2.1.

Cell Lines

Totally four different cell lines were investigated in this study (Table 1) including human breast cancer cell line BT474 that overexpresses HER2 protein, human breast epithelial cell MCF-10A with HER2 amplification (HER2+ control), MCF-10A without HER2 amplification (HER2− control), and BT474 with acquired lapatinib resistance. The controls were obtained by transducing MCF-10A cells with a retroviral vector-encoding human HER2 or with vector alone, as previously described.²⁴ To generate drug-resistant cells, HER2+ breast cancer cells BT474 were cultured in increasing concentrations of lapatinib, following published methods.^25–27 BT474 cells exhibit an ${\mathrm{IC}}_{50}$ to lapatinib of $\le 0.1\mu \text{M}$ (Ref. 14). At this time, BT474 lapatinib-resistant cells are growing in the laboratory in the presence of 2-μM lapatinib, whereas the parental BT474 serves as drug-sensitive cells. Both lapatinib resistant and sensitive cells were treated with lapatinib overnight (12 to 16 h) before being processed for RS measurements.

Table 1

Cell lines investigated in the current study. Raman spectra were collected from multiple batches of cell culture from each group of cells. The lapatinib sensitive and resistant breast cancer cells were treated with lapatinib before spectral acquisition.

All the cell lines were seeded in Petri dishes at a density of $1\times {10}^6\text{cell}{\mathrm{ml}}^{-1}$ and grown in a Dulbecco’s modified Eagle medium in a humidified 5% ${\mathrm{CO}}_2$ atmosphere at 37°C. The cells were scrapped free from the dish using a disposable cell scraper, transferred to a 15-ml centrifuge tube, and centrifuged at 1000 rpm for 3 min at room temperature. The supernatant was aspirated, and the cells were washed by three series of saline, followed by centrifugation at 1000 rpm for 3 min after each wash. After the last wash, the cell pellets were transferred to a quartz microscope slide for Raman measurement.

2.2.

Raman Spectra Collection

Raman spectra were acquired using a confocal Raman microspectroscopy (Ramascope Mark III, Renishaw Inc., Gloucestershire, United Kingdom) with a 785-nm diode laser (Innovative Photonic Solutions, Monmouth Junction, New Jersey). Each spectrum was acquired through a $50\times /0.75\mathrm{NA}$ objective (Leica Microsystems Inc., Buffalo Grove, Illinois, $50\times$ N Plan) using 30-s exposure time and two accumulations with a laser power of 30 mW on the sample. Raman shift of the collected spectra is 300 to $1800{\mathrm{cm}}^{-1}$, and the spectral resolution is $3{\mathrm{cm}}^{-1}$. Multiple spectra were acquired from each pellet. A summary of cell lines and the total number of spectra from each group is listed in Table 1. Background spectrum of the quartz slide was subtracted from each sample spectrum using the “automatic spectral subtraction function” in GRAMS/AI spectroscopy software (Thermo Fisher Scientific Inc., Waltham, Massachusetts). The resulting spectrum equals to the source sample spectrum minus a factored background. The final background factor is automatically determined by iteratively subtracting the background multiplied by various factors in the range of 0 to 2 in a step of 0.0001 until the residual reaches the tolerance of 1e-6. Fluorescence background was subtracted using an automated modified polynomial fitting method written in MATLAB (MathWorks, Natick, Massachusetts).²⁸ To account for inherent variation in intra- and intersample absolute signal intensities, the spectra were normalized to their respective mean intensity in the fingerprint range of Raman shift 700 to $1750{\mathrm{cm}}^{-1}$ (Ref. 29).

2.3.

Data Analysis and Discrimination Algorithm Development

Raman spectra were pooled and analyzed using high-dimension multivariate statistical method in $R$ package called glmnet. A fast and probability-based algorithm for high-dimension discrimination with lasso and elastic-net regularized generalized linear models (glmnet) was applied to classify the cell lines using $k$-fold cross-validation by $k=5$. Glmnet method includes both ridge-regression and lasso penalties for optimally reducing the dimensions through elastic-net. The glmnet package in $R$ allows the data to be fit to follow the probability distributions with multivariate statistical models such as linear, logistic and multinomial, Poisson, and Cox regression models. Here, we used logistic and multinomial model for the analysis purpose by the parameter family “multinomial.” The misclassification rate was set up as the optimal destination to find out the minimum misclassification rate for the best discrimination status under a specific selection of the parameter lambda value. The advantages of glmnet method are its flexibility to include a wide range of common situations and to work on large high-dimension dataset.^30,31

3.1.

Comparing HER2+ BT474 Breast Cancer Cells with HER2 Positive and Negative Controls

Raman spectra were collected from HER2+ breast cancer cells BT474, HER2+ control (MCF-10A/HER2), and HER2− control (MCF-10A). The mean spectra with $1\times$ standard deviation (shaded area) are shown in Fig. 1 with offset for clarity in display. Spectral characteristics at various wavenumbers were observed to be different between the cell lines. The spectral regions that showed statistically significant differences across groups are marked with peak positions, which mainly appear in the region of lipid (1304, 1338, and $1660{\mathrm{cm}}^{-1}$), protein (852, 872, 1003, and $1260{\mathrm{cm}}^{-1}$), and nucleic acids (721, 758, 781, and $1088{\mathrm{cm}}^{-1}$) bands.^32–34 The Raman bands centered in the region of 1440 to $1450{\mathrm{cm}}^{-1}$ and at $1304{\mathrm{cm}}^{-1}$ arise from ${\mathrm{CH}}_2$ deformation modes of protein and lipids^35,36 and thus will be assigned in this study along with other accompanying bands for both molecules.

Fig. 1

Mean spectra from HER2+ breast cancer cells. HER2− control (a), HER2+ control (b), and BT474 (c) with $1\times$ standard deviation (shaded area). For clarity in display, spectra were offset from the baseline. The scale bar indicates the intensity of Raman scattering unit. Selected spectral bands with statistically significant differences ($p<0.05$) among the cell lines are labeled with peak positions.

To further investigate the biochemical differences between the cell lines, spectra were subtracted to reveal the difference between HER2+ BT474 cancer cells and controls. The difference spectra were calculated by subtracting the normalized mean spectrum from HER2+ control from HER2− control [Fig. 2(a)] as well as BT474 from HER2− [Fig. 2(b)] and HER2+ [Fig. 2(c)] controls. Increased nucleic acid content in cancer cells is revealed by the positive difference at 752 and $1088{\mathrm{cm}}^{-1}$ that were assigned to the phosphodiester groups in nucleic acids ($p<0.05$, Figs. 2(b) and 2(c)].³⁴ Slightly enhanced nucleic acid content was also observed in the difference spectra between HER2+ and HER2− controls [$p<0.01$, Fig. 2(a)], indicating that HER2 overexpression leads to active cell proliferation. The intense positive difference at $1594{\mathrm{cm}}^{-1}$ [Figs. 2(b) and 2(c)] can be attributed to the ring $\mathrm{C}═\mathrm{C}$ stretch vibrations of nucleic acids,^37,38 in agreement with the elevation of nucleic acid content in HER2+ cells indicated by 752 and $1088{\mathrm{cm}}^{-1}$.

Fig. 2

Difference spectra of HER2+ control/HER2− control (a), HER2+ BT474 cancer cells/HER2− control (b), and HER2+ BT474 cancer cells/HER2+ control (c) with offset for display. The straight lines mark the original $y=0$ line for each difference spectrum. Selective spectral bands with significant differences ($p<0.05$) among the cell lines are marked with peak positions.

The difference spectra of BT474 to the controls [Figs. 2(b) and 2(c)] suggest decreased proteome with HER2 amplification by the strong negative spectral features at the amide III (1247 and 1270/1278) and $1438{\mathrm{cm}}^{-1}$ ($p<0.01$). The latter has previously been assigned to ${\mathrm{CH}}_2$ and ${\mathrm{CH}}_3$ deformations in the side chain of collagen in normal breast tissue.^33,39,40 The protein amide I band at $1650{\mathrm{cm}}^{-1}$ did not show consistent variation as amide III or the band at $1438{\mathrm{cm}}^{-1}$, which confounded the interpretation on proteome content in the current study. The Raman bands at 1003 and $1625{\mathrm{cm}}^{-1}$ arise from the ring structures of phenylalanine and tryptophan/tyrosine, respectively.^41,42 Both bands are more intense in cancer cells [Fig. 2(b) and 2(c), $p<0.01$), suggesting possible elevation in the expression of such aromatic amino acids rich proteins.

The positive difference at $1344{\mathrm{cm}}^{-1}$ in Figs. 2(b) and 2(c) arises from the wagging mode of ${\mathrm{CH}}_2$ and ${\mathrm{CH}}_3$ in lipids or proteins.⁴⁰ This enhancement along with the concurrent increase ($p<0.01$) at $1666{\mathrm{cm}}^{-1}$, which was assigned to the $\mathrm{C}═\mathrm{C}$ stretching in unsaturated lipids,⁴⁰ indicates elevated lipid content in HER2+ BT474 cancer cells. HER2+ control also demonstrates a slight increase in lipid composition than HER2− control evidenced by the positive difference at 1457 and $1663{\mathrm{cm}}^{-1}$ [Fig. 2(a), $p<0.05$].

The enhanced lipid content in breast cancer cells comes from exacerbated synthesis of fatty acid and phospholipids.⁴³ The lipogenesis regulated by fatty acid synthase is associated with upregulation of HER1/HER2 tyrosine kinase receptors in the cells.^44,45 Our study indicates augmented lipogenic process with HER2 overexpression in both cancer and control epithelial cells, suggesting that the excessive HER2 expression might play an important role in promoting the upregulation of fatty acid synthase. This is in agreement with a previous report that the HER2 overexpression stimulates the fatty acid synthase promoter.⁴⁶

The probability-based multivariate statistical analysis in glmnet was used to discriminate the cell lines and to identify specific spectral features that are important for classification. The confusion matrix for classification from the glmnet analysis is shown in Table 2. The rows of the table show the numbers of spectra measured for each cell line (true), whereas the columns list the number of spectra predicted to each group. The sensitivity and specificity of discrimination for each cell line are calculated by the ratio of true prediction (bold) to the sum of all true values in each row and to the sum of all predicted values in each column, respectively. The spectra from BT474 (131 out of 132) were discriminated from controls with 99.3% sensitivity and 100% specificity. HER2+ and HER2− controls were classified into the correct groups with 100% sensitivity and 98.8% and 100% specificities, respectively, indicating the potential of RS in identifying HER2+ breast cancer cells.

Table 2

The confusion matrix for the classification from glmnet analysis.

3.2.

Characterizing Lapatinib-Sensitive HER2+ BT474 Cancer Cells and Lapatinib-Resistant HER2+ BT474 Cancer Cells with Treatment

Mean Raman spectra from lapatinib-treated breast cancer cells ($N=104$) and lapatinib-resistant cells ($N=104$) consist of similar band contours with low variations within each group (Fig. 3). Statistical analysis of the spectra showed significant difference ($p<0.05$) in Raman bands, which are labeled with peak positions on Fig. 3. Phospholipid (1307, 1333, and $1657{\mathrm{cm}}^{-1}$), protein (855, 1156, and $1173{\mathrm{cm}}^{-1}$), and nucleic acids (721 and $781{\mathrm{cm}}^{-1}$) contribute to the major differences between the cell lines.

Fig. 3

Stacked mean Raman spectra from lapatinib-resistant (a) and lapatinib-sensitive (b) breast cancer cells with $1\times$ standard deviation (shaded area). Selective spectral bands with significant differences ($p<0.05$) among the cell lines are marked with peak positions.

Figure 4 shows the difference of spectra from lapatinib-resistant and lapatinib-sensitive cells. With lapatinib treatment, drug-resistant cancer cells have a higher phospholipid content with respect to the drug-sensitive cancer cells, as suggested by the positive differences at 778, 1364 and $1384{\mathrm{cm}}^{-1}$ (Fig. 4, $p<0.01$). Content of unstatuarated lipids appears decreased as indicated by the negative difference at $1651{\mathrm{cm}}^{-1}$ ($p<0.01$). The prominent negative differences at 1282, 1310, 1341, and $1680{\mathrm{cm}}^{-1}$ reflect decreased proteome in lapatinib-treated drug resistant cells ($p<0.05$). Significant positive differences are observed at 730 and $778{\mathrm{cm}}^{-1}$ that arise from the ring breathing of nucleotides,⁴⁰ indicating enhanced nucleic acid abundance in the drug-resistant cancer cells.

Fig. 4

Difference spectrum of lapatinib-resistant cells with respect to the lapatinib-sensitive cells with treatment. The straight line marks $y=0$ for reference. Selective spectral bands with significant differences ($p<0.05$) among the cell lines are marked with peak positions.

Treatment with lapatinib in drug-sensitive cells blocks HER2 phosphorylation and downstream PI3K-Akt signaling.^14,47 Conversely, in resistant cells, lapatinib inhibits HER2 phosphorylation but the PI3K-Akt activity is recovered,²⁷ which involves in the lipogenic pathway in breast cancer cells.⁴⁸ The greater lipid content indicates active endogenous lipogenesis in drug-resistant cancer cells despite the treatment.