2.1.

Cancer Cell Lines

The MCF10 model is a panel of isogenic human breast epithelial cell lines, namely MCF10A (nonmalignant), MCF10AneoT (premalignant), MCF10CA1h (malignant, but not metastatic), and MCF10CA1a (metastatic), each at a distinct stage of cancer progression.²¹ Spontaneously immortalized, nonmalignant breast epithelial MCF10A cells were developed from a human benign fibrocystic breast disease specimen. Introduction of the H-Ras oncogene in MCF10A cells gave rise to the premalignant MCF10AneoT cells which were serially passaged in immunocompromised mice resulting in the accumulation of tumor-promoting mutations. This led to the development of human tumor xenografts with varying degree of malignancy, ultimately giving rise to various MCF10CA1 clones. MCF10CA1h cell line consists of well-differentiated tumor cells that are malignant, but not metastatic, whereas MCF10CA1a line consists of poorly differentiated cells that are metastatic in immunocompromised mice.^21–26 These cell lines have been extensively characterized by cytogenetic and microarray analysis and have been used to study key signaling pathways and molecular and genetic events in tumorigenesis.^27–29 Since the cell lines in the MCF10 series are isogenic and represent the major stages of breast cancer progression, they were an ideal choice for this study. All the cell lines were obtained from Barbara Ann Karmanos Cancer Institute (Detroit, Michigan).

2.2.

Cell Lines Growth Conditions

MCF10A and MCF10AneoT cells were cultured and maintained in Dulbecco’s modified Eagle’s medium nutrient mixture F-12 HAM [with 15 mM 2-[4-(2-hydroxyethyl)piperazin-1-yl]ethanesulfonic acid (HEPES), ${\mathrm{NaHCO}}_3$, pyridoxine, and l-glutamine; Sigma] supplemented with 5% horse serum (Sigma), $100\mathrm{ng}/\mathrm{mL}$ choleratoxin (Sigma), $20\mathrm{ng}/\mathrm{mL}$ EGF (SAFC Biosciences, St. Louis, Missouri), $10\mu \text{g}/\mathrm{mL}$ insulin (Sigma, St. Louis, Missouri), and $0.5\mu \text{g}/\mathrm{mL}$ hydrocortisone (Sigma). MCF10CA1h and MCF10CA1a cells were grown in Dulbecco’s modified Eagle’s medium nutrient mixture F-12 HAM (with 15 mM HEPES, ${\mathrm{NaHCO}}_3$, pyridoxine, and l-glutamine; Sigma) supplemented with 5% horse serum. All the cells were cultured in 5% ${\mathrm{CO}}_2$ incubator at 37°C.

2.3.

Sample Preparation

For 2-D culture, the cells were cultured in a $25\text{-}{\mathrm{cm}}^2$ flask (Nunc, Rochester, New York) to reach 80% confluency. The cells were trypsinized and 0.5 mL of the cell suspension was added to 35-mm tissue culture dishes (Falcon, Franklin lakes, New Jersey) with $1.5\mathrm{cm}\times 1.5\mathrm{cm}$ gold slides placed inside the dish. Subsequently, 2.5 mL of culture medium was added and the cells were cultured to reach 80% confluency. By using this method, the cells were grown both on the surface of the gold slide and in the tissue culture dish. The slides were subsequently taken out and placed into a new tissue culture dish. The cells on the gold slide were rinsed twice with Gibco phosphate-buffered saline (PBS) at 37°C (pH 7.4, Life Technologies, Carlsbad, California) to remove the remaining culture medium and placed in 1.3 mL PBS to prevent the cells from drying out during measurement. The main steps involved in sample preparation are schematically illustrated in Fig. 1.

Fig. 1

Schematic of a two-dimensional (2-D) and three-dimensional (3-D) cell culture systems: (a) 2-D; (b and c) 3-D culture models (top) and 3-D cell sample preparation (bottom). 3-D cell lines were first cultured in a 24-well plate in an embedded model, washed, and a piece of gel was then taken, placed onto the gold slide for Raman measurement that followed.

In Fig. 1, the 2-D cell culture and the two common types of 3-D cell culture models are shown: Fig. 1(b) shows, an “embedded model” in which a mixture of cell suspension with liquefied Matrigel is prepared as a first step followed by the addition of culture media after the solidification of the Matrigel. The second method for 3-D cell culture preparation is the “on-top model,” Fig. 1(c), where solidified Matrigel is used as a substratum before adding the mixture of culture media and cell suspension. Here, the cells adhere to the surface of the Matrigel and initiate the formation of clusters on the surface. In this work, in order to keep the cell clusters alive during measurement, the embedded model was adopted as the surrounding Matrigel protects the specimen from drying out due to the evaporation of the liquid medium during measurement.

For 3-D cell culture, $2\times {10}^5\text{cells}$ were resuspended in 10 μL of growth medium and added to 150 μL Matrigel basement membrane matrix (BD Biosciences, San Jose, California) on ice. The mixture of cells and Matrigel was evenly layered in a 24-well tissue culture treated plate (BD Falcon, Franklin lakes, California) and incubated in a 5% ${\mathrm{CO}}_2$ incubator at 37°C for 30 min. After 30 min, 1.5 mL of prewarmed growth medium was gently layered from the side of the well onto the matrix/cell mixture. The cells were maintained in the 3-D culture at 37°C with 5% ${\mathrm{CO}}_2$ for 5 days with the medium replaced every 2–3 days. The cells were allowed to form 3-D spheroids and then imaged using the Ziess AxioObserver microscope with Axiovision software.

The cells were analyzed after 5 days in the 3-D culture. The culture medium was removed and the cells were washed twice with PBS to remove the remaining culture medium; PBS in the well was subsequently removed and a section of the Matrigel with 3-D cell spheroids was removed with a spatula. The Matrigel was placed on the gold slide for the spectral analysis, and the remaining Matrigel was kept in the incubator for further measurements.

2.4.

Data Collection by Raman Spectrometer

The Senterra Confocal Raman Spectrometer (Bruker Optics Inc., Billerica, Massachusetts), fitted with a $50\times$ objective was used. The 50 μm pinhole was activated to remove stray light from the out-of-focus plane. Cells grown in 2-D and 3-D cultures are shown in Figs. 2(a) and 2(b), respectively. A 633 nm He-Ne laser with 20 mW power was used for excitation and an integration time of 45 s was used for Raman spectra acquisition. During measurement, the cells attached to the gold slide displayed a spindle form, thus no tissue damage was observed under the laser conditions used. This can be explained by the fact that cells gradually display a round shape instead of the spindle shape when dying. The schematic of the optical layout is illustrated in Fig. 2(c).^30–34

Fig. 2

Schematic of sampling for 2-D and 3-D culture systems: (a) five spots were chosen for data collection in each 2-D cell, (b) 15 spots distributed throughout the spheroid were chosen in the 3-D spheroid, and (c) schematic of the optical layout of the Senterra Confocal Raman spectrometer: the laser is directed through a neutral density filter to excite the sample through the objective and the scattered photons are collected by the same objective. A pinhole is used to detect photons from the plane of focus.

For both 2-D and 3-D cell samples, three independent biological replicates were performed. In the 2-D cell cultures, 10 cells were measured in each cell, five measurements were made by choosing different spots (one is at the center) to ensure that all components of the cell were probed. For the 3-D cell sample, since the cells formed clusters, called spheroids, the samples were studied as spheroids rather than as single cells. In each biological replicate, four spheroids were chosen for analysis and 15 different spots were chosen across the spheroid (one center point and four spots surrounding this spot and 10 spots at the exterior) for measurement in each sample. The data collection and analysis for both 2-D and 3-D cultures are noted below.

2.5.

Data Pretreatment

Data analysis was done using R software. The raw spectra were first processed to accommodate the $500\text{−}1800{\mathrm{cm}}^{-1}$ range of the Raman shift which contains most of the biological fingerprints.^17–19 The spectra were baseline corrected by fitting polynomials using the R software to remove the effects of fluorescence and allow for the acquisition of spectra with band edges up to the theoretical baseline to visualize minor peaks. The spectra obtained were normalized to ensure that the area under the curve was unity, for further evaluation.

2.6.

Statistical Analysis

The statistical analysis of this study was conducted in three steps. First, PCA was applied to assess data variability and to detect outliers. These outliers were removed from the data set and consequently not included in the subsequent statistical techniques. In sequence, PLS-DA, a well-established classification technique was applied to the data to actually classify the samples based on their cell lines. Beyond the development of classification models, in a third step, an interpretation of the main differences between cells in different stages of cancer was attempted by elastic net,³⁵ providing an interesting visualization of specific compounds associated with this progression. Elastic net is a version of penalized least squares that combines both Ridge and Lasso regressions. In this technique, the wavenumbers that reveal an explicit relationship with the different stages of tumor progression are selected with their respective coefficient estimates that represent the extent of contribution of each of these wavenumbers to the correct classification of each cell line.

3.1.

Raman Spectra

Figure 3 shows the average Raman spectra from a total of 150 spectra for the 2-D [Fig. 3(a)] and 180 spectra for the 3-D [Fig. 3(b)] cell culture model in the range $500-1800{\mathrm{cm}}^{-1}$. Each spectrum was previously baseline corrected and normalized as described in Sec. 2. Standard deviation information is also provided.

Fig. 3

Mean Raman spectra of (a) 2-D and (b) 3-D cell cultures of benign, MCF10A (black), premalignant, MCF10AneoT (dark gray), malignant, MCF10CA1h (medium gray), and metastatic MCF10CA1h (light gray). Standard deviation is provided.

Despite the similarities between the 2-D and the 3-D spectra, especially in the region 800–1150, 1200–1350, and $1400-1700{\mathrm{cm}}^{-1}$, differences could be noted in the spectral window $500-700{\mathrm{cm}}^{-1}$, and peaks in the 2-D cell culture samples merged into a broad shoulder in the 3-D cell systems.

Spectra from the metastatic (MCF10CA1a) cell lines had a higher intensity than other cell lines in the region $850-1150{\mathrm{cm}}^{-1}$. This trend was observed in both the 2-D and 3-D cell culture samples, although it is more evident in the 2-D cells. In the 3-D system, it is evident that premalignant samples show higher intensity than other cell lines in the 1451 and $1650{\mathrm{cm}}^{-1}$ peaks, and nonmalignant samples show a higher intensity than malignant and metastatic at the $1650{\mathrm{cm}}^{-1}$ peak. An opposite trend was observed in nonmalignant and premalignant at the $1620{\mathrm{cm}}^{-1}$ shoulder, where the 2-D samples exhibited intensity in the following order; $\text{nonmalignant}>\text{premalignant}>\text{malignant}>\text{metastatic}$, while in the 3-D system the premalignant samples show higher intensity than nonmalignant. However, based only on the visual observation of the spectra, it is not possible to arrive at major conclusions, especially when considering the standard deviation between each cell line. In the 2-D model the minimum–maximum values of the standard deviation in the nonmalignant, premalignant, malignant, and metastatic cell lines were 0.06–1.22, 0.05–1.04, 0.07–0.57, and 0.05–0.57, respectively; in the 3-D models, these values were 0.03–0.21, 0.02–0.24, 0.02–0.18, and 0.02–0.31, respectively. A discussion on the main differences between the spectra of each cell line will be conducted in the next sections supported by statistical analysis.

3.2.

Exploratory Analysis and Outlier Removal by PCA

Using PCA, the original high-dimensional spectral data set containing chemical and biological characteristics of the cancer cell lines was projected to a low-dimensional space where the first principal components (PCs) explained most of the variance between the samples. Samples with $Q$ or $T2$ statistical values over the threshold were classified as outliers and removed from the data set. The procedure was repeated until no outliers were identified. For the 2-D cell culture models, 19 observations from the original data set containing 110 observations were identified as outliers. On the other hand, two from the original 48 observations from the data set of the 3-D cell culture models were identified as outliers and removed from the data set.

The resulting scatter plots obtained by PCA analysis are shown in Fig. 4. For the 2-D cell culture, the first two PCs clearly separated metastatic cells from the other cell lines [Fig. 4(a)]. The inclusion of the third PC provides a 3-D scatter plot that also discriminates malignant, but nonmetastatic from the rest of the group, as shown in Fig. 4(b). Thus, based on the variance of the spectral data from the 2-D cell culture, the two stages of cancer could be separated into two clusters while nonmalignant and premalignant were clustered together. In the case of the 3-D cell culture, the first two PCs displayed four clusters of the cell lines, although some overlap was observed [Figs. 4(c) and 4(d)].

Fig. 4

Three-dimensional cell culture model better discriminates the four stages of breast cancer; PC scores scatter plots: (a), (c), PC1 versus PC2 of 2-D (a) and 3-D (c) cell models, (b), (d), PC1versus PC2 versus PC3 of (b) 2-D, and (d) 3-D models.

3.3.

Classification Based on PLS-DA

Results from our exploratory analysis indicate that Raman spectra can provide sufficient information to develop classification models for different stages of cancer cells. Thus, PLS-DA was further applied to the spectral data with the purpose of obtaining classification models for assigning samples to categories. Based on leave-one-out cross-validation, five and six latent variables were chosen and employed in the development of the 2-D and 3-D cell culture models. Table 1 summarizes the classification results obtained. Accurate classification was obtained for both the 2-D and 3D cell culture models with the accuracy up to 100% for malignant and premalignant in both models. The sensitivity in the validation set for nonmalignant cells is lower in 2D models compare to the sensitivity of nonmalignant cells in 3D models. The overall classification error for the 3-D model is considered to be lower than the 2-D model, especially for the classification of nonmalignant cells.

Table 1

Summary of PLS-DA prediction results of the model developed for classification of cancer cell lines.

Due to the fact that relatively small number of samples was used in the development of the 3-D model, a lower number of latent variables would be preferred in order to guarantee the stability and robustness of the model. However, the number of latent variables employed reflects the complexity of the samples being analyzed, and a reduction in this number leads to an accuracy decay. Nevertheless, Table 2 shows the classification results for the 2-D and 3-D models constructed with one to three latent variables, respectively. Satisfactory results were obtained with only three latent variables for both the 2-D and 3-D models.

Table 2

PLS-DA prediction results of the models developed for the classification of 2-D and 3-D cancer cell lines (number of latent variables are included).

3.4.

Variable Selection by Elastic Net

Elastic net was applied in this study for visualization of the discriminating wavenumbers and to know what extent these wavenumbers contributed to the correct classification of each cell line. Multinomial logistic models were developed by varying the parameter $\alpha$. Perfect statistical classification was achieved for an $\alpha$ value of 0.5 in both the 2-D and 3-D systems. The coefficient estimates from the developed models were then extracted.³³

As demonstrated in Fig. 5, each stage of breast cancer was identified by specific coefficient estimates in both the 2-D and 3-D models; overall, 3-D cells exhibited fewer coefficients for each stage. Premalignant cells were identified by the prominent positive peak at $1228-1234{\mathrm{cm}}^{-1}$ and a negative peak at $1774{\mathrm{cm}}^{-1}$ in both 2-D and 3-D, but only in the 3-D model a prominent positive peak appeared at $612-614{\mathrm{cm}}^{-1}$. High negative coefficients for nonmalignant cells were observed at $1365{\mathrm{cm}}^{-1}$ in the 3-D model. Premalignant, malignant, and metastatic stages were identified by different coefficient estimates in the 2-D and 3-D cell cultures, although the premalignant state yielded a negative peak at $604-612{\mathrm{cm}}^{-1}$. A prominent negative peak for the premalignant cell line was observed at $1651{\mathrm{cm}}^{-1}$ in the 2-D model. Malignant cells were identified by a strong positive peak at $688{\mathrm{cm}}^{-1}$ in the 2-D model and strong negative peaks at 555 and $889{\mathrm{cm}}^{-1}$ in the 3-D and 2-D models, respectively. Higher peaks associated with the metastatic cell line were mostly observed in the 3-D model, with a strong positive peak at $1349-1358{\mathrm{cm}}^{-1}$ and negative peaks in the following regions: 732, 1006, 1236–1240, and $1578{\mathrm{cm}}^{-1}$.

Fig. 5

Elastic net coefficient estimates for $\alpha =0.5$ for (a) 2-D and (b) 3-D cell culture models. A peak indicates that the correct classification of spectra is associated with the corresponding spectral region. A positive peak indicates higher intensity than other cell lines; a negative peak indicates lower intensity.

We examined each of the nonzero coefficient estimates and, based on the existing work, tentative assignment of these coefficients to chemical compounds were made that relate to specific regions in the Raman spectrum (Table 3).

Table 3

Coefficient estimates obtained from elastic net and its tentative peak assignment from the literature (Refs. 32–40).

4.1.

Raman Spectroscopy for Breast Cancer Classification

Based on our results (Fig. 4 and Table 1), the methods and analysis procedures used for classifying 3-D models using PCA and PLS-DA has a similar trend as the 2-D classification, providing evidence that the methods developed are robust.³⁸ Based on PLS-DA, excellent results were achieved in the classification of metastatic and malignant cells (specificity 100%, sensitivity 100%); however, lower sensitivity was observed when classifying nonmalignant cells in the 2-D model compared to that of the 3-D system. Low sensitivity (60%) and high specificity (100%) for nonmalignant cells indicate the possibility of nonmalignant cells to be classified as premalignant cells in the 2-D model; however, malignant and metastatic stages would not be classified as nonmalignant. Higher sensitivity was achieved for nonmalignant cells in the 3-D model, which suggests that the change in the extracellular matrix environment and interaction plays a key role in the development of premalignancy in breast cancer.⁴⁰ Malignant stages are accurately classified in both the 2-D and 3-D models; suggesting that the malignant characteristic is preserved in both the 2-D and 3-D cell culture models.

4.2.

Variable Selection by Elastic Net

Although PLS-DA alone can provide satisfactory classification, this technique cannot reveal sparse variables that contribute to the classification of each cell line, since each of the latent variables used in the development of the PLS-DA models is a linear combination of several original variables, representing a problem in terms of variable selection. Elastic net was used in this study to solve the aforementioned problem, making the data more succinct and simpler, and providing a good interpretation of the model.⁴¹ More importantly, elastic net reveals the relationship between breast cancer stages and discriminating wavenumbers that give the most contribution to the correct classification of each cell line. This knowledge will be useful in understanding the underlying biomolecular processes in the breast cancer tumorigenesis. Based on our results, extracellular matrix plays an important role in the genesis of malignancy. Prominent coefficients indicate that nonmalignant cells exhibit a high content of polysaccharide and the lowest aromatic amino acid content (Table 3),^42,43–51 which is in agreement with the previous report which implicates lipid genesis in neoplastic transformation^52,50 and characteristic aromatic amino acid content during malignancy development.^53,51 High lipid and nucleic acid content are characteristics of metastatic cells, while high aromatic content and lower polysaccharide composition are characteristic of malignant cells.^42–53 These results suggest that lipid, polysaccharide, and aromatic amino acid constituents can serve as potential biomarkers in monitoring tumor progression.