Hyperspectral stimulated Raman scattering microscopy facilitates differentiation of low-grade and high-grade human prostate cancer

Current diagnosis of aggressive human prostate cancer (PCa) still highly relies on the histopathology. However, the subjective nature of histopathology leads to inevitable discordance among pathologists. Especially for the differentiation between Gleason score 6 (low-grade) and 7 (high-grade), the inter-observer discordance can be up to 40% [1]. Low-grade PCa are slowly growing, which can be left alone without any treatment, and only the high-grade ones require effective treatment. Therefore, there is an urgent clinical need for differentiating high and low grade PCa.

Altered lipid metabolism has been well-recognized as a hallmark of cancer [2]. Intracellularly, excess lipids from de novo synthesis and uptake are stored in lipid droplets (LDs) [3]. Moreover, LD accumulation has been found in many types of human cancers, including brain cancer, breast cancer, colon cancer and others [4–7]. To understand the role of LD in cancer progression, it is necessary to measure compositional information of individual LDs inside single cells. However, such information cannot be obtained by conventional methods, such as lipophilic dye staining, mass spectrometry and nuclear magnetic resonance spectroscopy.

Stimulated Raman scattering (SRS) microscopy enables rapid and sensitive imaging of biomolecules [8, 9]. Multi-color SRS microscopy was developed to achieve virtual histology for many types of cancers [10–14]. By integrating high-speed imaging capability of SRS microscopy and full spectral analysis capability of confocal Raman spectroscopy, our previous study analyzed individual LDs of interest in single cancer cells, and found aberrant cholesteryl ester (CE) accumulation in human PCa tissues [15]. This previous study shows the potential of CE as a marker for diagnosis of aggressive PCa. However, no significant difference in CE accumulation between the low-grade and high-grade PCa was found, primarily because this study only analyzed the composition of manually selected LDs without quantitative analysis of the whole field of view. To differentiate high-grade and low-grade PCa, it is necessary to quantitatively map CE distribution inside LDs of human PCa tissues.

In order to perform quantitative imaging of cholesterol in intact tissues, Wang et al developed a hyperspectral stimulated Raman scattering (HSRS) microscope based on the fingerprint vibrational signals [16]. However, the SRS signals in fingerprint region from LDs of cancer tissues are much weaker than those in C–H stretching region. Recently, Bae et al developed HSRS microscopy to identify different subtypes of glioblastoma pathologies [13]. Liao et al developed a HSRS microscope based on the C–H stretching vibrational signals to analyze lipid composition in LDs of single live cells [17]. By using this method, Li et al studied the lipid unsaturation level of LDs in ovarian cancer stem cells (CSCs) [18].

Here, we demonstrated quantitative mapping of CE percentage and unsaturation level of LDs in human prostate tissues by measuring C–H stretching Raman bands with a HSRS microscope. Thanks to high spectral resolution of our system, we found a characteristic cholesterol Raman band around 2870 cm⁻¹ (which has been reported by [19]), and the height ratio between this cholesterol band and the CH₂ stretching band around 2850 cm⁻¹ linearly increased as the CE molar percentage increased in the total neutral lipids. In the meanwhile, when triglyceride (TG) and CE have the same unsaturation level (the number of C=C bonds on the fatty acid chains), the height ratio between the =CH bond band around 3006 cm⁻¹ and the CH₂ stretching band around 2850 cm⁻¹ was inversely proportional to the molar percentage of CE present in the total neutral lipids. Thus, simply based on such height ratios, CE percentage and unsaturation level could be quantitatively mapped in PCa tissues. In total of 6 normal prostate tissues, 9 low-grade and 9 high-grade PCa tissues from patients, we found lipofuscin accumulation in all the normal prostate but none in PCa. While all the high-grade PCa tissues had LD accumulation, only 3 low-grade PCa tissues had LD accumulation. Notably, among all the PCa tissues with LD accumulation, no significant difference in LD amount was found between low-grade and high-grade PCa. Fortunately, both CE percentage and unsaturation level of the LDs were significantly different between low-grade and high-grade PCa. Furthermore, it was shown that CE percentage could differentiate low-grade and high-grade PCa with high sensitivity and specificity.

2.1. HSRS microscope

2.2. Chemicals preparation

2.3. Human prostate tissues preparation

2.4. MCR analysis

We used multivariate curve resolution (MCR) algorithm as described previously [20] to calculate lipid, protein and lipofuscin concentrations in figure 1. MCR is a bilinear model capable of decomposing a measured spectral image set D into concentration data profiles and spectra of input chemical components, represented by matrices C and S^T , as below:

$$\begin{align}D = C \cdot {S^T} + E,\end{align} \tag{ 1 }$$

Zoom In Zoom Out Reset image size

here, T means the transpose of matrix of S. E is the residual matrix or experimental error. The SRS spectral data matrix D is decomposed into the product of two smaller matrices C and S, C is the concentration matrix of each input component and S is the output spectra of all fitted components. The output concentration map of simulated chemical components at each pixel is represented to be percentage relative to the intensity of the MCR fitted spectrum. Given initial input of pure spectra from prior knowledge, an alternating least squares (ALS) algorithm calculates the concentration C and spectrum S by the equation iteratively. The results finally fit the data matrix D with the least relative error. Both concentration and spectra data profiles are served as constraints during the alternating least squares iteration.

2.5. CE percentage and unsaturation level analysis

Given that neutral lipids in LDs are predominantly triacylglycerol (TG) and CE, we conducted HSRS measurements of mixed emulsions containing cholestery oleate and glyceryl trioleate, as the standards of certain CE molar percentage, to calibrate the relation between CE percentage and Raman height ratio. The CE molar percentage was assigned as c, and so the TG molar percentage was (1 − c). The height ratio (r, I₂₈₇₀/I₂₈₅₀) between the Raman bands (2870 cm⁻¹/2850 cm⁻¹) was correlated with the CE molar percentage (c), as shown in the following equations:

$$\begin{align}r = \frac{{{I_{2870}}}}{{{I_{2850}}}} & = \frac{{\left( {1 - c} \right){I_{TG2870}} + c{I_{CE2870}}}}{{\left( {1 - c} \right){I_{TG2850}} + c{I_{CE2850}}}} \nonumber \\ & = \frac{{{I_{TG2870}} + \left( {{I_{CE2870}} - {I_{TG2870}}} \right)c}}{{{I_{TG2850}} + \left( {{I_{CE2850}} - {I_{TG2850}}} \right)c}},\end{align} \tag{ 2 }$$

$$\begin{equation}c = \frac{{{I_{TG2870}} - {I_{TG2850}}r}}{{\left( {{I_{CE2850}} - {I_{TG2850}}} \right)r - \left( {{I_{CE2870}} - {I_{TG2870}}} \right)}},\end{equation} \tag{ 3 }$$

where I_TG2870 represents the peak height of the Raman band (2870 cm⁻¹) for pure TG sample; I_TG2850 represents the peak height of the Raman band (2850 cm⁻¹) for pure TG sample;

I_CE2870 represents the peak height of the Raman band (2870 cm⁻¹) for pure CE sample; I_CE2850 represents the peak height of the Raman band (2850 cm⁻¹) for pure CE sample.

Based on the HSRS spectra of LD emulsion, the correlation between height ratio (r) and CE molar percentage (c) is expressed by the following equation:

$$\begin{equation}r = \frac{{0.814 + 0.190c}}{{1 - 0.0374c}},\end{equation} \tag{ 4 }$$

where r represents the height ratio between the Raman bands (2870 cm⁻¹/2850 cm⁻¹) and c represents CE molar percentage. Thus, by simply using SRS images at the Raman bands 2870 and 2850 cm⁻¹, CE percentage in LDs could be quantitatively mapped in PCa tissues.

Under the condition that TG and CE have only one C=C bond on each fatty acid chain (unsaturation level equals 1), the height ratio (r', I₃₀₀₆/I₂₈₅₀) between the Raman bands (3006 cm⁻¹/2850 cm⁻¹) can be fitted by a standard curve, as below:

$$\begin{equation}r{^{^{\prime}}} = - 0.0899c + 0.130,\end{equation} \tag{ 5 }$$

where r' represents the height ratio between the Raman bands (3006 cm⁻¹/2850 cm⁻¹) and c represents CE molar percentage. The unsaturation level of LDs (the average number of C=C bonds on each fatty acid chain) in tissues was then calculated by the following equations:

$$\begin{equation}{\text{unsaturation level}} = \frac{{{r_{tissue}}}}{{ - 0.0899c + 0.130}},\end{equation} \tag{ 6 }$$

where c represents CE molar percentage calculated by equation (4), and r_tissue represents the height ratio of the Raman bands (3006 cm⁻¹/2850 cm⁻¹) of LDs in tissues hyperspectral SRS images.

2.6. LDs area fraction analysis

2.7. Statistical analysis

We firstly demonstrated that HSRS imaging and MCR analysis were capable of label-free quantitative analysis of lipid composition and lipid area fraction in human PCa tissues in situ, including low-grade PCa (n = 9) and high-grade PCa (n = 9), and normal adjacent tissues (n = 6). We performed HSRS imaging of human prostate tissues at C–H stretching region (2800–3030 cm⁻¹). As shown in figure 1(a), substantial HSRS signals increased from the lipid-rich cell membranes and intracellular LDs. Meanwhile, weak HSRS signals in tissues also were derived from the lipid-poor cell nuclei. Morphologically, SRS images at 2850 cm⁻¹ provided information very similar to that of H&E staining images from the neighboring tissue slices.

The obtained HSRS image stacks were analyzed with above mentioned MCR algorithm. Based on the spectra of lipid, protein, and lipofuscin (a mixture of proteins, lipids, and pigments), MCR algorithm decomposed the stacks into concentration maps of each composition (figure 1(a)). Lipofuscin has a unique frequency vibration (2915 cm⁻¹) between the vibration frequency of protein CH₃ and lipid CH₂. As shown in figure 1(b), only lipofuscin but no LDs were found in normal prostate tissues, while only LDs but no lipofuscin were found in PCa tissues. Moreover, lipid area fraction was found to be significantly greater in high-grade PCa than that in low-grade PCa. Unfortunately, if ruling out the undetectable cases in low-grade PCa (6 out of 9), no significant difference was found between the rest low-grade cases and the high-grade ones (figure 1(c)). This makes difficult to differentiate high-grade PCa from low-grade PCa solely relying on the LDs fraction described in section 2.6. Thus, there is a need to further quantitatively map lipid composition of LDs in PCa tissues in situ.

In order to quantitatively map the lipid composition in LDs, we conducted HSRS experiments of C–H region in CE–TG mixture emulsions of different CE molar percentages. As shown in figure 2(a), The C–H stretching region covers CH₂ stretching band for all lipids (2850 cm⁻¹), characteristic band for cholesterol (2870 cm⁻¹), and =CH stretching band for unsaturation level (3006 cm⁻¹). As shown in figure 2(b), we found that the experimental data of height ratio I₂₈₇₀/I₂₈₅₀ matched well with the theoretical curve as described in equation (4).

Zoom In Zoom Out Reset image size

As shown in figure 2(c), the height ratio I₃₀₀₆/I₂₈₅₀ was negatively correlated with the CE molar percentage, which was linearly fitted as the equation (5). This finding suggests that, if LDs contain significant amount of CEs, the height ratio I₃₀₀₆/I₂₈₅₀ will not only be affected by unsaturation level but also the CE percentage as well. In figure 2(c), the fitted curve showed neutral lipids with monosaturate, the upper side of the fitted curve indicated neutral lipids with polyunsaturated fatty acid chains, and the lower side indicated neutral lipids with saturated fatty acid chains.

According to the quantitative analysis method describe above, CE percentage and unsaturation level of LDs were then quantitatively mapped in PCa tissues by simply using SRS images at the Raman bands 2870, 2850, and 3006 cm⁻¹. The molar percentage of CE was calculated with the height ratio I₂₈₇₀/I₂₈₅₀ of each pixel inside LDs areas by equation (4). The unsaturation level of lipids was calculated with the height ratio (3006 cm⁻¹/2850 cm⁻¹) of each pixel inside LDs areas was least squares fitted with by equations (5) and (6). As shown in figures 3(a) and (c), CEs and unsaturation lipids were highly abundant in the LDs in PCa tissues.

Zoom In Zoom Out Reset image size

As shown in figure 3(a), the distribution of CEs was heterogeneous, especially in low-grade PCa. Statistical analysis further showed that CE molar percentage was significantly higher in high-grade PCa compared to low-grade PCa (figure 3(b)). In addition, the amount of CEs in high-grade PCa was almost twice as much as the one in low-grade PCa. In addition, the unsaturation level of lipids was significantly higher in high-grade PCa compared with low-grade PCa (figure 3(d)).

Considering that the difference in CE molar percentage was more significant, CE molar percentage was finally used to achieve differentiation of low-grade PCa and high-grade PCa with higher specificity and sensitivity. As shown in figure 4, receiver operating characteristic curve (ROC) analysis showed that the area under classification (AUC) could be as good as 0.93 using CE percentage as the marker to distinguish low-grade and high-grade PCa. When setting the cutoff of CE percentage at 65%, 92% sensitivity and 83% specificity could be achieved.

Zoom In Zoom Out Reset image size

First of all, this work demonstrates that CE percentage and unsaturation level of LDs in human prostate tissues can be quantitatively mapped by measuring C–H stretching Raman bands with a HSRS microscope. Different with MCR based quantitative methods, Fu et al [21] used the height ratio between Raman bands around 3015 and 2965 cm⁻¹ to quantify CE percentage in LDs. This simplified process enables quantitative chemical imaging of CE and lipid unsaturation by multi-color picosecond SRS microscope, which would be a easier version of imaging setup for biomedical applications. Nevertheless, we notice that the height of the Raman band around 3015 cm⁻¹ is affected by both CE percentage in LDs (as shown in figure 2) and unsaturation level (=CH stretching, same as 3006 cm⁻¹ in our system) of LDs, it would be more accurate to use the height ratio between Raman bands around 2870 cm⁻¹ (specific cholesterol band) and 2850 cm⁻¹ (CH₂ stretching of total lipids) to quantify CE percentage in LDs in intact cancer tissues, especially when unsaturation level varies between CE and TG in LDs.

Secondly, our results suggest the potential of CE as a marker to differentiate low-grade PCa and high-grade PCa. Since the introduction of prostate specific antigen (PSA) screening, PCa has become the most widely diagnosed non-skin cancer in men in the United States [22]. A controversy exits about the usefulness of PSA screening due to high false positive rate. Histopathology works as a gold standard for differentiating low-grade and high-grade PCa. However, the subjective nature of histopathology leads to inevitable discordance among pathologists. Especially for the differentiation between Gleason score 6 (low-grade) and 7 (high-grade), the inter-observer discordance can be up to 40% [1]. Since only the high-grade PCa requires effective treatment while low-grade PCa can be left alone without any treatment, there is an need for an objective and accurate method to differentiate low-grade and high-grade PCa. It has been reported that aberrant CE accumulation occurs in many types of human cancers, including PCa, brain cancer, pancreatic cancer, liver cancer and others [3, 4, 7, 15]. In this study, our data shows that although no significant difference in LD amount was found between low-grade and high-grade PCa among all the PCa tissues with LD accumulation, both CE percentage and unsaturation level of the LDs were significantly different between low-grade and high-grade PCa. Furthermore, CE percentage could be used to differentiate low-grade and high-grade PCa with high sensitivity, high specificity, and big AUC from ROC analysis.

Third, the unsaturation level of lipids was significantly greater in high-grade PCa tissues compared with the low-grade PCa. Our data about the correlation between height ratio I₃₀₀₆/I₂₈₅₀ and CE percentage suggests that both CE percentage and lipid unsaturation level affects the height ratio I₃₀₀₆/I₂₈₅₀. Based on our calibration curve, we found that the LDs in low-grade PCa contained primarily monosaturated fatty acid chains, while the LDs in high-grade PCa contained more polyunsaturated fatty acid chains. Previously, Pan et al found an increase in unsaturated fatty acids in isolated LDs from cancer cells [23]. Li et al found significantly increased level of unsaturated lipids in CSCs compared with non-CSCs [18]. Thus, it is possible that increased lipid unsaturation is also a metabolic marker for advanced PCa and a potential target for PCa therapy.

This work was supported by the National Natural Science Foundation of China (No. 62027824 and No. 81501516 to S Yue), the Open Project Program of Wuhan National Laboratory for Optoelectronics (No. 2018WNLOKF026 to S Yue), the Fundamental Research Funds for the Central Universities (No. YWF-21-BJ-J-549 to S Yue), and the 'Excellent Hundred Talents' Program start-up fund from Beihang University (to S Yue). P W acknowledges support from the National Key Research and Development Program of China (2016YFA0201403), the National Natural Science Foundation of China (62075076).

The data that support the findings of this study are available upon reasonable request from the authors.