Voltage imaging reveals the dynamic electrical signatures of human breast cancer cells

Cell culture

We cultured MDA-MB-231 cells, a kind gift from the laboratory of Dr. Janine Erler, in high glucose DMEM (GIBCO, #41966029) supplemented with 5% FBS (Sigma, #F7524) and penicillin–streptomycin (Sigma, #P4333). For imaging, we plated 10k–30k cells on 12 mm collagen-coated glass coverslips (rat tail collagen, Sigma, #122-20). Cells were plated the afternoon prior to imaging. Imaging was performed in phenol red-free Leibovitz’s L-15 medium (Thermo Fisher, #21083027), except during the high-K⁺ control experiments where mammalian physiological saline (MPS) was used instead (described below).

MCF-10A cells were obtained from ATCC. They were cultured in DMEM/F12 (GIBCO, #31331) supplemented with 5% horse serum (GIBCO, #16050), 10 μg/ml insulin (Sigma, #I-1882), 20 ng/ml epidermal growth factor (Sigma, #E-9644), 100 ng/ml cholera toxin (Sigma, #C-8052), 500 ng/ml hydrocortisone (Sigma, #H-0888), and 100 mg/ml penicillin/streptomycin (GIBCO, #15070). Cells were confirmed to be mycoplasma-negative (e-Myco plus Mycoplasma PCR Detection Kit, iNtRON Biotechnology). The passage was carried out using 0.25% trypsin-EDTA (GIBCO) followed by centrifugation (1000 rpm, 4 min) and resuspension in a complete medium. Some sets of MCF-10A cells were cultured in transforming growth factor-β1 (TGF-β; Peprotech, #100-21) at 5 ng/mL in complete media for 48–169 h to drive EMT.

The following cell lines were cultured as part of the human breast cancer cell line panel investigated. MDA-MB-468 (a kind gift from the laboratory of Dr. George Poulogiannis, ICR), Cal-51 (a kind gift from the laboratory of Prof. Nicholas Turner, ICR), SUM-159, Hs578t (kind gifts from the laboratory of Dr. Rachel Natrajan, ICR), and MDA-MB-453 (a kind gift from the laboratory of Professor Claire Isacke, ICR) were cultured in high glucose DMEM (GIBCO, #41966-029), supplemented with 10% heat-inactivated FBS (GIBCO, #10500-064) and 1% Penicillin–Streptomycin (GIBCO, #15140-122). T-47D and BT-474, kind gifts from the laboratory of Prof. Nicholas Turner (ICR), were cultured in RPMI 1640 (GIBCO, #11835-063), supplemented with 10% FBS and 1% Penicillin–Streptomycin. Prior to imaging, 12 mm glass coverslips were coated with 50 μg/ml rat tail Type I collagen (Corning, #354236, lot 0295002) in 0.02 M acetic acid. Collagen-coated coverslips were incubated for 2 h at 37 °C. Coverslips were washed twice with PBS and allowed to dry at ambient temperature. All cell lines were passaged using 0.25% trypsin–EDTA (GIBCO), centrifuged (1000 rpm, 5 min), and resuspended in a complete medium. For imaging, 10k–30k cell dilutions were plated on prepared coverslips. Cells were allowed to incubate overnight prior to imaging for optimal attachment. All cell lines were confirmed to be mycoplasma-negative (e-Myco Plus Mycoplasma PCR Detection Kit, iNtRON Biotechnology).

Imaging

We prepared a 200 μM stock solution of the electrochromic voltage dye di-4-AN(F)EP(F)PTEA²⁹ (100 nmol aliquots, Potentiometric Probes) in L-15 solution. The stock solution was kept for a maximum of 3 days after dissolving. Immediately prior to imaging, we gently washed the coverslip-adhered cells three times with warmed L-15 before placing them under the microscope. The coverslip was weighted in place with a tantalum ring and submerged in the dye diluted in L-15 to a final concentration of 3 μM. The cells rested in this configuration for 15 min before imaging. During imaging, cells were maintained at a temperature between 30 and 37 °C by a homemade open-loop water perfusion system or by a closed-loop heated chamber platform (TC-324C, PM-1, Warner Instruments).

Our custom-built widefield epifluorescence microscope formed an image of the cells through a ×25 1.0 NA upright water dipping objective (XLPLN25XSVMP, Olympus) and 180 mm focal length tube lens (TTL180-A) onto a scientific complementary metal-oxide-semiconductor (sCMOS) camera (Orca Flash 4 v2, Hamamatsu). Imaging was performed with two-color sequential excitation and imaged in a single spectral channel. Fluorescence was ratiometrically excited in two channels resulting in opposite-direction voltage signals in the collected emission (Fig. 1c). LEDs were driven using a Cairn OptoLed (P1110/002/000). The first channel was illuminated with 405 nm LED (Cairn P1105/405/LED), filtered with a 405/10 nm bandpass (Semrock LD01-405/10), and combined with a 495 nm long pass dichroic (Semrock FF495-DI03) and an additional 496 nm long pass (Semrock FF01-496/LP). The second channel was illuminated with a 530 nm LED (Cairn P1105/530/LED), filtered with 520/35 nm filter (Semrock FF01-520/35), and combined with a 562 nm long-pass dichroic (Semrock FF562-DI03). Emission was collected through a 650/150 nm bandpass filter (Semrock FF01-650/150) onto the sCMOS camera (Fig. 1a). Images were acquired in Micromanager 2⁵⁵ with the Orca Flash 4’s ‘slow scan’ mode, using the global shutter and frame reset with 4 × 4 digital binning. Imaging was performed at 5 Hz. During every image period, a 3-ms-exposure frame illuminated with each LED was acquired in rapid succession (Fig. S4). Illumination intensities for each channel were approximately matched between each channel and adjusted to give a signal intensity of around 4000 counts/pixel in labeled cell membranes. Intensities were typically between 0.1 and 1.3 mW/mm² for the blue excitation and 1.5 and 3.4 mW/mm² for the green excitation.

Imaging protocols

We acquired each trial, consisting of sequences of 10,000 frames (5000 dual color-excited acquisitions) at different locations on the coverslip. Imaging locations were selected from confluent areas (median 975 cells/mm², interquartile range 605, 1247 cells/mm²). We acquired between 1 and 6 trials per coverslip, with each trial occurring in a distinct location. In tetrodotoxin (TTX) experiments, we first imaged 1–2 control trials without TTX (‘pre’ trials). We then added 1 mM of TTX citrate (Abcam, ab120055) in PBS stock solution to the imaging medium to achieve a final concentration of 1 or 10 μM. 1–4 trials were imaged in the presence of TTX (‘post’ trials). For certain 10 μM TTX experiments, following 1–2 trials acquired in the presence of TTX, we replaced the TTX-containing medium with regular dye-containing L-15 medium and imaged 1–2 trials in this condition (‘washout’).

In IbTx and apamin experiments, we first imaged 3 control trials without these agents (‘pre-’ trials. We then added 24.1 μM of either IbTx citrate (Tocris, #1086) or apamin citrate (Tocris, #1652) in L-15 to the imaging medium to achieve the final concentration of 100 nM. In separate experiments, we treated cells with either 100 nM apamin or IbTx for 45–60 min. The cells were then washed three times with L-15 and voltage dye added as described above to assess V_m dynamics following the washout of the KCa inhibitors.

High potassium wash-in trials were conducted in mammalian physiological saline (MPS⁴⁸), consisting of (in mM): 144 NaCl, 5.4 KCl, 5.6 d-glucose, 5 HEPES, 1 MgCl₂, 2.5 CaCl₂. Mid-trial, a high-potassium solution was washed in to depolarize the cells for validation of the voltage dye function. This solution was osmotically balanced consisting of (in mM): 49.4 NaCl, 100 KCl, 5.6 d-Glucose, 5 HEPES, 1 MgCl₂, 2.5 CaCl₂.

Patch-clamp voltage dye calibration

We assessed the range of fluorescence change expected for known changes in V_m through whole-cell voltage-clamp and simultaneous voltage imaging. Cells were imaged in phenol red-free Liebovitz’s L-15 medium at room temperature. Healthy, dye-labeled cells were selected and patched with pipettes between 3 and 10 MΩ. The pipette contained an Ag/AgCl bathed in intracellular solution (in mM): 130 K-Gluconate, 7 KCl, 4 ATP-Mg, 0.3 GTP-Na, 10 Phosphocreatine-Na, 10 HEPES. Voltage-clamp signals were amplified with a Multiclamp 700B (Molecular Devices) and digitized with Power 1401 (Cambridge Electronic Design) using Spike2 version 9.

Ratiometric imaging was performed as described above but at an increased rate of 100 frames/s. During each imaging trial, V_m was clamped for 1 s epochs at values varying between −60 and +30 mV in 10 mV increments. Fluorescent time courses were extracted from a cellular region of interest (ROI) around the patched cell for both excitation channels. The trials were bleach-corrected and converted to ΔF/F₀ using a linear fit to their time course. We calculated the average blue-to-green excited frame ratio (ΔR/R₀) across trials at each holding potential. A line was fit to ΔR/R₀ vs. V_m. The line gradient reflects the sensitivity of ΔR/R₀ to V_m (% change per 100 mV) for each cell (n = 12 cells).

Image processing

All data analysis was performed in Python 3 using NumPy⁵⁶, SciPy⁵⁷, Tifffile, Scikit-image⁵⁸, Scikit-learn⁵⁹, and Pandas⁶⁰. Figures were generated using MatPlotLib⁶¹. Our dual-color excitation scheme generated image time series interleaving blue and green light-excited frames (Fig. 3a). We subtracted the constant dark value from each frame and separated the time series into two color channels. We applied a pixel-wise high pass filter, rejecting signals slower than 0.01 Hz, to the separated time series. In particular, slowly varying signals (mainly bleaching) were removed from each channel by dividing the stacks pixel-wise by a temporally filtered version of themselves. The filter was a uniform filter of length 1000 points. The filter was symmetric (i.e. time point t0 was affected by points t > t0 and t < t0 equally). These filter-normalized time series were then divided, blue frames by green frames, to find the ratio image for each time point (Fig. 3a). Cells were segmented using CellPose⁶², using the default cytoplasmic segmentation model and approximate cellular diameter of 30 pixels. Additional segmentations of active cells that the Cellpose network did not identify were added by hand. The segmented ROIs were eroded with a single (1-round of binary) before extracting the ROI time courses to suppress the effects of movement at the cell edges. For each eroded ROI, we calculated the median of the pixel values at each time point. The mean of the time courses was subtracted and offset so that they were symmetric at about 1.

Event detection

We implemented an event detection algorithm to identify significant changes to ΔR/R₀ reflecting the fluctuation of V_m. We first calculated the time course of the intra-ROI pixel-wise standard deviation for each eroded ROI. We filtered the median (calculated as above) and standard deviation time courses with a σ = 3 point (i.e. 0.6 s) Gaussian filter. V_m fluctuation events were identified when the temporally filtered median pixel value diverged from 1, its time average value, by more than 2.5 times the temporally filtered standard deviation (Fig. 1h). Short events were removed and neighboring events merged by 2 iterations of binary opening and then 2 rounds of binary closing on the detected event array. Where events consisted of positive- and negative-going ΔR/R₀, they were split into entirely positive-going and entirely negative-going events.

Time series and event inclusion criteria

Cells were considered dying/dead and excluded from analysis where the raw pixel values increased in brightness by more than 25% during the acquisition in the blue channel, indicating loss of membrane polarization. Events arising from non-voltage-related changes in brightness were identified as simultaneous events and excluded from the analysis. In the MCF-10A image series, where events were very rare, imaging artefacts were identified and excluded were more than three events overlapped by more than 30% in time. In MDA-MB-231 data, events were excluded where more than 5 events overlapped by more than 50% in time. Following event detection, all active cell time series were then evaluated by visual inspection of processed videos to reject events caused by floating dust, focal shifts, or other apparent imaging artefacts. Events satisfying both the automated and manual quality-control measures were analyzed for event frequency, polarity, amplitude and duration.

In the feature-based analysis, we only included trials that completed the full 920 s imaging period (18,279 out of 18,752), and then applied the event detection algorithm described above to identify “active” cells (982 out of 18,279). We then excluded all-time series containing apparent imaging artefacts (dust, etc.). Following quality control and event detection, 297 MDA-MB-231, 28 MCF-10A, and 33 MCF-10A+TGF-β time series were admitted to the feature-based analysis pipeline described below (Fig. 4a, b).

V
_m time series clustering analysis

Complementing the event-based analysis, developed the Cellular DES Pipeline to classify the ROI-extracted time series according to its most salient dynamic features extracted by the Catch-22 algorithm³². The analysis realized with the Cellular DES Pipeline provides insight into the time series characteristics beyond simple event detection and quantification, enabling the classification of the heterogeneous V_m dynamics into like clusters.

From the admitted time series, we extracted 22 features from each cellular ROI’s median time series with the Catch-22 algorithm (Fig. 4c). After plotting the distribution of the individual features, around 80% (259/324) of cells shared the same value for the feature corresponding to the first minimum of the automutual information function, which we subsequently excluded from the feature list. We rescaled the raw feature values for the remaining 21 features between 0.0001 and 1 and applied the Box-Cox transformation to normalize their distributions. We then rescaled the normalized values to between 0 and 1 to ensure equal weighting into the clustering algorithms. To evaluate the number of dynamic electrical signature (DES) classes, we implemented hierarchical clustering and Gaussian Mixture Modeling (GMM) on the 21 normalized features. Both clustering algorithms generated clusters with silhouette coefficients, which measure subtype dissimilarity⁶³, decreasing to their lowest levels between 5 and 6 clusters (Fig. 4d). Based on these silhouette scores and on visual inspection of the time series, we chose to sort the time series into four types as this resulted in the most homogeneous classes. Based on the general pattern of each type, we named the DES classes: small blinking (blinking-s), waving, noisy and large blinking (blinking-l).

To select an exemplar time series from each DES class (Fig. 5), we performed Principal components analysis (PCA) and visualized the four classes in a 2-dimensional feature space. Each feature cluster occupies a unique area of the PC space. To identify the exemplar time series of each type, we calculated the components of each feature and drew a vector for each feature’s coefficients of PC1 and PC2 (Fig. 5). These vectors, therefore, point to the type exhibiting the corresponding features most saliently. To identify exemplar time series from each type, we sorted the time series according to the feature whose vector points to that type.

Statistics and reproducibility

Mean negative event (−VE) rates were compared by bootstrapped significance test, one-sided or two-sided, with the number of cells, the number of cultures (“slips”), and the resulting p-values specified in each case. The Supplementary Data contains the data used to generate all figures.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.