2.1.

Cell Culture

Macrophage-like cell lines J774A.1 and Raw264.7 are cultured on 5 cm culture dishes (BD Biosciences, San Jose, California) and immersed in Dulbecco’s modified Eagle’s medium (Nacalai Tesque, Kyoto, Japan) supplemented with 10% fetal bovine serum (Nacalai Tesque) and incubated at 37°C in a humidified atmosphere with 5% ${\mathrm{CO}}_2$. One day before experiments, cells are trypsinized with a solution containing 0.25% trypsin and 1 mM ethylenediaminetetraacetic acid (Nacalai Tesque) for 5 to 10 min at 37°C to detach them from the plastic surface, and plated on 3.5 cm quartz-bottom dishes (Fine Plus International, Kyoto, Japan) containing cultured medium and incubated as above. Just before measurements, the culture medium is replaced with a phosphate buffer saline (Nacalai Tesque) solution supplemented with D-glucose (5 mM) and ${\mathrm{MgCl}}_2$ (2 mM) by washing the dish three to four times. Cells are then observed at room temperature.

2.2.

Raman Acquisition and Imaging

All measurements are performed with a Raman imaging setup described in detail in Ref. 26. Briefly, cells are observed on an inverted microscope with a $60\times$ objective (NA 1.27, water immersion, Nikon, Tokyo, Japan). The excitation is performed with a continuous wave laser at 532 nm (Verdi V-6, Coherent, Santa Clara, California), with a power density of $480\mathrm{mW}/{\mu \mathrm{m}}^2$. A dual-axis galvano-mirror placed in the descanned region [GM2, see Fig. 2(a)] enables the selection of the location of measurement for point excitation.

For imaging, the measurement is performed in slit-scanning conditions, i.e., the signal emitted by a whole line is measured during one frame acquisition [see Fig. 2(a)], enabling parallel detection. The excitation line is generated by rapidly scanning the laser spot (at 100 Hz) in the vertical direction with the galvano-mirror located outside the descanned region of the microscope (GM1) to form a line on both the sample and the detector, leading to an average power density of $2.44\mathrm{mW}/{\mu \mathrm{m}}^2$.

The collected backscattered light is separated from the excitation laser with a long-pass dichroic mirror and sent into a 500 mm focal length Czerny-Turner spectrometer (Shamrock, Andor Technology, Belfast, Northern Ireland). The light dispersed by the grating ($300\mathrm{lp}/\mathrm{mm}$) is then measured with a low-noise CMOS camera (Orca-Flash 4.0, Hamamatsu Photonics, Hamamatsu, Japan) with an exposure time of 3 s.

Quantitative phase images are recorded with an interferometric setup based on off-axis digital holography.²⁸ Acquisition and reconstruction of the holograms are performed according to the description given in Ref. 26.

2.3.

Spectra Preprocessing

Before analysis, all spectra are baseline corrected by employing an estimated curve at large intervals based on low quantile values (i.e., it has a high probability of being smaller than the signal), which is then computed for all points with cubic spline interpolation.²⁹ Cosmic ray removal is then applied by considering that points having a value outside a range of three times the standard deviation in a window of 20 pixels are cosmic rays, which are replaced by the median value within this interval.

In the case of hyperspectral images, a background signal estimated by measuring an empty region of the substrate is also first subtracted from the spectra. The remaining baseline is then removed as described above.

2.4.

Multivariate Analysis

All multivariate analysis procedures are performed after spectra preprocessing and are implemented in the MATLAB^® environment (The MathWorks, Natick, Massachusetts). In the case of imaging data, pixels containing cell spectra are separated from the background in the preprocessed Raman image by creating an image from the C-H stretching region (2910 to $3020{\mathrm{cm}}^{-1}$). Cell spectra are then selected by applying a threshold on the image, as cells have a strong Raman response in this region. The pixels extracted from the three-dimensional data set are then concatenated into a two-dimensional matrix for multivariate analysis.

Vertex component analysis (VCA) is a method that extracts the endmembers of a spectra data set as the vertices of a simplex, and is applied here based on the implementation described by Nascimento and Bioucas Dias.³⁰

Principal component analysis (PCA) is a method that defines a new orthonormal basis representing the data which maximizes the variance of data. It is applied here by first mean-centering every variable in the data set and then factorizing the matrix through singular value decomposition, defined as

3.1.

Raman Imaging and Spectroscopy

Raman imaging is illustrated in Fig. 1(a), where a color-coded image of live J774A.1 cells can be extracted from the three-dimensional hyperspectral data by highlighting specific bands in the spectra [see Fig. 1(b)]. The general shape and content of the cells is highlighted in green with a C-H stretching band (2970 to $3000{\mathrm{cm}}^{-1}$), while lipids (2850 to $2880{\mathrm{cm}}^{-1}$) in red are essentially visible in the cytoplasm and in lipid droplets, making nuclei particularly visible by their absence. A specific protein, which provides a strong signal due to its resonance at 532 nm, cytochrome c, can also be seen in blue located in certain regions of the cytoplasm (1570 to $1600{\mathrm{cm}}^{-1}$). The corresponding image acquired by QPM, containing a larger field of view, is shown in Fig. 1(c), with a full dynamic range corresponding to 5.1 rad.

Fig. 1

(a) Color-coded contrast of a Raman image of J774A.1 cells (scale bar $10\mu \mathrm{m}$). Red: lipids (2850 to $2880{\mathrm{cm}}^{-1}$). Green: C-H stretching (2970 to $3000{\mathrm{cm}}^{-1}$). Blue: cytochrome c (1570 to $1600{\mathrm{cm}}^{-1}$). (b) Raw Raman spectra extracted from $5\times 5$ pixels square regions of the image in (a), in various parts of the cells. Spectral bands used to make the image in (a) are also represented. (c) Corresponding QPM image with a larger field of view.

In the imaging case, the laser power is distributed along a whole region, effectively reducing the amount of signal retrieved per pixel. Furthermore, despite the parallel detection of the slit-scanning configuration,³¹ the 3-s exposure per line employed during measurement means that the $400\times 200$ pixels image of Fig. 1(a) still requires 10 min for acquisition. The limitation in the amount of signal can be partially overcome by averaging together detected signals, as shown in Fig. 1(b), where spectra extracted as spatial averages from $5\times 5$ square regions within specific locations of cells are shown. However, the spectra still exhibit a significant amount of noise despite the spatial averaging. It is also interesting to note that even though highly contrasted images can be extracted [see Fig. 1(a)], the spectral differences between different locations in cells are tenuous.

As Raman imaging requires significant acquisition times and is technically challenging, a common measurement approach in single cell analysis is then to acquire spectra in point-mode acquisition, where a single spectrum from one point is measured with the detector acting as a line detector, as depicted in Fig. 2(a). This approach provides signals as shown in Fig. 3(a), where spectral features can be identified far more easily than in the case of specific pixels or averaged regions from within a Raman image, and are assumed to be representative of the molecular content of the whole cell.

Fig. 2

(a) Different approaches in scanning principles of the Raman microscope. Imaging is performed by slit-scanning with two scanning mirrors, with only one being in the descanned region. Fixed point measurement is done with no scanning, while hybrid scanning is performed with two descanned mirrors, resulting in a scan pattern at the sample but a nonscanned point at the slit plane, giving the same signal to noise as fixed point measurements. (b) Quantitative phase image of J774A.1 cells, with the square scanning region used for hybrid Raman imaging represented by the dashed square.

Fig. 3

Spectra extracted from both fixed point (a) and hybrid scanning (b) excitation, where five endmembers obtained by vertex component analysis are displayed, showing the variability of spectral profiles for each scanning type. The fingerprint region (500 to $1700{\mathrm{cm}}^{-1}$) has been scaled by a factor of 3 to improve visibility.

The important question then becomes how well a spectrum taken from a single location in the cell represents the overall cellular content. As illustrated in Fig. 1(a), different cell locations do possess different molecular contrasts. This is particularly important when considering, for example, the distribution of cytochrome c, which is sequestered in mitochondria and clearly not homogeneously distributed in the cytoplasm. The measured spectra may then vary significantly not only by the overall morphology of the cell (i.e., whether measured inside the cytoplasm, nucleus, etc.), but also by the local molecular content (for example, the mitochondrial distribution in the cytoplasm). These variations cannot easily be assessed unless a full Raman image is acquired, which is time-consuming, as noted. This implies that point-based measurement approaches may require a very large amount of samples in order to ensure that local variations within individual cells are not influencing the data analysis.

To circumvent this local variation, we propose a hybrid spectral point measurement based on a random scanning pattern within a predefined region in the target cell. In this case, the excitation beam is rapidly scanned in a square region within the cell body, as represented in Fig. 2(a), which is selected from the additional imaging modality [see Fig. 2(b)]. Importantly, in contrast to the full Raman imaging case, the scanning is performed here through an x-y galvano-mirror located in the descanned region of the setup (GM2), so that the emitted signal stays located as a single point at the slit of the spectrometer, still yielding a strong signal as in the standard fixed-point case, but with a measurement derived from a larger region of the cell, which is expected to provide information more suitable for the analysis of cellular changes. This approach is also more efficient than averaging pixels from a Raman image, since by optically averaging the emission on one location for acquisition, the read-out noise (one of the most significant noise sources in Raman imaging) is significantly reduced.

In order to reduce any systematic error in the measurement, the laser spot is scanned randomly within a square region of $\sim 10\times 10{\mu \mathrm{m}}^2$ with a scanning frequency of 200 Hz in both directions, as represented by the dashed square in Fig. 2(b). The size of the square is first chosen according to the size of the sample to measure, so that the region should be large enough to enclose a region that can be considered significant of the whole sample (e.g., large enough to contain parts of the cytoplasm and the nucleus in case of cells in culture), while being small enough that it is contained within the cell area and does not sample significant regions outside the cell. The measured region shape is here arbitrarily chosen as a square, but could be adjusted to other shapes by simply changing the scanning pattern. The random scan is important so that, even for varying exposure times, we do not change the effective sampling patterns within the cell. To accurately set the location of the square region in the cell, an additional modality is necessary. In our case, we are using simultaneous QPM based on digital holography, which provides additional useful information, although the scanning approach, denoted here as hybrid Raman imaging, could be generalized to include other types of imaging.

3.2.

Cell Molecular Content Represented with Point Spectra

The faster acquisition rates available using point rather than imaging measurements make it possible to measure a larger amount of samples, which is required for robust, statistically significant analysis of biological effects in cell populations. Measurements are performed on $\sim 100$ cells per data set to account for variations in cell condition. Furthermore, to also account for possible changes due to different substrates and culturing conditions, the measurements are spread over four dishes per data set, with $\sim 25$ measurements per dish. To ensure that the measurements are representative of the population, both procedures (point and hybrid) are measured in random locations in the cell. Since the measurement region of both hybrid and fixed-point measurements is small compared to the overall size of the cell, we position both square and point measurements randomly throughout the cell to avoid systematic errors in the measurements.

Typical point spectra measured on J774A.1 cells are shown in Fig. 3 for both fixed-point [Fig. 3(a)] and hybrid scanning [Fig. 3(b)], with the fingerprint region scaled for the sake of visibility. In order to estimate the variability of spectra within a data set, the extreme ones are displayed here by retrieving five endmembers extracted by VCA. Spectra are then normalized by the intensity of the peak at $1662{\mathrm{cm}}^{-1}$, to account for overall changes in the signal intensity.

The spectra in Fig. 3 clearly show that the fixed-point scanning approach provides more variability in the measurements, with some spectra emanating essentially from one molecular species, such as lipids (first curve), or with strong variations in the background signal, such as the last curve, where a significant contribution from the quartz substrate at $800\hspace{0.17em}{\mathrm{cm}}^{-1}$ is present. As expected, the spectra measured with the hybrid scanning approach seem far more similar to each other, as the measurement throughout a whole region reduces the influence of particular local content by optical averaging. It is worth remembering that in imaging mode, the spatial variations in the sample are part of the measured information, but single point Raman measurements carry the inherent assumption that variations are low so that they are representative of the whole sample. This implies that the two approaches have different assumptions on how molecular content is distributed in the cell. The large variability of fixed point measurements results from the spatial variation in the cell, whereas the hybrid scanning measurements are much more consistent, indicating that the hybrid scanning mode samples the molecular content more effectively.

In order to quantitatively compare the information content of each data set, we also compute the covariance matrix defined as $\mathit{C}=\mathit{X}{\mathit{X}}^T$, where $\mathit{X}$ is the $P\times N$ data matrix for $N$ measurements and $P$ data points in one spectrum. We then consider the rank of $\mathit{C}$, which provides the number of linearly independent dimensions contributing to the variance of the data. The rank of $\mathit{C}$ then is an estimation of the amount of information present. The ranks of the data sets are shown in Fig. 4(a) for both scanning approaches (fixed and hybrid) and for two cell types, J774A.1 and Raw264.7. Each rank is the average of at least three experiments, with the error bars representing the standard deviation.

Fig. 4

The hybrid scanning approach generates significantly more information on cell content. (a) Ranks of covariance matrices for measurements performed with both fixed and hybrid scanning, and for different cells types (J774A.1 and Raw264.7), where 104 cells were measured for each experiment. (b) Cumulative variance described by principal component analysis (PCA) for both fixed and hybrid scanning. All bars and curves are the average of at least three experiments, with error bars representing standard deviations.

Surprisingly, although fixed-point measurements seem immediately more diverse than hybrid ones (see Fig. 3), the information content is actually lower as the mean rank of the fixed-point covariance matrices is 19 and 26.7 for J774A.1 and Raw264.7 cells, respectively, compared to 89.3 and 97.3 for hybrid scanning, for a total of 104 dimensions. This demonstrates that despite the fact that they appear similar, hybrid scanning measurements possess more linearly independent dimensions, denoting a larger information content compared to the fixed-point case.

This aspect is also confirmed through another approach; when applying PCA to the J774A.1 data set, where by considering the contribution to the data variance of each principal component (PC) and plotting the cumulative variance [see Fig. 4(b)], it is possible to see that the data based on the fixed-point scanning can be represented with significantly fewer components. Three PCs already account for $>99\%$ of the data variance, while 47 PCs are required in the hybrid case. Applying PCA on the Raw264.7 cell lines produces similar results.

3.3.

Cell Line Classification Through Point Spectra

In order to test how the increased information available through hybrid scanning relates to an actual cellular analysis implementation, we investigate the possibility of distinguishing J774A.1 and Raw264.7 cell types based on their respective spectra by multivariate analysis. The cell line measurements, each composed of spectra from $\sim 100$ cells, are concatenated together to perform PCA on the whole set and estimate the separability of the two classes in the PCA space. Up to this point, there is no use of a priori knowledge of the cell types, and loading vectors are generated based on the inherent spectral variance in the data. Two loading vectors are then selected to visualize the scores of the data, as shown in Fig. 5. The best loading vectors for separating the cell lines are selected by choosing the ones that maximize the distance between the centroids of the two classes based on the a priori knowledge of the cell type. The two scanning methods (fixed point versus hybrid) are compared on the components best able to separate the cell lines since a one-to-one comparison on each PC is unsuitable since, as shown above, the number of relevant components is very different between the two approaches. The loading vectors used to compute the scores are shown in Figs. 5(a) and 5(c) for, respectively, the fixed-point and hybrid scanning. The corresponding scores are then displayed in Figs. 5(b) and 5(d), along with confidence intervals, shown as ellipses whose major axes contain 95% of the data points in their projection direction.

Fig. 5

Hybrid scanning improves the Raman-based separation between cell lines by multivariate analysis, as shown by PCA decomposition of data sets composed of J774A.1 and Raw264.7 cell lines. (a) and (c) Loading vectors used for displaying score plots, where centroid distances between the two classes are maximized. (b) and (d) Score plots showing the separation between the two classes, with 95% confidence regions, for fixed point [(a) and (b)] and hybrid scanning [(c) and (d)], respectively. Each score plot is representative of at least three experiments.

We can see that the hybrid scanning provides better separation, with the confidence regions being nearly fully separated, in contrast to the fixed-point case, where the whole distribution of J774A.1 cells is nearly fully enclosed within the confidence region of Raw264.7 data points. It is also possible to see that the data in the fixed-point case are less uniformly distributed within the confidence regions, as can be expected from their higher variability, identified in Fig. 3.

The loading vectors extracted by PCA for both measurement methods also present rather different characteristics. In the case of hybrid scanning, PC1 is always one of the two components providing the best separation with a spectrum similar to the global features of cells which can be seen by comparing it with the spectra of Fig. 3. On the other hand, PCA applied to fixed-point data always extracts fewer significant vectors (here PC3 and PC5). This is likely due to the higher variability in fixed-point spectra, so that the low numbered PCs are dominated by the variation within the group rather than between the cell lines.

A further point of note is the shape of the loading vectors, which indicate the spectral contrast providing the separation. PCA can produce loading vectors with both positive and negative features which can be difficult to interpret. The hybrid scanning loading vectors appear to have easily identified peaks and an overall shape that is more similar to raw spectra (for example, those shown in Fig. 1), which should make interpretation and later analysis simpler.

3.4.

Imaging and Classification

While the use of multivariate analysis for discrimination is common with point measurements in the literature, this analysis approach has also been used with imaging data.³² We can then extend our comparison of the separation power of Raman measurements to include imaging as well as fixed-point and hybrid scanning measurements. Prior to analysis, spectral data are first extracted separately from images of, respectively, J774A.1 and Raw264.7 cells (see Sec. 2.4), and then concatenated together, leading to a matrix containing $\sim 50,000$ spectra. The analysis is thereafter performed as in the previous section.

The PCA of imaging spectra is shown in Fig. 6, which presents a separation distribution comparable to hybrid scanning in Fig. 5(d), while the fixed-point separation is different from both imaging and hybrid results. This similarity in separation happens despite the large differences in how the data are measured: in the case of hybrid spectra, each data point represents one cell, for a total of $\sim 100$ cells per class; on the other hand, PCA of image data involves one spectra from each pixel in the image, with around four to eight cells in the field of view and with far more data points ($\sim \mathrm{25,000}$ per class). The data outside the 95% confidence region in the imaging case are also more spread than in the point measurements case, and the separation is less clear than for the hybrid scanning data. This can be explained by the fact that image pixels contain several types of spectra, leading, for example, to the image contrast in Fig. 1(a), which will induce more spread than optically averaged hybrid point spectra.

Fig. 6

Separation of cell lines by spectra extracted from pixels in Raman imaging by PCA decomposition. (a) Loading vectors that maximize the centroid distance, and the corresponding score plot. (b) The separation between the two classes with 95% confidence ellipses. For visibility, only 10% of the data points (randomly selected) are displayed on the graph. The figure represents three experiments.

Despite these differences, it is interesting to note the similarities between the two types of measurements, where the score plots show almost identical trends between imaging data and hybrid scanning. The extracted loading vectors, which maximize the class centroid distance, are very similar (PC1 and PC3 or 4), and in both cases, PC1 is selected for all considered experiments. PC1 presents general cell features in both cases, with a nearly identical C-H stretching, and a fingerprint region presenting general features of lipids and proteins in the imaging case [$1450{\mathrm{cm}}^{-1}$ $\delta ({\mathrm{CH}}_2,{\mathrm{CH}}_3)$, 1650 to $1660{\mathrm{cm}}^{-1}$ amide I)^33,34 along with cytochrome c features (753, 1130, $1575{\mathrm{cm}}^{-1}$),³⁵ while the hybrid PC1 also possesses additional features along with the same peaks.

Of special interest is the striking similarity of the second separating PC between the two measurement modes, where several identical spectral features can be identified, such as 1575 and $1650{\mathrm{cm}}^{-1}$ as discussed above. Several identical features are also present as compensating factors of previously discussed features (asymmetric peaks), such as 1118 to 1134 and $1440{\mathrm{cm}}^{-1}$. While some features in the imaging case [see Fig. 6(a)] cannot be definitely identified as peaks due to noise levels, it is interesting to note that they correspond to peaks in the hybrid scanning case, such as several features in the 1200 to $1350{\mathrm{cm}}^{-1}$ region. Both vectors also possess a strong asymmetric feature at 996 to $1014{\mathrm{cm}}^{-1}$, which can be attributed to phenylalanine ($1005{\mathrm{cm}}^{-1}$, aromatic ring breathing).³⁶

By comparison, fixed-point decomposition leads to vectors that are far less characteristic, where PCs of lower order are selected for separation. Furthermore, there are far fewer visible spectral features in the fingerprint region, apart from the contribution at $800{\mathrm{cm}}^{-1}$ in PC5, which can be attributed to the quartz substrate and should not, in principle, significantly contribute to the separation between the two cell lines. The results indicate that fixed-point measurement is the weakest of the three scanning approaches in terms of the separation ability between groups, the richness of spectral information, and the resulting ease of analysis of the separation vectors.