Bladder cancer cells shift rapidly and spontaneously to cisplatin-resistant oxidative phosphorylation that is trackable in real time

Genetic mutations have long been recognized as drivers of cancer drug resistance, but recent work has defined additional non-genetic mechanisms of plasticity, wherein cancer cells assume a drug resistant phenotype marked by altered epigenetic and transcriptional states. Currently, little is known about the real-time, dynamic nature of this phenotypic shift. Using a bladder cancer model of nongenetic plasticity, we discovered that rapid transition to drug resistance entails upregulation of mitochondrial gene expression and a corresponding metabolic shift towards the tricarboxylic acid cycle and oxidative phosphorylation. Based on this distinction, we were able to track cancer cell metabolic profiles in real time using fluorescence lifetime microscopy (FLIM). We observed single cells transitioning spontaneously to an oxidative phosphorylation state over hours to days, a trend that intensified with exposure to cisplatin chemotherapy. Conversely, pharmacological inhibition of oxidative phosphorylation significantly reversed the FLIM metabolic signature and reduced cisplatin resistance. These rapid, spontaneous metabolic shifts offer a new means of tracking nongenetic cancer plasticity and forestalling the emergence of drug resistance.


Results
SP and NSP cells were recovered by Hoechst exclusion staining and FACS from J82 and T24 bladder cancer cell lines as reported previously 7,12,13 . RNAseq and gene set enrichment analysis (GSEA) 25 revealed that OxPhos and mitochondria-encoded gene sets were most significantly enriched in SP cells across all metabolic gene sets in the Kyoto Encyclopedia of Genes and Genomes (KEGG) 26 (Fig. 1a, b). When all differentially expressed genes between SP and NSP were analyzed, OxPhos genes were predominantly overexpressed in SP (Fig. 1c). Furthermore, among the genes differentially expressed between SP and NSP, all the OxPhos genes were upregulated in SP (Fig. 1d). Having identified a significant transcriptional signature associated with OxPhos in SP cells, we turned to explore the relative metabolic states of SP and NSP subpopulations. First, we profiled central carbon metabolism in SP and NSP cells by measuring intracellular metabolite pool sizes using liquid chromatography-mass spectrometry (LC-MS) based metabolomics 27 . This analysis revealed that the tricarboxylic acid (TCA) cycle metabolites citrate/isocitrate and fumarate were significantly upregulated in the SP subpopulation (Fig. 2a), as were aconitate, malate, and oxaloacetate. Visualization of the signed log 10 FDR q-value on a metabolic pathway map showed that all TCA cycle metabolites except for alpha-ketoglutarate and succinate were upregulated in SP cells (Fig. 2b). To test whether the TCA cycle was significantly enriched relative to other pathways in SP cells, we performed metabolite set enrichment analysis (MSEA) using intracellular metabolite pool sizes. Indeed, MSEA Figure 2. Glutamine-derived TCA cycle is upregulated in SP population. (a) Volcano plot of intracellular metabolite pool sizes. Data represents average weighted log 2 fold change (SP/NSP) and FDR-corrected q-value from three independent experiments. (b) Metabolic pathway map representing signed log 10 FDR q-value on a color scale for intracellular pool sizes of metabolites comparing SP and NSP populations. Metabolites that were not measured are shown as small circles with grey color. Isomers that were not resolved by LC-MS are shown as diamonds. (c) Enrichment of TCA cycle pathway with MSEA analysis of intracellular metabolites pool sizes of SP and NSP populations. Metabolites were ranked based on the signed log 10 FDR q-value comparing SP and NSP populations. (d) Labelling patterns of aconitate and citrate/isocitrate from [U-13 C]-glutamine. M-values reflect the number of 13 C (5 total) remaining after α-ketoglutarate is modified in the TCA cycle. M4 is associated with conversion of α-ketoglutarate to succinyl-CoA, giving up one heavy carbon in the reaction, and is coupled to oxidative phosphorylation. M5 is associated with carboxylation of α-ketoglutarate to become isocitrate and retaining all 5 13 C atoms. This reversal of TCA is also known as "backwards" flux. * denotes p-value less than 0.0002, ** denotes p value less than 0.00004 by Student's t test. Results are means of three biological replicates with small error bars; raw data are provided in Supplemental Table S1. (e) Metabolic pathway map representing log 2 fold change of SP/NSP on a color scale for [U-13 C]-glutamine fractional contribution. The arrows show the two paths, M4 forward flux (blue arrow) and M5 backward flux (red arrow) that [U-13 C]-glutamine can take after becoming α-ketoglutarate and entering TCA cycle (black arrow). Metabolites that were not measured or had less than 1% fractional contribution are shown as small circles with grey color. Isomers that were not resolved by LC-MS are shown as diamonds. (f) Enrichment of TCA cycle pathway with MSEA analysis of [U-13 C]-glutamine fractional contribution of SP and NSP populations. Metabolites were ranked by the log 2 ratio of SP/NSP [U-13 C]-glutamine fractional contribution. (g) OCR in SP (red) is higher than NSP (blue) in the Seahorse analysis corresponding to the expected higher OxPhos metabolism of SP cells. *(b) and (e) were drawn using cytoscape: version 3.5.1. https:// cytos cape. org/ roadm ap. html. www.nature.com/scientificreports/ confirmed that the TCA cycle was the most enriched pathway in the SP population (Fig. 2c). In comparison, none of the glycolytic metabolites were significantly altered between SP and NSP cells, and MSEA did not indicate an enrichment of glycolysis in SP or NSP populations. Next, to better understand the differences in metabolic flux between SP and NSP subpopulations, we performed stable isotope tracing metabolomics. We first cultured cells with [U-13 C]-glutamine as a tracer to differentiate between oxidative metabolism and reductive carboxylation in the TCA cycle. Specifically, oxidative TCA cycle flux results in the loss of one heavy carbon and the labeling of downstream TCA cycle metabolites with only four heavy carbons (Fig. 2d, M4), whereas reductive carboxylation retains all five heavy carbons on isocitrate, aconitate, and citrate (Fig. 2d, M5) 28 . Isotopomer distributions for metabolites from the TCA cycle revealed that SP cells exhibit an increased percentage of M4 aconitate and citrate/isocitrate, suggesting increased oxidative metabolism (Fig. 2d). SP cells also had a complementary reduction in M5 aconitate and citrate/isocitrate percentages, suggesting reduced reductive carboxylation in SP cells. Additionally, we visualized the total fractional contribution of [U-13 C]-glutamine on a metabolic pathway map 29 and found that SP cells exhibited an increased contribution of glutamine-derived carbon to the TCA cycle, glutathione metabolism, and aspartate (Fig. 2e). Furthermore, MSEA revealed that the fractional contribution of glutamine to TCA cycle metabolites was significantly enriched relative to other metabolic pathways (Fig. 2f). We next labeled cells with [U-13 C]glucose and performed LC-MS metabolomics on flow-sorted SP and NSP populations. We again observed an increased fractional contribution of [U-13 C]-glucose to most TCA cycle metabolites in SP cells (Supplemental Fig. S2). Next, we measured the oxygen consumption rates (OCR) of SP and NSP cells using the Seahorse assay. Consistent with our metabolomic data, we found that SP cells exhibited a significantly increased rate of both basal and maximal OCR compared to NSP cells (Fig. 2g). Taken together, these data support that SP cells have increased levels of mitochondrial aerobic metabolism than NSP cells.

Scientific
Having established the differential metabolic states of SP and NSP cells, we asked whether cell transitions between these states could be tracked using two-photon FLIM, a non-invasive, non-interventive and label-free microscopy method that allows real-time tracking of metabolism in cells and tissues 21,30,31 . Here, we image NADH as the intrinsic fluorescent marker and measure its fluorescence decay times (lifetime) to monitor cellular energy metabolism. A short fluorescence decay lifetime is correlated with a high ratio of bound/total NADH, which corresponds to a lower ratio of free NADH/NAD+ and a more OxPhos metabolic state [21][22][23][24][30][31][32] . Supplemental Fig. S3 describes decay trends and analysis of NADH FLIM data in the lifetime-phasor. The phasor's graphical interface is an elegant and powerful tool for interacting with live-cell imaging data and for extracting meaning out of metabolic imaging.
First, FLIM signatures were measured in pure, FACS-sorted SP and NSP cells (Fig. 3a, b). Hoechst dye is necessary for FACS sorting of SP and NSP cells, but it has an emission spectrum that overlaps with NADH fluorescence emission. To eliminate this background noise, NADH FLIM signal was extracted by thresholding out the brightest pixels which corresponded to the nuclear Hoechst dye. The remaining fluorescent signal within the cytoplasm of the J82 cells is relatively photon-limited and therefore needed denoising and amplification of signalto-noise ratio (SNR). To accomplish this, we applied a recently published filtering method for FLIM phasors, Complex Wavelet Filter 33 . Using this approach, the FLIM phasor distribution of SP cells (Fig. 3a) was distinct from that of NSP cells (Fig. 3b). SP cells had a higher bound/total-NADH ratio (BTNR), 0.85 ± 0.05, as compared to the BTNR of NSP cells, 0.7 ± 0.07, connoting a greater OxPhos metabolic state in SP cells, consistent with our RNAseq, metabolomic, and functional observations. Having assigned differential FLIM signatures to SP and NSP cells, we were able to conduct cell fate tracking experiments for the first time without the constraints of prior Hoechst staining and FACS sorting, as we had originally set out to do. Accordingly, in subsequent experiments where FLIM alone was used to identify the aggressive, drug resistant phenotype, the SP-like cells are termed "OxPhos" cells, as distinct from NSP-like "Glycolytic" cells.
Next, we used FLIM to characterize and track single cell metabolic fates over 48 h (Fig. 3c), using the BTNR cut-off of 0.8 established in earlier experiments to assign OxPhos (> 0.8) versus Glycolytic (< 0.8) phenotype to cells of interest. When randomly examining 4 fields of view at 25× magnification (35 cells total), we found four examples of cells whose metabolic states shifted spontaneously from glycolytic to OxPhos and three other cells that shifted from OxPhos to glycolytic in that time frame (Fig. 3c). This bidirectional change is consistent with our prior plasticity findings 7,12,13 . To assess the effect of drug treatment on cell fates, we treated J82 bladder cancer cells with Cisplatin at IC50 (25 μM) and tracked them in culture for 48 h. In the first 13 h, metabolic FLIM signal was collected at 30-min intervals. During the 5.5-13 h window, we tracked one example of a cell that transitioned from Glycolytic to OxPhos, while its neighbor maintained its initial glycolytic state (Fig. 3d). For the remaining time frame (13-48 h), FLIM imaging intervals were expanded to every hour. After 48 h, cell survival (49% of initial cell count) was consistent with administration of an IC50 dose of Cisplatin, and the OxPhos population as a percentage of total surviving cells climbed to 42% versus 5% for the untreated cells (Fig. 3e). To further confirm that the FLIM signature is not cell line specific, we performed the same experiment in UM-UC-3 (Supplemental Fig. S4a-d) and T24 cells (Supplemental Fig. S4e-h) and observed a similar trend as early as 2 h after cisplatin treatment.
In order to functionally validate the link between the FLIM signature, metabolic state, and cisplatin resistance, we treated cells with phenformin, an inhibitor of mitochondrial complex I and OxPhos metabolism. Phenformin treatment caused a significant shift in the FLIM signature ( Fig. 3f-g), away from OxPhos and toward glycolytic metabolism. Consistent with this, phenformin treatment synergistically increased cell death in response to cisplatin (CDI 0.69, p = 0.048) (Fig. 3h), suggesting that interference with the rapid transition to OxPhos metabolism reduces cancer cells' resistance to cisplatin treatment. Similar results were obtained using UM-UC-3 and T24 bladder cancer cells (Supplemental Fig. S4i- www.nature.com/scientificreports/

Discussion
Tumor cells have long been known to exhibit unique adaptive metabolic profiles, most notably aerobic glycolysis (Warburg effect) 34 . Recent work has added a new layer of complexity to this model by demonstrating that tumor cells in fact retain functional mitochondria, which play an instrumental role in integrating signals and metabolically adjusting cellular activity in stressful conditions, termed reverse-Warburg or Crabtree effect 35 .
Mitochondrial DNA is now recognized as playing key driver roles in cancer progression and adaptability, and upregulated mitochondrial biogenesis and higher OxPhos levels have been reported for different cancer types and cancer stem cells 36 . For example, one recent study of transcriptional classification of IDH wild-type GBM defined a 'mitochondrial GBM' that exclusively relied on OxPhos and exhibited marked vulnerability to OxPhos inhibitors 37 . Another recent report showed that mitochondrial ATP powers drug efflux pumps that contribute to cancer cell drug resistance, a mechanism that can be mitigated by treatment with an OxPhos inhibitor 15 . Other studies of tumorigenic, drug resistant cancer stem-like cells have demonstrated distinct metabolic states for these subpopulations [38][39][40][41] with the preponderance of evidence suggesting an OxPhos phenotype. For example, www.nature.com/scientificreports/ side population cells in lung cancer 42 , sphere-forming and CD133þ cells in glioblastoma 43 and pancreatic ductal adenocarcinoma (PDAC) 44 , and ROS low quiescent leukemia stem cells 45 all exhibited preferential OxPhos energy production. Importantly, while prior studies showed that certain tumor types or cell subpopulations exhibited distinct metabolic states, the manner in which those states arose remained obscure. It might result from canonical selection processes, whereby clonal subpopulations with distinct metabolic features emerged over weeks or months of cell divisions or drug selection. Alternatively, these metabolic distinctions are part of a much more rapid and dynamic phenotypic shift, like the one described recently by our group and by others 7,8,12,13,46,47 . Here, for the first time to our knowledge, we observe directly that single cancer cells are indeed capable of rapid metabolic shifts over the course of hours, and that these transitions are dramatically amplified with exposure to chemotherapy. Transcriptional profiling of isogenic SP and NSP cells sorted from the same bladder cancer cell line identified the OxPhos pathway as the most highly upregulated in the aggressive, drug resistant SP cells relative to NSP cells. Consistent with this, LCMS comparison of the two cell subpopulations revealed that TCA cycle and oxidative metabolism was upregulated in SP, a finding that was functionally corroborated by Seahorse experiments showing a significantly higher oxygen consumption rate in SP cells. Assigning an OxPhos metabolic signature to SP cells enabled, for the first time, the prospect of tracking the transition to and from this aggressive, drug resistant state in real time.
We sought a systemic technique which would allow us to monitor the metabolic signatures of single cells quantitatively and simultaneously with minimal perturbation. Two-photon FLIM is a powerful non-invasive optical method that yields information on cell metabolism in cells and tissues by correlating the decay profile of endogenous autofluorescent light to a specific biomolecular source 21,30,31 By localizing and quantifying specific molecular components such as NADH without the need for ectopic tropic labels, it constitutes a powerful tool for tracking in-vivo metabolic changes associated with stem cell differentiation, cancer development and progression, response to chemotherapy, and other biological processes [22][23][24]48 . Moreover, the cellular and subcellular resolution of FLIM makes it ideal for tracking metabolic transitions of individual cells noninvasively over time.
We therefore performed FLIM analysis on sorted pure SP and NSP cells, and then on unsorted cells in culture over time, both with and without cisplatin treatment. As expected, based on the prior transcriptomic and metabolomic profiling of bulk cells, SP cells exhibited a more OxPhos FLIM signature that was readily distinguishable from the more Glycolytic signature of NSP cells. When we applied this FLIM profile to unperturbed cells (no Hoechst, no FACS sorting) growing in culture, we were able to identify OxPhos cells and Glycolytic cells and then observe them convert from one metabolic state to the other in real time. When cisplatin was administered, the more drug resistant OxPhos cells maintained their OxPhos metabolic FLIM signature, whereas more drug sensitive glycolytic cells transitioned in greater numbers towards a more drug resistant OxPhos state or died from the treatment. Conversely, when the mitochondrial Complex I inhibitor phenformin was applied, the FLIM signature shifted away from OxPhos and toward glycolytic metabolism, with a concomitant reduction in cisplatin resistance. These rapid mitochondrial-driven shifts in metabolic state and drug resistance are not consistent with alterations in nuclear or mitochondrial DNA and are more likely attributable to reversible epigenetic regulation 49 or post-translational modifications of mitochondrial proteins 49,50 , regulatory mechanisms affecting the electron transport chain that are the subject of ongoing investigation.
Collectively, our findings demonstrate that metabolic plasticity plays a key role in the rapid phenotypic shifts to and from a drug resistant state. Specifically, transition to a mitochondria-mediated OxPhos state can occur stochastically as part of the "bet-hedging" adaptation of cancer cell populations, and also in a more rapid and coordinated manner when drug selection is applied. Metabolic plasticity is a complex process involving multiple oncogenes (e.g. BCL2) and phenotypic shifts like EMT, as well as cross-talk with the tumor microenvironment 46,51 . For example, higher OxPhos in tumor cells has been shown to induce MHC-I expression, rendering cells more sensitive to cytotoxic T lymphocytes (CTL) mediated lysis but more resistant to NK cell-mediated lysis 51,52 .Collectively, these observations underscore the important role played by OxPhos and highlight this metabolic shift as a potential therapeutic target. Concomitant OxPhos inhibition with chemotherapy-as done with phenformin in this study-may curtail cancer cells' ability to evade toxicity when first exposed. Moreover, the ability to track single cells and groups of cells noninvasively with FLIM as they shift to a more drug resistant metabolic state can be readily applied to a broad array of tumor models to identify, recover, and further characterize cells as they undergo this transition, yielding new mechanistic insights and therapeutic targets to surmount drug resistance. www.nature.com/scientificreports/ Material availability. All unique/stable reagents generated in this study will be made available on request.

Methods details
Measurement of oxygen consumption rate (OCR). OCR was measured using Seahorse XF Cell Mito Stress Test (Agilent, Santa Clara, CA, USA) at the USC Leonard Davis School of Gerontology Seahorse Core. Briefly, SP and NSP cells were sorted by flow cytometry and seeded at 10 4 per well in 15 replicates. Assays were initiated by replacing the growth medium with 175 μL XF assay medium (specially formulated, unbuffered Dulbecco's modified Eagle's medium for XF assays; Seahorse Bioscience) supplemented with 2 mM sodium pyruvate and 25 mM glucose, pH 7.4. The cells were kept in a non-CO 2 -incubator for 60 min at 37 °C before placement in the Analyzer. The cells were treated with 1 μM oligomycin; 1 μM tri-fluorocarbonylcyanide phenylhydrazone (FCCP); and 0.5 μM mixture including rotenone and antimycin A according to the instruction. Seahorse XFe Wave Software (Agilent) was applied to analyze the data. All readings were normalized to DNA concentration using Hoechst dye. Average values from 15 replicates were plotted. Student t test were used for statistical analysis.
Flow cytometry. Hoechst staining and FACS analysis and sorting were conducted as described previously 12,13 . Briefly, J82 cells were trypsinized, counted, and resuspended in prewarmed 10% FBS DMEM media at a concentration of 10 6 /mL. Hoechst 33,342 was added at concentration of 5 μg/mL, incubated for 2 h in 37 °C water bath and gently inverted several times during the course of incubation. Cells were washed and resuspended in ice-cold DMEM media. 7-AAD used to discriminate dead cells was added to the cells at a final concentration of 2 μg/mL. Samples were incubated for at least 5 min on ice before FACS analysis and sorting (FACSAria and FACSLSR-II, BD Biosciences, both equipped with UV lasers).  www.nature.com/scientificreports/ a Luna 3 µm NH2 100 Å (150 × 2.0 mm) column (Phenomenex). The flowrate was 300 µl/min, and the gradient was from 15 A to 95% A in 18 min, followed by an isocratic step for 9 min and re-equilibration for 7 min. All samples were run in biological triplicate.
Metabolomic data analysis. Metabolites were detected and quantified as area under the curve based on retention time and accurate mass (≤ 5 ppm) using the TraceFinder 3.3 (Thermo Scientific) software. Raw data was corrected for naturally occurring 13 C abundance. Data was normalized to the sum of total signals in each sample. Pathway maps were made with Cytoscape. For each experiment, the p-value was calculated with a two tailed Student's t test. To evaluate combined significance from independent experiments, p values were combined with Fisher's method and then corrected for multiple hypothesis testing using the Benjamini-Hochberg method.
Metabolite set enrichment analysis (MSEA). Intracellular pool sizes or 13 C fractional contribution data was ranked based on signed log 10 FDR q-value comparing SP/NSP populations, and pathway enrichments were calculated using the unweighted statistic in the GSEA java applet. To perform GSEA, the gene expression data from RNAseq was ranked by the moderated t-test statistic calculated in limma. Metabolic pathways were downloaded from KEGG, and pathway enrichments were calculated using the unweighted statistic in the GSEA java applet.

Non de-scanned multiphoton fluorescence lifetime imaging (FLIM). Fluorescence lifetime images
were acquired with a Leica SP8 DIVE FALCON inverted microscope coupled to a tunable infrared laser system (Spectra Physics Insight3X). For image acquisition the following settings were used: image size of 1024 × 1024 pixels and pixel dwell time of 12.6 μs/pixel. FLIM images were acquired from 440 to 475 nm bandwidth on the Leica Non-Descanned hybrid detectors. FLIM data were acquired and analyzed using the Leica FLIM/FCS module. The excitation wavelength was 740 nm with an average power of about 0.6 mW. FLIM data was collected in 7 integrated frame repetitions. We also verified that any autofluorescence due to culture media would not interfere with either the FLIM or the spectral autofluorescent signature of cells.

FLIM data analysis
Every pixel of the FLIM image was transformed to one pixel in the phasor plot as previously described and reported in detail 21,30,31 . The coordinates g and s in the phasor plot were calculated from the fluorescence-intensity decay of each pixel of the image by using Fourier transformations. Analyses of phasor distributions were performed by cluster identification. Individual cells or regions of interest (ROI) within an image were segmented by hand and the selected image pixels are analyzed within the phasor plot. Segmentation resulted in distributions of pixels and centers of mass were determined for each ROI. Position of the center of mass for an ROI is a measure of the average BTNR for that ROI (Supplemental Fig. S3b).

Data availability
The datasets generated during the current study are available from the corresponding author on reasonable request. www.nature.com/scientificreports/