Raman spectroscopy reveals phenotype switches in breast cancer metastasis

Abstract

The accurate analytical characterization of metastatic phenotype at primary tumor diagnosis and its evolution with time are critical for controlling metastatic progression of cancer. Here, we report a label-free optical strategy using Raman spectroscopy and machine learning to identify distinct metastatic phenotypes observed in tumors formed by isogenic murine breast cancer cell lines of progressively increasing metastatic propensities. Methods: We employed the 4T1 isogenic panel of murine breast cancer cells to grow tumors of varying metastatic potential and acquired label-free spectra using a fiber probe-based portable Raman spectroscopy system. We used MCR-ALS and random forests classifiers to identify putative spectral markers and predict metastatic phenotype of tumors based on their optical spectra. We also used tumors derived from 4T1 cells silenced for the expression of TWIST, FOXC2 and CXCR3 genes to assess their metastatic phenotype based on their Raman spectra. Results: The MCR-ALS spectral decomposition showed consistent differences in the contribution of components that resembled collagen and lipids between the non-metastatic 67NR tumors and the metastatic tumors formed by FARN, 4T07, and 4T1 cells. Our Raman spectra-based random forest analysis provided evidence that machine learning models built on spectral data can allow the accurate identification of metastatic phenotype of independent test tumors. By silencing genes critical for metastasis in highly metastatic cell lines, we showed that the random forest classifiers provided predictions consistent with the observed phenotypic switch of the resultant tumors towards lower metastatic potential. Furthermore, the spectral assessment of lipid and collagen content of these tumors was consistent with the observed phenotypic switch. Conclusion: Overall, our findings indicate that Raman spectroscopy may offer a novel strategy to evaluate metastatic risk during primary tumor biopsies in clinical patients.

Keywords: Cancer metastasis; Phenotype switch; Raman spectroscopy; Random forests; TWIST.

Figure 1

Label-free Raman spectroscopy for identifying metastatic phenotypes. (A) The different steps of metastatic cascade accomplished by the tumors formed by each of the cell lines of the 4T1 tumor model employed in this study are shown. (B) The overview of Raman mapping of the tumors and spectral analysis is presented. MCR-ALS spectral decomposition and random forest classification using leave-one-mouse-out analysis were performed, where all the spectra from each mouse was excluded from training dataset and used as an independent test dataset. (C) The tumors formed by the 4T1 cells silenced for the expression of metastasis-promoting genes are employed for spectroscopic measurements.

Figure 2

Spectral differences between tumors of varying metastatic potential. (A) The mean (dark line) and 1 standard deviation (shaded region) of the Raman spectra collected from isogenic tumors of varying metastatic potential are shown. (B) A subset of constituent spectra derived using MCR-ALS decomposition of Raman spectral dataset that harbor features of lipids and collagen are plotted. The box and whisker plots show the variation of scores of (C) lipid-like and (D) collagen-like MCR-ALS components with metastatic potential of the tumors. Statistical significance as assessed by Wilcoxon rank-sum test p-value < 0.05 for each metastatic tumor group in comparison with non-metastatic 67NR group are denoted using asterisks and the corresponding effect sizes (r) show the magnitude of differences.

Figure 3

Supervised classification of metastatic potential and stage-specific phenotypes. The results for leave-one-mouse-out random forest classification are shown as heatmaps for the prediction of - metastatic phenotype of the tumors formed by the four cell lines with differential metastatic potential (A) and stage specific metastasis abilities to accomplish intravasation (B), extravasation (C), and metastatic growth (D). The true labels for the analysis (negative or positive for each step) in panels C-D are assigned based on known behaviors of these tumors in vivo. The top and bottom rows in each heatmap, respectively, show the true labels and predicted labels of the individual mice (columns) in each group. The labeled central rows in the heatmaps show the distribution of the predicted labels for spectra from each mouse into the classes in the training dataset. The overall class prediction for each mouse is obtained by thresholding on the prediction frequencies.

Figure 4

Stage-specific spectral markers of metastatic progression. The predictor importance estimates derived from random forest classification of Raman spectra based on their stage-specific metastatic abilities are shown for intravasation (A), extravasation (B), and metastatic growth (C). The five most important (non-neighboring) Raman features (cm^-1) are highlighted on these plots and presented as a Venn diagram (D) to visualize the overlap between different steps. Null sets are denoted by Φ.

Figure 5

Raman spectroscopic identification of metastatic phenotypes due to subtle alterations in gene expression. (A) Heatmap representation of random forest classification of tumors from the three 4T1-variant cell lines (knockdown and control). The classifier model was trained on data from the original tumor panel - 67NR, 168FARN, 4T07, and 4T1. The overall class prediction in the bottom row for each mouse is obtained by thresholding on the prediction frequencies of the four classes in the labeled intermediate rows. A comparison of the MCR-ALS scores between the knockdown and control tumors for the components resembling lipids (B) and collagen (C) is shown using box and whisker plots. Statistical significance as assessed by Wilcoxon rank-sum test p-value < 0.05 for each genetically altered tumor group in comparison with their respective control groups is denoted using asterisks and the effect sizes (r) show the magnitude of differences.

Figure 6

Differences between metastatic phenotype switches due to distinct alterations in gene expression. (A) Protein-protein interaction network of genes identified as overexpressed in the tumors accomplishing intravasation, extravasation, metastatic growth is shown. The network nodes are colored by their pathway membership and the interactions between nodes are colored by type, as listed in the respective legends. (B) The results of leave-one-mouse-out random forest classification of the spectra from tumors obtained by silencing TWIST1, FOXC2, and CXCR3 expression in 4T1 cells are shown. The overall class prediction in the bottom row for each mouse is obtained by thresholding on the prediction frequencies of the three classes in the labeled intermediate rows. (C) Representative photographs of metastatic lungs of mice harboring tumors obtained by silencing TWIST1, FOXC2, and CXCR3 expression in 4T1 cells are shown along with their corresponding controls.