2.1.

Raman Tweezer Configuration Setups

Figures 2a and 2b show diagrams of the optical arrangements for backscatter and forward-scatter configurations, respectively.

Fig. 2

Diagrams of the Raman tweezer configuration setups for (a) backscatter and (b) forward-scatter configurations. Key to abbreviations: M, mirror; EF, edge filter; BP, bandpass filter; BS, beam splitter; DM, dichroic mirror; DF, dichroic filter; RD, Raman differentiator; CCD, charge-coupled detector.

2.2.

Cell Preparation

Both PC-3 and MGH-U1 cells were cultured in Ham’s F12, 7% fetal calf serum (FCS), and $2\mathrm{mM}$ L-glutamine. LNCaP cells were cultured in RPMI 1640, 10% FCS, and $2\mathrm{mM}$ L-glutamine. All cultures were grown in T25 flasks at $37°\mathrm{C}$ in a humidified atmosphere of 5% $\mathrm{C}{\mathrm{O}}_2$ in air. Once the cells reached 70% confluence, they were displaced from the flask by addition of $1\mathrm{mL}$ of trypsin. This was followed by addition of $5\mathrm{mL}$ of the growth media. If required, a portion of these cells was used for further subcultures, while the rest were used for experiments. The cells used for experiments were dispensed into $1\mathrm{mL}$ aliquots (equivalent to $1∕6$ of a T25 flask) to falcon tubes and stored on ice until required for analysis. The $1\mathrm{mL}$ aliquots containing cells in growth media were centrifuged at $1500\mathrm{rpm}$ for $5\min$ . The growth media was then decanted off, leaving a cell pellet at the bottom of the falcon tube. $1\mathrm{mL}$ of phosphate-buffered saline solution (PBS) was added to the pellet and aspirated to ensure the cells were fully suspended in the PBS. All tissue-culture media were obtained from Invitrogen, Paisley, UK, and all other reagents were purchased from Sigma-Aldrich, Poole, UK.

2.3.

Cell Trapping and Raman Collection

The trapping and Raman collection methods varied slightly depending on configuration used.

2.4.

Data Processing

Following spectral acquisition, each spectrum was converted into ASCII files and transported into OMNIC software. The background was subtracted from each raw spectrum to obtain the actual cell spectrum (see Section 2.3.1 for background-collection procedure). A similar background-subtraction procedure was used by Chen ⁴³ and Hamden ⁴⁴ The spectral region $700–1750{\mathrm{cm}}^{-1}$ was used for analysis. Spectra were initially subjected to extended multiplicative scattering correction (EMSC), followed by vector normalization and mean centering. Vector normalization, mean centering, and principal components analysis (PCA) were carried out using Matlab (Mathworks, Inc) using built-in algorithms. EMSC was carried out using Unscrammbler (Camo Inc) software. All other statistical analysis was carried out using SPSS Release 11.0.0 (SPSS, Inc).

2.5.

EMSC

The EMSC method was developed by Martens and Stark to correct for additive (baseline) and multiplicative effects in near-infrared spectroscopy.⁴⁵ The superimposition of baselines on Raman spectra results in baseline shift, which requires removal as a preprocessing step. Recently, EMSC has found applications in correcting Raman spectra of biological samples.^{46, 47, 48, 49}

EMSC was developed from an earlier method, the multiplicative signal correction (MSC) method. In the MSC method, the additive and multiplicative effects are first estimated by least-squares regression with respect to a reference spectrum. The corrected spectrum is then obtained by subtracting every original value by the additive effect and dividing by the multiplicative effect. EMSC builds on MSC by correcting for wavelength-dependent spectral variations from sample to sample.

3.1.

Background Subtracted Spectrum

Figure 3 shows (i) a raw bladder (MGH-U1) spectrum, (ii) the background spectrum of PBS, and (iii) the resulting background subtracted spectrum. The raw spectrum shows an increase in intensity with decreasing wavenumber, which was not observed in the background spectrum and which is most probably caused by fluorescence. This results in a slight slope in the baseline of the background subtracted spectrum.

Fig. 3

Raman spectra of (i) bladder (MGH-U1) cell in PBS (raw spectrum), (ii) background spectrum of PBS, and (iii) background subtracted spectrum of bladder (MGH-U1) cell.

3.2.

Cell-Size Distribution

Histograms for MGH-U1 and PC-3 cell diameters are shown in Fig. 4a and 4b . Optical images of a typical MGH-U1 and PC-3 are displayed in Fig. 4c and 4d. The data show that PC-3 cells are larger than MGH-U1 with a mean diameter of $26.9\mu \mathrm{m}$ $(\pm 4.7\mu \mathrm{m})$ compared with $19.1\mu \mathrm{m}$ $(\pm 3.1\mu \mathrm{m})$ . A one-way ANOVA test shows that these means are not equal at 95% confidence level. Superimposed on each histogram is a normal distribution curve. Visual examination of the histograms and normal distribution curves suggests that PC-3 and MGH-U1 cells are normally distributed. However, the MGH-U1 histogram shows a slight skew of the distribution, with the low diameter end less normally distributed as observed with Chinese hamster ovary cells.⁴⁹ This observation is confirmed by the Kolmogrov–Smirnov (K-S) test, which tests for the probability that a collection of data is normally distributed. At a confidence level of 95%, the K-S test shows that the PC-3 cells are normally distributed but MGH-U1 are not. However, when the confidence level is increased to 98%, both cell types are normally distributed. The fact that bladder cells are smaller than PC-3 cells and that there is probably a minimum size of a cell that could exist might skew the MGH-U1 distribution at the low-diameter end. This could explain why PC-3 cells are more normally distributed than MGH-U1 cells. However, the results show that the cells were randomly selected for analysis without introducing any intentional or unintentional bias into the selection criteria. This is important because there was a concern that the small cells may trap preferentially to the larger ones, particularly the very large prostate cells, but the data show that this is not the case.

Fig. 4

Histograms showing distribution of cell diameter for (a) bladder (MGH-U1) and (b) PC-3. The normal distribution curve is also included. (c) Optical image of a typical trapped bladder and (d) PC-3 cell.

3.3.

Raman Spectra and PCA Plots

Figure 5a shows the vector-normalized mean spectra for prostate and bladder cells taken using the backscattering configuration, with the resulting PCA plot shown in Fig. 5b. The spectra represent the mean of 39 and 38 MGH-U1 and PC-3 spectra, respectively. The data in this PCA plot were obtained from two different cell cultures, i.e., two independent replicates for each cell line, and analyzed over two different days. No subclusters within each cell line are obvious in the PCA plot, which suggests that biological differences within a cell line are small compared with differences between cell lines and that cell culturing procedures are reproducible.

Fig. 5

(a) Mean vector-normalized spectra and (b) PCA plot from backscatter configuration. (c) Mean spectra and (d) PCA plot from forward-scatter configuration. Each point in the PCA plot represents one spectrum.

Figure 5c shows the vector-normalized mean spectra taken in the forward-scattering configuration, with the resulting PCA plot shown in Fig. 5d. The spectra represent the mean of 22 and 20 bladder and PC-3 spectra, respectively. In the PCA plot, each data point represents a single spectrum.

Mean standard deviations/(normalized counts) for both configurations were calculated over the wavenumber range $700–1750{\mathrm{cm}}^{-1}$ . For backscattering, this gave values of 0.015 and 0.032 for MGH-U1 and PC-3, respectively. For forward scattering, values of 0.023 and 0.010 for MGH-U1 and PC-3, respectively, were obtained.

The PCA plots show separation of the PC-3 and MGH-U1 cells for both configurations. Although the separation is not 100% in each case, the data suggests that PC-3 and MGH-U1 cells can be discriminated based on their respective spectra. Separation occurs mainly on the PC4 axis in each case, accounting for 4.6 and 3.4% of the total variance, respectively.

3.4.

Factor-Loading Plot Analysis

To determine which biochemical features were responsible for the discrimination seen in Fig. 5a and 5b, the loadings for the principal components (PC4 in each case) that separated the cells in each PCA plot were plotted together (Fig. 6 ). For the backscattered plot, the majority of PC-3 spectra have positive scores on PC4, while for the forward-scattered plot the majority of bladder spectra have positive scores on PC4. Observation of both plots together shows that many spectral features are anticorrelated to each other (i.e., have a negative loading in one plot but positive loading in the other). Biochemical moieties assigned to these features are likely to be responsible for discrimination in both configurations.

Fig. 6

Loading plots for forward-scatter (FS) and backscatter (BS) configurations. Vertical dashed gray lines link pairs of anticorrelated peaks. Numbers for each pair of anticorrelated peaks relate to band assignments given in Table 1.

Table 1

Assignment of loadings for both back- and forward-scattering configurations. The spectrum ID number for each pair of bands relates to the numbers in Fig. 6. Key to abbreviations: A, adenine; C, cytosine; G, guanine; T, thymine; U, uracil; Phe, phenylalanine; Trp, tryptophan; Tyr, tyrosine.

Spectral features relating to positive peaks for the forward-scattering configuration loading plot are more intense in MGH-U1 cells than PC-3, because MGH-U1 is highly correlated to PC4. Because spectral intensity is related to concentration, biochemical moieties assigned to these features may be present in relatively greater amounts in MGH-U1 cells than PC-3 cells. However, spectral features relating to negative peaks in the backscattering configuration loading plot are less intense in PC-3 than MGH-U1, because PC-3 is highly correlated to PC4. Therefore, biochemical moieties assigned to these features are likely to be present in relatively smaller amounts in PC-3 cells than MGH-U1 cells.

Table 1 shows positions of peaks in the factor-loading plots, with tentative assignments for moieties that are likely to be responsible for discrimination in both configurations. The Raman band assignments were obtained from Ref. 35, 44, 50, 51, 52, 53, 54, 55, 56. The table shows a consistent assignment of Raman bands in the loading plot. This is highlighted by the fact that protein and nucleic acid moieties mainly have positive loadings for the forward-scattering configuration, while all lipid and carbohydrate moieties have positive loadings for the backscattering configuration. For example, bands assigned to spectrum peaks 15, 17, 18, and 20 are lipid moieties, and all have positive loadings for backscattering configuration. Spectrum peaks 12, 13, and 14 are assigned to nucleic acid and proteins, and these have positive loadings for the forward-scattering configuration.

From these assignments, proteins and nucleic acids can be considered to be more abundant in MGH-U1 than PC-3 cells, while lipids and carbohydrates are more abundant in PC-3 cells. Nucleic acids and proteins are mainly found in the cell nucleus, while lipids are largely distributed within the cytoplasm and cell membrane. Therefore, differences in these biochemical amounts between PC-3 and MGH-U1 may be a consequence of differences in the nucleus-to-cytoplasm ratio between these cells. It is possible that this could be linked to cell size, as the proportion of nucleus and cytoplasm probed is likely to vary with cell size. The role of the nucleus-to-cytoplasm ratio in the spectral discrimination of cell types was investigated by Romeo ⁵⁷ Determining if this postulation holds, by measuring the relative nucleus area to cytoplasm area in both cell types, is beyond the scope of this work, because the nucleus could not be clearly differentiated in the cell images obtained during data acquisition.

3.5.

Effect of Cell Size on PCA Separation

From the histograms in Fig. 2a and 2b, it was shown that PC-3 cells are generally larger than MGH-U1 cells. Therefore, PCA separation of the PC-3 from MGH-U1 cells may be due to cell size. It is also possible that discrimination may be due to differences in cell biochemistry but as an indirect consequence of the cell size. This is because a small cell has a relatively large surface area to total cell volume compared with a larger cell. Therefore, a smaller cell may contain more membrane biomolecules (membrane lipids and proteins) relative to nonmembrane biomolecules compared with a larger cell.^{58, 59} Thus, it could be possible to explain discrimination of PC-3 from MGH-U1 on the basis of both cell biochemistry and cell size.

To examine the effect of cell size on discrimination, scatter plots of PC score versus cell size (diameter) were generated [Fig. 7a and 7b ]. PC4 score was chosen, because most of the separation occurs on this component [see Fig. 5b and 5d]. Observation of Fig. 7a (backscatter) shows a tentative relationship between cell size and PC4 score, with the score increasing with increasing cell diameter. The majority of MGH-U1 cells have a negative score and are generally smaller than the PC-3 cells, which generally have a positive score. For the forward-scattering plot [Fig. 7b], a relationship between score and size is even more tenuous. The bladder cells distributed significantly more on the score axis than on the cell-size axis. For PC-3 cells they are distributed more on the cell-size axis than score axis. This would suggest that there is little or no relationship between score and cell size.

Fig. 7

Scatter graph showing correlation between cell size and PC4 score for (a) backscatter and (b) forward scatter.

From these plots, no firm conclusions can be made on whether cell discrimination is due to cell size. To further this investigation, we obtained 10 Raman spectra of LNCaP cells using the forward-scattering configuration (parameters and data processing were the same as for the other cells). The mean diameter of these cells was $19.8\mu \mathrm{m}$ ( $\pm 1$ standard deviation), and they were therefore closer in size to MGH-U1 than PC-3 cells. The spectra obtained for the LNCaP cells were added to those from the PC-3 and MGH-U1 cells, and the whole data set was reanalyzed as described in Section 2.4.

The resulting PCA plot incorporating LNCaP cells is shown in Fig. 8a , and a scatter plot of PC4 score versus cell size for LNCaP, MGH-U1, and PC3 cells is shown in Fig. 8b. Note that each time new data is added to the PCA, the resulting PCs may change. Thus, this new PC4 is not necessarily the same as the one in Fig. 5d and may correspond to different biochemical features.

Fig. 8

(a) PCA of spectra collected using the forward-scattering configuration incorporating LNCaP cells. (b) Relationship between PC4 score and cell size for bladder (MGH-U1), PC-3, and LNCaP cells.

Figure 8a shows that the majority of LNCaP cells are separated well away from MGH-U1 and PC-3, with the latter two cell types clustering close together. This result suggests that cells of similar size can be discriminated, as LNCaP cells do not cluster close to MGH-U1 cells. The lack of a relationship between discrimination and cell size is also tenuously confirmed by the scatter plot of the new PC4 score versus cell size [Fig. 8b], in which the majority of LNCaP cells, which are similar in size to MGH-U1 cells, are separated on the PC4 axis, suggesting cell size is not the dominant factor responsible for discrimination. However, the influence of cell size on the discrimination of cells has not been conclusively determined, as two LNCaP cells have clustered with MGH-U1. Thus, the influence of cell size on the separation of cells need to be investigated further by studying more cell lines.

3.6.

Consideration of Photoinduced Damage

It is likely that the laser powers used for this study induce some degree of photodamage to the cell, principally, at these wavelengths, photo-oxidation. A study by Puppels showed that damage to human lymphocytes occurs at the $514.5\mathrm{nm}$ wavelength for powers of $5\mathrm{mW}$ ,⁶⁰ and furthermore, Wood showed that prolonged exposure at $632.8\mathrm{nm}$ using as little as $1\mathrm{mW}$ could induce changes in the spectra of human erythrocytes.⁶¹ It was therefore important to examine the impact of photoinduced damage on our results, even though our ultimate aim is to classify the cells and thus the fate of the cell is not necessarily important.

To study the influence of exposure, we trapped a PC-3 cell in backscattering configuration for an extended period of time, taking spectra every $30\mathrm{s}$ . Every spectrum obtained for the first $3\min$ , together with spectra obtained for 5, 10, 15, 20, and $25\min$ exposure time, was added to those from the PC-3 and MGH-U1 cells [same data as in Fig. 5b], and the whole data set was analyzed as described in Section 2.4. Figure 9a shows the resulting PCA plot, with the additional time-course spectra highlighted in blue. This shows that all the additional PC-3 spectra have positive PC4 scores and have clustered with the PC-3 spectra (in red) obtained from the normal exposure time of $30\mathrm{s}$ .

Fig. 9

(a) PCA plot incorporating additional PC-3 time-course spectra. (b) Expanded region of the PCA plot in which the additional time-course spectra are located. Each data point in (b) is labelled with its respective exposure time (in minutes).

The region of the PCA plot in which the additional time-course spectra are located is expanded, as shown in Fig. 9b. Each data point is labeled with its respective exposure time (in minutes). This figure clearly shows a trend of increasing exposure time from left to right on the PC3 axis (increasing positive PC3 score with increasing time).Therefore, the PC3 axis could represent an increase in cell photodamage for these spectra.

However, the most important point of both Fig. 9a and 9b is that the separation between the spectrum obtained at $0.5\min$ (normal exposure) and spectra obtained at subsequent longer exposures is smaller than the separation between PC-3 (red) and MGH-U1 (black) spectra. This is most notable for spectra obtained during the first $3\min$ ; however, as exposure time increases, photodamage becomes more pronounced and separation increases. Nevertheless, the data spread of all the time-course spectra is still less than that for the PC-3 data set as a whole.

These findings demonstrate that exposing cells for an extended period of time, at least up to $3\min$ has little effect on the PCA results, shown in Fig. 5b. Therefore, it can be concluded that the impact of photoinduced damage on our results for exposures of $30\mathrm{s}$ at a power of $27.5\mathrm{mW}$ can be considered minimal.

It is important to stress, however, that we cannot rule out the fact that photodamage may occur with these powers at exposure times of less than $30\mathrm{s}$ . Unfortunately, signal-to-noise levels at reduced powers or exposure times would make analysis of the photodamage difficult. What we can say is that if photodamage does occur within the first $30\mathrm{s}$ it does not prevent identification/classification of the cells and therefore does not affect the conclusions drawn from the data.