Adipose-derived stromal cells (AD-MSCs) were obtained from lipectomies and liposuctions (healthy donors, no diabetic donors) upon written informed consent of the donors, following the guidelines approved by the Kantonale Ethik Kommission (KEK) Zurich Swiss (KEK-ZH: StV 7-2009) and international ethical guidelines (ClinicalTrials.gov; Identifier: NCT01218945). The stromal vascular fraction (SVF) isolated from human fat tissue was obtained with the consent of the patient according to Swiss ethics (BASEC-Nr.: 2019-01504).
Cells and cell culture
Twenty-two human adipose tissue samples (100–600 g) were obtained from lipectomies and liposuctions (healthy donors, no diabetic donors) . AD-MSCs were isolated from fat tissue, with the consent of the donors according to Swiss (KEK-ZH: StV 7-2009) and international ethical guidelines (ClinicalTrials.gov; Identifier: NCT01218945) . The extraction procedure was performed according to . AD-MSCs were cultured in Dulbecco’s modified Eagle’s medium (DMEM) (PAN Biotech) supplemented with 10% of fetal bovine serum (FBS) (Biowest), 1% of antibiotics (100× penicillin, 100× streptomycin) (Biowest), and 1% l-glutamine 200 mM (Sigma) (called AD-MSC medium). Medium was changed every 3 days, and cells were passaged with 1× Trypsin-EDTA (Life Technologies) for 5 min at 37 °C when cells were about 80% confluent. Cells were incubated at 37 °C in an atmosphere with 95% humidity and 5% CO2.
In vitro differentiation of human AD-MSCs
For osteogenic differentiation, AD-MSCs were seeded at a density of 1.6 × 104 cells/cm2 in Nunc™ 24-well plates (Thermo Fisher Scientific) or at a density of 1 × 104 in 96-well plates (TPP). For adipogenic differentiation, cells were cultured at a density of 1.6 × 104 cells/cm2 in Nunc™ 24-well plates (Thermo Fisher Scientific). Differentiation was started 24 h after seeding with StemPro® Osteogenesis Kit or StemPro® Adipogenesis Kit (Gibco/Life Technologies). For chondrogenic differentiation, cells were cultured at a density of 5 × 103 cells/cm2 in a Nunc™ 24-well plate (Thermo Fisher Scientific) and differentiation was induced at the 4th day of culture using the StemPro® Chondrogenesis Kit (Gibco/Life Technologies). All media were changed every 4 days.
Assessment and classification of trilineage differentiation potential
Differentiation assessment via specific staining was performed for all three differentiation lineages after 14, 17, and 21 days of differentiation. For Alizarin Red S (Sigma) staining, cells were washed with PBS and fixed with 4% (v/v) formaldehyde (Sigma) for 30 min at RT. Upon washing twice with ddH20, Alizarin Red S solution (0.7 g Alizarin Red S diluted in 50 ml ddH2O at pH = 4.2) was added for 20 min at RT. Afterwards, cells were washed four times with ddH2O, dried, and stored in the dark until image acquisition. For Oil Red O (Sigma) staining, cells were washed once with PBS and fixed with 10% (v/v) formaldehyde (Roth) for 1 h at RT. Afterwards, cells were washed twice with ddH2O, rinsed twice with 60% (v/v) 2-propanole (Sigma) in ddH20, and dried. Oil Red O working solution (0.15 g Oil Red O in 50 ml 60% (v/v) 2-propanole in ddH2O) was added for 10 min at RT. After four ddH2O washing steps, cells were dried and images were directly taken. For Alcian Blue 8GX (Sigma) staining, cells were washed with PBS and then fixed with 4% (v/v) formaldehyde (Sigma) for 20 min at RT. Afterwards, cells were washed twice with ddH2O and incubated for 3 min with 3% (v/v) acetic acid (Merck Millipore) in ddH20. Alcian Blue solution (0.1 g Alcian Blue 8GX in 100 ml of 3% acetic acid in ddH20 at pH = 2.5) was given for 1 h at RT. Cells were washed four times with ddH2O, dried, and stored in the dark until image acquisition. Images of the entire wells at days 14, 17, and 21 of differentiation were acquired with Cytation 5 imaging reader (BioTek). Quantification of differentiation was performed according to , and subsequent classification of AD-MSC into “good,” “bad,” and “intermediate” differentiating lines was performed applying the interquartile range distribution. We defined cell lines present in the 4th quartile as “good,” lines present in the 2nd and 3rd as “intermediate,” and lines in the 1st quartile as “bad.”
Isolation of the stromal vascular fraction
Stromal vascular fraction (SVF) was isolated from human fat tissue with the consent of the patient according to Swiss ethics (BASEC-Nr.: 2019-01504) and according to . Briefly, lipectomies were cut in small pieces and extensively washed with PBS. Enzymatic digestion was performed with 0.075% collagenase I (Gibco) at 37 °C for 45 min in a rotating disk. The reaction was neutralized with AD-MSC medium and centrifuged at 850g for 10 min. For lysis of the red blood cells, the pellet was incubated for 10 min at RT in 160 mM NH4Cl and then extensively washed with PBS. The SVF was then filtered through a 100-μm filter nylon mesh and was either directly processed for FACS sorting followed by osteogenic differentiation, or frozen in AD-MSC medium supplemented with 10% DMSO (Sigma).
Fluorescence activating cell sorting (FACS)
AD-MSC lines were washed with PBS and stained with ALP-APC (R&D) (1/50) and CD73-FITC (Biolegend) (1/160) for 25 min at 4 °C. Upon washing, the cell fractions (controls sorted, ALP+/CD73+, ALP−/CD73high, ALP−/CD73low) were sorted with a FACS BD Aria III 5L and seeded in Nunc™ 96-well plates (TPP) at a density of 1.2 × 104 cells/cm2 for osteogenic differentiation. Controls sorted were unstained cells processed through the FACS and collected without sorting specific subpopulations. Differentiation was induced 24 h after seeding. Freshly isolated SVFs were washed with PBS and stained with ALP-APC (R&D) (1/50), CD73-FITC (Biolegend) (1/160), and CD45-PE (Biolegend) (1/160) for 25 min at 4 °C. SVF fractions (controls sorted, CD45−/ALP+/CD73+, CD45−/ALP−/CD73high, CD45−/ALP−/CD73low) were sorted with FACS BD Aria III 5L and plated in vitro at a density of 1 × 104 in 96-well plates (TPP) for osteogenic differentiation. All media were changed every 4 days.
Immunohistochemistry and immunofluorescence
Paraffin-embedded samples of human fat tissue were selected for immunohistochemical and immunofluorescence analysis. Samples were deparaffinized with xylene and rehydrated by an increasing ethanol gradient for hematoxylin and eosin (H&E) staining. Target retrieval was performed using the PT Link (DAKO) at pH solution 9.0 (DAKO). Immunohistochemistry staining was performed using a Dako Autostainer Link 48. Primary antibodies used were as follows: rabbit monoclonal ALP (Abcam, 1/200), mouse monoclonal CD73 (Abcam, 1/200), mouse monoclonal CD31 (DAKO, 1/200), and the appropriate EnVision HRP secondary antibody (EnVision HRP rabbit or mouse, DAKO, 1/500) according to the manufacturer’s instruction. Immunofluorescence was performed using a Dako Autostainer Link 48 with the following antibodies: rabbit monoclonal ALP (Abcam, 1/200), mouse monoclonal CD73 (Abcam, 1/200), Alexa Fluor 488 goat anti-rabbit IgG (Thermo Fisher, 1/200), and Alexa Fluor 546 goat anti-mouse IgG (Thermo Fisher, 1/200) according to the manufacturer’s instruction. Sections were visualized with LEICA DM6600 with a × 20 magnifying objective lens.
Mass cytometry antibody panel and staining procedures
The antibody panel consisted of 31 monoclonal anti-human metal-conjugated antibodies, which included cell surface, cytoplasmic, and transcription targets (Table S1). When possible, already metal-conjugated antibodies were purchased from Fluidigm; otherwise, antibodies were conjugated in-house with isotopically pure lanthanide metals according to the commercially available MaxPar Antibody Labelling Kit (Fluidigm). Labeled antibodies were stored at 4 °C in antibody stabilizer solution (Candor Bioscience). Titration of each antibody was performed on a one-to-one mix of cells consisting of PBMCs (peripheral blood mononuclear cells), HEK (human embryonic kidney cells 293), Hela (cervical cancer cells), Jurkat (human T lymphocyte cells), Saos2 (sarcoma cells), Nalm6 (B cell precursor leukemia cells), SHSY5S (neuroblastoma cells), and human AD-MSCs. These different cell lines, which we called MIX, were chosen in order to have for each marker a positive and a negative control cell type. Sample staining was performed as described in the MaxPar Cell Surface, MaxPar Cytoplasmic/Secreted Antigen, and MaxPar Nuclear Target protocols (Fluidigm) with minor changes. Briefly, cells were first subjected to cell surface antibody staining, followed by cytoplasm staining, and nuclear staining. For the cytoplasmic and intranuclear staining, cell fixation steps were shortened to 10 min. Cells were then resuspended in 4% paraformaldehyde (Electron Microscopy Sciences) and stored at 4 °C until acquisition. In the day of CyTOF acquisition, cells were washed with MaxPar Fix and Perm Buffer (Fluidigm) containing Cell-ID Intercalator-IR (Fluidigm) and incubated at RT for 1 h. Cells were washed with ddH2O and then diluted in ddH2O with 10% EQ Calibration Beads (Fluidigm) at 1 million cells/ml before acquisition with CyTOF 2 mass cytometer (Fluidigm).
Mass-tag cellular barcoding
For all CyTOF experiments, the Cell-ID 20-Plex Pd Barcoding Kit (Fluidigm) was used following the manufactural instructions. In short, 1 million cells per condition and per line were washed with PBS and then incubated with Cell-ID Cisplatin (Fluidigm) for 10 min at RT. Afterwards, cells were fixed with MaxPar Fix Buffer (Fluidigm) for 10 min at RT, washed with MaxPar Barcode Perm Buffer (Fluidigm), and incubated with the appropriate barcode for 30 min at RT. Finally, cells were washed with Cell Staining Buffer (Fluidigm) and combined depending on the CyTOF experiment in one or more tubes before antibody staining. Depending on the planned CyTOF experiment, a specific barcoding strategy was developed in order to minimize technical bias and highlight biological differences.
Barcoding strategies for the osteogenic differentiation experiments
For this differentiation experiment, we had a total of 102 samples. Thus, having only 20 different barcodes available, we distributed the barcoded samples into 6 tubes (Tables S2). In each tube when possible, there was one “good,” one “bad,” and one “intermediate” line for all the collected time points. The 17 AD-MSC lines cultured under osteogenic condition were collected during the first 5 days (day 0, day 1, day 2, day 3, day 4) of differentiation. At each day, the samples were barcoded, pooled into the appropriate tube, and stored at 4 °C until day 4. At day 4, a unique antibody master mix was prepared and distributed into the six tubes. In order to monitor tube-to-tube variations, we added to each of the six tubes twice the MIX (PBMCs, HEK, Hela, Jurkat, Saos2, Nalm6m, SHSY5S, AD-MSCs) for a total of 102 samples (Tables S2). Stability of the barcoded samples stored at 4 °C during the four collection days was extensively proved in preliminary tests (data not shown).
Barcoding strategy for prediction of differentiation potential in five new AD-MSC lines
Five not yet characterized AD-MSC lines (new AD-MSCs) together with 9 already characterized AD-MSC lines (reference) were collected in their undifferentiated state (day 0). Next, together with one MIX, they were all barcoded according to the barcode plan (Table S3) and pooled into one single tube for antibody staining and CyTOF acquisition as described above.
Barcoding strategy for the passage experiment
AD-MSCs F28, F22, F5, and F14 were cultured in AD-MSC medium in Nunc™ 6-cm plates (Thermo Fisher Scientific) and passaged when 90% confluence was reached. This was repeated from passage 3 (p3) to passage 20 (p20). At each passage, part of the cells was frozen in AD-MSC medium supplemented with 10% DMSO (Sigma). All AD-MSC lines from p3 to p20 were thawed the same day and barcoded according to the barcode plan (Table S4). All barcoded passages from the same cell line were pooled into one tube. Each tube contained also twice a MIX as a control. Cells were stained with the antibody panel following the protocols mentioned above and then processed in CyTOF2 (Fluidigm).
Mass cytometry data analysis
Mass cytometry data.fcs files collected from each set of samples were normalized using the executable MATLAB version of the Normalizer tool  and concatenated using the .fcs concatenation tool from Cytobank. Individual samples were debarcoded using the executable MATLAB version of the single-cell debarcoder tool .
Quantification of the staining of the triplicates of undifferentiated cells (control) and cells cultured with differentiation medium (differentiation) is presented as mean ± s.d. Quantification of the triplicates of the staining of the FACS sorted subpopulations is presented as mean ± s.d. For statistical analyses, the one-way ANOVA Dunnett’s multiple comparisons test was used to compare the ALP+/CD73+ population with the other sorted fractions within the same day as well as for comparing the percentage of ALP+, CD73+, and CD271+ cells in the “good” category for each day with the same day of the “intermediate” and “bad” ones. *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, and **** p ≤ 0.0001. Pearson’s correlation was used to determine the correlation between the ALP frequency measured by CyTOF at days 0, 1, 2, 3, and 4 with the staining intensity measured at days 14, 17, and 21 for the osteogenic differentiation lineage.
Mass cytometry measurements were transformed using the inverse hyperbolic sine (arcsinh) function with a cofactor of 5 and subsequently median-centered on a per-marker basis.
CellCNN was trained with the objective to classify “good” versus “bad” AD-MSC lines from their corresponding mass cytometry measurements at day 0 (undifferentiated state). Training examples (multi-cell inputs) comprised 2000 cells, sampled uniformly at random from the original mass cytometry samples. In total, we sampled 1000 training examples per class (“good” or “bad” cell lines). For the top-k pooling layer, we considered values of k such that the ratio of k over the multi-cell input size would be one of [0.5%, 1%, 3%, 5%]. The remaining CellCNN parameters were set to their default values.
Defining the selected cell subpopulation
The default CellCNN filter interpretation analysis was performed to define and characterize the selected cell subpopulation. Initially, learned filters were clustered and a single representative filter was retained from each cluster. As a second step, a score was derived for each representative filter, measuring how well this filter alone can classify the validation samples. Only one representative filter achieved a positive score, and this filter was used to define the selected cell subpopulation (i.e., cells with positive score with respect to that filter) in individual mass cytometry samples at d0, d1, d2, d3, and d4.
Mass cytometry data that support the findings of this study are available on request from the corresponding author [P.C.].