Altered Mitochondria Functionality Defines a Metastatic Cell State in Lung Cancer and Creates an Exploitable Vulnerability

Lung cancer is the most lethal cancer type, causing over 145,000 deaths per year in the United States alone (1). Metastatic cancer spread is a major reason for cancer-related deaths. However, despite the importance of targeted therapies directed against advanced-stage cancer, treatment options for metastatic cancer remain limited. Genomic analysis of human lung adenocarcinoma, which is a major lung cancer sub-type, has led to the identification of targetable oncogenic alterations, including EGFR mutations and ALK translocations (2–7). Despite the clinical value of targeted therapies directed toward these cancer-specific alterations, many tumors do not contain genomic alterations that can be targeted for therapeutic benefit. Furthermore, even when targeted therapies exist, responses are often short lived. Hence, there remains a critical need to better understand the metastatic cancer state and uncover targetable pathways that are specific to late-stage cancer.

Global gene expression analysis of human tumors has uncovered changes in gene expression programs that predict patient outcome. These studies have contributed to our understanding of the mechanisms that lead to cancer progression and drug resistance (8). Unfortunately, these studies have only rarely identified key therapeutic targets as gene expression does not necessarily implicate function. Alternatively, functional genomics focuses on unbiased screening coupled with functional readouts to uncover pathways that could have therapeutic potential (9). Extensive efforts to quantify the effect of pooled shRNA and sgRNA libraries across panels of human cell lines have revealed important mediators of cancer cell survival (10, 11). However, the relatively small number of cell lines from individual cancer types, extensive genomic heterogeneity, and derivation from primary tumors and metastases of different origins often render these studies underpowered to identify molecular features that uniquely distinguish the nonmetastatic from the metastatic cell state. These complications suggest that more defined models may have increased sensitivity and specificity to uncover novel aspects of valuable therapeutic targets.

To better understand the biological forces that drive lung cancer progression, genetically engineered mouse models of human lung adenocarcinoma have been generated with conditional alleles to express oncogenic Kras and inactivate the p53 tumor suppressor (12–14). Notably, this Kras^G12D;Trp53^KO lung adenocarcinoma model recapitulates the genetic events and histological appearance of early- and late-stage human tumors (3, 15). Thus, this model system represents a useful in vivo and ex vivo tool to identify novel regulators of lung adenocarcinoma progression and metastasis (16).

We previously generated cell lines from nonmetastatic primary lung tumors (T_nonMet) and metastases (Met) from the Kras^G12D;Trp53^KO mouse model (14, 16). These T_nonMet and Met cell lines maintain their differences in metastatic ability, exhibit dramatic differences in gene expression programs, and analysis of these cell lines has contributed to the discovery of several regulators of the metastatic process (14, 17–19). Expression of individual regulators and the overall gene expression program can predict human lung adenocarcinoma patient outcome, suggesting that these cell lines recapitulate at least some aspects of the metastatic state of human lung tumors (14, 17–19). This unique set of cell lines from early-stage nonmetastatic tumors and late-stage metastases provides an opportunity to investigate whether cancer progression creates unique vulnerabilities that can be exploited therapeutically.

To systematically uncover genes that are specifically required to sustain the growth and/or survival of cancer cells in the metastatic cell state, we performed a pooled genome-scale lentiviral-shRNA library screen. A targeted secondary in vivo screen validated top candidates, including a highly significant enrichment for mitochondria and mitochondria ribosome gene-sets associated with metastasis specific lethality. Analyses of mitochondrial membrane potential, reactive oxygen species (ROS) levels, mitochondrial respiration, glycolytic, and fuel utilization assays, as well as electron microscopy confirmed dysfunctionality of mitochondria as a discriminating feature of the metastatic state. These observations highlight a metastasis-specific vulnerability that can be therapeutically exploited with pharmacological targeting to block metastasis in vivo.

Selected methods are detailed in this section. An extensive version of this section is available as Supplementary Methods within the Supplementary Material of this article.

Kras^LSL-G12D, Trp53^flox and Rosa26^LSL-tdTomato mice have been described previously (12, 16). Tumors were initiated by intratracheal infection of mice with lentiviral vectors expressing Cre-recombinase as previously described (20). All animal studies and procedures were performed according to and approved by the Stanford Institute of Medicine Animal Care and Use Committee. Animals of both sexes were used.

The 368T1 and 394T4 cell lines were generated from primary lung tumors, the 238N1 and 482N1 cell lines were generated from lymph node metastases, the 404M1, 2691M1 cell lines were generated from distant organ metastasis, the 889PF from disseminated cancer cells within the pleural cavity, and the 579DLN cell line was generated from a distant lymph node metastasis of lung-adenocarcinoma-bearing Kras^LSL-G12D/+;Trp53^flox/flox (Kras^G12D;Trp53^KO) Kras^LSL-G12D/+;Trp53^flox/flox;Rosa26^LSL-Tomato/+ (Kras^G12D;Trp53^KO;R26^Tomato) mice (14, 16). All cell lines were regularly (∼monthly) confirmed to be Mycoplasma negative. 293T cell lines were obtained from the ATCC (ATCC Cat# CRL-3216, RRID:CVCL_0063) and identity was regularly confirmed by STR analysis. All cell lines were passaged at least once after thawing before use in subsequent experiments. The average time in culture was 4 weeks (range between 2 and 6 weeks), then a new vial was thawed.

Total RNA was isolated from 1 × 10⁶ cells of each line (368T1, 394T4, 238N1, and 482N1) using the Qiagen RNeasy mini kit. For each sample, 1 μg of total RNA was used for library construction. The integrity and quality of RNA were assessed before library construction using an Agilent Bioanalyzer. RNA-sequencing libraries were prepared using the Illumina TruSeq RNA kit according to the manufacturer's instructions. High-throughput sequencing was performed on Illumina HiSeq. For analysis of the RNA-seq reads, we performed differential gene and transcript expression analysis using TopHat (TopHat, RRID:SCR_013035) and Cufflinks (Cufflinks, RRID:SCR_014597). RNA-seq reads were separately aligned to the mouse genome (mm9) and the aligned RNA-seq reads were assembled into transcripts. Annotated transcripts were obtained from the UCSC genome browser (http://genome.ucsc.edu, UCSC Genome Browser, RRID:SCR_005780) and the Ensembl database (Ensembl, RRID:SCR_002344). Transcript abundances were measured in Fragments Per Kilobase of exon per Million fragments mapped (FPKM). Finally, Cuffdiff (Cuffdiff, RRID:SCR_001647) was used to define differential expression. The data have been deposited in NCBI's Gene Expression Omnibus [(GEO), RRID:SCR_005012] and are accessible through GEO Series accession number GSE159169 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE159169).

Lentiviral particles from the 100K mouse shRNA pool (obtained from Broad Institute: The RNAi Consortium; ref. 21) were transduced at low MOI (0.3; each shRNA was stably integrated into an average of 600 cells) into 4 cell lines (368T1, 394T4, 238N1 and 482N1; 3 replicates per cell line). Virus transduction and cell line passaging were performed according to the method described by Cheung and colleagues (10) with minor modification. In brief, the virus and cell mixture was seeded into a 12-well plate at approximately 2 mL per well, and cells were transduced by spinoculation at 1,000 × g for 2 hours with 12 μg/mL polybrene at 30°C. After 24 hours, the 12 wells from each replicate transduction were pooled, and the combined cells were transferred into a 15-cm dish. Cells were selected with 4 μg/mL puromycin starting day 4 post-transduction. For all subsequent passages, 7.2 × 10⁷ cells per replicate were carried over. The remaining cells for all passages were collected, resuspended in 1 mL of PBS, and stored at −20°C for genomic DNA isolation. Passaging for each cell line was continued for at least 22 population doublings. Puromycin selection was maintained for the entire experiment.

To identify the shRNA sequences enriched in screened cells, genomic DNA was isolated from cells collected at each timepoint, and hairpin sequences were quantified using Illumina DNA sequencing (22).

We applied the RNAi Gene Enrichment Ranking (RIGER), Kolmogorov–Smirnov (KS), and Weighted Sum (WS) algorithms to evaluate candidate genes from the pooled shRNA screens. KS is nonparametric and calculates gene scores from a collection of shRNA phenotype profiles, similar to the method described by Luo and colleagues (11). WS algorithms takes the combined sum of the first and second best ranks for hairpins for each gene. The best ranking hairpin is given a weight of 0.25 and the second best ranking hairpin is given a weight of 0.75. The weighted values (WS score) are determined by the sum of the first and the second weighted ranks, and genes are ranked by this WS score (https://software.broadinstitute.org/GENE-E/extensions.html). We considered all three ranking approaches to generate a list of top 50 candidate genes (Supplementary Fig. S2A) whose knockdown was found to impact only the Met cell lines in the screen (238N1 and 482N1), or only the Non-Met cell lines (368T1 and 394T4), or each cell line individually. Heatmaps for each set of genes were generated using the heatmap.2 function from the R package gplots (V 3.0.3). Clustering (depicted in Fig. 1C and D) was performed according to https://software.broadinstitute.org/GENE-E/doc.html. Each box represents the measure of global correlation between the entire dataset for one cell line versus another.

A list of candidate genes was generated from the initial screen as described above. In addition, control genes (12 “lethal to all” and 5 “advantageous to all,” as well as 32 inert genes) were chosen (Supplementary Fig. S2A). For the secondary validation screen, a pooled library of 512 TRC shRNA clones was produced by cherry picking these clones from the sequence-validated TRC shRNA library used in the first screen. Each gene was targeted by 2 shRNAs.

The viral packaging and collection procedure were established on the basis of the manufacturer's guidelines (https://portals.broadinstitute.org/gpp/public/resources/protocols). In brief, 293T cells were plated in ten 15-cm dishes (Falcon/Corning) at 80,000 cells per cm² in DMEM (Gibco) supplemented with 10% FCS (Invitrogen), 2 mmol/L l-glutamine (Invitrogen), and penicillin–streptomycin (Invitrogen). The next day, for each plate separately, a Packaging-Plus Mix was prepared consisting of 6 μg of shRNA library, 10 μg psMD2.G, 20 μg psPAX2 and 60 μL PLUS Reagent (Invitrogen) in a total volume of 1.3 mL OptiMEM (Invitrogen) and incubated for 15 minutes. To obtain a transfection mix, 1.4 mL of OptiMEM containing 90 μL of Lipofectamine (Invitrogen) was added to the Packaging-Plus Mix and incubated for another 15 minutes before adding dropwise to the plated 293T cells. The medium was changed 24 hours post-transfection. The medium from both the 48 and 72 hours harvests after transfection was combined and filtered through a 0.45-μm PES filter (Nalge Nunc). Lentivirus aliquots were stored at −80 °C until used for transduction of the selected cell lines.

To perform large-scale infections, 3.6 × 10⁶ target cells (MOI 0.3; each shRNA was stably integrated into an average of 2,000 cells) were transduced for each replicate. The virus-cell mixture was split across a 12-well plate at 2 mL per well. A spin infection was performed by centrifugation at 1,000 × g for 2 hours at 30°C. After 20 hours, the 12 wells from each replicate infection were pooled, and the combined cells were transferred into a 15-cm dishes. Transduced cells were selected with 4 μg/mL puromycin from day 4 after infection. For all subsequent passages, 1.2 × 10⁶ cells per replicate were carried over. The remaining cells for all passages were collected, resuspended in 1 mL of PBS, and stored at −20 °C for genomic DNA isolation. Passaging for each cell line was continued for at least 22 population doublings. Puromycin selection was maintained for the entire experiment.

To validate genes that were specifically required for the survival or growth in vivo, 3 × 10⁶ puromycin-selected transduced cells (238N1, 482N1) were injected into back flanks of 3 NSG-recipient mice. Mice were analyzed 4 weeks after transplantation. Tumors were disaggregated to single-cell suspensions by mincing with razor blades and treated with collagenase IV, dispase, and trypsin at 37°C for 30 minutes. Genomic DNA was isolated from cells collected from each tumor, and hairpin sequences were quantified using Illumina DNA sequencing as described above.

To normalize the read count of each hairpin, raw read counts were divided by total counts of each sample (per million). Hairpins with normalized count values <10 for across all early time point samples were eliminated from downstream analyses. Changes in shRNA representation over time were calculated by average read counts of late time points minus early time points.

To discover pathways driving pan- or metastasis-specific lethality top candidate genes identified in the screen were mapped to their direct human ortholog using information provided in the Mouse Genome Database (23) at www.informatics.jax.org. For Gene Set Enrichment Analysis (GSEA) analysis (ref. 24; SeqGSEA, RRID:SCR_005724), a rank-ordered list of candidate genes was analyzed using GSEA v2.2.0 software available from the Broad Institute (http://www.broad.mit.edu/gsea) and gene ontology (GO)-term analysis from The Database for Annotation, Visualization and Integrated Discovery (DAVID; https://david.ncifcrf.gov/, DAVID, RRID:SCR_001881). Only genes with a human homolog were used for this analysis.

For cell sorting, primary tumors and metastases were dissociated using collagenase IV, dispase, and trypsin at 37°C for 30 minutes. After dissociation, the samples were continually on ice, in contact with ice-cold solutions, and in the presence of 2 mmol/L EDTA and 1 U/mL DNase to prevent aggregation. Cells were stained with antibodies to CD45 (BioLegend Cat# 103133, RRID:AB_10899570), CD31 (BioLegend Cat# 102423, RRID:AB_2562186), F4/80 (BioLegend Cat# 123131, RRID:AB_10901171), and Ter119 (BioLegend Cat# 116233, RRID:AB_10933426) to exclude hematopoietic and endothelial cells (defined in combination as Lineage^negative). To analyze mitochondria membrane potential, cells were stained with 200 nmol/L Mitotracker deep red (Invitrogen) for 15 minutes at 37°C. A fixed number of GFP-positive control cells was spiked into each sample for normalization to account for staining artifacts due to different cell numbers in the samples. DAPI was used to exclude dead cells. FACSAria sorters (BD Biosciences) were used for analysis. To analyze ATP levels, equal numbers of cells were analyzed with the reversed CellTiter-Glo Luminescent Cell Viability Assay (Promega) according to the manufacturer's instructions, using an ATP standard curve for calibration.

Six- to 10-week-old NSG mice of similar weights, randomized for both male and female animals, were used for cell transplantation experiments. For the subcutaneous injections, 5 × 10⁴ 889PF or 579DLN cells were injected into both flanks of NSG recipient mice. Treatment of mice was started 7 days after transplantation with either 100 mg/kg phenformin or vehicle by oral gavage or 9 days after transplantation with either 1 mg/kg doxycycline or vehicle by intraperitoneal injection once daily. Mice were analyzed 3 or 4 weeks after transplantation, respectively. For the intravenous injections 579DLN cells were treated with either phenformin (200 μmol/L) or vehicle for 48 hours in vitro. Then cells were trypsinized and 1 × 10⁵ cells were injected intravenously into the lateral tail vein. Starting at the day of injection, mice were treated daily either with 100 mg/kg phenformin or with vehicle control by oral gavage. For the short-term intravenous experiment, 579DLN cells were treated with either phenformin (200 μmol/L), doxycycline (15 μg/mL), FCCP (2 μmol/L), or vehicle for 48 hours in vitro. Then cells were trypsinized and 5 × 10⁵ cells were injected intravenously into the lateral tail vein of immunocompetent 129/Bl6 F1 mice. Mice were sacrificed either three or seven days after injection and lungs were harvested, weighed, and analyzed by fluorescence microscopy and flow cytometry. During analysis of the animals, the performing scientist was blinded toward experimental groups. No statistical method was used to predetermine sample size. The Stanford Institute of Medicine Animal Care and Use Committee or the Landesamt für Natur, Umwelt- und Verbraucherschutz des Landes Nordrhein-Westfalen (LANUV) approved all animal studies and procedures, respectively.

Graphs and statistics were generated using the GraphPad Prism software (GraphPad Prism, RRID:SCR_002798). Significance, where indicated, was calculated using the Wilcoxon test for non-normally distributed data. Significance was determined as a P value <0.05. No statistical method was used to predetermine sample size.

To systematically uncover genes that are specifically required for survival or growth of metastatic cancer cells, we performed a pooled lentiviral-shRNA library screen in two T_nonMet (368T1 and 394T4) and two Met (238N1 and 482N1, both from lymph node metastases) lung cancer cell lines derived from the Kras^G12D;Trp53^KO mouse model (Fig. 1A and B). This screen included >92,000 shRNAs targeting approximately 20,000 genes, thus, enabling a broad interrogation of cancer cell states and having the potential to uncover genes essential for growth of metastatic cells. Importantly, both T_nonMet and Met lines share similar oncogenic and tumor-suppressor alterations that likely increases our ability to detect significant effects and implicate candidate genetic targets. To identify genes that are essential for each cell line, we used high-throughput sequencing of the shRNAs in the cancer cell populations before and after 22 population doublings (“early time point” and “late time point,” respectively, Fig. 1A). As anticipated, many shRNA targets were lethal to all cell lines, which led to clustering of the “early time point” samples away from the “late time point” samples (Fig. 1C; Supplementary Data Table S1). Among “late time point” samples, the diversity between Met cell lines was greater when compared with each other than the diversity among T_nonMet cell lines (Fig. 1D). shRNAs that were underrepresented in all cell lines were greatly enriched for genes with known essential pathway functionality (Supplementary Fig. S1A–S1D). Importantly, this genome-scale screen identified genes specifically required for the survival or growth of cancer cells in either the nonmetastatic state or the metastatic state as well as for individual cell lines (Fig. 1E; Supplementary Fig. S1E; Supplementary Data Table S2).

Comparative analyses using KS, WS, and best-ranking shRNA algorithms revealed a list of shRNAs and genes most prominently differentially affecting the growth of T_nonMet versus Met cell lines (Supplementary Data Tables S3–S5). To investigate the expression of these genes in the T_nonMet (368T1 and 394T4) and the Met (238N1 and 482N1) cell lines, we also performed RNA-sequencing based gene expression analysis (Supplementary Data Table S6). We incorporated pathway-level analysis based on identified shRNA effect and expression levels to define candidates for further functional validation. On the basis of these analyses, we performed a secondary screen using 2 of the initially used shRNAs targeting each of 240 top metastasis-specific lethal candidate genes in two T_nonMet cell lines (368T1 and 394T4) and four metastasis derived cell lines (238N1 and 482N1 from lymph node metastases, as well as 404M1 and 2691M1 from distant metastases). This secondary screen included positive control shRNAs that were lethal to all cell lines, control shRNAs (inert) that should have no effect, and shRNAs that were found to provide a competitive growth advantage to all cell lines (Supplementary Fig. S2A). This focused secondary screen was performed in the same manner as the initial screen and generated a robust readout of the effect of candidate gene knockdown, with high reproducibility among replicates (Supplementary Fig. S2B). All the control shRNAs performed as anticipated: Lethal shRNAs had reduced representation in all cell lines, advantageous shRNAs gained representation in all cell lines, and inert shRNAs had no effect. Importantly, many of the candidate metastasis-specific lethal shRNAs had reduced representation specifically in metastasis-derived cancer cells, validating our findings from the initial screen (Fig. 2B–E; Supplementary Fig. S2C and S2D; Supplementary Data Table S7). Furthermore, in parallel with this in vitro screen, we also transplanted two of the Met cell lines (238N1 and 482N1) transduced with the lenti-shRNA validation pool subcutaneously into mice to assess the effect of each shRNA on growth in vivo (Supplementary Data Table S8). The data obtained from this in vivo screen were consistent with the results obtained in vitro, further validating our screening results (Fig. 2F and G).

Our primary and secondary shRNA screens, including data from both in vitro and in vivo experiments, uncovered several top candidate genes whose shRNA-mediated suppression was specifically detrimental to cancer cells in a metastatic state (Fig. 3A and B; Supplementary Fig. S3A; Supplementary Data Table S9). To prioritize hits and investigate cellular programs specifically required for the growth of Met cells, we performed gene set enrichment and GO-term analyses of our initial screening data. Interestingly, both analyses revealed mitochondria (mt), the mitochondria ribosome and mitochondria ribosomal subunit gene sets as the top gene sets that, when targeted, were specifically lethal to Met cells (Fig. 3C). Indeed, RNA-Sequencing of T_nonMet and Met cell lines revealed that mitochondria-encoded genes were more highly expressed in Met cells whereas most of the nuclear-encoded mitochondrial ribosome proteins were more highly expressed in T_nonMet cells (Supplementary Fig. S3B).

We initially tested whether this distinct response to mt-targeting shRNAs might be attributable to mitonuclear imbalance. Mitonuclear imbalance is the stoichiometric imbalance between nuclear genome and mt genome-encoded oxidative phosphorylation proteins (25). Interestingly, RNA-Seq on primary tumors and metastases isolated from the Kras^G12D;Trp53^KO mouse model (16) did not uncover significant differences in the overall expression of mt-encoded genes (Supplementary Fig. S3C), and western blot analysis of T_nonMet and Met cells for nuclear (Atp5a, Uqcrc2, Sdhb) and mt-encoded (mt-Co1) proteins of the respiratory chain complex did not identify differences in expression levels or patterns (Supplementary Fig. S3D). Thus, mitonuclear imbalance is unlikely to have driven the distinct response in our shRNA screen.

To further interrogate whether the metastatic cell state is characterized by altered mitochondrial functionality, we generated fluorescently labeled T_nonMet and Met cells and performed in vitro competition assays. Treatment with ethidium bromide (which impairs mitochondria DNA replication and transcription; refs. 26–28), high concentrations of doxycycline (which inhibits mitochondrial translation; refs. 29, 30), phenformin (a complex I inhibitor of the respiratory chain that causes lactic acidosis, a hallmark of mitochondrial impairment; refs. 31, 32), or FCCP (carbonylcyanide p-trifluoromethoxyphenylhydrazone, a protonophore that dissociates electron transport from ATP synthesis and allows transport to proceed without synthesis; ref. 33) were each consistently more detrimental to Met cells compared with T_nonMet cells, as the Met/^TnonMet ratio was decreasing in each treatment. Furthermore, also under more physiological low oxygen conditions (1% oxygen) all treatments affected Met cells more than T_nonMet cells, though to a lesser extent (Fig. 4A–D; Supplementary Fig. S4A–S4D). Thus, these pharmacological data were consistent with differences in mitochondrial functionality between primary tumor-derived and metastasis-derived lung cancer cells and suggest that mitochondria are a critical and pharmacologically targetable Met cell-specific vulnerability.

To determine whether the differential sensitivity to mitochondria-targeted therapy is indeed due to differences in mitochondria functionality, we performed Seahorse assays for oxygen consumption rate and mitochondrial function. Met cells (n = 3 cell lines per group) had reduced spare respiratory capacity and coupling efficiency compared with T_nonMet cells (Fig. 4E–G). Furthermore, higher ECAR/OCR (extracellular acidification rate/oxygen consumption rate) ratio and higher ECAR levels pointed to a higher use of glycolysis as primary metabolic pathway in Met cells (Fig. 4H; Supplementary Fig. S4E). In addition, Seahorse assays for fuel dependency revealed that, whereas fatty acids do not appear to play a significant role as a fuel source in either T_nonMet or Met cells, Met cells showed a higher dependency on glycolysis-derived pyruvate and were less able to use glutamine (Fig. 4I–L; Supplementary Fig. S4F and S4G). Analysis of relative ATP levels produced in galactose-containing media revealed no significant difference in the capacity to perform oxydative phosphorylation between Met and T_nonMet cells (Fig. 4M), and the ratio of ATP production in glucose versus galactose-containing media was also not significantly different (Fig. 4N). However, comparing relative growth rates of cells growing in glucose or galactose-containing media revealed that T_nonMet cells were less affected when solely being able to perform oxydative phosphorylation instead of glycolysis compared with Met cells (Fig. 4O). Taken together, these results suggest that metastatic lung cancer cells have a reduced usage rate of glutamine as fuel during oxidative phosphorylation, reduced fuel flexibility and might primarily utilize glycolysis to produce ATP but are not completely dependent on it.

Furthermore, we performed flow cytometry to assess mitochondria membrane potential in T_nonMet and Met cell lines using the JC1 dye. Met cells in vitro had significantly more depolarized mitochondria, indicating lower membrane potential and functionality (Fig. 4P). However, treatment with any of the three mt-targeting compounds (phenformin, ethidiumbromide, or doxycycline) did not reveal significant differences in the relative change of depolarization levels between T_nonMet and Met cells (Supplementary Fig. S4H and S4I). This observation was further validated by analysis of ROS production. Met cells had higher baseline ROS levels in steady-state culture, yet under treatment conditions the relative changes in ROS level between Met and T_nonMet cells did not differ (Fig. 4Q; Supplementary Fig. S4J and S4K). Similarly, neither treatment with doxycycline nor ethidium bromide revealed a differential effect on T_nonMet or Met cell mitonuclear imbalance (Supplementary Fig. S4L). Finally, no distinct effects of phenformin on major signalling pathways between Met and T_nonMet cells were detectable by western blot analysis (Supplementary Fig. S4M), and neither phenformin, nor doxycycline or ethidiumbromide significantly altered proliferation of either Met or T_nonMet cells (Supplementary Fig. S4N-S4Q). Interestingly, migratory capacity was generally higher in T_nonMet than Met cells, independent of normoxic, hypoxic (1% O₂) or acidic (pH 6.5) conditions (Supplementary Fig. S4R–S4U), and we confirmed that Met cells might be partially reprogrammed toward a more “stem-like” phenotype as they all could form and grow spheres in three-dimensional culture whereas only one of three T_nonMet cells was able to grow (Supplementary Fig. S4V and S4W). Collectively, these results indicate that the net-effect of mt-targeting drugs is the same in both cancer states, but the steady-state of mt function is lower in Met cells, rendering them more susceptible to treatment.

To determine whether this altered mitochondrial functionality is also present in vivo, we initiated lung tumors in Kras^G12D;Trp53^KO;R26^Tomato mice, sorted Tomato^positive cancer cells from primary tumors and distant metastases, and measured the mitochondrial membrane potential using Mitotracker deep red. Cancer cells isolated from metastases had significantly more depolarized mitochondria than cancer cells from primary tumors, indicating a higher fraction of dysfunctional mitochondria (Fig. 5A), even though ATP-production was not significantly affected (Fig. 5B). Indeed, electron microscopy analysis of tissue sections from primary tumors and pleura metastases from Kras^G12D;Trp53^KO;R26^Tomato mice revealed that although the total number of mitochondria is the same in primary tumors and metastases, the number of abnormal mitochondria (characterized by loss of density and bridging of cristae, indicating a disturbed respiratory chain setup) is significantly higher in metastasis cells (Fig. 5C–E). These results indicate that the altered mitochondria functionality discovered in Met cells in vitro is also present in metastatic cancer cells in vivo.

To further explore whether this change in mitochondria functionality of Met cells creates a specific and pharmacologically addressable vulnerability in vivo, we used a subcutaneous transplantation mouse model that recapitulates several aspects of the metastatic cascade (16). In this model, Met cells are grown as primary subcutaneous tumors and the cancer cells metastasize spontaneously to the lung and liver. We treated mice with established primary tumors (7 days after transplantation) with either vehicle or 100 mg/kg phenformin orally once daily for 3 weeks and then quantified primary tumor weight and size, lung weight, and the number and size of Tomato^positive metastases in the lungs. Although primary subcutaneous tumors were of similar size and weight in both the control and phenformin-treated groups, the number of metastases in the lungs of phenformin-treated animals (as assessed by both macroscopic fluorescent stereomicroscope and microscopic IHC analysis) was reduced by approximately 3-fold (Fig. 5F–H). Furthermore, we repeated this experimental setup and treated animals with established primary subcutaneous tumors (9 days after transplantation) with either vehicle or 1 mg/kg doxycycline i.p. once daily for 3 weeks. Here, primary subcutaneous tumors were slightly reduced in weight in doxycycline-treated animals, yet in 6 of 8 mice the number of Tomato^positive cancer cells in the lungs was strongly decreased compared with control-treated animals (Supplementary Fig. S5A–S5C). This indicates that mitochondria-targeting treatments reduce metastatic ability in vivo that is consistent with the Met-specific effect of mt-targeting compounds in vitro.

To further investigate the effect of phenformin on the later stages of the metastatic cascade in vivo, we intravenously injected mice with Met cells that were pretreated with phenformin for 48 hours in vitro and subsequently continued dosing the mice with phenformin for 7 days. Tomato^positive cancer cells were reduced approximately 3-fold in the lungs of phenformin-treated animals compared with control-treated mice (P = 0.0002). However, proliferation rates of cancer cells in the lungs did not differ between the groups (Fig. 5I–K). We then expanded the experimental setup into four groups: Met cells were either pretreated with phenformin or vehicle in vitro and subsequently injected intravenously into recipient mice that received either further treatment with phenformin or vehicle for 3 days in vivo. Interestingly, pretreatment of Met cells with phenformin before intravenous injection significantly reduced the number of Tomato^positive cancer cells in the lungs compared with controls, independent of whether the recipient mice were further treated with phenformin or vehicle (Supplementary Fig. S5D and S5E). This is consistent with a cancer cell autonomous effect of phenformin on metastatic ability, that was not due to induction of apoptosis or anoikis (Supplementary Fig. S5F and S5G). To further validate that other mitochondria-targeting compounds can impact metastatic ability, we intravenously injected Met cells pretreated in vitro with either phenformin, doxycycline, FCCP or vehicle for 48 hours. After 3 days, lungs were harvested and analyzed by flow cytometry. Importantly, all three mitochondria-targeting agents reduced the number of Tomato^positive cancer cells in the lungs of recipient mice (Supplementary Fig. S5H and S5I). Taken together, mitochondrial function is significantly altered in metastatic cancer cells and this “Achilles heel” can be therapeutically exploited for benefit in vivo.

Loss-of-function genetic screens have been widely applied as a powerful discovery tool in biology (34). We chose to take this approach using a high-throughput shRNA screen to functionally characterize the metastatic cell state and to potentially identify novel therapeutically exploitable targets that are specifically lethal to metastatic lung adenocarcinoma cells. For this, we used T_nonMet and Met cell lines that were isolated from the well-established autochthonous Kras^G12D;Trp53^KO mouse model of lung adenocarcinoma and are distinguished by their metastatic ability (14, 16–19). These cell lines have fewer confounding variables such as different origins and genetic diversity, thus enabling us to identify genes that are differentially important for maintaining growth and/or viability of metastatic cells.

Here, we identify mitochondria as a cellular structure with a distinct functionality in nonmetastatic primary tumor cells and metastatic cancer cells. We found that mitochondria in Met cells are functionally altered, both in vitro and in vivo. Mitochondria in Met cells have a disturbed morphology and lower membrane potential. In vitro, we found that this modified mitochondria functionality leads to reduced spare respiratory capacity and coupling efficiency, independent from specific cellular signaling. In addition, we found that treatment with mitochondria-targeting drugs results in Met and T_nonMet cells being affected at similar absolute levels, further strengthening our conclusion that mitochondria are structurally and functionally compromised in steady-state metastatic cancer cells and that upon further targeting are less able to compensate.

At first glance, it seems counterintuitive that the disruption of a vital cell organelle such as the mitochondria would lead to cells that are seemingly better adapted for metastasis. However, mitochondrial dysfunction has been observed in multiple types of cancer cells during the metastatic process (35). Mitochondria dysfunction can either describe changes in mitochondria metabolism or refer to changes in mitochondria dynamics. The former is a well-known and thoroughly studied topic (36), with the most prominent alteration being the Warburg effect, where dysfunctional mitochondria generate ATP by glycolysis even in the presence of oxygen (37). With the balance between glycolysis and oxidative phosphorylation disturbed, metabolite levels change and can promote hypoxia, enhancing further tumor progression (35, 36, 38). In addition, elevated levels of glycolysis increase the acidity of the tumor environment, further promoting invasion and metastasis by breaking down the extracellular matrix and driving apoptosis of normal cells (39, 40). On the other hand, recent efforts have focused on mitochondria structure and dynamics: Alterations in mitochondrial proteostasis, bioenergetics, and trafficking is associated with extensive bioenergetic defects, including reduced oxygen consumption, decreased ATP generation, and aberrant production of ROS that have been suggested to promote tumor progression and metastasis by directly affecting tumor cell motility (41). Most recently, it was shown that mitochondrial dynamics are dysregulated in a variety of cancers, which in turn leads to specific and therapeutically exploitable sensitivities (42).

In our study, metastatic cancer cells show a higher dependency on glycolysis-derived pyruvate, primarily use glycolysis, and are less able to use glutamine, as indicated by a reduced rate of glutamine consumption as fuel during oxidative phosphorylation. Yet, although recent studies have shown that the monocarboxylate transporter-1 (MCT1) is responsible for providing lactate as a fuel of the TCA cycle in lung cancer (43) and both MCT1 and MCT2 have been found in melanoma and breast cancer to be important for their metastatic ability, respectively (44, 45), we did not identify either MCT1 or MCT2 as metastasis-specific hits in our screen. This could be either due to different metabolic functions in this tumor entity, due to an upregulation of these genes during the metastatic process when mitochondria become dysfunctional, or due to a metastasis-independent function of MCT1 and MCT2 so that their respective depletion is also lethal for nonmetastatic tumor cells.

Classical antitumor approaches have largely focused on targeting DNA replication or DNA damage repair and the interruption of cell division to induce apoptosis (46). However, these processes are not cancer-specific and their therapeutic inhibition leads to severe side effects. Finding metastasis-specific vulnerabilities, even if those vulnerabilities are related to conserved cellular functions, could allow for lower dosing and reduced patient side effects. Because of the importance of mitochondria during carcinogenesis and disease progression, researchers have started focusing on mitochondrial metabolism as a more selective target to induce cancer cell death. Using this concept, multiple clinical studies have been, or are being, conducted mainly around the dual α-ketoglutarate dehydrogenase and pyruvate dehydrogenase inhibitor CPI-613 and a compound called DCA, that inhibits pyruvate dehydrogenase kinase (PDK) and activates pyruvate dehydrogenase (47). Yu and colleagues (48) repurposed the FDA-approved arthritis drug leflunomide, which promotes mitochondrial elongation through its function as an agonist of the fusion GTPase MFN2 in pancreatic cancer.

We used phenformin as a proof-of-concept tool to specifically target Met cells in vitro and in vivo. Phenformin is an antidiabetic drug belonging to the class of biguanides. Because of its side effect of lactic acidosis it was withdrawn from the market in most countries in the 1970s. In contrast with metformin, phenformin has a lipophilic structure that allows a rapid entrance to the cell and mitochondria without transporters. Once assimilated in the mitochondria, phenformin inhibits the mitochondrial complex I in a manner similar to metformin (49). The biguanide-induced lactic acidosis results from insulin deficiency that leads to protein as well as fat catabolism and therefore accumulation of pyruvate. Because biguanides such as phenformin and metformin inhibit the pyruvate carboxylase, pyruvate cannot be metabolized in gluconeogenesis. In addition, pyruvate cannot be catabolized through the citric acid cycle because of the massive amount of acetyl-CoA derived from the fatty acid oxidation. Thus, the increased pyruvate is metabolized to lactate (31). Because of its higher affinity for mitochondrial membranes and the side effect of severe lactic acidosis, it was hypothesized that phenformin could have a higher anticancerous effect than metformin (49, 50). Because Met cells have a higher dependency on pyruvate it seems logical that they are more sensitive to phenformin treatment, as we observed both in vitro and in vivo. As many shRNA “hits” in the initial and secondary screen were against mitochondria ribosomal proteins (Mrp) and other mitochondria-essential genes and reduced proliferation and/or induced cell death, choosing phenformin as therapeutic option to target mitochondria with a more specific mode of action may be a more valid therapeutic option.

Phenformin was proposed previously as a specific treatment for Lkb1-deficient but not p53-deficient lung tumors (51). In our system, both primary tumor and metastasis-derived cells are p53-deficient and we did not observe an effect of phenformin on tumor cell growth in vitro or in vivo. Instead, we showed that phenformin specifically targets the mitochondria in metastatic cancer cells that reduces early metastatic ability. In recent years, it has been discussed whether the most prominent mitochondrial alteration in cancer—the Warburg effect—is caused by mitochondria that have been damaged over the course of tumor development, or whether the Warburg effect and the functionally altered mitochondria are a mere consequence of a favorable adaptation to proliferate under hostile conditions. Likewise, it remains unclear whether the change in mitochondria functionality in Met cells observed in our study are required for their metastatic ability or if they are a result of the collateral damage that occurs during the metastatic process. Nevertheless, our findings provide novel insight into the metastatic cell state, identifying underlying functional differences and potentially therapeutic intervention opportunities for lung adenocarcinoma.

C.-H. Chuang reports funding from Stanford University and American Lung Association during the conduct of the study. P. Dujardin reports grants from Cusanuswerk e.V. during the conduct of the study. D. Yang reports grants from Stanford Graduate Fellowship and Tobacco-Related Disease Research Program Dissertation Award during the conduct of the study. M.M. Winslow reports grants from the DOD CDMRP Lung Cancer Research Program and the NIH during the conduct of the study. B.M. Grüner reports grants from DFG—German Research Foundation, German ministry for economics, The Hope Funds for Cancer Research, and Pancreatic Cancer Action Network/AACR during the conduct of the study. No disclosures were reported by the other authors.

C.-H. Chuang: Conceptualization, formal analysis, investigation, methodology, writing-original draft, writing-review and editing. M. Dorsch: Investigation, writing-review and editing. P. Dujardin: Investigation, writing-review and editing. S. Silas: Formal analysis, writing-review and editing. K. Ueffing: Investigation, writing-review and editing. J.M. Hölken: Investigation, writing-review and editing. D. Yang: Investigation, writing-review and editing. M.M. Winslow: Conceptualization, supervision, funding acquisition, writing-original draft, writing-review and editing. B.M. Grüner: Conceptualization, supervision, funding acquisition, investigation, methodology, writing-original draft, writing-review and editing.

We thank Rosanna Ma, Santiago Naranjo, Eva Hahn, and Dagmar Thyssen for technical assistance; The Stanford Shared FACS facility for technical support; Sean Dolan and Alexandra Orantes for administrative support; Sven-Thorsten Liffers for help with bioinformatic illustrations; Verena Jendrossek, Johann Matschke, and Mihaela Keller for support with hypoxia experiments; Former and current members of the Winslow, Grüner, Schramm, and Siveke laboratories for helpful comments. We particularly thank Drs. John Doench and David Root from the Broad Institute Genetic Perturbation Platform and the Functional Genomics Consortium for providing critical reagents and technical expertise. The TEM experiment described was supported by ARRA award number 1S10RR026780-01 from the National Center for Research Resources (NCRR). C.-H. Chuang was funded by a Stanford Dean's Fellowship and an American Lung Association Fellowship. B.M. Grüner was a Hope Funds for Cancer Research Fellow supported by the Hope Funds for Cancer Research (HFCR-15-06-07). This work was supported by an Emmy Noether Award from the German Research Foundation (DFG; GR4575/1-1 to B.M. Grüner), DOD CDMRP Lung Cancer Research Program LC130773 and NIH-R01-CA175336 (to M.M. Winslow). P. Dujardin is recipient of PhD fellowship from Cusanuswerk e.V. S. Silas was supported by a Stanford Graduate Fellowship and the HHMI International Student Research Scholarship. D. Yang was supported by a Stanford Graduate Fellowship and by a TRDRP Dissertation Award (24DT-0001).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.