Recent advances in animal modeling, imaging technology, and functional genomics have permitted precise molecular observations of the metastatic process. However, a comprehensive understanding of the premetastatic niche remains elusive, owing to the limited tools that can map subtle differences in molecular mediators in organ-specific microenvironments. Here, we report the ability to detect premetastatic changes in the lung microenvironment, in response to primary breast tumors, using a combination of metastatic mouse models, Raman spectroscopy, and multivariate analysis of consistent patterns in molecular expression. We used tdTomato fluorescent protein expressing MDA-MB-231 and MCF-7 cells of high and low metastatic potential, respectively, to grow orthotopic xenografts in athymic nude mice and allow spontaneous dissemination from the primary mammary fat pad tumor. Label-free Raman spectroscopic mapping was used to record the molecular content of premetastatic lungs. These measurements show reliable distinctions in vibrational features, characteristic of the collageneous stroma and its cross-linkers as well as proteoglycans, which uniquely identify the metastatic potential of the primary tumor by recapitulating the compositional changes in the lungs. Consistent with histological assessment and gene expression analysis, our study suggests that remodeling of the extracellular matrix components may present promising markers for objective recognition of the premetastatic niche, independent of conventional clinical information. Cancer Res; 77(2); 247–56. ©2016 AACR.

While local breast cancers are largely responsive to current therapeutic strategies, treatments to permanently eradicate metastasis are yet to be developed. Consequently, nearly all breast cancer-related deaths today result from metastatic disease that involves distant organs (1). The distribution of metastases is a non-random process with each tumor type manifesting a characteristic pattern of metastatic involvement in distant vital organs (2, 3). Stephen Paget's “Seed and Soil” hypothesis originally shifted the attention from a sole focus on the behavior of primary tumor cells to the important role of the stroma at the secondary site (4, 5). Seeking to understand the basis of metastasis organotropism, his seminal hypothesis postulated that a receptive microenvironment at the secondary organ (soil) is crucial to the engraftment of circulating tumor cells (seed). This also provided a conceptual framework for later observations in experimental metastasis assays that cancer cells derived from a distant site display enhanced metastatic ability to that specific organ (6). Yet, it is only with recent advances in animal metastasis assays, genomic profiling, and real-time imaging techniques that the molecular components that drive organ-specific metastasis have been specifically probed. Translation of the preclinical findings on the metastatic microenvironment into a clinical test, however, has not yet been realized.

Building on the seed and soil hypothesis, emerging evidence suggests the formation of a premetastatic niche (7, 8), i.e., collective changes at the target metastasis sites prior to the arrival of the first tumor cells. This niche development in the preferred metastatic sites appears to be driven by soluble growth factors secreted by the primary tumor and recruitment of tumor-associated cells (9). The priming of the secondary organs was initially attributed to the localization of hematopoietic bone marrow progenitor cells expressing vascular epithelial growth factor receptor 1 (VEGFR-1) due to VEGF being secreted by the primary tumor (7). Exosomes secreted from primary tumors have also been reported to play a significant role in mobilizing these progenitor cells to the premetastatic sites (10). The recruitment of tumor-associated cells provides an increased availability of chemokines, growth factors, matrix degrading factors, and adhesion molecules that initiate the metastatic cascade (8, 9). This process is reported to be accompanied by remodeling of the extracellular matrix (ECM) in the premetastatic niche, notably through the upregulated expression of matrix metalloproteinases (MMP; ref. 11), transformation of local fibroblasts, and focal expression of fibronectin. For instance, a recent series of investigations have revealed that lysyl oxidase (LOX), an enzyme secreted by hypoxic tumor cells, modulates the ECM in premetastatic sites by cross-linking collagen fibrils, thereby making it more receptive to further myeloid cell infiltration (12, 13).

While promising, these findings also highlight the need for further research to reveal a holistic picture of the premetastatic stage that trigger (or inhibit) engraftment and proliferation. This, in turn, demands molecular-specific and quantitative analytical tools that can provide direct readouts from multiple biomolecules without necessitating individual labeling. Such a tool would inform if and how the compositional contributors of the stromal microenvironment in metastatic sites are changing in response to a spontaneously disseminating primary tumor—but prior to the arrival of tumor cells. Vibrational spectroscopy offers a promising tool to meet these demands, owing to the wealth of intrinsic molecular information (that obviates the need for imaging probes), extensive multiplexing capability, and facile readout (14–17).

Spontaneous Raman spectroscopy, in particular, has emerged as an attractive technique for the diagnosis of cancers with high specificity and free of interobserver variability (18). Based on inelastic scattering of light arising from the interactions with the tissue being analyzed, Raman spectroscopy affords subcellular signal localization and can easily be extended to in vivo approaches (19, 20). Recently, its ability to discern pathologies in advance of their clinical manifestations has also been shown (21). Malins and colleagues elegantly demonstrated the early detection sensitivity of vibrational spectroscopy in a study, where spectral changes in the DNA of primary tumor were noted 57 days prior to the appearance of histologic changes (22). We hypothesized that the utility of Raman spectroscopic information could also be extended to identifying the premetastatic niche, due to the unique structural and chemical changes associated with the evolving soil. Important clues also come from a recent report by Kwak and colleagues, demonstrating the utility of infrared (IR) spectroscopic imaging in predicting cancer recurrence by exploiting molecular features of the tumor microenvironment (23), and our recent observation that lymph nodes in mice with metastatic tumor xenografts displayed an increased collagen I density (24). Consistent with these recent literature reports, we suspected that the collagen architectural modifications, in part, preceded the seeding of metastatic cancer cells. Because Raman spectra report vibrational features characteristic of collagen and its cross-linking moieties as well as glycoproteins, our goal in this study was to identify Raman spectral patterns that are able to detect characteristic molecular changes in the premetastatic niche.

Here, we have investigated lungs from mouse models that recapitulate spontaneously disseminating breast cancer cells of low and high metastatic potential and exploited the molecular basis of Raman spectroscopy to probe the premetastatic niche (Fig. 1). Raman spectroscopic mapping measurements revealed subtle, but consistent, changes in the vibrational features of ECM components of the lungs, in particular in their collagen fiber matrix and proteoglycan content. The definition of the premetastatic adaptations in spectral terms facilitated the development of a decision algorithm, which accurately differentiates lungs in mice with metastatic MDA-MB-231 tumor xenografts from that in mice with MCF-7 xenografts and normal controls. A continuous model of ECM modifications, based on the metastatic potential of the primary tumor, is proposed to explain the differential signatures—in the confirmed absence of any tumor cells in the lungs. This model is in agreement with observations from Masson's trichrome staining and gene expression analysis performed on microarray data of premetastatic lung samples from mice harboring breast tumor xenografts. Taken together, this study demonstrates the potential of Raman spectroscopy as a rapid, objective, and label-free tool in the recognition of premetastatic changes. We envision that our findings here will also accelerate the use of Raman spectroscopy in identifying distinct biochemical signatures in organ-specific niches, thereby enabling a better understanding of organotropism.

Figure 1.

Raman spectroscopic profiling of premetastatic lungs. A, Mouse models, orthotopically xenografted with human breast cancer cells of different metastatic potential (MCF-7 and MDA-MB-231), were used to study stromal adaptations in the lung prior to seeding of tumor cells. B, Representative in vivo brightfield (left) and fluorescence (right) images of mouse growing a tdTomato-expressing breast tumor xenograft. C, Mean Raman spectra (with the shadow representing ±1 standard deviation) acquired from lungs of normal mice, and premetastatic lungs of MCF-7 and MDA-MB-231 xenografted mice are shown.

Figure 1.

Raman spectroscopic profiling of premetastatic lungs. A, Mouse models, orthotopically xenografted with human breast cancer cells of different metastatic potential (MCF-7 and MDA-MB-231), were used to study stromal adaptations in the lung prior to seeding of tumor cells. B, Representative in vivo brightfield (left) and fluorescence (right) images of mouse growing a tdTomato-expressing breast tumor xenograft. C, Mean Raman spectra (with the shadow representing ±1 standard deviation) acquired from lungs of normal mice, and premetastatic lungs of MCF-7 and MDA-MB-231 xenografted mice are shown.

Close modal

Tissue preparation and histopathology

Six-week-old female athymic nu/nu mice (NCI, MD) were orthotopically inoculated with 2 × 106 cells of the human breast cancer cell lines MDA-MB-231 (n = 3), or MCF-7 (n = 3) in their fourth right mammary fat pad, as detailed in our previous article (25). For comparison, control mice (n = 3) without tumor cell implantation were used in the study. Cell lines were obtained from the ATCC and stably transfected with a construct containing cDNA of tdTomato as described in our previous report (24). Cell lines tested negative for mycoplasma and were authenticated using short tandem repeat (STR) profiling prior to inoculation in mice. Cell lines were maintained in RPMI 1640 (Sigma Aldrich) supplemented with 10% fetal bovine serum (Sigma Aldrich) and 1% penicillin–streptomycin (Sigma Aldrich) in a humidified incubator at 37 °C/5% CO2. Prior to implantation of MCF-7 cells, mice were supplemented with 17β-Estradiol (Innovative Research of America, SE#121, 0.72 mg/pellet, 60-day release) in their neck region (26). Primary tumor size was monitored, and mice were sacrificed within 8 to 12 weeks of cell implantation when primary tumors grew to approximately 500 mm3 in volume. Control mice were also sacrificed in this timeframe. Freshly excised lungs of mice were cleaned in phosphate buffered saline (PBS) and fixed in formalin for 24 hours. Formalin-fixed lung tissue samples were rinsed thoroughly in excess PBS to remove any residual formalin before acquiring Raman spectra. Following spectral acquisition, tissues were stored in 70% ethanol and sent to JHU Histology Services for paraffin embedding and serial sectioning, after which one of the sections was used for haematoxylin and eosin (H&E) staining. The unstained slides were utilized in our laboratory to perform Masson's trichrome staining for collagen as detailed in our previous study (24). The Institutional Animal Care and Use Committee at the Johns Hopkins University School of Medicine approved the protocol of this study.

Acquisition of Raman spectra

Formalin-fixed lung specimens were rinsed in PBS, flattened, and placed on a clean aluminum block. There was no interference of the tissue Raman spectrum from the aluminum substrate, which also ensured a consistent probe-tissue imaging distance. A custom-built portable, fiber-probe–based Raman spectroscopy system was used for spectral acquisition (27). Briefly, an 830-nm diode laser (500 mW maximum power, Process Instruments) was used to excite the sample. A lensed fiber-optic Raman bundled contact probe (Emvision LLC) having a diameter of 2 mm (and an estimated tissue sampling volume of 1 mm3) was used to deliver the excitation beam through its central fiber and collect the back-scattered light through an annular ring of optical fibers. The scattered light was directed to a spectrograph (Holospec f/1.8i, Kaiser Optical Systems). The spectra were then recorded using a thermoelectrically cooled CCD camera (PIXIS 400BR, 20 × 20 μm pixels, 1,340 × 400 array, Princeton Instruments). The laser power at the lung tissue samples was maintained at around 15 mW in this study and the tissue was kept moist throughout the period of laser exposure by intermittent addition of PBS. A total collection time of 10 seconds (10 accumulations of 1 second each to prevent CCD saturation) was used for acquisition of each spectrum. Spectroscopic mapping was pursued to overcome the limitations of conventional fiber probe-based point spectroscopy that only examines a small area of tissue and suffers from undersampling. Wide area mapping, over the entire lung specimen, was performed by scanning the optical probe using a pair of motorized translation stages (travel range: 13 mm, T-LS13M, Zaber Technologies Inc.) in each orthogonal direction. Zaber console (open-source software) was used to control the raster scan through the PC serial ports. The mapping protocol also ensured the collection of sufficient spectra (approximately 300 spectra per mouse) for the development of robust classification models.

Data analysis

The Raman instrument was wavenumber-calibrated using 4-acetamidophenol (Tylenol) spectra. Raman spectra recorded from mouse lungs were restricted to the fingerprint wavenumber region (500–1,850 cm−1) for analysis and normalized to lie between 0 and 1 in order to remove the effects of potential differences in laser power at the sample. Principal component analysis (PCA) was used to reduce the dimensionality of the spectral dataset to a few dimensions characteristic of the maximum variance in the dataset (28). This transformation converts the set of spectral recordings into a set of values of linearly uncorrelated variables that form an orthogonal basis set. The spectral dataset of each mouse model was subjected to PCA using the statistical toolbox of MATLAB 2015b (Mathworks) to obtain principal component (PC) scores and loadings that highlight the spectral features characteristic of the class. The use of these key patterns (PCs) enhances sensitivity of the analysis by not focusing on small differences in Raman signatures that may arise from natural variation or sampling.

To visualize the differences among the classes, radial visualization maps were plotted using the Radviz tool of Orange data mining toolbox (29). Here, we utilized the scores of select PCs obtained from subjecting the entire spectral dataset to PCA. Guided by the Vizrank algorithm, the PCs were chosen to maximize class separation. In the radial visualization plot, the scores of a spectrum determine the position of the corresponding data point relative to the PC pivots. Partial least squares discriminant analysis (PLS-DA), a supervised classification technique based on partial least squares regression, was used to create decision models from the acquired Raman spectra for identifying the premetastatic niche (30). PLS-DA-derived classification models were built and trained using a leave-m-out cross-validation approach that utilizes randomly chosen training data consisting of 60% of the data of each class and test data constituted by the remaining 40% of the spectra. Randomized equalization of classes was implemented prior to PLS-DA model development to avoid skewing the model through disproportionate class sizes. Multiple iterations of class equalization and splitting into testing and training sets (10 × 100) were performed to obtain average performance metrics of the PLS-DA derived classification models.

Collagen quantification of Masson's trichrome stained tissue slides was achieved using FIJI (Image-J-based open-source software; ref. 31) and MATLAB (Mathworks). The color deconvolution feature provided by FIJI was used to extract an 8-bit frame (dense collagen presence = 0 and no collagen presence = 256) corresponding to the color, indicative of collagen content in the trichrome stains. The color was defined by average RGB values of pixels in a small user-selected region of interest (ROI) chosen in the image. Using in-house MATLAB code, the intensity of the pixels was converted to obtain a measure of collagen density in each frame. The data were averaged over the entire lung tissue section with n >35 fields of view (FOV) per class, where each FOV was approximately 1.75 mm × 1.33 mm. Statistical significance of differences across the classes was evaluated using the Student t test. A conventional criterion of P value less than 0.05 was used to consider differences as statistically significant.

Microarray dataset

The gene expression microarray dataset GSE62817 from the Gene Expression Omnibus (GEO) of the National Center for Biotechnology Information (http://www.ncbi.nlm.nih.gov/geo/) was used in this study (32). This dataset contains gene expression data from premetastatic lungs of BALB/c mice injected with tumor cells into their fourth mammary fat pad. In particular, 67NR (nonmetastatic) and 4T1 (metastatic) breast carcinoma cell lines were used and lung tissue was collected when the tumors reached a volume of 50 mm3. Control mice with no tumor cell injections were utilized for comparison. Briefly, RNA was extracted using a Qiagen kit, and Affymetrix microarrays (Mouse 430-v2) were used to analyze the expression profile of tissue samples. The heat map was generated using Gene-e matrix visualization and analysis software (http://www.broadinstitute.org). We used the moderated F-test statistic for selecting relevant genes. Consistent with the number of different groups and number of samples per group in the dataset, a threshold F-test statistic of 2.53 (corresponding to α = 0.125 level of significance) was used.

Lung was selected as the target organ in the current pilot study, as it offers a favorable site for spontaneous dissemination of breast cancer and is the most commonly studied metastatic site in animal models (9, 33). Primary orthotopic MDA-MB-231, and also eventually MCF-7, breast tumor xenografts used in our study preferentially metastasize to the lungs (34, 35). Spectroscopic mapping of the lungs, as opposed to a limited number of discrete point measurements, was pursued to encompass a large FOV with high spectral contrast. This would also account spectroscopically for the intrinsic biological variation in lung tissue that could otherwise suppress the subtle differences expected from premetastatic adaptations. Figure 1C shows average Raman spectra recorded from lung samples of control mice (control) as well as mice bearing MCF-7 (MCL) and MDA-MB-231 (MDL) tumor xenografts. The spectra shown here were background subtracted for the tissue autofluorescence component. While gross visual inspection reveals limited spectral variations, we reason that a subset of pixels (representing specific molecular moieties) has predictive power that is lost in examining the average value of the spectra across the lung specimen. In an effort to focus on elucidating the differentiating biochemical characteristics, we used PCA. To preserve the subtle spectral features, we performed PCA on the normalized spectra recorded from the specimen without background subtraction. For comparison, the results obtained following fifth-order best-fit polynomial based autofluorescence background removal have also been provided alongside (and in Supplementary Information).

Consistent differences in Raman spectra reflect biochemical changes in premetastatic lungs

Figure 2 shows the first 7 PC loadings in order of spectral variance for each of the three classes, control, MCL, and MDL. The first few PCs in each class are evidently influenced by the broad tissue autofluorescence signal; the characteristic Raman features are more prevalent in PCs 4 through 7. The PCs derived from the spectra belonging to the lungs of control mice exhibit notable Raman features at 859 cm−1 (C–C stretch of proline in collagen), 1,003 cm−1 (C–C stretching vibration of the aromatic ring in the phenylalanine side chain), 1,442 cm−1 (CH2 deformations in lipids), 1,592 cm−1 (tentatively attributed to carbon particles) and 1,653 cm−1 (amide-I feature of proteins with potential contributions of C=C stretching in lipids) with a weaker peak at 1,304 cm−1 (in-plane CH2 twisting modes of lipids). These features are concordant with prior observations in the literature (36–38). Table ST1 (Supplementary Information) lists the prominent peaks observed in the PCs and their characteristic band assignments.

Figure 2.

PCA of the acquired Raman spectra. A, PC loadings derived from spectra of lungs from control mice, i.e., bearing no tumor xenograft. B, PC loadings derived from spectra of lungs belonging to mice bearing MCF-7 xenografts (labeled as MCL in the text). C, PC loadings derived from spectra of lungs belonging to mice with MDA-MB-231 xenografts (labeled as MDL in the text). Dotted and dot–dashed lines highlight collagen and proteoglycan features, respectively.

Figure 2.

PCA of the acquired Raman spectra. A, PC loadings derived from spectra of lungs from control mice, i.e., bearing no tumor xenograft. B, PC loadings derived from spectra of lungs belonging to mice bearing MCF-7 xenografts (labeled as MCL in the text). C, PC loadings derived from spectra of lungs belonging to mice with MDA-MB-231 xenografts (labeled as MDL in the text). Dotted and dot–dashed lines highlight collagen and proteoglycan features, respectively.

Close modal

Visual inspection of the PC loadings shows an enhancement of the 859 cm−1 peak, which can be attributed to collagen, for the MDL specimen in comparison with MCL and control as well as a new peak at 917 cm−1 (C–C stretch of proline ring; ref. 17) for the non-control samples. These spectral differences suggest a positive correlation of collagen density in the lung specimens with the presence of a primary tumor xenograft and, importantly, with its metastatic potential. Previous studies have discussed the role of collagen in the premetastatic niche and have shown evidence of collagen cross-linking and the creation of a metastatic growth permissive fibrotic microenvironment at secondary sites, which was mediated by LOX secreted by hypoxic tumors (39, 40). Inhibition of LOX synthesis in human breast cancer cells has been shown to reduce the accumulation of CD11b+ myeloid cells in premetastatic organs of mice with orthotopic tumors and prevent metastasis (12). Another pertinent peak was observed at approximately 1,061 cm−1 in the MCL and MDL PCs, which is known to be a key spectral marker for proteoglycans (41, 42). This finding offers an intriguing insight into the nature of molecular modifications in the premetastatic niche, particularly in light of the study of Gao and colleagues. This study demonstrated that myeloid cells in premetastatic lungs (recruited by primary tumor derived secretory factors) aberrantly expressed versican, an ECM proteoglycan (43). Versican stimulated mesenchymal-to-epithelial transition of metastatic tumor cells by reducing phospho-Smad2 levels, which led to elevated cell proliferation and accelerated metastases. In fact, lung metastasis in mouse models was found to be significantly impaired through knockdown of versican, reinforcing the importance of proteoglycan content as a premetastatic site marker. Furthermore, the gradual increase in the prominence of proteoglycan marker in PCs with increasing metastatic potential is in agreement with the seminal report of Kaplan and colleagues, which showed that recruitment of bone marrow–derived cells is correlated to the aggressiveness of the primary tumor (7). On the other hand, a significant suppression of the peaks at approximately 1,302 cm−1 and 1,442 cm−1 was noted with a smaller reduction in the intensity of the 1,653 cm−1 feature. Because the former two peaks are characteristic of lipids and the latter also has lipid contributions, one can reasonably infer a relative reduction in the lipid content corresponding to spectra from lungs of mice bearing primary tumor xenografts.

Given the large dimensionality of the spectral data, however, it is challenging to judge whether the differences across the classes are significant from visual inspection of the PC loadings alone. To observe these differences better, we used radial visualization plots that map the scores of multiple PCs onto a two-dimensional space for the purpose of clustering. Figure 3 shows a representative radial visualization plot constructed by using PCs derived from a randomized spectral selection with 300 points per class (control, MCL, and MDL). These were chosen from the total set consisting of approximately 900 spectra/class, which in turn were constituted by approximately 300 spectra acquired from spatially distinct points in the lung lobes of each mouse. Supplementary Fig. S1 shows the corresponding radial visualization map after subtraction of tissue autofluorescence background. In order to obtain informative projections of the class-labeled data, the VizRank algorithm was used to grade the PCs by their ability to visually discriminate between classes (44). Evidently, there are pronounced differences in the Raman spectra acquired from lung specimens of control, MCL, and MDL mouse models, most likely owing to differential priming through factors secreted by the primary tumor. The presence of a small overlap of clusters from control and MDL mice indicates a limited development of the premetastatic niche in some of the latter cases and requires further analysis, as detailed in the ensuing paragraphs. While the PC score-based plot offers a satisfactory tool for preliminary data exploration, it does not provide quantitative information about the potential of Raman spectroscopy in recognizing the class (metastatic potential) and in understanding how the lung(s) of an individual mouse responds to the primary tumor xenograft.

Figure 3.

Visualization of spectroscopic differences due to premetastatic adaptations. Radial visualization plot showing clusters formed by spectra recorded from lung samples of sacrificed mice bearing MDA-MB-231 and MCF-7 breast cancer xenografts as well as controls without xenografts.

Figure 3.

Visualization of spectroscopic differences due to premetastatic adaptations. Radial visualization plot showing clusters formed by spectra recorded from lung samples of sacrificed mice bearing MDA-MB-231 and MCF-7 breast cancer xenografts as well as controls without xenografts.

Close modal

Thus, we used partial least squares–discriminant analysis (PLS-DA)-based classification models for translating the spectroscopic measurements in the premetastatic lungs to identification of the type of primary tumor xenograft. We used an equal number of spectra belonging to each class (control, MCL, and MDL) and their class labels to train the classification algorithm. To ensure robustness, we evaluated the classifier by testing on a separate validation dataset as detailed in the Data analysis section. Average correct rates of prediction of 90.1%, 97.7%, and 78.4% (95.4%, 95.6%, and 75.1% after autofluorescence background subtraction) were obtained for the spectra belonging to control, MCL, and MDL, respectively. The relevant confusion matrix of the reference and predicted labels is shown in Supplementary Table ST2. The lower correct classification rate for MDL spectra in both the cases is in agreement with the overlap of the MDL and control clusters observed on the radial visualization plot in Fig. 3. 

In order to understand the root cause of the MDL spectra misclassifications, we repeated the former analysis by leaving one mouse out of the dataset each time (Table 1 and Supplementary Table ST3 after autofluorescence background subtraction). Removing mouse MD #3 (arbitrary numbering of mice used for tabulating results) yields near-perfect classification accuracy indicating significantly lesser premetastatic adaptations in the lungs of this animal. Furthermore, removing mouse MD #3 also improved the classification rate of spectra belonging to control mice due to enhanced contrast in the training data. Notably, removal of any other mouse from the classification protocol did not result in as significant a change in the accuracy levels. This reinforces the fact that the improvement observed on removal of mouse MD #3 data was not due to overtraining of the model on smaller numbers, as otherwise similar enhancements would have been noted in all the other cases. The inadequate priming of the MD #3 lungs is also supported by application of Chauvenet's criterion to the set of classification rates obtained for the MDL class (Table 1). The latter results in designation of MD #3 as the sole outlier in the group due to its significant deviation from the mean by more than the maximum allowable number of standard deviations (τmax = 1.96 for a sample size of n = 10). Application of Chauvenet's criterion also facilitates determination of individual sample eligibility for training the PLS-DA classifier. The spectroscopic measurements, thus, capture the inherent variability in metastasis, which is commonly regarded as an inefficient process that only a subset of tumor cells can successfully navigate (45, 46) and is known to exhibit sporadic occurrence across a cohort of animals.

Table 1.

Correct classification rates (%) of the PLS-DA–derived model using leave-one-mouse-out protocol

Correct classification rate (%)Chauvenet's criterion for MDL (n = 10; τmax = 1.96)
Mouse excludedControlMCLMDLτ = |xi − xmean|/σResult
None 90.1 97.7 78.4 0.26 Retain 
MD #1 81.2 97.0 75.7 0.61 Retain 
MD #2 83.8 96.8 76.2 0.54 Retain 
MD #3 100.0 98.6 99.4 2.54 Eliminate 
MC #1 88.8 98.0 78.5 0.24 Retain 
MC #2 89.6 97.3 77.6 0.36 Retain 
MC #3 88.8 98.3 77.3 0.40 Retain 
Control #1 87.2 96.4 73.1 0.96 Retain 
Control #2 92.9 97.3 80.2 0.01 Retain 
Control #3 92.9 97.4 86.4 0.81 Retain 
Correct classification rate (%)Chauvenet's criterion for MDL (n = 10; τmax = 1.96)
Mouse excludedControlMCLMDLτ = |xi − xmean|/σResult
None 90.1 97.7 78.4 0.26 Retain 
MD #1 81.2 97.0 75.7 0.61 Retain 
MD #2 83.8 96.8 76.2 0.54 Retain 
MD #3 100.0 98.6 99.4 2.54 Eliminate 
MC #1 88.8 98.0 78.5 0.24 Retain 
MC #2 89.6 97.3 77.6 0.36 Retain 
MC #3 88.8 98.3 77.3 0.40 Retain 
Control #1 87.2 96.4 73.1 0.96 Retain 
Control #2 92.9 97.3 80.2 0.01 Retain 
Control #3 92.9 97.4 86.4 0.81 Retain 

NOTE: MD and MC refer to mouse models with MDA-MB-231 and MCF-7 tumor xenografts, respectively.

Finally, we conducted a negative control study to verify that the predictive power of the developed algorithms was not driven by potential spurious correlations in the spectral dataset (47). For this validation study, we assigned random class labels to the spectra irrespective of their true class origins and used the PLS-DA-derived classification models after similar splitting of the data into training and test sets. This resulted in an average correct classification rate of 33.3% with a standard deviation of 1.4% (and 33.6% with a standard deviation of 1.4% after background subtraction) for 1,000 iterations. The significantly low rate of correct classification (consistent with the likelihood of random selection of the true class label, 1/3) underscores the absence of chance correlations in the developed model.

Histologic assessment of the premetastatic niche in mice lungs

Due to their high metastatic potential and preference for metastasis to lungs, orthotopic MDA-MB-231 xenografts are frequently used to replicate breast cancer metastasis and organotropism (33, 48). Aggressive subpopulations of MDA-MB-231 are often derived through multiple rounds of in vivo selection and reimplantation and have been recently reported to result in macro-metastases to the lungs in 100% of all tested mice (35). In our study, we observed no cancer cell seeding in lungs of mice bearing MDA-MB-231 tumor xenografts (time of sacrifice: 8–12 weeks post orthotopic tumor inoculation). Prior optical tracking studies by Winnard and colleagues showed that orthotopically implanted MDA-MB-231 cells reached lungs only after ∼15 weeks of implantation in SCID mice (34). They also observed the absence of distant metastases after 8 weeks, consistent with the time period of sacrifice in our study. MCF-7 cells, often classified as nonmetastatic (49), were likewise not expected to engraft in the lungs within this 8- to 12-week time frame. However, it is noteworthy that MCF-7 cells are known to eventually metastasize to lungs in immunodeficient mice such as NSG (35).

Here, the lung tissue sections from each mouse were H&E stained to check for the onset (or the lack thereof) of metastasis. Also, to histologically examine the differences in collagen content across the classes, serial sections were processed with Masson's trichrome stain. Figure 4 shows representative images of H&E and Masson's trichrome stained lung sections belonging to each class (control, MCL, and MDL). The H&E images corroborate the lack of any metastatic lesions in the lung specimens. The Masson's trichrome-stained sections were used for quantification of the mean collagen density for each class (detailed in Materials and Methods). Figure 5A shows the mean bar plot that highlights the differences in collagen density for control mice and mice bearing MCF-7 and MDA-MB-231 xenografts. We observe that the metastatic potential of the primary tumor is positively correlated with the collagen density in the premetastatic niche. Yet, the differences in the mean collagen density values between MCL and MDL samples do not reach statistical significance (P < 0.05). Based on our spectroscopic findings, we suspected that the lung specimens of mouse MD #3 may possibly skew the collagen density values of the MDL set. Accordingly, we recalculated the values by removing the images of the lungs of this mouse, as shown in Fig. 5B. With this modification, the differences among each pair of classes were found to be statistically significant. This improvement of contrast in collagen density corresponds well with our spectroscopic findings and reflects the biochemical sensitivity of the vibrational spectroscopic data.

Figure 4.

Histologic assessment of premetastatic lungs shows stromal changes. Top (A–C) and middle (D–F) panels display representative microscopic images of H&E- and Masson's trichrome–stained slides at ×5 and ×10 magnifications, respectively. The H&E-stained sections confirm the absence of tumor cell seeding in the lungs of controls. Masson's trichrome stain delineates collagen fibers in the extracellular matrix and is quantified through image processing, as shown in G–I. Left (A, D, and G), lung sections derived from control mice; middle (B, E, and H) and right (C, F, and I), lung sections from mice bearing MCF-7 (nonmetastatic) and MDA-MB-231 (metastatic) tumor xenografts, respectively. The scale bars in the top and middle panels represent 1,000 and 500 μm, respectively.

Figure 4.

Histologic assessment of premetastatic lungs shows stromal changes. Top (A–C) and middle (D–F) panels display representative microscopic images of H&E- and Masson's trichrome–stained slides at ×5 and ×10 magnifications, respectively. The H&E-stained sections confirm the absence of tumor cell seeding in the lungs of controls. Masson's trichrome stain delineates collagen fibers in the extracellular matrix and is quantified through image processing, as shown in G–I. Left (A, D, and G), lung sections derived from control mice; middle (B, E, and H) and right (C, F, and I), lung sections from mice bearing MCF-7 (nonmetastatic) and MDA-MB-231 (metastatic) tumor xenografts, respectively. The scale bars in the top and middle panels represent 1,000 and 500 μm, respectively.

Close modal
Figure 5.

Quantification of collagen fiber density in premetastatic lungs. A, Bar plot showing mean and standard deviation of collagen density across the three classes (with all mice included) along with pairwise Student t test P values. B, Bar plot showing mean and standard deviation of collagen content across the three classes (after exclusion of MDA-MB-231 xenograft bearing mouse displaying atypical Raman data) along with pairwise Student t test P values.

Figure 5.

Quantification of collagen fiber density in premetastatic lungs. A, Bar plot showing mean and standard deviation of collagen density across the three classes (with all mice included) along with pairwise Student t test P values. B, Bar plot showing mean and standard deviation of collagen content across the three classes (after exclusion of MDA-MB-231 xenograft bearing mouse displaying atypical Raman data) along with pairwise Student t test P values.

Close modal

In light of the spectroscopic identification of stromal adaptations, we further sought to investigate the genetic underpinnings of premetastatic priming of lungs. We performed gene expression analysis on publicly available microarray data (GSE 62817) to determine markers in premetastatic lungs in response to primary breast tumors of divergent metastatic potential (32). Specifically, the data included gene expression levels corresponding to the lungs of normal mice (n = 5) as well as premetastatic lungs of mice injected with nonmetastatic 67NR breast carcinoma cells (n = 5) and with metastatic 4T1 breast carcinoma cells (n = 4). Seeking to isolate genes relevant to our study, we restricted our search to genes encoding for key stromal constituents and significantly overexpressed in premetastatic lungs of 4T1 tumor–bearing mice. Figure 6 shows the heatmap representing expression levels of these genes along with corresponding moderated F-statistic. Premetastatic lungs of the 4T1 tumor bearing mice demonstrate a selective upregulation of genes related to ECM constituents, notably collagen, fibronectin, versican, and glypican. Importantly, each of these ECM components exhibits a decreasing gradient of values from 4T1 to 67NR and then to control cases. The differential expression of stromal genes in response to primary tumor development can, thus, help explain our observations of discernible biochemical alterations in premetastatic lungs of mice bearing MCF-7 xenografts, even though these cells rarely metastasize in the mouse model used.

Figure 6.

Gene expression changes in premetastatic lungs as a function of metastatic potential of primary tumor. Microarray gene expression data heat map was obtained by analyzing the publicly available dataset GSE62816 on the Gene-e data visualization and analysis platform. The sample cohort includes lungs of mice bearing breast tumor xenografts of different metastatic potential. Total RNA was isolated from the premetastatic lungs and hybridized on an Affymetrix Mouse Genome 430 2.0 Array. Genes that are relevant to spectral markers identified in the current study and overexpressed in response to the metastatic potential of the primary tumor were analyzed. Moderated F-value of 2.53 was set as the criterion for inclusion.

Figure 6.

Gene expression changes in premetastatic lungs as a function of metastatic potential of primary tumor. Microarray gene expression data heat map was obtained by analyzing the publicly available dataset GSE62816 on the Gene-e data visualization and analysis platform. The sample cohort includes lungs of mice bearing breast tumor xenografts of different metastatic potential. Total RNA was isolated from the premetastatic lungs and hybridized on an Affymetrix Mouse Genome 430 2.0 Array. Genes that are relevant to spectral markers identified in the current study and overexpressed in response to the metastatic potential of the primary tumor were analyzed. Moderated F-value of 2.53 was set as the criterion for inclusion.

Close modal

Taken together, our findings suggest that remodeling of the ECM, such as an increase in collagen and proteoglycan content, occurs in response to primary tumor–derived factors, which precedes the actual seeding of tumor cells at the distant metastatic site. The data in this study support a continuous premetastatic niche formation model from primary tumors with low and high metastatic potential, rather than discrete premetastatic adaptations that are representative of the highly metastatic model alone. This would also imply that premetastatic adaptations are a necessary condition for further progression but not predictive of the eventual success of metastases.

In conclusion, the current study proposes Raman spectroscopy as a label-free molecular-specific tool for detection of premetastatic adaptations in the stromal environment. Using breast cancer metastasis to the lungs as the paradigm, we have demonstrated that Raman spectroscopy accurately detects changes in the ECM of premetastatic lungs, which correlate with the metastatic potential of the respective primary tumor xenograft. We identified spectral markers corresponding to collagen and proteoglycan that offer molecular insights into the formation of the premetastatic niche while also facilitating objective detection. The data presented here are unique and complementary to other microenvironment profiling methods such as genomic assays and mass spectrometry. While breast cancer metastasis to the lungs has been chosen for the current study, it should be noted that this approach can be extended to study the development of premetastatic niches at any secondary target organ from primary breast and non-breast malignancies.

We envision that the use of Raman spectroscopic imaging in conjunction with further biochemical assays will offer detailed mechanistic insights into premetastatic niche formation and evolution. As such, this offers a unique research tool that combines microenvironment and cellular profiling through nonperturbative, multiplexed measurements of proteins, nucleic acids, lipids, and metabolites. Building on the ability to detect such subtle changes in tissue composition, and as discussed in recent reports (18, 23), we anticipate that Raman spectroscopic imaging can, with further refinement, facilitate surgical margin assessment in tissue conserving surgery and provide prediction of tumor recurrence. Integration of Raman spectroscopy with minimally invasive biopsy needles can also permit real-time, in situ detection of malignancies (19, 50).

No potential conflicts of interest were disclosed.

Conception and design: S.K. Paidi, A. Rizwan, K. Glunde, I. Barman

Development of methodology: A. Rizwan, I. Barman

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S.K. Paidi, A. Rizwan, C. Zheng, M. Cheng

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S.K. Paidi, A. Rizwan, C. Zheng, K. Glunde

Writing, review, and/or revision of the manuscript: S.K. Paidi, A. Rizwan, K. Glunde, I. Barman

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M. Cheng

Study supervision: K. Glunde, I. Barman

S.K. Paidi and I. Barman acknowledge the JHU Whiting School of Engineering Startup Funding. C. Zheng acknowledges the support of the National Construction of High Quality University Projects of Graduates from the China Scholarship Council (CSC; Grant No. 201406170141). A. Rizwan, M. Cheng, and K. Glunde acknowledge the support of NIH R01 CA154725.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
Sporn
MB
. 
The war on cancer
.
The Lancet
1996
;
347
:
1377
81
.
2.
Hess
KR
,
Varadhachary
GR
,
Taylor
SH
,
Wei
W
,
Raber
MN
,
Lenzi
R
, et al
Metastatic patterns in adenocarcinoma
.
Cancer
2006
;
106
:
1624
33
.
3.
Lee
YNM
. 
Breast carcinoma: Pattern of metastasis at autopsy
.
J Surg Oncol
1983
;
23
:
175
80
.
4.
Paget
S
. 
The distribution of secondary growths in cancer of the breast
.
The Lancet
1889
;
133
:
571
3
.
5.
Fidler
IJ
,
Kripke
ML
. 
Metastasis results from preexisting variant cells within a malignant tumor
.
Science
1977
;
197
:
893
5
.
6.
Fidler
I
. 
Selection of successive tumour lines for metastasis
.
Nature
1973
;
242
:
148
9
.
7.
Kaplan
RN
,
Riba
RD
,
Zacharoulis
S
,
Bramley
AH
,
Vincent
L
,
Costa
C
, et al
VEGFR1-positive haematopoietic bone marrow progenitors initiate the pre-metastatic niche
.
Nature
2005
;
438
:
820
7
.
8.
Hiratsuka
S
,
Watanabe
A
,
Aburatani
H
,
Maru
Y
. 
Tumour-mediated upregulation of chemoattractants and recruitment of myeloid cells predetermines lung metastasis
.
Nat Cell Biol
2006
;
8
:
1369
75
.
9.
Psaila
B
,
Lyden
D
. 
The metastatic niche: Adapting the foreign soil
.
Nat Rev Cancer
2009
;
9
:
285
93
.
10.
Costa-Silva
B
,
Aiello
NM
,
Ocean
AJ
,
Singh
S
,
Zhang
H
,
Thakur
BK
, et al
Pancreatic cancer exosomes initiate pre-metastatic niche formation in the liver
.
Nat Cell Biol
2015
;
17
:
816
26
.
11.
Hiratsuka
S
,
Nakamura
K
,
Iwai
S
,
Murakami
M
,
Itoh
T
,
Kijima
H
, et al
MMP9 induction by vascular endothelial growth factor receptor-1 is involved in lung-specific metastasis
.
Cancer cell
2002
;
2
:
289
300
.
12.
Erler
JT
,
Bennewith
KL
,
Cox
TR
,
Lang
G
,
Bird
D
,
Koong
A
, et al
Hypoxia-induced lysyl oxidase is a critical mediator of bone marrow cell recruitment to form the premetastatic niche
.
Cancer Cell
;
15
:
35
44
.
13.
Erler
JT
,
Bennewith
KL
,
Nicolau
M
,
Dornhöfer
N
,
Kong
C
,
Le
QT
, et al
Lysyl oxidase is essential for hypoxia-induced metastasis
.
Nature
2006
;
440
:
1222
6
.
14.
Matousek
P
,
Stone
N
. 
Development of deep subsurface Raman spectroscopy for medical diagnosis and disease monitoring
.
Chem Soc Rev
2016
;
45
:
1794
1802
.
15.
Fernandez
DC
,
Bhargava
R
,
Hewitt
SM
,
Levin
IW
. 
Infrared spectroscopic imaging for histopathologic recognition
.
Nat Biotechnol
2005
;
23
:
469
74
.
16.
Walsh
MJ
,
Reddy
RK
,
Bhargava
R
. 
Label-free biomedical imaging with mid-IR spectroscopy
.
IEEE J Sel Top Quantum Electron
2012
;
18
:
1502
13
.
17.
Stone
N
,
Kendall
C
,
Shepherd
N
,
Crow
P
,
Barr
H
. 
Near‐infrared raman spectroscopy for the classification of epithelial pre‐cancers and cancers
.
J Raman Spectrosc
2002
;
33
:
564
73
.
18.
Kong
K
,
Rowlands
CJ
,
Varma
S
,
Perkins
W
,
Leach
IH
,
Koloydenko
AA
, et al
Diagnosis of tumors during tissue-conserving surgery with integrated autofluorescence and raman scattering microscopy
.
Proc Natl Acad Sci U S A
2013
;
110
:
15189
94
.
19.
Jermyn
M
,
Mok
K
,
Mercier
J
,
Desroches
J
,
Pichette
J
,
Saint-Arnaud
K
, et al
Intraoperative brain cancer detection with raman spectroscopy in humans
.
Sci Transl Med
2015
;
7
:
274ra19
.
20.
Matthäus
C
,
Krafft
C
,
Dietzek
B
,
Brehm
BR
,
Lorkowski
S
,
Popp
J
. 
Noninvasive imaging of intracellular lipid metabolism in macrophages by raman microscopy in combination with stable isotopic labeling
.
Anal Chem
2012
;
84
:
8549
56
.
21.
Pandey
R
,
Paidi
SK
,
Kang
JW
,
Spegazzini
N
,
Dasari
RR
,
Valdez
TA
, et al
Discerning the differential molecular pathology of proliferative middle ear lesions using raman spectroscopy
.
Scientific reports
2015
;
5
:
13305
.
22.
Malins
DC
,
Anderson
KM
,
Gilman
NK
,
Green
VM
,
Barker
EA
,
Hellström
KE
. 
Development of a cancer DNA phenotype prior to tumor formation
.
Proc Nat Aca Sci U S A
2004
;
101
:
10721
5
.
23.
Kwak
JT
,
Kajdacsy-Balla
A
,
Macias
V
,
Walsh
M
,
Sinha
S
,
Bhargava
R
. 
Improving prediction of prostate cancer recurrence using chemical imaging
.
Scientific Reps
2015
;
5
:
8758
.
24.
Rizwan
A
,
Bulte
C
,
Kalaichelvan
A
,
Cheng
M
,
Krishnamachary
B
,
Bhujwalla
ZM
, et al
Metastatic breast cancer cells in lymph nodes increase nodal collagen density
.
Scientific Reps
2015
;
5
:
10002
.
25.
Rizwan
A
,
Serganova
I
,
Khanin
R
,
Karabeber
H
,
Ni
X
,
Thakur
S
, et al
Relationships between LDH-A, lactate, and metastases in 4T1 breast tumors
.
Clin Cancer Res
2013
;
19
:
5158
69
.
26.
Yarden
RI
,
Lauber
AH
,
El-Ashry
D
,
Chrysogelos
SA
. 
Bimodal regulation of epidermal growth factor receptor by estrogen in breast cancer cells
.
Endocrinology
1996
;
137
:
2739
47
.
27.
Paidi
SK
,
Siddhanta
S
,
Strouse
R
,
McGivney
JB
,
Larkin
C
,
Barman
I
. 
Rapid identification of biotherapeutics with label-free raman spectroscopy
.
Anal Chem
2016
;
88
:
4361
8
.
28.
Jolliffe
I
. 
Principal component analysis
.
Wiley Online Library
; 
2002
.
29.
Demšar
J
,
Curk
T
,
Erjavec
A
,
Gorup
C
,
Hocevar
T
,
Milutinovic
M
, et al
Orange: Data mining toolbox in python
.
J Mac Lear Res
2013
;
14
:
2349
53
.
30.
Brereton
RG
.
Applied chemometrics for scientists
.
John Wiley & Sons
; 
2007
.
31.
Schindelin
J
,
Arganda-Carreras
I
,
Frise
E
,
Kaynig
V
,
Longair
M
,
Pietzsch
T
, et al
Fiji: An open-source platform for biological-image analysis
.
Nat Meth
2012
;
9
:
676
82
.
32.
Kowanetz
M
,
Wu
X
,
Lee
J
,
Tan
M
,
Hagenbeek
T
,
Qu
X
, et al
Granulocyte-colony stimulating factor promotes lung metastasis through mobilization of Ly6G+Ly6C+ granulocytes
.
Proc Natl Acad Sci U S A
2010
;
107
:
21248
55
.
33.
Minn
AJ
,
Gupta
GP
,
Siegel
PM
,
Bos
PD
,
Shu
W
,
Giri
DD
, et al
Genes that mediate breast cancer metastasis to lung
.
Nature
2005
;
436
:
518
24
.
34.
Winnard
PT
,
Kluth
JB
,
Raman
V
. 
Noninvasive optical tracking of red fluorescent protein-expressing cancer cells in a model of metastatic breast cancer
.
Neoplasia
2006
;
8
:
796
806
.
35.
Iorns
E
,
Drews-Elger
K
,
Ward
TM
,
Dean
S
,
Clarke
J
,
Berry
D
, et al
A new mouse model for the study of human breast cancer metastasis
.
PLoS One
2012
;
7
:
e47995
.
36.
Huang
N
,
Short
M
,
Zhao
J
,
Wang
H
,
Lui
H
,
Korbelik
M
, et al
Full range characterization of the Raman spectra of organs in a murine model
.
Opt Ex
2011
;
19
:
22892
909
.
37.
Frushour
BG
,
Koenig
JL
. 
Raman scattering of collagen, gelatin, and elastin
.
Biopolymers
1975
;
14
:
379
91
.
38.
Czamara
K
,
Majzner
K
,
Pacia
M
,
Kochan
K
,
Kaczor
A
,
Baranska
M
. 
Raman spectroscopy of lipids: A review
.
J Raman Spectrosc
2015
;
46
:
4
20
.
39.
Joo
YN
,
Jin
H
,
Eun
SY
,
Park
SW
,
Chang
KC
,
Kim
HJ
. 
P2Y2R activation by nucleotides released from the highly metastatic breast cancer cell MDA-MB-231 contributes to pre-metastatic niche formation by mediating lysyl oxidase secretion, collagen crosslinking, and monocyte recruitment
.
Oncotarget
2014
;
5
:
9322
34
.
40.
Cox
TR
,
Bird
D
,
Baker
AM
,
Barker
HE
,
Ho
MW
,
Lang
G
, et al
LOX-mediated collagen crosslinking is responsible for fibrosis-enhanced metastasis
.
Cancer Res
2013
;
73
:
1721
32
.
41.
Kunstar
A
,
Leijten
J
,
van Leuveren
S
,
Hilderink
J
,
Otto
C
,
van Blitterswijk
CA
, et al
Recognizing different tissues in human fetal femur cartilage by label-free raman microspectroscopy
.
J Biomed Opt
2012
;
17
:
116012
.
42.
Gamsjaeger
S
,
Klaushofer
K
,
Paschalis
EP
. 
Raman analysis of proteoglycans simultaneously in bone and cartilage
.
J Raman Spectrosc
2014
;
45
:
794
800
.
43.
Gao
D
,
Joshi
N
,
Choi
H
,
Ryu
S
,
Hahn
M
,
Catena
R
, et al
Myeloid progenitor cells in the premetastatic lung promote metastases by inducing mesenchymal to epithelial transition
.
Cancer Res
2012
;
72
:
1384
94
.
44.
Leban
G
,
Zupan
B
,
Vidmar
G
,
Bratko
I
. 
Vizrank: Data visualization guided by machine learning
.
Data Min Kno Dis
2006
;
13
:
119
36
.
45.
Weiss
L
. 
Metastatic inefficiency
.
Adv Cancer Res
1990
;
54
:
159
211
.
46.
Wong
CW
,
Lee
A
,
Shientag
L
,
Yu
J
,
Dong
Y
,
Kao
G
, et al
Apoptosis: An early event in metastatic inefficiency
.
Cancer Res
2001
;
61
:
333
8
.
47.
Arnold
MA
,
Burmeister
JJ
,
Small
GW
. 
Phantom glucose calibration models from simulated noninvasive human near-infrared spectra
.
Anal Chem
1998
;
70
:
1773
81
.
48.
Lu
X
,
Kang
Y
. 
Efficient acquisition of dual metastasis organotropism to bone and lung through stable spontaneous fusion between MDA-MB-231 variants
.
Proc Nat Aca Scis
2009
;
106
:
9385
90
.
49.
Rizwan
A
,
Cheng
M
,
Bhujwalla
ZM
,
Krishnamachary
B
,
Jiang
L
,
Glunde
K
. 
Breast cancer cell adhesome and degradome interact to drive metastasis
.
NPJ Breast Cancer
2015
;
1
:
15017
.
50.
Barman
I
,
Dingari
NC
,
Saha
A
,
McGee
S
,
Galindo
LH
,
Liu
W
, et al
Application of raman spectroscopy to identify microcalcifications and underlying breast lesions at stereotactic core needle biopsy
.
Cancer Res
2013
;
73
:
3206
15
.