Abstract

   Liver metastasis (LM) confers poor survival and therapy resistance
   across cancer types, but the mechanisms of liver-metastatic
   organotropism remain unknown. Through in vivo CRISPR-Cas9 screens, we
   found that Pip4k2c loss conferred LM, but had no impact on lung
   metastasis or primary tumor growth. Pip4k2c-deficient cells were
   hypersensitized to insulin-mediated PI3K/AKT signaling and exploited
   the insulin-rich liver milieu for organ-specific metastasis. We
   observed concordant changes in PIP4K2C expression and distinct
   metabolic changes in 3,511 patient melanomas, including primary tumors,
   LMs and lung metastases. We found that systemic PI3K inhibition
   exacerbated LM burden in mice injected with Pip4k2c-deficient cancer
   cells through host-mediated increase in hepatic insulin levels,
   however, this circuit could be broken by concurrent administration of
   an SGLT2 inhibitor or feeding of a ketogenic diet. Thus, this work
   demonstrates a rare example of metastatic organotropism through
   co-optation of physiological metabolic cues, and proposes therapeutic
   avenues to counteract these mechanisms.

Introduction

   Metastasis is a major determinant of mortality in patients with
   cancer^[112]1, and metastatic patterns have important clinical and
   therapeutic implications^[113]2. Liver metastasis (LM) occurs
   frequently across different cancer types, including melanoma, and
   carcinomas of the colon, pancreas and breast^[114]3 and confers a poor
   prognosis and reduced response rates to modern cancer therapies, such
   as immune checkpoint blockade (ICB), when compared to response rates in
   other common metastatic sites, such as the lung^[115]3–[116]5. The
   underlying mechanisms for these clinical phenotypes are poorly
   understood. This is in part due to difficulties in modeling of salient
   molecular underpinnings of metastatic organotropism pre-clinically.

   Despite extensive efforts in a variety of disease contexts, divergent
   somatic mutations that fully explain organ-specific metastasis have not
   yet been identified^[117]6,[118]7. While this may be due to the limited
   power of individual studies and extensive disease heterogeneity, it is
   likely that other biological factors underly cancer organotropism.
   Indeed, recent work in melanoma patients, for example, revealed a
   higher rate of chromosomal instability, a neural-like cell state and
   metabolic adaptation enriched in the brain metastases compared to
   extra-cranial metastases^[119]8. Whether similar organ-specific
   mechanisms exist for other sites is unknown.

   The liver is the major metabolic organ, where a multitude of cues are
   generated, received, and integrated, to regulate the supply of systemic
   nutrients, such as glucose and fatty acids^[120]9. Additionally, the
   liver has a dual blood supply via the hepatic artery and the portal
   vein, each of which may be a route for metastatic spread. There is also
   increasing evidence for hepatic modulation of local and peripheral
   immunity^[121]10,[122]11. These and other distinct organ features may
   contribute to the liver-metastatic niche in melanoma, and other
   cancers.

   To identify determinants of liver-metastatic organotropism in an
   unbiased fashion, we leveraged a syngeneic murine melanoma model, and
   performed large-scale CRISPR-Cas9 screens. We identified loss of
   Pip4k2c as driver of liver- but not extra-hepatic metastasis. We showed
   that this organotropism is due to co-optation of the insulin-rich liver
   milieu, to which Pip4k2-deficient cells are hypersensitized to, and
   unraveled downstream signaling, and metabolic and transcriptomic
   adaptations enriched in liver metastases in mouse and patient tumors.
   We observed a paradoxical increase in LM burden in animals treated with
   systemic PI3K inhibitors and showed that this was due to host-mediated
   responses that exacerbated the insulin-gradient, thereby promoting
   metastatic selectivity of Pip4k2c-deficient cells to the liver.
   Combination of PI3K inhibition with either systemic SGLT2 inhibition or
   feeding animals a ketogenic diet, reduced LM burden, while having
   little effect on other metastatic sites or primary tumor growth.
   Analyses and additional functional validation in prostate cancer and
   colorectal cancer models indicate a similar role of Pip42kc in
   conferring liver metastasis, suggesting that the proposed biological
   and therapeutic principles may extend beyond melanoma.

Results

CRISPR screens reveal determinants of liver metastasis

   To identify drivers of liver metastasis, we used a syngeneic melanoma
   model, HCmel12, which was derived from Hgf/Cdk4^R24C mice that
   spontaneously develop melanoma^[123]12 and recapitulates the potential
   for liver metastasis and therapeutic response patterns seen in patients
   with metastatic melanoma^[124]13,[125]14. We performed a large-scale in
   vivo CRISPR-Cas9 screen in Cas9-expressing HCmel12 cells, perturbing
   713 kinases in the mouse genome with 2,852 single-guide RNAs (sgRNAs)
   (four per target gene, and 100 non-targeting control guides using the
   “Brie” library^[126]15) ([127]Fig. 1a, [128]Extended Data Fig.
   1a–[129]c). The edited cell pool (Cas9 + Brie), as well as
   Cas9-expressing cells without sgRNAs and parental cells, were injected
   via tail vein or subcutaneously implanted in C57BL/6 (B6) animals. In
   parallel, edited cell lines were maintained in vitro and collected at
   several time points for the duration of the experiment ([130]Methods).
   Dissection of animals ([131]Methods) revealed that the Cas9 + Brie
   group had a significantly higher rate of animals with liver metastasis
   ([132]Fig. 1b–[133]d), higher metastasis count per animal ([134]Fig.
   1e), and increased liver metastatic burden compared to the Cas9 and
   parental groups ([135]Fig. 1f), while there was no difference in lung
   metastatic burden across these groups ([136]Fig. 1g). This suggested
   that loss of kinases targeted by the Brie sgRNA library enhanced the
   liver metastatic potential of HCmel12 melanoma cells.

Figure 1. In vivo CRISPR-Cas9 screen identifies drivers of liver metastasis.

   Figure 1.
   [137]Open in a new tab

   a, Experimental design of in vivo CRISPR-Cas9 screen. b, Photographs of
   representative livers removed from animals in different groups as
   indicated. Yellow arrowheads indicate pigmented and non-pigmented liver
   metastases. c, Hematoxylin and Eosin staining of representative
   sections of livers from indicated groups, scale bar 2mm, 40x
   magnification (n=2 independent experiments). d, Stack bar plots
   indicate fraction of animals bearing liver metastases across different
   experimental groups. e,f Liver metastasis count (e) and calculated
   liver metastasis disease burden (f) across experimental groups. g, Lung
   metastasis count in corresponding animals from (f); (d-g), n=9 mice per
   group; bars represent mean ± s.e.m.; One-Way ANOVA Tukey’s multiple
   comparisons test. Data is representative of two independent experiments
   (d-g). Significance levels as indicated on top of each comparison. h,
   Exemplary distribution of sgRNA read fractions in three animals from
   the in vivo CRISPR-Cas9 metastasis screen. Pie charts show the fraction
   of the most abundant sgRNAs (sgRNAs with >2% of total reads) in each
   individual liver metastatic lesion next to each animal. i, Summary of
   fractions of enriched sgRNAs across all lesions from (h) within the
   individual mouse (larger pie charts below each animal).

   Before identifying perturbations that may alter liver metastatic
   potential, we first assessed the sgRNA library complexity within the
   plasmid pool, edited cells, transplanted primary tumors, and distant
   metastasis (liver, lung and lymph nodes). Compared to the plasmid pool,
   cell line input, and primary tumors, which showed preserved and
   co-correlated library diversity, sgRNA diversity decreased
   substantially in metastatic lesions ([138]Extended Data Fig. 1d–[139]e)
   across different sites, while samples from the same metastatic sites
   showed stronger co-correlation ([140]Extended Data Fig. 1d). In line
   with this, principal component analysis (PCA) of sgRNA representations
   revealed clades of metastases and transplanted tumors/cell lines along
   the PC1 (explaining 49.7% of variability) and PC2 (5.4 % of
   variability) ([141]Extended Data Fig. 1f). Next, we identified sgRNAs
   targeting essential genes (e.g., Plk1, Cdk7 and Cdk12) that were
   depleted after 14 days of editing in cells prior to injection into
   animals ([142]Extended Data Fig. 2a–[143]c, [144]Supplementary Table
   1). Additionally, sgRNAs targeting Cdk4, a known dependency in HCmel12,
   were also significantly depleted (log2FC=-10.206, FDR<0.06). Compared
   to edited cells prior to injection, primary tumors showed significant
   depletion of multiple genes, including Jak1 and Jak2, suggesting that
   the Jak/Stat pathway may be important for tumor growth, as absence of
   intact Jak1/Jak2 has been implicated in immune evasion^[145]16. Stk11
   was also depleted in primary tumors; loss of Stk11 (encoding for Lkb1)
   promotes metastasis and confers therapy resistance^[146]17. The most
   strongly enriched perturbation involved in primary tumor growth was
   loss of Dapk3 (log2FC=4.3693, FDR<0.007) ([147]Extended Data Fig. 2b,
   [148]Supplementary Table 2), a putative tumor suppressor gene whose
   deleterious mutations or loss occurs in up to ~7% of melanomas^[149]18.
   Together, these data suggested that the conducted screen was robust and
   yielded known and unknown biology, with sgRNA diversity reducing along
   the metastatic progression axis. This observation suggested that a
   small fraction of perturbations may determine metastatic
   potential^[150]19.

   We next sought to identify sgRNAs associated with metastatic potential
   to the liver. We used three criteria to nominate such putative drivers:
   sgRNAs that (1) are enriched with an FDR <0.1 and/or (2) account for
   >2% of sequencing reads within a given lesion and/or (3) are enriched
   in two or more biological replicates. We identified 80 sgRNAs that meet
   one (65), two (17) or all (2) of these criteria ([151]Supplementary
   Table 3). Among the genes targeted by these sgRNAs are those involved
   in cell motility (Tesk1, Pkn1, Mylk4)^[152]20–[153]22, nucleotide
   homeostasis and metabolism (Prps1, Prps2, Adck2, Nme8)^[154]23,
   modulation of insulin sensing (Trib1, Pip4k2c)^[155]24–[156]26, or
   function as tumor suppressors (Fgfr3)^[157]27, such that the loss of
   these genes would promote hallmarks of metastasis. The “top hit” in
   this screen by all applied metrics was loss of Pip4k2c which encodes
   Phosphatidylinositol-5-Phosphate 4-Kinase Type 2 Gamma. Specifically,
   loss of Pip4k2c was the most enriched perturbation in five of 14 (~36%)
   investigated liver metastatic lesions. In all of these lesions, sgRNAs
   targeting Pip4k2c accounted for the majority of sequencing reads, while
   in four of these five lesions, this perturbation accounted for >98% of
   sequencing reads ([158]Fig. 1h,[159]i, [160]Extended Data Fig. 2d).
   Furthermore, loss of Pip4k2c was the only perturbation (at an FDR
   <0.06), that was highly specific to liver metastasis ([161]Extended
   Data Fig. 2e). These findings suggest that Pip4k2c loss may uniquely
   promote liver metastasis.

Pip4k2c loss sensitizes cells to insulin-mediated signaling

   Pip4k2c is part of the Pip4k family that consists of three isoforms in
   both mice and humans^[162]24. The role of these kinases is poorly
   understood, but recent work suggests that Pip4k’s, and the gene product
   of Pip4k2c, suppress Pip5k activity in response to insulin. This
   results in reduced production of phosphatidylinositol-4–5-bisphosphate
   (PI (4,5) P[2]), the substrate of PI3 Kinase (PI3K), which converts PI
   (4,5) P[2] to phosphatidylinositol-3-4-5-triphosphate (PI (3,4,5) P[3])
   to subsequently activate AKT. Thus, we hypothesized that loss of
   Pip4k2c would lead to hyperactivation of the PI3K/AKT pathway in
   response to insulin. We hypothesized that because the liver is the
   organ with the highest insulin concentration (outside of the pancreas,
   where it is produced), Pip4k2c-loss mediated PI3K/AKT
   hypersensitization to insulin in tumor cells could encourage formation
   of an organ-specific metastatic niche.

   To test whether Pip4k2c promotes PI3K/AKT hypersensitivity to insulin
   in melanoma cells, we generated HCmel12-Pip4k2c-KO cell lines
   (Pip4k2c^KO) using four different sgRNAs, including the top enriched
   sgRNA in our in vivo screen ([163]Extended Data Fig. 3a–[164]b). We
   next determined the impact of Pip4k2c^KO on downstream PI3K/AKT pathway
   activation in response to insulin. Compared to parental cells, insulin
   stimulation of Pip4k2c^KO resulted in higher and prolonged induction of
   phosphorylated AKT in both murine HCmel12 and the human melanoma model
   A375 ([165]Fig. 2a,[166]b, [167]Extended Data Fig. 3c–[168]e),
   increased phosphorylation of insulin receptor (INSR) ([169]Extended
   Data Fig. 3f) and showed increased cell migration capacity, while the
   proliferation rate was not altered ([170]Extended Data Fig. 3g–[171]j).
   Bulk RNA-sequencing of parental compared to Pip4k2c^KO cells stimulated
   with insulin revealed enrichment of mTOR pathway activity
   ([172]Extended Date Fig. 3k, [173]Supplementary Table 4). We rescued
   Pip4k2c^KO with one of two open reading frames (ORFs): full-length
   (Pip4k2c^Rec) or allosteric-domain deficient (aa69–75 (VMLLPDD →
   EIFLPNN)) (Pip4k2c^AD), which was previously implicated in Pip4k2c
   function^[174]24 ([175]Extended Data Fig. 3l–[176]n). While Pip4k2c^AD
   continued to demonstrate high baseline and insulin-mediated Akt
   phosphorylation, Pip4k2c^Rec had lower p-Akt levels in both human and
   murine melanoma models ([177]Extended Data Fig. 3l–[178]n). This
   suggests that lack of the allosteric domain, but not the kinase domain
   of Pip4k2c, is sufficient to confer hypersensitization to insulin; this
   finding is consistent with prior work showing that Pip4k2c has very
   poor kinase activity^[179]24. Treatment with pan-PI3K inhibitor
   GDC-0941 (pictilisib) (with and without insulin stimulation)
   ([180]Extended Data Fig. 3l–[181]n) completely abrogated baseline Akt
   phosphorylation in cells with all of these genotypes, while insulin
   partly rescued PI3K activity ([182]Extended Data Fig. 3l–[183]n), which
   is consistent with a prior report^[184]28. We also tested PI3K alpha
   inhibitor BYL-719, which overall showed lower efficacy but similar
   results for insulin-mediated bypass activation ([185]Extended Data Fig.
   3l–[186]n).

Figure 2. Loss of Pip4k2c enhances sensitivity to insulin and promotes liver
metastasis, but not lung metastasis.

   Figure 2.
   [187]Open in a new tab

   a,b, Immunoblots showing phosphorylated and total protein of key
   signaling nodes of pAkt^S473 and pAkt^T308 in WT or Pip4k2c^KO cells in
   (a) HCmel12 and (b) A375 melanoma cell models with (+) or without (-)
   insulin (250ng/ml) stimulation. c,d, Liver metastasis disease burden
   (in mm^3) and lung metastases count following tail vein injection of
   parental and Pip4k2c KO HCmel12 melanoma in C57BL/6 mice. e,f, Liver
   and lung metastasis burden following tail vein injection of parental
   and PIP4K2C KO A375 melanoma cells in NSG mice. g, Incidence of
   macroscopic liver metastasis in animals following orthotopic injection
   of parental or Pip4k2c KO CT26 colorectal cancer cells in BALB/c mice.
   h, Comparison of tumor weight of primary tumors from experiment shown
   in (g); (c-f) n=10 mice per group; (g-h) n=4–5 mice per group; mean
   ±s.e.m. One-Way ANOVA Tukey’s multiple comparisons test for (c,d) and
   Mann-Whitney Two-tailed test for (e-g). Data are representative of
   three (c,d), two (e,f) or one (h,g) independent experiments.
   Significance levels as indicated on top of each comparison.

Loss of Pip4k2c confers liver-metastatic organotropism

   To validate the role of Pip4k2c loss in vivo, we next injected either
   parental (Pip4k2c^WT), Pip4k2c^KO, Pip4k2c^Rec, or Pip4k2c^AD HCmel12
   melanoma cells via tail vein in B6 mice. Pip4k2c^KO resulted in a
   significantly increased LM burden that was rescued by Pip4k2c^Rec, but
   not Pip4k2c^AD ([188]Fig. 2c). Importantly, none of these perturbations
   altered lung metastatic burden ([189]Fig. 2d). Consistently, loss of
   PIP4K2C in human melanoma cell line A375 conferred significantly
   increased metastasis to the liver, but not to the lung ([190]Fig.
   2e,[191]f, [192]Extended Data Fig. 3o). To determine whether Pip4k2c
   loss plays a role in LM development in other cancers that also
   frequently metastasize to the liver, we generated Pip4k2c KO in a
   colorectal cancer cell line (CT26) and orthotopically implanted these
   or parental cells into the colonic submucosa of BALB/c mice. Primary
   tumors arising from parental CT26 resulted in LM in 20% while
   Pip4k2c^KO cells gave rise to LM in 75% of animals ([193]Fig. 2g).
   Notably, Pip4k2c^KO gave rise to smaller primary colorectal tumors
   ([194]Fig. 2h), thus, indicating that higher rates of LM development
   were due to increased invasive capacity, as observed in melanoma models
   ([195]Fig. 2c,[196]e). Together, these results suggest that loss of
   Pip4k2c hypersensitizes cancer cells to insulin mediated PI3K/AKT
   hyperactivation, and thereby mediates metastasis specifically to the
   liver, but not to the lung.

Systemic PI3K inhibition enhances LM through host effects.

   Since PI3K-i could abrogate PI3K/AKT signaling in vitro ([197]Extended
   Data Fig. 3m–[198]n), we reasoned that systemic treatment with the same
   drug would reduce LM burden in vivo. We injected parental or Pip4k2c^KO
   cells and treated animals with either vehicle control or GDC-0941
   ([199]Methods). Surprisingly, we found that liver-metastatic burden was
   significantly increased in animals bearing Pip4k2c^KO tumor cells,
   while lung metastatic burden was not altered ([200]Fig. 3a,[201]b). We
   reasoned that this paradoxical increase in LM burden in Pip4k2c^KO
   tumor bearing mice in response to PI3K inhibition may be explained by
   the effects of the drug on the host, which may promote liver metastasis
   in an insulin-dependent manner. Indeed, treatment with GDC-0941
   resulted in a significant increase in blood glucose levels that peaked
   at 60 minutes after administration and was sustained for at least 90
   minutes ([202]Fig. 3c). Accordingly, the levels of C-peptide, a short
   amino-acid polypeptide of pro-insulin and equimolar surrogate for
   insulin production, were significantly increased in GDC-0941 treated
   animals with either Pip4k2c^WT or Pip4k2c^KO bearing tumors ([203]Fig.
   3d, [204]Extended Data Fig. 3p), suggesting that these differences in
   metastatic burden are explained by host responses to this treatment. To
   assess whether GDC-0941 induced glucose and insulin spikes influenced
   metastatic tumor cell sensitivity to insulin in situ, we treated
   Pip4k2c^KO liver-metastasis bearing mice with GDC-0941 or vehicle
   control. After 60 minutes when blood glucose concentration peaks
   following PI3K-i treatment, we performed [F18] fluorodeoxyglucose micro
   positron-emission tomography (FDG μPET). We found that Pip4k2c^KO liver
   metastases had a 3-fold increase in FDG uptake in PI3K-i treated mice
   ([205]Fig. 3e, [206]f). Lastly, we directly measured homeostatic levels
   of insulin in livers and lungs, confirming significantly higher levels
   in the hepatic environment ([207]Fig. 3g). If high insulin
   concentration in the liver is an important cue for liver-metastatic
   organotropism, we reasoned that increasing insulin levels in the lung
   should increase the rate of metastasis to this organ. To test this, we
   subcutaneously implanted mice with slow-release insulin pads ([208]Fig.
   3h), which continuously deliver high systemic levels of insulin,
   followed by injection of cancer cells via tail vein, and determined
   liver and lung metastatic burden. While insulin delivery resulted
   increased metastatic burden was increased in both liver and lung
   ([209]Fig. 3i), there was a significantly elevated lung-metastatic
   potential compared to control pad-implanted animals ([210]Fig. 3j,
   [211]k), suggesting that some disease burden was diverted to the lungs
   with higher local insulin levels. Together, these results suggest that
   Pip4k2c^KO cells efficiently co-opt the insulin-rich liver milieu for
   metastasis, and this can be further enhanced by host responses to PI3K
   inhibition.

Figure 3. Systemic feedback loop induced by PI3K inhibitor treatment in vivo
promotes liver organotropism in Pip4k2c-deficient tumors in an
insulin-dependent manner.

   Figure 3.
   [212]Open in a new tab

   a, Liver metastasis disease burden (in mm^3) upon tail vein injection
   of parental and Pip4k2c KO HCmel12 melanoma cells with and without
   GDC-0941 treatment (daily, 5 days a week, 100mg/kg). b, Lung metastasis
   burden (metastasis count) in corresponding animals from (a). n = 10
   mice per group; mean ±s.e.m; c, Blood glucose levels (in mg/dL) in mice
   following treatment with GDC-0941 or vehicle control over time
   (minutes). D, C-peptide levels (in pM) in corresponding animals from
   (c); n = 10 mice per group; mean ±s.e.m; 2-way ANOVA Tukey’s multiple
   comparisons test for (a-b) and Mann-Whitney Two-tailed test for (c-d).
   Significance levels as indicated. E, f, Quantification of FDG-μPET
   (18F-Fluorodeoxyglucose in per cent of injected dose (ID) per gram of
   (e)) and corresponding liver metastasis uptake (f) 60 min after
   GDC-0941 administration or vehicle control. Yellow arrows indicate
   individual liver metastatic lesions. Cardiac FDG-uptake is visualized
   (asterisk); n=3 mice per group, mean ±s.e.m; unpaired two-sided-t-test
   (e). g, Insulin levels (in pg/ul) in liver and lung tissues from
   non-tumor bearing mice. n=10 mice per group, mean ±s.e.m; Mann-Whitney
   Two-tailed test. h, Schematic illustrating the insulin pad experiment
   in mice; i, j, Percentage of metastasis surface area in liver (i) and
   lung (j) of animals implanted with a control pad or insulin pad, and
   followed by injection of tumor cells via tail vein in NSG mice. k,
   Ratio of lung over liver metastatic burden from experiment in (i,j);
   n=5 mice per group; mean ± s.e.m; Mann-Whitney Two-tailed test; Data
   are representative of three (a-d), two (e-k) independent experiments.
   Significance levels as indicated on top of each comparison.

Human and murine LM exhibit distinct metabolic adaptations

   To further examine cancer cell states in liver compared to lung
   metastases, we performed single-cell RNA-sequencing (scRNA-seq) of
   eight mice with concurrent liver and lung metastases. Following
   rigorous quality control ([213]Methods, [214]Supplementary Tables 4 and
   [215]5), a total of 95,606 cell transcriptomes were included for
   downstream analyses. We identified tumor cells by their expression of
   melanocytic lineage markers, including Mitf, Mlana, Dct, Tyr, and
   Tyrp1. Differential gene expression and subsequent GSEA of melanoma
   cells revealed strong enrichment of metabolic pathways in cells
   isolated from liver metastases, including regulation of oxidative
   phosphorylation ([216]Fig. 4a, [217]Extended Data Fig. 4a–[218]g). In
   support of the observations in melanoma, we analyzed another bulk
   RNA-seq data set of a prostate cancer model^[219]29 that develops
   metastasis to multiple organs, including to liver, lung, and bone, and
   find the same metabolic pathway enrichment in liver compared to lung
   metastases ([220]Fig. 4b).

Figure 4. multi-omics analysis of human and mouse liver and lung metastasis.

   Figure 4.
   [221]Open in a new tab

   a, Top enriched pathways (adjusted p value < 0.05, top 5 pathways per
   gene set ranked by NES) of cancer cells isolated from concurrent liver
   (n=8 specimens, pink) vs lung (n=8 specimens, purple) metastases in
   mice injected with HCmel12 melanoma cells. b, Pathway enrichment (FDR
   q-value < 0.05) within Arriaga et al., cohort comparing liver (n=7
   specimens, pink) vs. lung (n=14 specimens, purple) metastases in a
   prostate cancer metastasis model. c, Enriched pathways (FDR q-value <
   0.05) in liver (n=78 patients, Met500; n=15 patients, LM15, pink) vs
   lung (n=16 patients, Met500, purple) metastases across tumors. d,
   Enriched pathways (adjusted p value < 0.05) in liver (n=364) vs lung
   (n=743) metastases in patients with melanoma tumors; FDR value for GSEA
   was obtained by Benjamin-Hochberg correction. e,f, Expression of
   PIP4K2C in melanoma patient tumors comparing primary cutaneous tumors
   (n=2,404 patients) vs liver (n=364) (e), and liver (n=364 patients) vs
   lung (n= 743 patients). Mann-Whitney Two-tailed test. g, Expression of
   Pip4k2c from the Arriaga et al., cohort comparing liver (n=7 specimens)
   vs lung (n=14 specimens), Wilcoxon rank-sum test. For all boxplots, n
   refers to the number of samples, and p refers to the p-value. The
   center line indicates the median, the box limits denote the first and
   third quartiles, and the whiskers indicate the lowest or highest data
   points at the first quartile minus or plus 1.5 times the interquartile
   range; Significance levels as indicated. h, Cumulative results of
   selected top differentially abundant metabolites in animals with liver
   and/or lung metastases following injection with either parental
   (PIP4K2C^WT) or PIP4K2C^KO A375 human melanoma cells; n=5 mice per
   group; unpaired two-sided-t-test. Lactate (p=2.36727E-05), Pyruvate
   (p=0.03), Citrate (p=0.25), a-Ketoglutarate (p=0.18), Succinate
   (p=0.004), Fumarate (p=0.04), Malate (p=0.04), Hexose (p= 6.72275E-06).

   To determine the relevance of these phenotypic differences in a patient
   context, we performed RNA-sequencing of melanoma LMs and combined these
   with publicly available data ([222]Methods)^[223]30, totaling 252
   metastasis transcriptomes, including 93 liver metastases
   ([224]Supplementary Tables 4 and [225]5). Comparison of LM
   transcriptomes with lung metastases revealed strong enrichment of
   metabolic pathways in LM, including TCA cycle, OxPhos, and mTORC1
   signaling ([226]Fig. 4c, [227]Extended Data Fig. 4 h–[228]j, and
   [229]Supplementary Tables 4 and [230]5). These results were maintained
   when including only samples with high purity, reducing potential
   transcriptome contamination by healthy hepatocytes ([231]Extended Data
   Fig. 4j,[232]k, [233]Supplementary Tables 4 and [234]5). Additionally,
   our scRNA-seq pathway analyses were exclusively performed on tumor
   cells yielded comparable results. To comprehensively determine the role
   of these observations in a large real-world data set, we analyzed 3,511
   cutaneous melanoma patients, including 2404 primary cutaneous melanomas
   (PCMs), 364 LMs and 743 lung metastases ([235]Supplementary Table 4).
   For all specimens, tumor regions were enriched through macro-dissecting
   and profiled using RNA-sequencing and exome sequencing (either
   whole-exome or 592 gene panel DNA-sequencing) at Caris Life Sciences
   (Phoenix, AZ). We found strong enrichment of metabolic pathways,
   including TCA cycle and OxPhos in LMs compared to PCMs and lung
   metastases ([236]Fig. 4d), which mirrored our observations in mouse
   models ([237]Fig. 4a,[238]b) and publicly available patient data
   analyzed here ([239]Fig. 4c). Overall, these transcriptomics data
   suggest that cancer cells in the liver enrich for key metabolic
   pathways when compared to other metastatic sites, both in melanoma and
   other cancers. In the same patient cohort, we found that PIP4K2C, but
   PIP4K2A or PIP4K2B, had a lower expression in LMs compared to PCMs and
   lung metastases (p<0.0001 and p=0.08, respectively) ([240]Fig. 4e,
   [241]f. [242]Extended Data Fig. 4l. [243]m). Additionally, comparison
   of genomics of LMs with either PCM or lung metastases in these patients
   showed distinct distribution of common oncogenic driver mutations,
   including BRAF, NRAS and NF1, and revealed an enrichment of the less
   commonly mutated (in melanoma) gene PIK3CA encoding for p110α, the
   catalytic subunit of PI3-kinase in LMs compared to PCMs and lung
   metastases (4.27% vs. 1.21%, p=0.003, and 4.27% vs. 2.34%, p=0.19,
   respectively) ([244]Extended Data Figure 4n). Similarly, we found lower
   expression of Pip4k2c, but not Pip4k2a/b, in LMs compared to other
   metastatic sites in the Arriaga et al., prostate cancer metastatic
   mouse model data ([245]Fig. 4g, [246]Extended Data Fig. 4o).

   We next sought to directly measure potential differences in metabolites
   between liver and lung metastases. For this purpose, we injected A375
   cells (parental or Pip4k2c^KO) and collected liver and lung metastases
   ([247]Methods). We performed mass spectrometry on five liver and five
   concurrent lung metastases from animals injected with Pip4k2c^KO cells,
   and three liver and five lung metastases from mice injected with
   Pip4k2c^WT cells. Principal component analysis (PCA) demonstrated that
   across and within each genotype, disease site (liver vs. lung) was the
   most important driver of variability in metabolic profiles
   ([248]Extended Data Fig. 5a,[249]b, [250]Supplementary Table 6).
   Several key metabolites of the TCA cycle, including a-ketoglutarate,
   succinate, fumarate and malate were found at higher abundances in liver
   compared to lung metastases ([251]Figure 4h, [252]Extended Data Fig.
   5c–[253]e), consistent with the findings indicated from our
   transcriptomics analyses ([254]Figure 4a–[255]d).

INSR is necessary for liver-metastatic organotropism

   Our data suggest that co-optation and metabolic adaptions to the
   insulin-rich liver milieu are strongly associated with liver-metastatic
   organotropism. To determine whether insulin acting on cancer cells is
   necessary for both liver-metastatic organotropism and increased disease
   burden in response to systemic PI3K inhibition, we transduced parental
   or Pip4k2c^KO cells with a doxycycline(dox)-inducible short hairpin RNA
   (shRNA) targeting the insulin receptor (Insr) to generate otherwise
   isogenic, conditional Insr knockdown (Insr^shIR) or
   Pip4k2c^KO/Insr^shIR cells ([256]Fig. 5a). As expected, we found a
   significantly impaired induction of Akt phosphorylation in response to
   insulin in both dox-induced Insr^KD and Pip4k2c^KO/Insr^shIR cells,
   compared to their non-dox treated counterparts. Across all conditions,
   GDC-0941 effectively abrogated p-AKT ([257]Extended Data Fig. 6a).
   Next, we injected mice with dox-treated or untreated
   Pip4k2c^KO/Insr^shIR tumor cells followed by treatment with vehicle or
   GDC-0941. Insr knockdown alone resulted in LM reduction compared to
   non-Dox treated animals (albeit not statistically significant), and in
   mice treated with GDC-0941, Insr knockdown in Pip4k2c^KO tumor cells
   resulted in a significantly reduced LM burden ([258]Fig. 5b). Across
   all conditions, there was no change in lung metastasis burden ([259]Fig
   5c). Overall, these results suggest that while GDC-0941 effectively
   abrogates Akt phosphorylation in vitro, host responses to the drug and
   subsequent insulin-dependent circuitry increase metastasis to the
   liver, but not in the lung.

Figure 5. Genetic, dietary or pharmacological strategies disrupting systemic
host effects of PI3K-inhibition reduces metastatic liver organotropism.

   Figure 5.
   [260]Open in a new tab

   a, Immunoblot probing Insr in Pip4k2c^KO/Insr^shIR cells prior to use
   for in vivo experiments (in b-c). b, Liver metastasis disease burden
   (in mm^3) following tail vein injection of Pip4k2c^KO/Insr^shIR without
   dox, GDC-0941 alone, dox-induced Insr knockdown alone, and combination
   of dox-induced Insr knockdown and GDC-0941. c, Lung metastasis burden
   (metastasis count) in corresponding animals from (b); n=10 mice per
   group. d, Blood glucose levels (in mg/dL) measured in mice (n=7 mice
   per group) following treatment with indicated diet or drugs (30 minutes
   post-administration) or vehicle control. e, C-peptide levels (in pM)
   following treatment with indicated combinations of GDC-0941 with SGLT2
   inhibition or animals fed with ketogenic diet (n=11 mice per group). f,
   Liver metastasis disease burden (in mm^3) following tail vein injection
   of Pip4k2c^KO cells with drug treatments (SGLT2i, GDC-0941), ketogenic
   diet or combinations thereof in indicated groups. g, Lung metastasis
   burden (metastasis count) in corresponding animals from (f); n=10 mice
   per group, mean± s.e.m.; h, Tumor growth of subcutaneously injected
   HCmel12 parental or Pip4k2c KO melanoma cell lines following no
   treatment, treatment with GDC-0941 alone, or combination of GDC-0941
   plus SGLT2 inhibition; n=5 mice per group, mean± s.e.m. i, Tumor
   diameter at day 24 in indicated treatment and genotype groups from (h).
   Box plots denote the minima, maxima and median in (h). Statistical
   significance was determined using One-Way ANOVA Tukey’s multiple
   comparisons test for (b-h). Data are representative of three (b-g) and
   two (h, i) independent experiments. Significance levels as indicated.

Combination interventions reduce LM burden

   We reasoned that abolishing the PI3K-i-induced forward loop could
   reduce liver metastatic burden. To test this, we treated animals with
   an inhibitor of sodium-glucose co-transporter 2 (SGLT2i), that inhibits
   glucose re-uptake in the kidney, thereby, preventing spikes in insulin
   levels, or we fed animals a ketogenic diet that depletes glycogen
   storage (Methods). Peak blood glucose and plasma C-peptide levels were
   strongly reduced when combining GDC-0941 with SGLT2i or a ketogenic
   diet ([261]Fig. 5d,[262]e). In Pip4k2c^KO LM-bearing animals,
   combination of either SGLT2i or a ketogenic diet with GDC-0941
   significantly reduced liver metastasis compared to animals receiving
   GDC-0941 alone ([263]Fig. 5f). Notably, combination of ketogenic diet
   with GDC-0941 also significantly reduced lung metastatic burden, while
   SGLT2i plus GDC-0941 selectively targeted liver-metastatic disease
   ([264]Fig. 5g), suggesting that addition of SGLT2i and a ketogenic diet
   to PI3K-i could circumvent GDC-0941-induced increase in LM.
   Importantly, compared to parental cells, loss of Pip4k2c had no
   differential impact on growth of subcutaneously implanted tumors or
   responses to GDC-0941 or GDC-0941 plus SGLT2i ([265]Fig. 5h,[266]I,
   [267]Extended Data Fig. 6b). This is consistent with the initial
   CRISPR-Cas9 screen and suggests PI3Ki-based therapeutic combinations
   are critical for abrogating insulin-mediated circuits that are most
   pronounced in LM ([268]Fig. 6).

Figure 6. Schematic of Liver metastatic organotropism.

   Figure 6.
   [269]Open in a new tab

   Model illustrating increased liver metastatic burden in animals
   injected with Pip4k2c^KO cells, (1) effects of systemic PI3K
   inhibition, and strategies to overcome compensatory feedback loop and
   increased liver metastatic burden with either (2) ketogenic diet or (3)
   SGLT2-inhibition.

Discussion

   Not all metastatic sites are created equal. Compared to other sites,
   metastasis to the liver confers a poorer prognosis across multiple
   cancer types and reduced responses to modern cancer
   therapies^[270]4,[271]31,[272]32. Mechanisms that determine a cancer
   cell’s metastatic fate towards one organ site over another are poorly
   understood and are not sufficiently accounted for by a model of serial,
   divergent acquisition of somatic mutations^[273]33,[274]34. The fact
   that cancer types with different oncogenic dependencies and distinct
   lineages exhibit a shared preponderance for metastasis to a specific
   organ - for example, liver metastasis in both melanoma and colorectal
   cancer, suggests that target-organ specific adaptations occur and may
   be determined by the required metabolic requirements and phenotypic
   plasticity^[275]35.

   Dissecting the molecular underpinnings of LM has been challenging due
   to limited pre-clinical models that recapitulate the liver-metastatic
   cascade in the presence of an intact immune system. Recent studies
   using intra-hepatic or portal/splenic vein injections of cancer cells
   revealed important but narrow aspects of LM biology and
   growth^[276]36,[277]37. Here, we leveraged a syngeneic melanoma model
   that has the potential to develop LM upon tail vein injection, thus,
   recapitulates several bottle-necks in the LM cascade. We performed
   functional screens and find that cancer cell intrinsic loss of Pip4k2c
   promotes melanoma cell metastasis by co-opting an insulin-rich liver
   milieu via activation of the PI3K/AKT pathway ([278]Fig. 6), without
   altering primary tumor growth or their metastatic potential to other
   organ sites (e.g., lung). Importantly, this mechanism may apply for
   both routes of dual blood supply to the liver, as we demonstrate
   increased LM via systemic circulation in our melanoma model and a
   prostate cancer model, and via portal vein, using a colorectal cancer
   model implanted into the colonic submucosa.

   Recent studies demonstrate that hyperinsulinemia, either in the context
   of insulin resistance or in response to systemic PI3K inhibition,
   increases tumor growth of subcutaneously implanted patient-derived
   xenograft models through inhibition of host
   PI3K^[279]28,[280]38,[281]39. Our study extends on this observation and
   shows that insulin-mediated tumor growth disproportionally occurs in LM
   over primary tumors or other metastatic sites, in line with higher
   insulin levels in the liver. These effects could be abrogated by
   reducing the insulin gradient between the liver and other organs
   through systemic delivery of insulin. Furthermore, combination of
   systemic PI3K inhibition with either inhibition of SGLT2 or a ketogenic
   diet could disrupt the effects of host-mediated increase in insulin
   levels and reduce LM, while having little or no effect on lung
   metastasis burden. These findings have potential clinical relevance in
   specific disease contexts. For example, patients with hormone receptor
   positive breast cancer, for which PI3K inhibitor alpelisib is
   FDA-approved and ongoing studies test the benefit of adding SGLT2
   inhibition or a ketogenic diet ([282]NCT05090358), may particularly
   benefit from such strategies, when LM are present. While activating
   mutations of PIK3CA are frequent in breast cancer and serve as the
   basis for the use of PI3K inhibitors in breast cancer, our study
   suggests that this oncogenic dependency may not be necessary to select
   patients toward PI3K inhibitor-based combination treatment strategies,
   thus, expanding the potential pool of patients who may benefit from
   such therapies.

   Our study has potential limitations: Fist, loss or reduced expression
   of Pip4k2c likely represents one example of resulting
   hypersensitization to insulin, and may have effects beyond regulation
   of this pathway that we did not measure^[283]40,[284]41. Our initial
   screen only targeted kinases, thus, more extensive unbiased approaches
   are necessary to characterize the entire landscape of potential drivers
   of LM development. Second, analysis of human data showed that PIK3CA
   mutations are infrequent in melanoma, but enriched in melanoma LM,
   thus, suggesting that genomic activation of the pathway may directly
   drive LM in a portion of patients, independent from regulation by
   Pip4k2c or its isoforms. Importantly, even in the presence of PIK3CA
   mutations, insulin is still required to activate the pathway. Third,
   transcriptomics analyses of human data may suffer from contamination by
   normal hepatocytes in LM, and respective parenchymal and infiltrating
   cells in other analyzed organ sites. However, it is unlikely that this
   factor had a major effect, as we observed similar pathway enrichment in
   mouse metastases profiled by single-cell RNA-sequencing, where analyses
   were restricted to malignant cells. Lastly, further work is required to
   determine the role of niche cells and the tumor-microenvironment in
   liver-metastatic organotropism.

Methods

Ethics statement

   The research performed in this study complies with ethical regulations.
   Human specimen research was approved by the Institutional Review Board
   at Columbia University Irving Medical Center (#AAAO5706). All mouse
   experiments were performed under IACUC approved animal protocols at
   Columbia University (AABE6570) and Harvard Medical School (AABQ9616).
   The maximum subcutaneous tumor size permitted by both IACUCs is 20mm
   and was never exceeded. Source data is provided for all in vivo
   experiments.

Mice

   6–8 -week-old C57BL/6J (Charles River: 027), NOD/SCID/IL2gR (NSG) (JAX:
   005557) and/or BALB/c (JAX: 000651) female mice were obtained from
   Charles River and/or Jackson Labs. Mice received a normal chow diet
   (PicoLab Rodent 20 5053 laboratory Diet St. Louis, MO) or a ketogenic
   diet (Thermo-Fisher AIN-76A) where indicated, with free access to
   drinking water. Diet composition is described in the [285]Supplementary
   Table 7. Mice were housed under pathogen-free conditions. Sex was not
   considered in the study design.

Cell culture

   Mouse melanoma cell line HCmel12 was kindly provided by Prof. Thomas
   Tüting, Magdeburg, Germany. Melanoma cells were grown in ‘complete’
   RPMI-1640 medium (Gibson Bioscience) as described previously^[286]13.
   RPMI 1640 supplemented with 10%FBS (Gibco^™ 10437028), 10mM
   non-essential amino acids, 1mM HEPES (all from Life Technologies), 20
   μM 2-mercaptoethanol (Sigma). Mouse colorectal carcinoma cell line CT26
   was cultured in RPMI 1640 medium supplemented with 10% FBS. A375 and
   HEK293 cells were purchased from ATCC and grown in DMEM supplemented
   with 10% FBS. Cell lines were routinely tested for mycoplasma
   contamination using Lonza MycoAlert Kit (Cat. # LT07–318).

Generation of a Cas9-expressing HCmel12 melanoma cells

   Cas9 lentivirus was generated from HEK-293T cells transfected with
   pLX-311Cas9v.2 plasmids (Addgene #118018) and packaging plasmids pMD2.G
   (Addgene, #12259) and psPAX2 (Addgene, #12260) using TransIT-LT1
   Transfection reagent (Mirus). HCmel12 melanoma cells were transduced
   with Cas9 lentiviral particles followed by blasticidin selection
   (1μg/ml) for 48 hours. To assess Cas9 cutting efficiency in
   HCmel12-Cas9 expressing cells, cells were transduced with pXPR_011
   (Addgene #59702; gift from John Doench and David Root). Lentivirus
   expressing EGFP, puromycin resistance and EGFP-targeting sgRNA. Cas9
   activity activity was estimated based on the fraction of EGFP-negative
   cells 7 days later using FACS Canto flow cytometry (BD Biosciences) and
   analyzed with FlowJo software (TreeStar, v10 for Mac). The fraction of
   EGFP-negative was used to estimate Cas9 efficiency.

Propagation of Kinome sgRNA libraries and Lentivirus production

   Mouse Kinome CRISPR Knockout Library (Brie) was purchased from Addgene
   (Addgene, #75316), that includes 2,852 sgRNAs targeting 713 kinase
   genes and 100 control non-targeting sgRNAs ([287]Supplementary Table
   3). For library production, HEK293T cells were seeded in DMEM + 10% FBS
   in T175 flasks one day prior. At 80% confluency, cells were transfected
   with the Brie library and the packaging/envelope plasmids pMD2.G
   (Addgene, #12259) and psPAX2 (Addgene, #12260) using TransIT-LT1
   Transfection reagent (Mirus). After 6 hours media was replaced with
   DMEM + 20% FBS, and virus was collected after 48 hours, filtered, and
   stored at -80 °C.

In vivo CRISPR-Cas9 screen

   HCmel12-Cas9 melanoma cells were transduced at a MOI < 0.1 with Brie
   lentivirus at ≥1000 cells/construct per infection replicate (~3x10^7
   pre replicate). Cells were spin-infected at 1000 x g, 37°C for 2 hours
   and incubated for 16 hours followed by a media change. On day 2 post
   transduction the cells were put into Blasticidin selection (1 μg/ml)
   for 7 days and Puromycin (1 μg/ml) for the remainder of the experiment.
   On day 14, 3x10^6 of library-transduced HCmel12-Cas9 cells were
   injected into the right flank or lateral tail vein of the mice. In
   parallel cells were kept in culture for 21, 28, 35, 42, 49,52 days.
   Tumor growth and signs of sickness were monitored over time. Mice were
   euthanized after 4 weeks and tumors and metastases were evaluated for
   metastasis presence, lesions were dissected for gDNA isolation, sgRNA
   amplification, and sequencing. Genomic DNA was extracted using Qiagen
   DNeasy Blood & Tissue Kit (69506, Qiagen). sgRNA amplification and
   sequencing library construction were performed as described
   previously^[288]15, and sgRNA library was sequenced on a Illumina
   HiSeq2500 Instrument. CRISPR–Cas9 screens were analyzed and visualized
   using software packages MAGeCK and MAGeCKFlute R (v4.0.3)^[289]42.

Generation of KO, knockdown and gene-overexpressing cell lines

   Individual tumor cell KOs were generated using Cas9 ribonucleoproteins
   (RNPs) nucleofection as described previously^[290]43. Target sequences
   were designed using CHOP-CHOP tool ([291]https://chopchop.cbu.uib.no)
   or sequences were derived from the original kinome library (sequence
   information in [292]Supplementary Table 8). For insulin receptor
   knockdown experiments doxycycline-inducible shRNA was used as described
   previously^[293]28 (Sequence information in Extended Data Table 8.
   Briefly, knockdown of insulin receptor was achieved using a
   doxycycline-inducible shRNA strategy. For generation of miR-E shRNAs,
   97-mer oligonucleotides were purchased coding for predicted shRNAs
   using Splash RNA, [294]http://splashrna.mskcc.org/27. Oligonucleotides
   were PCR amplified using the primers miRE-Xho-fw
   (5′-TGAACTCGAGAAGGTATATTGCTGTTGACAGTGAGCG-3′) and miRE-Eco-rev
   (5′-TGAACTCGAGAAGGTATATTGCTGTTGACAGTGAGCG-3′). To generate full-length
   and N-terminal mutant Pipk4k2c overexpression constructs, open-reading
   frames were cloned into pLX305 using Gateway cloning
   ([295]Supplementary Table 8). Human pDONR223-PIP5K2C (Addgene, #23450)
   and N-terminal mutant human PIP4K2C, ORF were cloned into pLX307 using
   Gateway cloning and Gibson cloning, respectively. All vectors were
   sequence verified using Sanger sequencing followed by whole plasmid
   sequencing.

Proliferation assay

   1000 cells were seeded in 96-well plates. After 1, 2, or 3 days of
   proliferation, cells were fixed in 4% PFA and stained with DAPI.
   Timepoint zero was used as a reference control. Images were acquired on
   IVOS microscope, and analysis was performed using Cell Profiler
   software ([296]https://cellprofiler.org).

Transwell migration assay

   Migration was assessed using transwell assays. 2×10^4 melanoma cells
   pre-treated with insulin were placed in the upper compartment of a
   transwell chamber with 8.0 μm pore size (Corning, 3422) containing 300
   μl (top) and 700 μl media in the lower chamber. Cells were led to
   adhere and transmigrate through the membrane at 37 °C 5% CO2. Wells
   treated with no insulin served as controls. After 16 hours,
   transmigrated cells on the lower surface of the membrane were stained
   with CFSE, fixed with 4% paraformaldehyde and the mean number was
   counted in 5 high power fields using Inverted Microscope (Olympus).

Immunoblots

   Cells were lysed in 1x Laemmli buffer (10,000 cells per μl) followed by
   5min incubation at 95°C. Lysates were separated by 10% SDS–gel
   electrophoresis and transferred to a PVDF membrane (IB24001, Fisher
   Scientific). Membranes were blocked for 1 h (11921673001, Roche
   blocking buffer) in Tris-Buffered Saline, pH 7.6 (TBS, S196830–2,
   Agilent Technologies) and probed with primary antibodies overnight at
   4°C followed by 1h incubation in secondary antibodies at RT. Proteins
   were detected using the Odyssey CLx Imaging system (LI-COR
   Biosciences). Antibody information is provided in the Reporting Summary
   of the paper. Membranes were stripped using Stripping Buffer (46430,
   Thermo Scientific). Full blots corresponding to the immunoblots shown
   in the main and [297]supplementary figures are provided as a Source
   File.

RNA isolation, RNA sequencing and analysis

   2x10^6 parental and Pip4k2c^KO HCmel12 melanoma cells were plated in
   6-well plates FBS-free media overnight, with three biological
   replicates per condition, total 12 samples. The next day cells were
   treated with insulin (125ng/ml) for four hours ([298]Supplementary
   Table 4). RNA was extracted using Qiagen kit (74136) and sequenced on a
   HiSeq2500 (Illumina). Transcripts were quantified with Salmon^[299]44
   using the mouse reference genome release GRCm38.97. The tximeta R
   package (v.4.3)^[300]45 was used to import transcript-level counts and
   aggregate them by gene. DESeq2 package was applied to generate
   normalized gene counts. Gene set enrichment analysis (GSEA) for pathway
   genes was identified using BROAD java GSEA software with normalized
   gene counts as input ([301]https://www.gsea-msigdb.org/gsea/index.jsp).

In vitro drug treatment

   Cells were washed 1x in PBS, placed in FBS-free media overnight, and
   treated for 1h with PI3K inhibitors (Pictilisib/GDC-0941 (S1065) and
   Alpelisib (S2814) from Selleck Chemicals) as indicated with insulin for
   15 min before collection for immunoblotting. To induce insulin receptor
   knockdown cells were treated with 1μg/ml doxycycline (D9891–5G, Sigma)
   for 48 hours, washed 1x with PBS, and subjected to signaling studies.

Animal metastasis and drug treatment studies in vivo

   5x10^5 HCmel12 mouse melanoma cells, 1x10^5 A375 human melanoma cells
   were injected into the lateral tail vein. Mice were monitored over
   three to four weeks. When mice showed signs of sickness, 20% of weight
   loss, impaired activity, hunching, and limb paralysis, animals were
   euthanized, and organs were harvested for evaluation on metastasis
   presence. For in vivo drug treatment studies, Pictilisib (GDC-0941,
   HY-50094) and Canagliflozin (JNJ 28431754, HY-10451) were purchased
   from Medchem Express (Monmouth Junction, NJ). The day after tumor cells
   injection mice were given vehicle control or treated (daily, 5 out of 7
   days) with GDC-0941 (100 mg/kg) or SGLT2i (6 mg/kg) prepared in ‘CMC’
   solution, 0.5% carboxy-methyl-cellulose sodium (CMC, C9481–500G, Sigma)
   in distilled water supplemented with 0.2% Tween-80 (BCBW1801, Sigma)
   for the duration of the experiment unless otherwise stated. To maintain
   Insr knockdown in vivo, mice bearing Pip4k2c^KO cells with
   dox-inducible Insr shRNA were administered 3 mg/kg of doxycycline
   (Sigma, D3072) intraperitoneally daily, for three to four weeks.
   Ketogenic diet was started 10 days prior to tumor cell injection where
   indicated. In experiments combining GDC-0941 and SGLT2 inhibitor
   (Canagliflozin), the latter was administered 60 min prior to
   GDC-0941^[302]28 administration. 2x10^5 tumor cells were injected
   subcutaneously and tumor growth was monitored over time. When tumors
   were palpable mice were treated with GDC-0941 (100mg/kg) alone or in
   combination with SGLT2i (6mg/kg), 3 times a week until the end of the
   experiment. Mice were euthanized and tumor growth was evaluated.

Endoscopy-guided orthotopic tumor transplantation and injections

   Orthotopic injections of CT26 cells was performed as described
   previously^[303]46. Briefly, cell lines were split 2 days prior. On the
   day of injection, cells at 80–90% confluency were harvested by TrypLE
   Express enzyme (Gibco, 12604013). Tumor cells were resuspended in PBS +
   10% Matrigel at 2x10^6/ml. 100μl was injected into the colon sub-mucosa
   of each BALB/c recipient mouse by optical colonoscopy using a Hamilton
   syringe (7656–1) and a custom 33G needle (Hamilton, custom made similar
   to 7803–05, 16”, Pt 4, Deg 12). Successful injections were confirmed by
   observing large bubbles in the colon mucosa. After 14 days mice were
   euthanatized and organs evaluated for metastasis presence. Optical
   colonoscopy was performed two weeks later to assess tumor formation
   using a Karl Storz Image 1 HD Camera System, Image 1 HUB CCU, 175-Watt
   Xenon Light Source, and Richard Wolf 1.9mm/9.5 Fr Integrated Telescope
   (part number 8626.431). One injection was performed per mouse.

Implantation of Insulin Pads in NSG mice

   NSG mice were implanted with 2x3 mm slow-release insulin pellets
   (LinBit - LinShin, Canada) on the dorsum under the skin. Under
   isoflurane anesthesia, mice were shaved, the surface of the skin
   treated with a 70% isopropyl alcohol and a lidocaine-based cream (e.g.
   Emla). The insulin pellet was inserted using 12G trocar and the skin
   opening is sealed with a drop of Vetbond. After a day, 1x10^5 tumor
   cells were injected via lateral tail vein and tumor metastatic burden
   was evaluated after four weeks.

Blood glucose, C-peptide and tissue insulin measurements

   10 μl of blood was taken from the tail of mice before treatment and at
   indicated time points (15, 30, 60, 90, 120,180, 240 min) and blood
   glucose was measured using a Prodigy AutoCode Glucometer. At the end
   point >100 μl of blood was drawn from the mice into heparin coated
   tubes (07980, Stem Cell Technologies). Blood was centrifuged (10,000g
   for 10 min at 4 °C), and plasma was stored at −80 °C. Mouse and human
   C-peptide levels in plasma and mouse insulin levels in liver and lung
   tissue were measured using commercial ELISA kits (Mouse C-peptide
   ELISA: 80-CPTMS-E01 (ALPCO); STELLUX Chemiluminescence Human C-peptide
   ELISA: 80-CPTHU-CH01 (ALPCO); Salem, NH; Insulin: Invitrogen cat.
   #EMINS).

Tissue collection

   Liver and lung tissue were collected immediately after mice were
   sacrificed, and snap-frozen in liquid nitrogen. For tissue protein
   isolation, 5mg of powdered liver and lung tissue were homogenized in
   300μl Mammalian Protein Extraction Reagent (catalog number: 78501)
   buffer with proteinase inhibitors (catalog number: 1861281) and
   incubated at 4 °C for 2 hours with rotating, followed by 20min spin at
   20.000g, 4 °C. Supernatant was collected and stored at -80°C.

FDG-PET

   C57/BL6 mice were injected with Pip4k2c^KO melanoma cells. After 3
   weeks mice were scanned for liver metastasis presence using PET with
   [F18] fluorodeoxyglucose (FDG) tracer. At the time of peak blood
   insulin feedback 60 min post-GDC-0941 injection animals were injected
   IV with FDG. 60 min later mice were placed into a 4-mouse bed and 30
   min static PET images were acquired using a PET scanner (Siemens,
   Munich, Germany) followed by micro CT (MILabs, Houten, Netherlands) on
   the same bed for anatomical references. Regions of interests (ROI) were