (Ethics Committee of Basel, EKBB, 2019-00816).

In vitro assays
Arrayed screen with primary hepatocytes or AML12 cells.Plates (96 or 384 well) were coated with the Collagen-I Cell Culture Surface Coating Kit (ScienCell Research Laboratories) according to the manufacturer's instructions.Primary mouse hepatocytes from Alb-cre;dCas9-SPH mice were isolated by perfusion as described above.After two washes in isolation buffer, the hepatocyte pellet was further purified by density separation according to a published protocol 77 .In brief, the pellet was resuspended in 10 ml isolation buffer and 10 ml Percoll solution (9 ml Percoll, 1 ml 10× PBS), then mixed thoroughly by inverting the tube several times.After centrifugation at 200g for 10 min at 4 °C, the hepatocytes were resuspended in isolation medium (supplemented with 1% penicillin-streptomycin) and plated at high density (15,000 hepatocytes per well in 96-well plates, 5,000 hepatocytes per well in 384-well plates).The same plating density was used for AML12 dCas9-SPH cells, which were generated introducing doxycycline-inducible dCas9-SPH into the Rosa26 safe-harbour by recombinase-mediated cassette exchange 78,79 and kept in culture with 2 μg ml −1 doxycycline.The next day, primary Alb-cre;dCas9-SPH hepatocytes or AML12-SPH were transfected with SB100X and transposon vectors harbouring sgRNAs against selected gene targets using Lipofectamine 3000 (Thermo Fisher Scientific).For every target, three sgRNAs were independently cloned and amplified into transposon vectors, and then pooled before transfection.Three wells were transfected for each target, and three wells were left untransfected.The next day, the transfection efficiency was estimated on the basis of GFP fluorescence.AKPS sLP-mCherry organoids were dissociated into single cells as described above, then 50 cells were seeded per well.After 5 days, colony formation was assessed by microscopy.
Treatment with recombinant mouse and human plexin B2.Recombinant human plexin B2 (5329-PB-050, Biotechne) and mouse plexin B2 (6836-PB-050, Biotechne) were reconstituted at 100 μg ml −1 in PBS.Human or mouse organoids were dissociated into single cells as described above, mixed with recombinant plexin B2 in growth medium, then seeded into 384-well plates at a density of 50 cells per well, in the absence or presence of 5,000 human or mouse hepatocytes.Colony formation was scored by microscopy.Where indicated, cultures were supplemented with 50 μM RAC1 inhibitor NSC23766 or 1 ng μl −1 anti-plexin B2 monoclonal antibody (67265-1, Proteogenic).The EdU-incorporation assay was performed using the Click-iT EdU Cell Proliferation Kit for Imaging, Alexa Fluor 647 dye (Invitrogen).In brief, 3 days after treatment of PDOs with rhPlexin B2, half of the culture medium was removed and replaced with 2× EdU-containing medium for 1 h, then the manufacturer's instructions were followed.
Multiplexed immunofluorescence and quantification.Multiplexed immunofluorescence was performed on the Comet instrument (Lunaphore) with the following antibodies: anti-cleaved caspase 3 (Cell Signaling, 9661), anti-CD68 (Abcam, ab125212), anti-CD4 (Abcam, ab183685), anti-Ki-67 (Abcam, ab15580), anti-E-cadherin (Cell Signaling, 3195), anti-α-SMA (1:1,000, Sigma-Aldrich, A2547), anti-CD146 (Abcam, ab75769).The fields of view (FOVs) containing individual liver metastases were cropped and saved using the HORIZON software (Lunaphore).Each condition (sgNT or sgPlxnb2 OE) had a minimum of five FOVs representing five different lesions taken from two mice.The individual FOVs were analysed in FIJI.In brief, each channel was thresholded manually, followed by application of a median filter for signal smoothing and filling of holes.Each image was overlaid with its corresponding thresholded image to verify the accuracy of the thresholding.The region corresponding to tumour within a FOV was demarcated as a ROI and the area covered by a specific antibody signal was quantified as the number of pixels within the thresholded image with respect to the total number of pixels within the ROI.For quantification of dividing tumour cells, signals from both Ki-67 and ECAD were used: the overlap between the two signals was calculated, then the area of dividing tumour cells was determined as Ki-67 + pixels within the overlap area over the area occupied by the nuclei of all cells (calculated using DAPI as a marker).

Quantification of metastatic foci and lesion area
H&E sections were imaged on the Leica DMi8 inverted microscope, equipped with a FLEXACAM C1 12 MP CMOS camera and analysed using QuPath software 81 .Whole-tissue area and single-liver metastases were manually isolated, producing a measure for whole-section area, metastatic area (μm 2 ) and metastasis number per section.Two non-consecutive sections quantified per animal, and a mean was calculated for the number of metastatic foci per liver section.

In situ hybridization
Single-molecule in situ hybridization.Custom DNA smFISH probes for Plxnb2 were designed in house and synthesized by Biosearch Technologies containing a 3′ amine reactive group (a list of probes is provided in Supplementary Table 3).All of the probes were pooled and labelled with AlexaFluor 594 dye according to a previously published protocol 82 .Mouse tissues were collected and fixed with 4% PFA in PBS for 3 h followed by overnight incubations in 30% sucrose, 4% PFA in PBS at 4 °C.Fixed tissues were embedded in Tissue-Tek OCT Compound (Sakura, 4583).Tissue sections (8 μm) were sectioned onto poly-l-lysine-coated coverslips, allowed to adhere by drying at room temperature for 10 min, followed by 15 min fixation in 4% PFA and overnight permeabilization in 70% ethanol.Probe hybridization was performed according to a previously published protocol 82 .Images were acquired using a ×63 oil-immersion objective with NA = 1.4 on the Leica THUNDER Imager 3D Cell Imaging system, equipped with a Leica LED8 Light engine and Leica DFC9000 GTC sCMOS camera.For quantification, 3-4× FOVs covering the entire width of the tissue were acquired for each sample and the images were processed using the Thunder deconvolution algorithm.Maximum-intensity projections of the processed images were rendered using ImageJ.Dot counting to determine the transcript numbers for each FOV was performed with FISHQuant 83 using the automatic thresholding function and the cell number was determined by segmenting and counting the nuclei using CellPose 84 .Spot counting and nucleus numbers were manually verified to ensure correctness.The average number of spots per cell was then measured by dividing the number of spots within the FOV by the number of nuclei.
Multiplexed in situ hybridization (Molecular Cartography).Probe design, sample preparation imaging and processing were conducted as previously described 85 .The analysis of the data, including cell segmentation, cell type annotation and portal versus central area annotation was described previously 86 .Visualizations were generated in ImageJ using genexyz Polylux tool plugin from Resolve BioSciences.

RT-qPCR
RNA extraction from fresh organoids, cells or liver tissue was performed using the Qiagen RNeasy purification kit.Then, 1 ng of total RNA was reverse transcribed using the cDNA synthesis kit (Takara Bio) according to the manufacturer's instructions.Expression of genes of interest was quantified with primers listed in Supplementary Table 3, by RT-qPCR using the Applied Biosystems SYBR Green Kit monitored by the QuantStudio3 system (Applied Biosystems).The samples were analysed in technical triplicates and the average cycle threshold values were normalized to Gapdh using the ∆∆C T method 87 .
Analysis.Reads were demultiplexed with Bcl2fastq v.2.20.0.422 (Illumina) and quality-checked with FastQC 89 .Adaptors were trimmed with cutadapt 57 .Data were processed using the zUMIs (v.2.9.4) platform to convert reads to count matrices per sample.Differential gene expression analysis was performed using edgeR 90 .GSEA was performed using the Bioconductor package fgsea with the default parameters on genes ranked by log[fold change] 91 .The Gene Ontology Biological Process and Hallmarks gene set collections from the Molecular Signatures Database were imported into R using the package msigdbr 92 .Cell type composition was estimated for significantly up-and downregulated genes in Enrichr 93 using Tabula Muris 94 as a reference (odds ratio test).

scRNA-seq
Library preparation.AKPS organoids were dissociated into single cells and incubated with 2 μg ml −1 rmPlexin B2 or vehicle in culture medium for 2 h at 37 °C.Cells were filtered, counted and loaded onto the GemCode Single-cell Instrument (10x Genomics).Libraries were generated according to the manufacturer's instructions from the Chromium Next GEM Single Cell 3′ end Reagent Kits v1.1 protocol.The quality and quantity of all of the libraries were assessed using the dsDNA high-sensitivity (HS) kit (Life Technologies, Q32854) on the Qubit 4 fluorometer (Thermo Fisher Scientific) and using the high-sensitivity D1000 reagents and tapes (Agilent, 5067-5585, 5067-5584) or high sensitivity D5000 reagents and tapes (Agilent, 5067-5593, 5067-5592) on the TapeStation 4200 system (Agilent Technologies).Paired-cell sequencing was performed for all libraries using the NovaSeq SP Reagent Kits (100 cycles).
Analysis.BCL files were demultiplexed using Bcl2fastq v.2.20.0.422 from Illumina, then single-cell count matrices were generated using Cell Ranger (v.5.0.0, 10x Genomics) with GRCm38 v.2020-A gene code.Datasets were integrated and processed using Seurat.Downstream analysis was conducted in R (v.4.1.0)using the Seurat (v.4.0.3197) package.The Seurat objects (rmPlexin B2-treated and control) were merged and cells with <100 or >2,500 detected genes were excluded.After log-normalization, the data were scaled regressing for mitochondrial reads, and principal component analysis was performed based on the 2,000 most variable features.Clustering and UMAP visualization were performed using ten principal components and a resolution of 0.2 for the shared nearest-neighbour clustering algorithm.Cluster markers were computed using the FindAllMarkers function, and GSEA was performed as described above.KLF4-target genes were obtained from the CHEA Transcription Factor Binding Site Profiles database 95 and computed using the AddModuleScore function.EMT and MET signatures were obtained from the GO Biological Process dataset.A list of all signatures and gene sets is provided in Supplementary Table 3.

snRNA-seq and snATAC-seq
Nucleus extraction and library construction.Combined profiling of gene expression and chromatin accessibility was performed from fresh frozen OCT-embedded livers.For each sample (2 sgNT and 2 sgPlxnb2 OE livers), three 50 μm liver sections were transferred into a prechilled gentleMACS C-tube (Miltenyi) and homogenized in the gentleMACS Octo Dissociator with 2 ml nucleus extraction buffer (Miltenyi).The nucleus suspension was filtered through a 70 μm SmartStrainer into a DNA-low-binding 5 ml tube (Eppendorf) and centrifuged at 150g for 3 min, at 4 °C.The pellet was resuspended in 5 ml 1% BSA in PBS and strained through a 30 μm SmartStrainer into a new tube.After centrifugation, the pellet was washed again in 5 ml 1% BSA in PBS.Nuclei were resuspended in 500 μl 1% BSA in PBS, counted and visually inspected.16,000 nuclei per sample were profiled using the Chromium Single Cell Multiome ATAC + Gene Expression kit (10x) according to manufacturer's instructions.Libraries were quality controlled and sequenced as described above.
Analysis.BCL files were demultiplexed using Bcl2fastq v.2.20.0.422 from Illumina, then single-nucleus count matrices were generated using Cell Ranger Arc (10x Genomics) with GRCm38 v2020-A gene code.RNA and chromatin profiles of the four datasets were integrated with Signac 96 (v.1.12.0) using the FindIntegrationAnchors function.Ambient RNA was removed with the decontX package 97 (v.1.0.0), then cell types were annotated based on the RNA profile.Tumour cells were subsetted and DEGs were calculated using the FindMarker function, and GSEA was performed as described above.Chromatin peaks were called with the CallPeaks function, then differentially open peaks and motifs were identified using the AddPeaks, FindPeaks and FindMotifs functions.

Extended Data Fig. 2 | See next page for caption.
Extended Data Fig. 2 | In vivo CRISPR-a screen identifies hepatocyte-derived factors that regulate metastatic seeding.a, Two-tailed Pearson correlation of sgRNA abundance in three independent library batches (Lib_1, Lib_2 and Lib_3).b, sgRNA coverage per mouse, calculated by dividing the number of sorted cells by 997 (sgRNAs) (n = 7 AlbCre;dCas9-SPH mice and n = 5 nonCre littermate controls).c, Proportion of reads mapped (left), zero count sgRNAs (middle), and Gini index (right) of sgRNA libraries amplified from distal (mCherry -, green) and proximal (mCherry + , red) hepatocytes across three independent batches and recipient AlbCre;dCas9-SPH mice (n = 7).d, Volcano plots showing aggregated results for all AlbCre;dCas9-SPH mice (left) and nonCre littermate controls (right).Red dots indicate seeding-promoting factors enriched in the proximity of metastases, green dots indicate seeding-suppressing factors depleted in the proximity of metastases.LogFC and adjusted p value calculated with unpaired robust rank aggregation (α-RRA) from the MAGeCK algorithm 60 .e, H&E staining (left) of a human CRC liver metastasis sample used for Visium spatial transcriptomics (10x Genomics) (1 representative image shown of 2 analysed replicates).After spot deconvolution, spots are annotated according to the predominant cell types, indicated in a colour-coded map (middle).The annotation is used as input for edge spot selection, identifying metastasisproximal and distal hepatocytes as well as metastasis core vs. edge (right).f, DEGs between metastasis-edge and metastasis-core spots.Two-tailed unpaired Wilcoxon test.Bonferroni correction for multiple testing.g, Top 20  active hepatocyte ligands and their regulatory potential on DEGs at metastatic edge (analysis conducted with Nichenet 65 ).h, Predicted interactions between top-scoring hits of the screen (proximal hits in red in top plot, distal hits in green in bottom plot) and cognate LRs whose expression is altered (logFC > |0.25|, adjusted P < 0.01, two-sided Wilcoxon test) between epithelial cells in liver metastases (LM) vs. matched primary tumours in two independent datasets of metastatic CRC (Wang et  Extended Data Fig. 3 | In vitro screen and Plexin B2 overexpression.a, Schematic representation of workflow for mutational analysis.1) Frequently mutated genes in hepatic metastases vs. primary tumours in a publicly available dataset 20 are filtered for LR-encoding genes.2) Cognate interaction partners are identified by cell-cell interaction prediction tools and filtered for hepatocyte expression and 3) intersected with SRFs.b, Experimental workflow for in vitro arrayed screens. 1) primary AlbCre;dCas9-SPH hepatocytes or AML12 dCas9-SPH are plated on collagen-coated plates and 2) transfected with sgRNA-encoding transposon vectors and SB100X, prior to 3) seeding of dissociated AKPS sLP-mCherry organoids.Data pooled from n = 3 mice.e, Immunohistochemistry for SEMA4A, C, D, and G in healthy human colon and CRC (n = 4) obtained from the Human Protein Atlas (https://www.proteinatlas.org/).f,g, Averaged expression of SEMA4A, 4C, 4D, 4G in human primary CRC (epithelial cells only, KUL dataset includes 5 patients 18 ).Cells grouped as high-relapse cells (HRCs) vs. other cells, or iCMS2 vs. iCMS3 cells.h, Averaged expression of SEMA4A, SEMA4C, SEMA4D, SEMA4G (green) projected on epithelial cells of the Samsung dataset and blended with expression of KLF4-target genes or of the MET signature (red).i, Expression of the core HRC signature in AKPS organoids treated with or without rmPlexin B2. j, Averaged expression of SEMA4A, SEMA4C, SEMA4D, SEMA4G in epithelial cells from CRC liver metastases (LM) or matched primary tumours in two independent datasets of metastatic CRC (Wang et

Fig. 1 |
Fig. 1 | Screening tumour-hepatocyte interactions in a mosaic liver.a, Schematic of the screen.DTCs interact with hepatocytes harbouring seedingpromoting or seeding-suppressing perturbations, influencing local metastatic outgrowth and therefore sgRNA enrichment in metastasis-proximal hepatocytes.b, Genes ranked by proximal enrichment score.Seeding-promoting factors are enriched in GFP + mCherry + hepatocytes (red).Suppressing factors are enriched in GFP + mCherry − hepatocytes (green).Top-scoring SRFs are listed on the right.c, The log-transformed fold change (FC) of individual sgRNAs (vertical lines) for two suppressing and two promoting factors across all mice (n = 7) and library batches (n = 3).d, GSEA of SRFs.The dot size indicates the gene set size.e, Interaction analysis in a human CRC liver metastasis.22 SRFs expressed by metastasis-proximal hepatocytes (orange) are predicted to interact with LRs on tumour cells at the metastatic leading edge (turquoise).f, Predicted interactions between SRFs and LRs that are frequently mutated in liver metastases.CNV status is shown in orange (amplified), yellow (mutation) or blue (deleted).g, Representative fluorescence micrograph showing co-culture of Alb-cre;dCas9-SPH hepatocytes overexpressing SRFs (GFP + ) and AKPS sLP-mCherry colonies (mCherry + , arrowheads).Scale bar, 100 μm.h-k, The CFU per well of AKPS sLP-mCherry organoids co-cultured with AML12 dCas9-SPH cells overexpressing SRFs (n = 3 wells; h), AKPS sLP-mCherry organoids co-cultured with AML12 cells overexpressing Plxnb2 (n = 10 wells; i), AKPS sLP-mCherry organoids treated with rmPlexin B2 (n = 3 wells; j) and PDOs treated with rhPlexin B2 (n = 5 wells; k).Statistical analysis was performed using ordinary one-way analysis of variance (ANOVA) with Dunnet's correction for multiple testing (h-j) or two-tailed unpaired t-tests (k).Data are mean ± s.d.For b-d, results are pooled from three independent experiments.

Fig. 3 |
Fig. 3 | Plexin B2 induces epithelialization of liver metastases.a, Uniform manifold approximation and projection (UMAP) of AKPS organoids treated with or without rmPlexin B2.The dots represent single cells, coloured by treatment or transcriptional clusters (0-3).Significantly enriched gene ontology (GO) terms for clusters 0 and 1, and the cluster composition per sample are shown.b, Representative fluorescence micrograph and quantification of ECAD (green) and Ki-67 (magenta) immunoreactivity in AKPS metastases in Plxnb2-OE and control livers.DAPI (nuclear counterstain) is shown in blue.Scale bar, 100 μm.The dots represent individual liver metastases (n = 5) pooled from two mice per group.c, UMAP of tumour cells in Plxnb2-OE and control livers (RNA profile).The dots represent single nuclei coloured by condition.d, GSEA in tumour cells in Plxnb2-OE versus control livers.e, Representative fluorescence micrograph and quantification of F-actin in AKPS metastases in Plxnb2-OE and control livers.Scale bar, 20 μm.The dots represent apical F-actin segments in metastatic glands (n = 18) pooled from two mice per group.f, UMAP analysis of tumour cells in Plxnb2-OE and control livers (ATAC profile).Dots represent single nuclei coloured by condition.g, Enriched transcriptionfactor-binding motifs in differentially open peaks in tumour cells in Plxnb2-OE versus control livers.P adj , adjusted P. h,i, Representative fluorescence micrograph and quantification of KLF4 (h) or ZEB1 (red) and ECAD (grey) (i) immunoreactivity in AKPS metastases in Plxnb2-OE, Plxnb2-KO and control livers.DAPI (nuclear counterstain) is shown in blue.Scale bar, 50 μm.Inset: KLF4 + or ZEB1 + nuclei (arrows).The asterisk indicates the background.The dots represent individual nuclei in metastases pooled from n = 3 mice per condition.Statistical analysis was performed using ordinary one-way ANOVA with Dunnet's correction for multiple testing (h and i) and two-tailed unpaired t-tests (b and e).Data are mean ± s.d.(c, d and f-i) pooled from two independent experiments, n = 2 mice per condition.
All of the experiments were performed in male and female mice aged 6-16 weeks.LSL-dCas9-SPH (dCas9-SPH, 031645) and Plxnb2 flox/flox mice (036883) were obtained from The Jackson Laboratory.Alb-cre (003574), LSL-Cas9 (Cas9, 024858) and mT/mG mice (007576) were obtained from a local live mouse repository.LSL-dCas9-SPH mice were obtained in 2019 and initially exhibited spontaneous urination, as also reported on the Jackson Laboratory website.The phenotype disappeared when the mice were outcrossed to B6J mice and crossed with the Alb-cre strain.Chow and water were available ad libitum, unless specified otherwise.All of the mice were in the B6J background and maintained under a 12 h-12 h light-dark schedule.Mice were housed and bred under specific-pathogen-free conditions in accredited animal facilities.At the experimental end point, mice were euthanized by raising the CO 2 concentrations.Approved humane end points (a palpable tumour with a diameter of more than 1.5 cm, or weight loss of more than 15%) were not exceeded.All of the experimental procedures were performed in accordance with Swiss Federal regulations and approved by the Cantonal Veterinary Office.Genotyping was performed by Transnetyx genotyping services.