2.1.

Cell Preparation

Breast and prostate cancer cells were grown directly on coverslips (MatTek 35 mm glass bottom dishes no. 1, poly-d-lysine coated) until they were $\sim 85\%$ confluent. They were subsequently treated with MPA at $10\mathrm{nM}$ (T47D cells and MDA-MB-231) or with R1881 hormone at 10 nM (LNCaP and PC3) or the vehicle (ethanol) as a control. After 4 days of treatment, the cells were fixed with 4% formaldehyde for 10 min and rinsed with phosphate buffered saline (PBS). For each cell sample, 28 to 32 individual cells were measured with 10 to 20 Raman spectra acquired per cell. Data were acquired with a custom built confocal Raman microscope equipped with an automated scanning stage to acquire point Raman spectra at different sample positions. The cancer cells were provided by the Tissue Culture Core Facility at the University of Colorado, Denver.

2.2.

Experimental Setup

2.3.

Raman Spectra Processing

Micro-Raman data were processed in MATLAB using custom routines and the bioinformatics toolbox. The acquired Raman spectra contain not only the desired signal from the sample, but also a background signal containing cosmic rays and the signals from the glass coverslip and PBS solution. In order to remove the various background signals, we performed the following data processing steps: (1) cosmic rays removal,²⁵ (2) data smoothing using a 5-point moving average filter, (3) background removal by subtracting an average of several point Raman spectra acquired off of the cell sample (glass and PBS only), (4) baseline correction using bioinformatics tool routines (msbackadj regression method with linear interpolation), (5) data normalization by the area under the curve, and (6) wavenumber calibration using polystyrene reference spectra.

2.4.

Least-Squares Fitting

To fit the Raman experimental data, we assume that each LD spectra can be described by a linear combination of the spectra of pure components. In this case, we wish to solve for the coefficients of the linear equation: $\mathbf{d}=\sum _i{c}_i{\mathbf{s}}_{\mathbf{i}}$ where $\mathbf{d}$ is the experimental LD spectrum, $c_i$ are coefficients, and ${\mathbf{s}}_{\mathbf{i}}$ are the pure substance spectra. To solve, we performed a least-squares fit of the data using built-in functions in MATLAB (Mathworks Inc., Natick, Massachusetts) in the Curve Fitting toolbox, constraining the coefficients to positive values. The standard deviation was calculated by fitting each individual Raman spectra acquired for a given sample, assuming a normal distribution in fit coefficients. Because the Raman signal is linearly dependent on concentration, the coefficients determine the relative concentrations of the pure species in the LDs.

2.5.

Principal Component Analysis

Dimensionality reduction methods such as principal component analysis (PCA) and linear discriminant analysis (LDA) have been widely applied for classification of Raman spectra.²⁶ PCA itself is not optimal for data classification, partly because PCA builds biases that are uncorrelated. The class label does not exist in the PCA definition. The eigenvectors of the spectra covariance matrix are calculated and those that explain greater than 99% of the data variability are selected for further processing.

The first principal component (PC1) accounts for the most variance among the data. We find we can elucidate differences between similar spectra if we graph different PCs against each other. When we think of a better algorithm to find discriminant directions, then a good example is the Fischer’s LDA.²⁷ This model is based on finding the line that best separates the classes. The variance between groups is maximized and the variance within a group is minimized according to Fisher’s criterion. Therefore, a combination of PCA and LDA gives a better performance. The first $N$ features (eigenvectors) extracted from PCA are projected onto a lower dimensional space using LDA to form a spectral clustered representation. The overall test error is evaluated and the precision, sensitivity, and specificity are consequently calculated.

2.6.

Lipid Size Distribution Analysis

TPAF and CARS images were processed in ImageJ for enumeration and size analysis of intracellular LDs. All images were taken under similar acquisition conditions; therefore, we used the same parameters for each image when performing image thresholding, filtering, edge detection, and counting. Using the three-dimensional object counter function in ImageJ [with a height of 0 for two-dimensional (2-D) images] we generated the statistics for LD counts and surface voxel number (size). We further imported the statistical data into MATLAB to generate plots of LD size distribution.

3.1.

Lipids Biomarkers by Micro-Raman Spectroscopy

Mean Raman spectra, Fig. 2, were obtained for low-wavenumber (850 to $1800{\mathrm{cm}}^{-1}$) and high-wavenumber (2750 to $3050{\mathrm{cm}}^{-1}$) regions for two breast cancer cell lines, T47D and MDA-MB-231, treated with MPA or ethanol (vehicle, V), and two prostate cancer cell lines, LNCaP and PC3, treated with R1881 or ethanol (vehicle, V) for 4 days. For each sample, 28 to 32 individual cells were measured with 10 to 20 spectra measured for each cell. Figure 2 shows that significant spectral differences between treated (MPA and R1881) and untreated (V) cells exist in both wavenumber ranges for the cell lines LNCaP and T47D. Differences were minimal for the cell lines MDA-MB-231 and PC3, which did not respond to the treatment. The specific Raman peaks responsible for the differences observed in the spectra are readily assigned to known lipid vibrational modes (Table 1), and are sensitive to differences in lipid composition within cytoplasmic LDs. The analysis of these spectral features is described below.

Fig. 2

Average Raman spectra of treated versus vehicle (control) cancer cells. Panels (a) and (b) are the low- and high-wavenumber regions, respectively, of spectra acquired from the breast cancer cell lines T47D and MDA-MB-231. Panels (c) and (d) are low and high wavenumber spectra, respectively, of prostate cancer cell lines LNCaP and PC3. Also displayed at the bottom of each plot are the difference spectra (mean subtraction). The most pronounced changes between spectra of treated and control cell samples occurred for T47D and LNCaP cell lines in the 1250 to $1300{\mathrm{cm}}^{-1}$ region and the peak at $1439{\mathrm{cm}}^{-1}$ (panels a and c) as well as the high-wavenumber region (panels b and d). Raman spectra of treated and control samples for MDA-MB-231 and PC3 cell lines did not show these pronounced differences indicating minimal response to hormone treatment.

Table 1

Band assignments of cancer cell lipids.28,29

3.2.

Lipid Composition Analysis

LDs in cells consist of a neutral lipid core (primarily TAGs and cholesteryl esters) enclosed by a phospholipid membrane. TAGs consist of a glycerol molecule joined by an ester bond to three fatty acid molecules. The Raman signal from cellular LDs occurs primarily from the chemical bonds, $\mathrm{C}─\mathrm{O}$, $\mathrm{C}─\mathrm{C}$, $\mathrm{C}═\mathrm{C}$, and $\mathrm{C}─\mathrm{H}$.^28–33 Unsaturated fatty acids contain more $\mathrm{C}═\mathrm{C}$ bonds (represented by the $1655{\mathrm{cm}}^{-1}$ band), while saturated fatty acids contain more ${\mathrm{CH}}_2$ groups and therefore have larger Raman peaks for those vibrational modes associated with ${\mathrm{CH}}_2$.

In order to identify significant Raman spectral features associated with intracellular lipid composition, we studied various Raman peak intensity ratios corresponding to particular lipid vibrational modes (Table 2). The degree of unsaturation and chain length of fatty acid methyl esters can be well determined by analysis of Raman spectra.³¹ Intracellular LDs have more complex spectra than pure fatty acids (FA); however, we can use similar analysis to infer relative changes in lipid composition in our cancer cell samples. The band intensity at $1655{\mathrm{cm}}^{-1}$ varies linearly with the number of $\mathrm{C}═\mathrm{C}$ bonds and for pure FAs can determine a molar unsaturation value. Alternatively, the degree of unsaturation of a pure sample can be characterized by the ratio of $\mathrm{C}═\mathrm{C}$ bonds to ${\mathrm{CH}}_2$ groups. This ratio has been shown to have a near-linear relationship with the mass unsaturation, the number of $\mathrm{C}═\mathrm{C}$ bonds per gram. The mass unsaturation can be calculated by the ratio of the $1655{\mathrm{cm}}^{-1}$ ($\mathrm{C}═\mathrm{C}$ stretching band) and $1440{\mathrm{cm}}^{-1}$ [$\delta ({\mathrm{CH}}_2)$ scissor], or $1294{\mathrm{cm}}^{-1}$ [$\delta ({\mathrm{CH}}_2)$ twist]. Note that these ratios vary with both the degree of unsaturation and chain length. In our measurements on pure fatty acid samples (Table 3), the mass unsaturation value of oleic acid ($18:1$) is half that of linoleic acid ($18:2$) and is 88% that of palmitoleic acid ($16:1$), as expected.

Table 2

Summary of Raman characteristics of intracellular lipid droplets.

Table 3

Summary of Raman characteristics of pure unsaturated fatty acids.a

The ratio of total unsaturated fatty acids ($3015{\mathrm{cm}}^{-1}$, olefinic $═\mathrm{CH}$ stretching) to the total amount of fatty acids ($2851{\mathrm{cm}}^{-1}$, ${\mathrm{CH}}_2$ symmetric stretch) termed TUFA/TFA³⁴ has been reported and proposed as a molecular marker for normal and tumor cells. This ratio shows similar trends for pure fatty acid samples as an indicator of the number of $\mathrm{C}═\mathrm{C}$ to ${\mathrm{CH}}_2$ species. We observe that this ratio decreases upon treatment for T47D-MPA and LNCaP-R1881 cancer cells, indicating increased saturated lipid content. Indeed, the ratio values for mass unsaturation and TUFA/TFA all indicate that for the hormone responsive cancer cell lines (T47D and LNCaP), there is an increase in saturated lipid content after treatment while the cell lines that are not hormone responsive (MDA-MB-231 and PC3) show less relative change in these ratio values upon treatment.

The ester carbonyl vibrational mode $\nu (\mathrm{C}═\mathrm{O})$ position has been shown to determine the length of a fatty acid chain.³¹ A shift to higher wavenumbers for longer chain length fatty acids has been well characterized for fatty acid methyl esters. In our results in cancer cells, the peak position was found to vary mainly between different cell lines, but with minimal variation upon hormone treatment of the cell lines. Another measure of chain length is the ratio of $[\delta ({\mathrm{CH}}_2)\text{scissor}]/\nu (\mathrm{C}═\mathrm{O})$ and the ${\mathrm{CH}}_2/{\mathrm{CH}}_3$ ratio. Again, we observe more differences between the different cell lines than changes associated with hormone treatment.

3.2.1.

Least-squares fitting

In order to further quantify the composition of the LDs, we performed a least-squares fit utilizing the full experimental Raman spectrum as opposed to ratios of values at particular bands. The fit assumes the experimental spectra are a linear combination of spectra of pure fatty acid components, cholesterol, and cytoplasm. A cytoplasm component is included in the fit because some amount of cytoplasm signal will be recorded if the LD size is smaller than the detection volume of the Raman microscope. The Raman spectra of pure samples in the low- and high-wavenumber regions are shown (Fig. 3).

Fig. 3

Pure fatty acids, cholesterol, and measured cytoplasm Raman spectra, (a) and (b) are the low- and high-wavenumber regions. The pure substances are, from the bottom to the top, oleic acid (Sigma, $>99.5\%$ GC), palmitoleic acid (Sigma, $>98.5\%$, GC liquid), linoleic acid (Sigma, $>99\%$), palmitic acid (Sigma, $>99\%$), stearic acid (Sigma, $>98.5\%$, Grade I, GC), and cholesterol (Sigma, $>99\%$). The palmitic and stearic acids were melted at 70°C.³⁵

We performed fits for low- and high-wavenumber regions for hormone responsive cells (T47D and LNCaP) to compare changes in intracellular lipid composition upon hormone treatment. Results for low wavenumber fits are shown in Figs. 4(a) and 4(b). The fit coefficients correspond to relative concentrations of the given pure substance in the LDs. The low wavenumber data resulted in fits with lower standard deviations. This is likely due to the fact that the low wavenumber spectra have more discrete peaks that can be better separated in the fitting routine. We performed fits with and without the cytoplasm and cholesterol spectra and did not observe any significant differences in the relative concentrations of the fatty acids. When including cholesterol spectra in our least squares fit, we observed a trend of reduced cholesterol concentration in the LDs of between 5% and 10%. We could not state accurately the relative concentrations of cholesterol compared to fatty acids because we were comparing Raman spectra in the solid state to estimate the contribution in the LDs, which may have different density values. It will take further investigation to determine if accurate cholesterol concentrations can be obtained from Raman spectra results.

Fig. 4

Relative concentrations of fatty acids from least squares fit of experimental Raman measurements of lipid droplets (LDs) for T47D (a) and LNCaP (b) cell lines with (MPA and R1881) and without (vehicle) hormone treatment. Standard deviations are shown. (Abbreviations: OA, oleic acid; POA, palmitoleic acid; LOA, linoleic acid; Palm, palmitic acid; and Stear, stearic acid).

As an additional metric, we looked at the relative concentrations between the saturated fatty acids (palmitic and stearic) and unsaturated fatty acids (oleic, palmitoleic, and linoleic). This can be represented as a percent saturation (namely saturated FA concentration divided by total FA concentration). For T47D, the percent saturation was $63.2\pm 7.1$ for vehicle (control) and $65.2\pm 8.3$ for MPA treated. For the LNCaP cell line, the percent saturation was $60.0\pm 5.8$ for vehicle (control) and $65.2\pm 8.3$ for R1881 treated cells.

Results for T47D indicate an increase in the relative amount of saturated fatty acids (palmitic and stearic) to unsaturated in the LDs of MPA-treated cells. These results are in agreement with our previous report,¹² where treatment with MPA leads to increased palmitate accumulation in triglycerides analyzed by chromatography methods. We observed a similar trend for the prostate cancer cell line LNCaP treated with R1881.

The decrease in cholesterol content with MPA or R1881 treatment was unexpected but may suggest that excess palmitic and oleic acids are preferentially shunted to triglycerides instead of cholesteryl esters, a concept already proposed in other models of LD accumulation.³⁶ The meaning of this change in cholesterol distribution in cancer cells treated with hormones is currently unknown, but can open up interesting avenues for further investigations into the role of cholesterol in cancer malignancy.

3.3.

Cancer Cell Line Classification

Multivariate statistical analysis is usually employed for Raman spectra classification of different types of tissue samples for diagnostic sensitivity and specificity assessment. We combined PCA and LDA to obtain a performance of the cancer cells’ discrimination as a function of various lipid signatures. PCA reduces the dimensionality of the data set and builds up the variables necessary for the LDA for a better explanation of the differences among various classes. Fischer’s linear discriminant algorithm is used for determining the parameters of LDA. The Fischer algorithm uses a linear function to project the $n$-dimensional data onto a line and to maximize the ratio of the variances between each of the two classes to the variance within each class.²⁷ We performed PCA analysis for low and high wavenumbers to obtain the eigenvectors with the largest eigenvalues for the greatest variations. We used half of the sample sets specified in Table 4 as the actual group and the other half for the predicted group. For the T47D breast cancer cell line, we chose the first 8 to 10 eigenvectors and for the LNCaP prostate cancer cell line, we used the first 8 to 12 eigenvectors. We imported the eigenvectors from PCA into the LDA classifier to form a cluster representation of the components. For the T47D treated versus untreated breast cancer cells, large differences in Raman spectra were observed in the low-wavenumber region giving a classifier accuracy, sensitivity, and specificity of 100% [Table 4(a)]. For the high-wavenumber region, the sensitivity was 100% and specificity 85.71%. The test error was 0.071. In the case of LNCaP treated and untreated groups, the overall accuracy for low-wavenumber region was 87.33% and the test error was 0.133. We calculated an accuracy of 96.88%, sensitivity of 100%, and specificity of 90.91% for the high-wavenumber region [Table 4(b)]. Both classifiers have $p<0.05$.

Table 4

Confusion matrix (a) for breast cancer cell lines using LDA classifier in the 850 to 1800  cm−1 wavenumber region and (b) for prostate cancer cell lines using LDA classifier in the 2750 to 3030  cm−1 wavenumber region.

3.4.

Label-Free Imaging of Lipid Dynamics Using CARS and TPAF

We recorded simultaneous TPAF and CARS images for each of the hormone responsive cancer cell lines, T47D (Fig. 5) and LNCaP (Fig. 6), with and without hormone treatments. The CARS signal in the epi-direction provides excellent contrast for the intracellular LDs while autofluorescence allows individual cells to be identified. The CARS images were postprocessed with custom routines developed in ImageJ to detect the centroid of each LD and calculate the diameter. We noticed from these images that T47D treated with MPA, Fig. 7(a), and LNCaP treated with R1881, Fig. 7(b), contain an increased number and size of LDs than the T47D and LNCaP treated with ethanol. Although we only used 2-D images for our analysis, we were able to well sample the intracellular LDs because cells were grown in a single layer on the coverslip. With 2-D analysis, we quantify the increase in number and size of LDs upon the treatment for both cell lines. We conclude that changes in LD size and quantity can be determined from CARS imaging and that this provides a noninvasive means to study lipogenesis in cancer cells.

Fig. 5

Autofluorescence and CARS images of treated versus vehicle control breast cancer cells. Panels (a) T47D-MPA autofluorescence image, (b) T47D-MPA CARS image, (c) T47D-V autofluorescence image, and (d) T47D-V CARS image, $\text{field}\text{-}\text{of}\text{-}\text{view}=211\times 211{\mu \text{m}}^2$ (scale bar 30 μm).

Fig. 6

Autofluorescence and CARS images of treated versus vehicle control prostate cancer cells. Panels (a) LNCaP-R1881 autofluorescence image, (b) LNCaP-R1881 CARS image, (c) LNCaP-V autofluorescence image, and (d) LNCaP-V CARS image, $\text{field}\text{-}\text{of}\text{-}\text{view}=211\times 211{\mu \text{m}}^2$ (scale bar 30 μm).

Fig. 7

The size distribution of LDs from CARS-B images. The number of lipids was normalized to number of total cells counted from two-photon images, T47D-MPA—# cells 339, T47D-V—# cells 597, LNCaP-R1881—# cells 490, and LNCaP-V—# cells 492.