Liquid biopsy-based single-cell metabolic phenotyping of lung cancer patients for informative diagnostics

Abstract

Accurate prediction of chemo- or targeted therapy responses for patients with similar driver oncogenes through a simple and least-invasive assay represents an unmet need in the clinical diagnosis of non-small cell lung cancer. Using a single-cell on-chip metabolic cytometry and fluorescent metabolic probes, we show metabolic phenotyping on the rare disseminated tumor cells in pleural effusions across a panel of 32 lung adenocarcinoma patients. Our results reveal extensive metabolic heterogeneity of tumor cells that differentially engage in glycolysis and mitochondrial oxidation. The cell number ratio of the two metabolic phenotypes is found to be predictive for patient therapy response, physiological performance, and survival. Transcriptome analysis reveals that the glycolytic phenotype is associated with mesenchymal-like cell state with elevated expression of the resistant-leading receptor tyrosine kinase AXL and immune checkpoint ligands. Drug targeting AXL induces a significant cell killing in the glycolytic cells without affecting the cells with active mitochondrial oxidation.

Fig. 1

Platform and metabolic markers of the on-chip metabolic cytometry. a The working flow of the single-cell on-chip metabolic cytometry assay on pleural effusion samples based upon triple fluorescence staining (2-NBDG/C12R/CD45). Metabolically active tumor cells are retrieved individually for DNA and RNA sequencing, as well as other functional assays. b Top, a picture of the PDMS microwell chip; Bottom, the bright field and fluorescence composite image of a representative block in the microchip (scale bar, 30 μm). The fluorescence signals of CD45 (Cy5), 2-NBDG (FITC), and C12R (TRITC) are shown in red, green, and blue, respectively. Putative metabolically active tumor cells are circled in green (for 2-NBDG^high) and blue (for C12R^high). c Co-location of mitochondria (mitotracker green) and C12R in A549 cells after 15 min incubation (scale bar, 30 μm). d C12R signal assessed within a set of abundant cellular reducing agents using 60 min incubation time, showing the minimal contribution to the measured signal from other common cellular reducing agents. Quadruplets were used to determine the error bars (n = 4 independent experiments, mean ± SD). e On-chip metabolic cytometry of A549 cells treated with DMSO control and rotenone. The data are represented as scatter plots and bar columns (control n = 2469 cells, rotenone n = 2501 cells, mean ± SD). f Relative 2-NBDG (n = 8433, 3498, 18148, 18987, and 17248 cells from left to right, respectively), ECAR (n = 3 independent samples), C12R (n = 9576, 3673, 10298, 14625, and 11428 cells from left to right, respectively), and OCR (n = 3 independent samples) readout changes of A549 cells in response to a set of metabolic inhibitors with respect to DMSO control (mean ± SD). For ECAR and OCR measurements, three replicates were performed in each experiment and each replicate represented the average of ten cycles of ECAR and OCR measurements (Also see Supplementary Fig. 2d). Source data are provided as a Source Data file

Fig. 2

Metabolic phenotyping of rare disseminated tumor cells in MPE. a Illustration of study design and distribution of patient driver oncogene mutations. b Scatter plot generated from the OMC reports 2-NBDG and C12R fluorescence intensity of all CD45^neg cells in MPE sample from P1. The histograms of 2-NBDG and C12R intensities of CD45^pos leukocytes (red) in MPE are shown on the top and right to generate cut-offs for identification of 2-NBDG^high and C12R^high cells (black dots). 2-NBDG^high and C12R^high cells are gated out by five and three standard deviations above mean of CD45^pos leukocytes, respectively. CD45^neg/2-NBDG^low/C12R^low cells are displayed in blue dots. c Representative images of four subsets of CD45^neg cells as well as CD45^pos leukocytes. The images are overlaid by a bright-field image and three fluorescence images (CD45: red; 2-NBDG: green; C12R: blue, scale bar, 30 μm). d The number of CD45^neg, metabolically active cells that are categorized into three subsets across 32 LADC MPE samples. The oncogenic driver mutation associated with each sample is listed. e Relative viability of the two metabolically active cell populations in response to 2-DG and phenformin, respectively, with respect to the DMSO control (n = 2 independent samples, mean ± SD). (*P < 0.05; **P < 0.005). Source data are provided as a Source Data file

Fig. 3

Partial least square – discriminate analysis (PLS-DA) for patient MPE samples. a The correlation between the N/R ratios and patient performance (ECOG scores upon follow-up) among 29 patient MPE samples. The Spearman correlation coefficient and the p value are labeled. b Confusion matrix of the predicted patient response from the leave-one-out cross validation of the PLS-DA model. The overall accuracy of the model prediction is 86.21%. c Variable Importance of the Projection (VIP) of the explanatory variables. VIP values represent the predictive capacity of the various clinical measurements. These values are obtained from the single-component fit and error bars represent 95% confidence intervals. A border line is plotted to identify the VIPs that are greater than 0.8 for identifying the variables that are not (VIP < 0.8), moderately (0.8 < VIP < 1), or highly influential (VIP > 1) (# 2-NBDG: number of 2-NBDG^high cells per 500,000 input cells; # of C12R: number of C12R^high cells per 500,000 input cells; Conc. MA: concentration of metabolically active cells assessed by number of metabolically active cells in 10 mL MPE; # DP: number of double positive cells per 500,000 input cells). d The ROC curve based on the PLS-DA model prediction with an area under the curve (AUC) of 0.952. Source data are provided as a Source Data file

Fig. 4

Therapy response and survival of patients predicted by the MPE metabolic phenotyping. a, b N/R ratios calculated by metabolic phenotyping of MPE samples can predict patient therapy response profiles evaluated by RECIST criteria upon follow-up for all the patients a and for EGFR-mutant patient b. The different metabolic phenotypes of MPE samples that are classified by the boundary lines at N/R = 0.5 and 2 can segregate patient response. The green dots denote patients who were on therapies at the MPE draw and metabolic phenotyping. The orchid dots denote newly diagnosed patients who were receiving first-line therapies after the MPE draw and metabolic phenotyping. The diamond dots represent the patients who were dead before the follow-up. PR: partial response, SD: stable disease, PD: progressive disease. c Kaplan–Meier survival curves for all the patients (left), newly diagnosed patients (middle), and EGFR-mutant patients (right) evaluated from effusion collection in groups of different metabolic phenotypes, respectively. Glycolysis phenotype, balanced phenotype, and mitochondrial oxidation phenotype are denoted as N/R ≥ 2, 0.5 < N/R < 2, and N/R ≤ 0.5, respectively. Vertical bars indicate patients censored at the cut-off date or loss of follow-up. The median survival time was significantly longer for patients in the mitochondrial oxidation group (not reached) and balanced group (not reached) than for glycolysis group (3.60 months, 95% CI: 1.93–7.03 months for overall, 2.43 months, 95% CI: 1.93–3.03 months for newly diagnosed, and 4.03 months, 95% CI: 1.93–10.10 months for EGFR-mutant) in all three cases. (Log-rank statistical test, *P < 0.05) d Maximum-intensity-projection PET images (left), axial CT (middle) and fused PET/CT images (right) showed variable FDG uptake in the pulmonary lesions across the three patients. e Normalized 2-NBDG uptake of 2-NBDG^high tumor cells in MPE samples overlaid with the SUVmax values from the PET images of the three patients. The 2-NBDG signals have been normalized to their respective cut-offs determined by 2-NBDG uptake of leukocytes for each patient (n = 40, 35, and 145 for P15, P25, and P30 respectively, mean ± SD). Source data are provided as a Source Data file

Fig. 5

Molecular signatures associated with each metabolic phenotype. a Copy-number plot across the chromosomes for 2-NBDG^high (green) and C12R^high (blue) cells for three patient MPE samples, highlighting their genome-profile similarity between the two metabolic phenotypes. Blue indicates deletion and red amplification. b Venn Diagram of the differentially expressed genes (DEGs) between two metabolic phenotypes across five patient MPE samples. A total of 179 DEGs are shared among five patients. c Log2 fold change of gene expression levels between the two metabolic phenotypes (2-NBDG^high vs C12R^high) across five patients. Representative genes involved in glycolysis, mitochondrial ETC complexes, and fatty acid oxidation are listed. d Enrichment of the DEGs up-regulated in each metabolic phenotype and shared by at least 4 out of 5 patients against 5 representative public databases by Enrichr. The top two entries ranked by the combined scores from each database are plotted (see also Supplementary Data 5 for p values and enrichment scores). e Gene set enrichment analysis of 5 relevant pathways across the five patients (P10, P12, P14, P17, and P24 from bottom to top respectively). NES denotes normalized enrichment score (*P < 0.05 and FDR q < 0.25). f Log2 fold change of gene expression levels between the two metabolic phenotypes (2-NBDG^high vs C12R^high) across five patients. Representative epithelial and mesenchymal markers as well as immune checkpoint ligands and MHC Class I-associated HLA genes are listed

Fig. 6

AXL as a potential drug target for the glycolytic phenotype. a Immunofluorescence staining shows detectable AXL receptor expression levels only on the 2-NBDG^high cells in the MPE sample of P29 (scale bar, 10 μm). b Cell viability of 2-NBDG^high cells upon AXL inhibition across three patient MPE samples with different genotypes (n = 3 independent experiments, mean ± SD, two-tailed Student’s t-test, *P < 0.05). c Cell viability of C12R^high cells upon AXL inhibition across three patient MPE samples with different genotypes (n = 3 independent experiments, mean ± SD). No statistically significant difference in cell viability is observed between drug and DMSO treated cells. Source data are provided as a Source Data file