A Multimodal Imaging Framework to Advance Phenotyping of Living Label-free Breast Cancer Cells

Summary

This protocol presents a multimodal approach combining Raman spectroscopy and tomographic phase microscopy for label-free, non-invasive phenotyping of living human breast cancer cells. It includes a user-friendly, human-bias-free analysis pipeline for rapid, quantitative morpho-chemical analysis, providing detailed molecular and structural insights with applications in cancer research and cell biology.

Abstract

We present multimodal confocal Raman micro-spectroscopy (RS) and tomographic phase microscopy (TPM) for quick morpho-chemical phenotyping of human breast cancer cells (MDA-MB-231). Leveraging the non-perturbative nature of these advanced microscopy techniques, we captured detailed morpho-molecular data from living, label-free cells in their native physiological environment. Human bias-free data processing pipelines were developed to analyze hyperspectral Raman images (spanning Raman modes from 600 cm^-1 to 1800 cm^-1, which uniquely characterize a wide range of molecular bonds and subcellular structures), as well as morphological data from three-dimensional refractive index tomograms (providing measurements of cell volume, surface area, footprint, and sphericity at nanometer resolution, alongside dry mass and density). By systematically breaking down the rich single-cell details that RS and TPM deliver, we demonstrate, in a quantitative manner, the advantage of such a multimodal microscopy method for phenotyping tumoral breast cancer cells. Our tools also provide further insight into the subcellular information without the use of any labels. Finally, we study and discuss any unique or correlated information that RS- and TPM-derived datasets feature. We believe our tools and quantitative data analysis pipelines can revolutionize phenotyping tasks in biomedical research when in need of a rapid and non-perturbative method on living cells in culture, with the potential for future translation into clinical and diagnostic applications.

Introduction

Cell characterization and phenotyping are pivotal aspects of biomedical research¹. Being able to characterize cells at both the morphological and molecular levels is paramount to understanding cell function, differentiation, and pathology^1,2,3. It is particularly of interest in the field of oncology, where accurate cell profiling is necessary for tumor diagnosis, determining disease stage, and ultimately identifying therapeutic targets². Traditional cell characterization methods cannot provide temporal dynamics and are usually based on invasive techniques, such as staining or labeling, which may interfere with the native state of the cells and introduce potential biases within the data⁴. Therefore, non-invasive strategies to provide accurate and thorough phenotyping of living cells, in their native, unperturbed state, are increasingly in demand.

In this context, label-free multimodal imaging methods have revolutionized cell biology research by combining complementary modalities to capture both morphological and molecular information^2,5. These approaches allow for the integrated and non-invasive characterization of the cells that individual traditional techniques may not always be able to provide^2,5,6. For instance, though microscopy techniques such as fluorescence microscopy have been widely used in morphological and functional cell imaging, they possess inherent limitations. Fluorescence methods often suffer from photobleaching, phototoxicity, and low quantities of targetable molecules, which restrict the analytical potential⁴. In addition, these approaches often require the use of exogenous labels or dyes, which may disrupt the native cell state or lead to artifacts⁷.

Multimodal imaging platforms, which combine the strengths of multiple imaging modalities, surpass these limitations by enabling the acquisition of different types of data from living cells without the need for labeling. The integration of Raman spectroscopy (RS) with other imaging modalities like phase contrast microscopy or tomographic phase microscopy (TPM) has been a promising strategy to overcome the limitations of single techniques. Indeed, multimodal imaging provides a deeper understanding of cellular phenotypes, especially in complex biological conditions and pathologies like cancer, where molecular and structural properties are crucial for accurate characterization^2,5.

Label-free RS is well-suited for chemical imaging of cells⁸. By detecting molecular vibrational modes, RS provides biochemical fingerprints that can reveal in-depth information about the molecular composition of cells^9,10,11. Raman spectra acquired from biological samples contain information about a variety of biomolecules such as lipids, proteins, nucleic acids, and metabolites, which play essential roles in understanding the molecular nature of cells^9,12. One of the main strengths of RS is that it is non-invasive, allowing for the characterization of living cells in their natural state without altering their physiological condition.

Despite its many advantages, RS also presents some limitations. Indeed, although RS provides detailed molecular information, it does not allow for quantitative information on the morphology or structure of cells². This is where complementary imaging techniques like TPM play a crucial role.

TPM is a form of quantitative phase imaging (QPI)^13,14, and it enables the visualization of high-resolution, label-free quantitative information about cell morphology^15,16. TPM differs from traditional microscopy techniques in that it does not require contrast agents; instead, it operates by measuring phase shifts resulting from variations in the refractive index (RI) of the sample¹⁷. Phase shifts provide information about the structural properties of the cell, including its volume, surface area, shape, and internal structure^2,18. TPM can achieve high spatial resolution, to the nanometer level, and hence is an ideal technique for recording high-resolution morphological details of living cells. Further, the ability to reconstruct 3D RI tomograms makes it possible to visualize subcellular structures in their vivo environment¹⁸, providing more accurate information compared to 2D projections. However, one of the main limitations of TPM is its lack of chemical specificity, as it primarily provides information on the cell's morphology and phase distribution, without being able to differentiate between distinct chemical components or identify biochemical features.

The combination of RS and TPM in a multimodal technique provides advantages in cell phenotyping. While RS allows for rich molecular information, TPM adds a structural aspect, allowing cellular morphology to be analyzed in considerable detail. Recently, the combination of RS with TPM has been successfully demonstrated in several studies focused on cancer research^2,19,20. For instance, previous work has reported that co-registering Raman spectroscopy with holographic phase microscopy data can provide complementary molecular and morphological information, allowing distinguishing among different cell types and tumor development stages². These investigations have highlighted the power of multimodal imaging to enhance the accuracy of cell phenotyping, particularly in the context of cancer, where the molecular and structural characteristics of cells are closely linked to their malignancy and behavior^2,20.

Despite the recent progress in multimodal imaging, there are still challenges to be addressed, particularly in data processing and analysis²¹. The large amounts of data that multimodal imaging techniques yield can be cumbersome to analyze, and straightforward procedures must be employed to extract meaningful information. To counter this challenge, user-friendly data processing pipelines must be established to handle the complexity of the data and minimize human bias in the analysis. Such pipelines are critical for the achievement of accuracy and reproducibility of the results and the possible translation of these approaches into clinical and diagnostic applications.

The goal of the present work is to illustrate the potential of merging RS with TPM for label-free, high-speed, and non-perturbative phenotyping of live single cancer cells. By merging these two robust imaging approaches, we provide a comprehensive and quantitative characterization of the molecular and morphological properties of human breast cancer cells (MDA-MB-231). The work also emphasizes the importance of establishing human-bias-free data processing pipelines to ensure that the extracted information from the multimodal datasets is correct and reliable. Here, we seek to establish a robust platform for cancer cell phenotyping that can be generalized to other types of cells. The versatility and sensitivity of this approach make it a potential candidate for a wide range of biomedical applications, from fundamental cell biology studies to diagnostics.

Protocol

1. Sample preparation

1. Seed MDA-MB-231 cells onto quartz Petri dishes.
   1. Plate MDA-MB-231 human breast cancer cells onto 3.5 cm quartz glass-bottom Petri dishes.
   2. Incubate the cells at 37 °C in a 5% CO₂ environment until fully attached to the surface (at least 12 h after seeding).
2. Wash cells and prepare for measurement.
   1. Remove the culture medium from the cells once they have adhered to the Petri dish.
   2. Wash the cells gently with phosphate-buffered saline (PBS) to remove residual medium.
   3. Add fresh cell culture medium and keep the cells covered with the medium throughout the entire measurement procedure to ensure their viability.
      NOTE: As culture medium, this study employed Phenol red-free DMEM/F-12 supplemented with 5% Fetal Bovine Serum (FBS) and penicillin (100 IU/mL)-streptomycin (100 µg/mL).

2. High-throughput Raman microscope system

NOTE: Technical details of the home-built Raman microscope employed for the current work concerning live-cell imaging of MDA-MB-231 cells were described previously²². Briefly, the utilized custom-built high-throughput Raman microscope system features a 785-nm Ti-Sapphire laser pumped by a 532-nm laser (Figure 1A). The 785-nm Raman excitation laser is coupled to an inverted IX83 fluorescence microscope body, with excitation directed to the sample through a UPLSAPO60× 1.2 NA water-immersion objective using a 749-nm dichroic mirror. A polarizer tunes the average laser power at the back port of the microscope. Bright-field imaging is enabled by coupling white light in a similar manner. The system switches between Raman and bright-field modes by changing dichroic filters automatically. The system combines galvo-mirror-based and stage scanning for high-throughput imaging. A custom MATLAB script interacts with open-source microscope control software, a data acquisition (DAQ) board, and signal detection systems to automate the data acquisition process. The backscattered Raman light is collected by the same objective, while visible light is filtered out by a 785 nm dichroic mirror and directed to an NIR spectrograph and a charge-coupled device (CCD) detector. This setup enables multi-dimensional imaging, capturing bright-field, fluorescence, and Raman signals across various positions and z-stacks^2,22.

1. Raman imaging procedure
   1. Preparation of the sample and preservation of physiological conditions
      1. Install the sample into the onstage incubator to maintain physiological conditions. Ensure the incubator is set to 37 °C, with real-time feedback regulation for temperature control.
      2. Maintain humidity levels within the chamber at >95% using the internal peripheral bath unit filled with distilled water. Set up the CO₂ gas controller to mix 5% CO₂ with 95% air and supply it to the incubator chamber.
   2. Automated water immersion management
      1. Switch on the home-built automated water-immersion feeder to provide water to the objective lens at a flow rate of 2 µL/min using syringe pumps and syringe needles glued to the objective tip.
      2. Regularly check the water levels to avoid evaporation and ensure consistent imaging conditions throughout the measurements.
   3. Raman spectroscopy data acquisition
      1. Turn on the pump laser and adjust its power to attain a laser power at the sample plane of 75 mW.
      2. Open the Micro-manager open-source microscope control software. On the Configuration settings, select BF and click Live on the left side of the software window in order to visualize the bright-field image. Select the desired single cell from which the Raman spectrum will be acquired. Then click Stop to end the bright-field visualization.
      3. Set the Z-coordinate of the Raman imaging plane by maximizing the signal counts on the CCD camera.
      4. To do so, open the CCD camera control software (LightField); on the Experiment Setting, set the Exposure Time to 1.5 s. Then, click Run on the Experiment toolbar, and turn on the laser beam.
      5. Then, adjust the Z position to maximize the signal counts in the fingerprint region from 600 cm^-1 to 1800 cm^-1 to optimize the visualization of the Raman peaks associated with biological signatures. Then turn off the laser and click Stop in the Experiment bar of the CCD camera software, and exit the software.
         NOTE: The maximized Raman signal counts on the CCD camera were observed to be in correspondence with the position along the Z axis coinciding with +8 µm with respect to the imaging plane not showing any trace of cell-related Raman peaks, due to a focal spot falling fairly inside the quartz. This value may vary according to cell type and status.
   4. Start the data acquisition process using the MATLAB script, which will automatically handle Raman channels in multi-dimensional measurements and save the hyperspectral Raman data.
   5. Set the field of view to 50 × 50 µm², 40 × 40 pixels, with each pixel area corresponding to 1.25 × 1.25 µm². Set the exposure time for each Raman spectrum equal to 1.5 s.
      NOTE: Perform wavenumber calibration regularly using acetamidophenol samples to annotate known Raman peaks. Calibrate the Raman shift axis through quadratic interpolation.
      CAUTION: Handle lasers with care. Always use appropriate protective eyewear rated for the laser wavelength during setup and operation.

3. RS data analysis

1. Post-process Raman hyperspectral maps via RamApp web-based tool (https://ramapp.io/):
   1. Log in to the web-based tool.
   2. Import the hyperspectral data. On the My Analyses tab, click on New and upload the .mat file, containing the spatial and spectral data variable and the x-axis variable (i.e., Raman shift values).
   3. To truncate the spectrum to the specified region(s) in the Raman shift, in the Crop and rotate drop-down menu, click Spectral crop and define the inferior border and the superior border of the Raman shift region equal to 600 cm^-1 and 1800 cm^-1, respectively.
      NOTE: By performing this spectrum truncation, the broad and high-amplitude peak related to the quartz material of the Petri dish is excluded from the analysis.
   4. To correct the Raman spectra for cosmic rays, within the Denoising options, click on Despike, select Z-score as a Method, and set the Threshold equal to 8.
      NOTE: The Despike preprocessing function helps to address the issue of pixels containing spikes in their signal, which result from cosmic rays.
   5. Check the option Use first difference to find outliers using the first discrete difference of intensities along the spectrum, and check the option Correct spikes. The spectrum of each bad pixel will be replaced by the median spectrum of its neighboring pixels.
   6. Click on Run to correct the spikes.
   7. To remove noise from the maps corresponding to each wavenumber, within the Denoising options, click on Smooth Map, select Median filter, and set the size of the side of the square that is considered when applying the filter on each pixel of the map equal to 3 pixels. Then click Run.
   8. To apply a filter to smooth the spectrum of each pixel, click on Smooth spectrum within the Denoising selections. Select the Savitzky-Golay filter as the Method and set the length of the filter window equal to 7 and the polynomial order of the filter equal to 2. Then, click Run.
      NOTE: A Savitzky-Golay filter requires specifying the length of the filter window (a positive odd integer) and the order of the polynomial.
   9. To remove the average background signal from the foreground, open the Optical substrate removal panel and click Identify substrate. Set k-means as a Method and the number of clusters equal to 2.
      NOTE: This function segments the cells (the "foreground") and their substrate using a clustering approach. You will then be able to remove the average background signal from the foreground or restrict some steps to only work on foreground pixels. The number of clusters to be selected may vary depending on the specific specimen.
   10. In advanced options, check Morphological cleaning to remove some small blobs from the foreground (e.g., spurious cell parts, etc.). Then click Run.
   11. In the Optical substrate removal panel, to subtract the average background signal, click Remove substrate, then select Global average (mean of the central 95% of values) and click Run.
       NOTE: K-means is typically faster, though non-deterministic (however, since the random seed is fixed, the same result will be obtained in each run, all else being equal). The number of clusters can be increased beyond 2 if certain parts of the foreground are still recognized as background. The cluster with the lowest average signal is defined as the background.
   12. To correct for the baseline fluorescence, open the Baseline panel and click Correct baseline. Select asPLS (Adaptive Smoothness PLS) and set λ equal to 5000000. Then, click Run.
       NOTE: There are 7 available methods that can be selected, 5 of which are based on penalized least squares (PLS). In these PLS methods, a smoothing parameter (λ) is necessary. A higher value of λ will result in smoother baselines.
   13. To normalize each Raman hyperspectral map by dividing it by its Frobenius norm, click Miscellaneous and select the Normalize option. Then, select Frobenius as the Method, and click Run.
       NOTE: The Frobenius norm is utilized to normalize the entire data matrix.
   14. To obtain a false-color image illustrating the spatial distribution of a subcellular component, on the Images panel on the right side, on the Intensity Image, click the Open menu symbol. Then click the Edit image option to open the editing panel.
   15. On the editing panel, select Single band, then define the spectral range (e.g., 715−725 cm^-1 for the cytoplasm). Select Double color on the Color map, define the color, the intensity threshold, and the opacity, then click Confirm.
   16. To download a single-point spectrum, click on the pixel of interest on the image to select it. Then, on the Spectral legend panel, in the Cursor section, click Download spectrum. Repeat the same procedure for different subcellular components.
   17. To export the foreground and substrate Raman spectra, on the Spectral legend panel, in the Substrate identification section, click Download spectrum for both the Foreground and Substrate sections.
       NOTE: In the previously described protocol for RS data analysis, preprocessing steps and univariate analysis were described to generate false-color images illustrating the spatial distribution of key subcellular components, Raman intensity spectra from both foreground and substrate, and single-point spectra of subcellular structures. While downstream analysis is often tailored to the specific biological question, this study now highlights representative multivariate and classification methods accessible via the Map analysis section of the web-based tool employed. These include Cluster analysis, Principal Component Analysis (PCA), Multivariate Curve Resolution (MCR), N-FINDR, and Spectral fitting. Spectral fitting enables the quantification and comparison of spectral features using distance-based approaches such as the Pearson Correlation Coefficient or Spectral Angle Mapper. Cluster analysis can be conducted using algorithms like k-means, Mini-batch k-means, or hierarchical agglomerative clustering, facilitating segmentation of hyperspectral data based on spectral similarity. PCA can be applied to reduce data dimensionality and highlight variance patterns, while MCR allows decomposition of mixed signals into pure component spectra. The N-FINDR algorithm aids in extracting endmember spectra by first applying PCA to reduce dimensionality, making it particularly suitable for hyperspectral datasets. These tools provide a robust framework for downstream exploration of biological patterns encoded in Raman spectral data.

4. TPM measurement protocol

1. Prepare the microscope setup.
   1. Place a drop of distilled water onto the objective lens of the microscope. Position the sample on the microscope's translation stage. Adjust the sample position to align it with the objective lens.
2. Align the objective and condenser lenses.
   1. Turn on the microscope and launch the TomoStudio imaging software. Click the microscope icon on the toolbar. After initialization, click Configuration. Select LiveCell in the Job panel and PBS as the Medium.
      NOTE: The RI value in the Medium panel is automatically set to 1.337 when PBS is selected as a medium.
   2. Access the Calibration tab in the control panel of the imaging software. Set the axial positions of the objective and condenser lenses by selecting Focus and Surface, respectively.
   3. Click Scanning mode to manually adjust the lenses, ensuring the illumination patterns are concentrated in the center of the field of view and appear almost still.
3. Locate the sample.
   1. Click Normal mode and adjust the translation stage to position the cell in the field of view. Focus on the sample by adjusting the axial position of the objective lens until the sample boundary becomes nearly invisible on the screen.
   2. Adjust the translation stage to locate a region without cells, ensuring an unobstructed view.
4. Calibrate the system.
   1. Select Calibrate in the imaging software to capture multiple 2D holograms at varying illumination angles.
5. Acquire the 3D RI tomogram.
   1. Adjust the translation stage to center the cell within the field of view. Navigate to the Acquisition tab and select 3D Snapshot to capture the tomogram of the cell.
      NOTE: The acquisition time for a 3D RI tomogram of a single cell is significantly shorter (in the order of seconds), compared to the acquisition time required by RS data; thus, the adoption of an onstage incubator for TPM is not typically required. However, if measurements on a large number of cells are to be performed, or in specific conditions, dedicated commercial incubation chambers providing physiological conditions are available and can be integrated into the TPM system.

5. TPM data analysis

1. Post-process TPM maps using TomoStudio software.
   1. To visualize the holographic tomograms, select the data on the Data Navigation panel, right-click the data, and click Open. The Research tab will automatically open for further analysis. On the Data Manager panel, click on RI Tomogram.
   2. In the Volume Visualization panel, select RI and draw four rectangular color boxes within the RI canvas.
      NOTE: Each color box allows associating, on the tomogram shown on the Display panel, a specific color to cell components characterized by the RI range identified by the box in the RI canvas. After drawing each color box, the related information automatically appears on the table below the RI canvas.
   3. In the panel below the RI canvas, set the minimum and maximum values of RI Range for each color box (i.e., first color box: MinRI = 1.3450 and MaxRI = 1.3499, second color box: MinRI = 1.3506 and MaxRI = 1.3571, third color box: MinRI = 1.3586 and MaxRI = 1.3650, fourth color box: MinRI = 1.3688 and MaxRI = 1.3866). Associate an opacity value and a color (by double-clicking on each box and selecting a color) with each box. This allows for the visualization of the tomograms according to the 3D RI distributions.
   4. Click Save to save the defined RI Ranges and color boxes.
      NOTE: The RI ranges should be adjusted depending on the specific samples.
   5. To attain quantitative descriptors of cell morphology (i.e., volume [V], surface area [S], projected area [A], mean RI, density [D], dry mass [DM], and sphericity [Φ]), utilize the Analysis interface.
      NOTE: Details on the theoretical calculation of the quantitative parameters can be found here^2,23.
      1. In the toolbar, click Analysis. On the Analysis panel on the right, select Manual in the Segmentation panel and set 1.3450 as RI Threshold. Then click Apply.
         NOTE: Make sure that the values of the RII (fl/pg) and Baseline RI are respectively 0.19 (the refraction increment, which is widely established by previous investigations²⁴) and 1.3370 (RI of PBS employed for RI of PBS background calibration). The imaging software performs a volumetric segmentation of the target cell according to the set RI threshold and computes the morphological indexes (V, S, A, mean RI, D, DM, and Φ) and reports them in an output table in the Measurement panel.
      2. Click Save to save the calculated morphological indexes.
         NOTE: In addition to the employed software, several open-source tools can be used for quantifying morphological traits from TPM data. CellProfiler (https://cellprofiler.org/) provides tools for image analysis, including segmentation and feature extraction, which are compatible with phase-contrast images. Fiji (https://fiji.sc/) is an open-source software with plugins for segmentation and the computation of morphological parameters²⁵. Napari (https://napari.org/) is a multi-dimensional image viewer with support for integrating machine learning models for segmentation and analysis. Ilastik (https://www.ilastik.org/) offers machine learning-based segmentation tools, and DeepImageJ (https://deepimagej.github.io/) integrates deep learning models with ImageJ/Fiji for advanced analysis. Besides, MorphoLibJ (https://imagej.net/plugins/morpholibj) provides tools for 2D and 3D morphological operations in Fiji. Cellpose3²⁶ is another deep learning-based tool for cell segmentation that can be applied to phase-contrast imaging.

Results

We described a workflow for the label-free morpho-chemical characterization of living cells based on confocal RS and TPM. We tested the ability of the proposed workflow to attain the molecular composition and morphological parameters of the human breast cancer cell line MDA-MB-231 (Figure 1).

MDA-MB-231 cells were seeded onto quartz glass-bottom Petri dishes. A custom-built confocal Raman microscope (Figure 1A), operating at 785-nm co...

Discussion

We presented a multimodal approach and straightforward analysis pipeline integrating two advanced, label-free imaging techniques: RS and TPM. The described workflow enabled rapid, non-invasive morpho-chemical phenotyping of living human breast cancer single cells (MDA-MB-231) without the need for external labeling.

RS is a non-destructive technique that allows for the characterization of the biochemical composition of biological samples through the detection of the vibrational modes of molecul...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
CCD camera	Princeton Instruments	PIXIS 100BR eXcelon	Detector for Raman scattering signal.
CCD camera control software	Princeton Instruments	LightField
Continuous-wave Ti-Sapphire laser (wavelength: 785 nm)	Spectra-Physics	3900 S	It is used for Raman microscopy excitation.
Dichroic mirror	ThorLabs	DMSP750B	749-nm dichroic mirror, deflects 785-nm Raman excitation to sample, filters out fluorescence and brightfield signals.
Fetal Bovine Serum	Thermo Fisher Scientific	A5256701
Fluorescence microscope body	Olympus	IX83
Laser (wavelength: 532 nm)	Spectra-Physics	Millennia eV	Pump laser for 785-nm Ti-Sapphire laser system.
MATLAB software	MathWorks	2023a
MDA-MB-231 cell line	ATCC	HTB-26
NIR spectrograph	Andror	Holospec HS-HSG-785-LF	It is used for spectral analysis of backscattered light.
Objective lens (60× 1.2 NA water-immersion)	Olympus	UPLSAPO60×
Onstage incubation chamber	Tokai Hit	stxg-welsx-set
Open-source microscope control software	µManager	Micro-Manager-2.0
Penicillin-Streptomycin (10,000 U/mL)	Thermo Fisher Scientific	15140122
Phenol free DMEM/F-12	Thermo Fisher Scientific	21041025
Phosphate buffered saline	Gibco	10-010-023
Quartz glass-bottom Petri dishes	Waken B Tech	SF-S-D12
Raman data analysis software	RamApp		Web-based tool for Raman hyperspectral data processing.
sCMOS camera	Hamamatsu Photonics	Orca Flash 4.0 v2
Syringe pump	Chemyx	Fusion 720
Tomographic phase microscope	Tomocube, Inc.	HT-2H	Commercial system for obtaining quantitative phase images.
TPM data analysis software	Tomocube, Inc.	TomoStudio	Software used for analyzing tomographic phase microscopy (TPM) data.