Study design and data analysis workflow
To establish our Raman “spectromics” approach we employed two widely used and robust mouse models in cardiovascular research: a) acute injury: myocardial ischemia/reperfusion injury (MI) by transient ligation of the left anterior descending artery (LAD) as described previously16 and b) chronic injury: continuous infusion of Angiotensin II in atherosclerosis-prone apolipoprotein E (ApoE)-deficient mice, leading to cardiac hypertrophy and fibrosis17 (Methods). The model of cardiac hypertrophy was selected to examine myocardial remodeling during chronic damage, while the MI model was suitable to explore infiltratory cells inside the infarcted heart and its inflammatory micro-milieu.
The workflow starts with scanning deparaffinized formalin-fixed, paraffin-embedded (FFPE) or cryosections with a confocal Raman microspectroscope (Fig. 1a). Scans were acquired at 1 µm/pixel resolution, with each pixel being represented by an individual Raman spectrum, composed by wavenumbers of Raman shift and corresponding intensities. The resulting data matrix with more than 43,000,000 data points was filtered and subsetted subsequently (Fig. 1b). We analyzed the “whole spectrum” with wavenumbers from 300 to 3000 cm−1 and compared it to the biological “fingerprint spectrum” between 400 and 1800 cm−1, as defined previously18. Technical outliers and paraffin peaks as residues in FFPE tissue sections were filtered out, as detailed in the Methods section.
Spatially-unaware Raman spectra show distinct clustering
Originally intended as exploratory analysis, we applied pipelines used for single-cell analyses to Raman spectroscopy datasets. While single-cell transcriptomic data provide information about genes and their transcript counts, Raman spectroscopy supplies data about wavenumbers and their corresponding intensities. Specific wavenumbers are a result of energy differences between incident and scattered photons and unique for individual molecules. Thus, wavenumbers can be interpreted like a gene transcript, but defining the actual material texture and molecular composition rather than genomic activity.
To test the hypothesis, we used Raman spectra acquired from an area of subendocardial fibrosis within a section from the hypertrophy model (Fig. 1d) and analyzed the dataset with the single-cell analysis tool Seurat19,20. Outlier-corrected whole spectrum data was integrated into the Seurat workflow (Supplementary Fig. 1b) to generate Uniform Manifold Approximation and Projection (UMAP) plots (Fig. 1c). We observed a distinct segregation of pixels into clusters, as we expect similar spectra of the same cardiac substructure to be clustered together. The segregation also displayed a dependence on the number of dimensions used. In contrast to transcriptomics, Raman spectra variances between clusters are characterized by a high number of small changes. However, using more than 20 Principal components (PCs) did not further improve clustering, as Seurat’s JackStraw- and ElbowPlot Analysis (Supplementary Fig. 3a–c) underlined.
The cluster analysis was carried out for both fingerprint spectrum (Fig. 1f) and whole spectrum (Supplementary Fig. 2). Interestingly, there was no difference in observed cluster numbers and an overall comparable clustering result. Consequently, the more precise fingerprint spectra were applied for further analyses. Figure 1g shows the average Raman spectrum for each identified cluster. We concluded that repurposing of established tools for genetic analysis to Raman spectral data is applicable and feasible.
Cluster characteristics, spatial decoding and biological assignment
We sought to investigate whether the detected cluster can be translated to cardiac structures or compartments. For this purpose, we performed differential “expression” analysis (Fig. 2a) which can be repurposed to identify typical peaks in Raman spectra within a cluster. Clusters 0 and 1 were enriched in assignments to 1005, 1457 and 1640–1668 cm−1 typically found in myocardium10,11,21. Cluster 2 was identified as area where no tissue covered the glass microscope slide. This cluster showed very low intensities overall with an enhancement between 400–600 cm−1 and 1000–1200 cm−1, resulting from background signal of glass. Cluster 3 demonstrated comparable peak intensities to cluster 4, but at a lower overall intensity and differences in relative peak ratios. Cluster 4 can be identified as collagen signature, showing characteristic peaks at 858, 940, 1249 and 1680 cm−1 22. Cluster 5 was characterized by peaks at 889, 1056, 1302 and 1416 cm−1 potentially resulting from spectral artefacts obtained by the removal of paraffin residues during preprocessing. Most relevant peaks expressed by cluster 6 were found at wavenumbers of 784, 1096, 1376 and 1580 cm−1 often reported for nucleic acids23,24. Cluster 7 demonstrated the overall strongest peak intensities, especially in the region between 1200 and 1700 cm−1 with peaks at 1123, 1343, 1401, 1605, and 1637 cm−1. An overview of the most relevant peaks and their molecular assignment is provided in Supplementary Table S1.
Analysis of the spectral assignments of the clusters retrieved by the established unsupervised transcriptomics workflow enabled to identify biologically relevant spectral signatures of myocardium, collagens and cells and was further capable of separating them from preprocessing artefacts and background signals. Furthermore, the relative contribution of each component to the overall composition of the tissue could be determined (Fig. 2b). Overlaying density estimate contour lines on the UMAP plot (Fig. 2c) showed that cluster 0 and 1 to transition continuously into each other, while cluster 2, 6, and 7 showed greater compactness and hence homogeneity within the clusters found.
Next, we aimed to explore how specific peaks are distributed over the UMAP plot and specifically how these peaks spatially characterize the actual scanned sample. Cluster-specific peaks from the average spectrum and the differential “expression” analysis performed before were selected and the corresponding intensities were plotted in a color-coded fashion on a UMAP plot containing pixels sorted by UMAP projection, and on the actual image, where these pixels are sorted in their correct spatial relationship (Fig. 2d). Peaks for univariate intensity-based images were chosen exemplarily at 1457, 480, 858, and 1605 cm−1 for clusters 0, 2, 4, and 7 to localize myocardium, glass, collagens and blood-derived cells based on the spatial distribution along with molecular assignments to amide II and C-C backbone vibrations of proteins, characteristic glass peaks, hydroxyproline and hemin. The 858 cm−1 Raman shift allowed for a distinct fit to collagen assignments, when comparing the intensity image (Fig. 2d) to the Picrosirius Red Staining of an adjacent section (Fig. 1d, right). Similar effects were found for the peak at 1605 cm−1, which demonstrated nuclear morphologies in the intensity images and could be assigned to blood cells. Interestingly, peak assignments to 1605 cm−1 were predominantly found in viable myocardium and not scar tissue. Whether different cell types can also be identified based on different spectra will be investigated later in this study.
To get back to the spatial dimension of the Raman scan, each pixel of the original Raman image was color-coded in its corresponding cluster color (Fig. 2e). The image showed clear spatial compartmentalization and strong similarities to the Hematoxylin and Eosin (H&E) and Picrosirius Red Staining (Fig. 1c), as well as the manually assigned Raman true component analysis (TCA) using commercially available software (Fig. 1e, Methods). Figure 2f illustrates the transition from the unsupervised clustering to supervised ground truth cluster assignment of the Raman image by a combination of the molecular interpretation of the retrieved spectral clusters as well as morphological/spatial assignments. Strikingly, more than one cluster was found that could be assigned to distinct cardiac compartments. Two clusters were observed in the fibrotic tissue area, from which cluster 3 seems to surround cluster 4. Cluster 4 corresponds to the fibrotic area correlating well with the fibrosis staining, whereas cluster 3 was only detected by spectroscopic imaging and suggests representing a pre-fibrotic boundary around the actual fibrosis. The unsupervised analysis method also found two clusters allocated to myocardium: Interestingly, cluster 0 and 1 displayed a clear spatial organization reaching from far (cluster 0) to close (cluster 1) proximity to the fibrotic area. We hypothesized that this spatial pattern is a result of cardiac remodeling and hypertrophy when approaching subendocardial fibrosis. Cluster 1 could represent a transition zone between healthy myocardium (cluster 0) and fibrosis (cluster 4). As these results were found using unbiased, spatially-unaware clustering of Raman data, we concluded that our “Raman spectromics” approach could represent a sensitive method for identification of tissue substructures and hidden molecular patterns.
Integration of spatial information confirms tissue clustering and myocardial zonations
To this point analysis of Raman spectra has been performed on pixels detached from their spatial context. By integrating spatial information about the position of the pixel inside the Raman scan image, we sought to validate the spatial-unaware clustering results shown before, as well as to reduce clustering artifacts which occurred before. As spatial transcriptomics have demonstrated previously, the additional spatial information can help address the analytical challenges of sparsity and noise by smoothing over adjacent pixels25, which are more likely to have a similar genetic – or in our case Raman – fingerprint.
To analyze the dataset with spatial resolution, we translated a bioinformatics tool for spatial transcriptomic analysis into the “Raman world”. BayesSpace provides a clustering method that uses a t-distributed error model to identify spatial clusters25. It implements a fully Bayesian model with a Markov random field (Fig. 3a), which hypothesizes that pixels belonging to the same type of cell or tissue should be closer to each other. We integrated Raman spectroscopy data into the BayesSpace workflow (Supplementary Fig. 1b) to cluster the previous sample of subendocardial fibrosis into 5 clusters for the fingerprint spectrum, or 4 clusters using the whole spectrum dataset (Supplementary Fig. 5). UMAP analysis of spatially-aware Raman data showed a completely different projection than without spatial resolution, with one cluster segregating far away from the others (Fig. 3b). This cluster could be easily identified as scan area without biological tissue. Notably, limiting the analysis from whole spectrum to the fingerprint area substantially improved clustering specificity and thus allocation of myocardial structures (Supplementary Fig. 5). In summary, integrating spatial information of pixels containing spectral information allowed us to reproduce cluster analysis performed by spatially-unaware algorithms from above. On one hand spatially-aware clustering can reduce clustering artifacts by smoothing over adjacent pixels, but on the other hand, it could decrease cluster sensitivity and increases inclusion of outliers, which were “sorted out” by spatially-unaware clustering through assignment to a dedicated cluster. However, when benchmarking the accuracy for detection of fibrotic spectra, spatially-aware clustering using BayesSpace outperformed spatially-unaware clustering (Supplementary Fig. 6). Further analyses indicate that especially in the larger space the benefits of spatially-aware clustering come into play, e.g. in reducing artificial clustering (Supplementary Fig. 7).
Strikingly, performing spatially-aware clustering reproduced our previous finding of an additional myocardial subcluster between healthy myocardium and fibrotic subendocardium. We hypothesized that this cluster represents a transition zone of myocardium undergoing heavy remodeling, implying degradation and synthesis of cellular and interstitial components. To identify the underlying differences between healthy and remodeling myocardium, we performed Principal Component Analysis (PCA) between the two myocardial subclusters identified using both spatially-aware clustering via BayesSpace (Fig. 3f) and spatially-unaware clustering (Fig. 3g) via Seurat. However, the clustering by PCA was not as pronounced as by UMAP plotting. A shift of the myocardial remodeling cluster towards negative PC1 values was demonstrated for both input data (Fig. 3h). Major Raman peaks identified by the PC1 loadings were shown at 758, 823, 853, 1005 and in the amide I region between 1605−1650 cm−1, suggesting differences in single amino acids such as tryptophane (758 cm−1), tyrosine (823 and 853 cm−1) and phenylalanine (1005 cm−1). The differences found in single amino acids as well as glycogen (853, 1022−1025 cm−1) assignments imply differences in protein composition and metabolism during cardiac remodeling. Moreover, alterations in the shape of the Amide I band towards broader shoulders in the region between 1605−1650 cm−1 indicate changes in protein secondary structure towards ß-sheet structures26,27,28 potentially driven by myocardial remodeling.
To challenge the robustness of our untargeted spatial Raman spectromics approach, we reproduced clustering of healthy and remodeling myocardium on further sections of cardiac hypertrophy (Supplementary Fig. 8). We conclude that both spatially-unaware and -aware clustering algorithms applied on Raman spectroscopy data allows for the detection of compartments and zonations in histological sections at subcellular resolution.
Deciphering intra- and intercluster heterogeneity by repurposing pseudotime and spatial trajectories
As shown before, unsupervised clustering algorithms could find myocardial zonations, which classical staining could not uncover. However, strict borders in the same tissue of myocardium must be critically questioned, as changes in the sense of remodeling tend to be of rather continuous nature. Hence, clustering into groups reduces important biological information by neglecting heterogeneity within a cluster. We aimed to visualize the transition from healthy to remodeling myocardium by deciphering intra- and intercluster heterogeneity within the two myocardial zonations found by spatially-unaware and spatially-aware cluster analysis.
To this end, we repurposed the bioinformatics tool Monocle29,30, which performs pseudotime analysis on single-cell data. Pseudotime is a measure of how much progress an individual cell has made through a process such as cell differentiation, activation or throughout life. These cellular processes underlie changes in gene expression, leading to differential expression despite being of same cell type29. By translating this method to Raman microspectroscopy, we aimed to uncover tissue heterogeneity and homogeneity as wells as spatial correspondence of pseudotime trajectories. Instead of ordering cells with gene expression changes along a time axis, we sought to order pixels with Raman spectra along an axis and let this ordering be driven by changes in peak intensities of specific wavenumbers. We expected pixels with similar spectra to be close together and a continuous spectral change when moving along the pseudotime track. When there are too many different intensity changes that do not allow ordering on the main axis, a branch is constructed and spectra are placed here. Hence, branching is a result of strong spectral heterogeneity.
Monocle introduced an algorithm to learn the sequence of gene expression changes each cell must go through as part of a dynamic biological process30. In our case, we analyzed changes in Raman spectra instead of gene expression. Fingerprint spectrum data of the subendocardial fibrosis section used before was integrated into the Monocle environment (Supplementary Fig. 1b). Looking at the UMAP projection with overlayed pseudotime trajectory, we discovered a highly branched track in the region of myocardial clusters (Fig. 4a). This result is suggestive for a high heterogeneity within the myocardial spectra of our Raman analysis. To further elaborate this finding, we picked a region in the center of the Raman image where pixels were approximately equally assigned to healthy (cluster 0, light blue) or remodeling (cluster 1, dark blue) myocardium. These spectra were subsequently analyzed using the DDRTree method provided with Monocle (see Methods), which ordered the selected pixels/spectra along a track with several color-coded branches and subclusters (Fig. 4b). Unexpectedly, the algorithm found the branches to be located nearly exclusively within the assigned pixels of remodeling myocardium (Fig. 4c). This result underlines the dynamics of remodeling of myocardium, while healthy myocardium is not affected by these changes.
We concluded that molecular dynamics are the result of vigorous degradation and synthesis of cardiac matrix during remodeling and can be illustrated by an unsupervised approach of translating pseudotime trajectories.
Spatial trajectories towards fibrosis uncover molecular dynamics
To fully reveal the spatial organization of the tissue dynamics found, we employed spatial trajectories approximating the fibrotic regions, in order to examine the underlying molecular pattern changes. Linear trajectories towards subendocardial fibrosis were generated and the log2-normalized intensities of the fingerprint spectrum along this trajectory were plotted, with focus on pixels assigned to cluster 0 (healthy myocardium), 1 (remodeling myocardium), 3 (pre-fibrotic boundary) and 4 (fibrosis) (Fig. 4d) or just cluster 0 and 1 (Fig. 4i). Exemplarily, intensity shifts for hydroxyproline from collagen (858 cm−1) and cytochrome c in cardiomyocytes (1318 cm−1) were plotted in Fig. 4e, f. The increase of the collagen band indicates a pre-fibrotic remodeling of myocardium close to the region of subendocardial fibrosis und confirms our previous finding of a pre-fibrotic boundary around the actual fibrosis area. Excitingly, the alteration of cytochrome c intensity reflects the passage through healthy myocardium, remodeling zone and fibrotic area with alterations in mitochondrial metabolism at each level. The findings of a distinct myocardial subcluster of remodeling myocardium, as well as metabolic alterations approaching the fibrotic regions were reproduced in n = 4 individual mice with cardiac fibrosis (Fig. 4g, h). We validated our findings by exploring molecular dynamics along further trajectories, provided in the supplementary information (Supplementary Fig. 11).
The pseudotime analysis of the two myocardial subclusters in Fig. 4c demonstrated high molecular alterations in the remodeling subcluster (cluster 1) in contrast to that attributed to the healthy myocardium (cluster 0). To translate these findings to the spatial context, we analyzed intra-cluster heterogeneity along the trajectory, by filtering all pixels assigned to cluster 0 and 1 (Fig. 4i). Remarkably, we found high dynamics (i.e., intensity incline and/or decline) especially within cluster 1 of remodeling myocardium but not cluster 0 (Fig. 4j). These results support our previous findings of strong branching of the pseudotime trajectory, which was observed mainly within cluster 1. Peaks that showed strong dynamics along the spatial trajectory in cluster 1 were identified at 722, 1339, and 1569 cm−1 and correlate to different vibrational modes of proteins such as the O-C-N bending (722 cm−1), CN stretching and NH bending (1569 cm−1) as well as C-C stretching of the protein backbone (1339 cm−1) especially found in α-helical protein structures (Fig. 4j and Supplementary Fig. 11b)31. The latter finding was consistent with our previous finding of changes in protein secondary structure towards ß-sheets in remodeling myocardium. To quantify the spectral dynamics along all wavenumbers and also individual samples, the maximum variability of the local polynomial regression fitting (loess) curve and the derivation from the loess curve were calculated (see Methods). Especially quantification of the “curviness” of the intensity shift along the trajectory using the derivation from the loess curve showed significant differences within one sample (p < 2.2e–16, Fig. 4k, right) and also as reproduced in 4 individual mice (p = 0.0458, Fig. 4l, right).
High-dimensional characterization of metabolic alterations in myocardial infarction by spatial trajectories and multimodal Raman-MALDI imaging
To challenge our findings derived from remodeling myocardium, we transferred our strategy of spatial trajectories to a murine MI model with transitory ischemia and consecutive reperfusion for 24 h. This results in an infarct area, surrounded by a hypoperfused area at risk with heavy cell infiltration, transitioning into healthy myocardium which was not affected by the coronary occlusion32 (Fig. 5a). We sought to spatially reconstruct molecular changes when approaching the infarcted area, with focus on metabolic alterations. As results of ischemia, myocardium behind the ligated coronary artery undergoes a rearrangement of energy metabolism, which is – among others – characterized by a shift of glucose metabolism from oxidative phosphorylation to enhanced glycolysis33.
We created a spatial trajectory targeting the ligation site, which could be clearly identified in the H&E staining (Fig. 5a). Intensity changes of the fingerprint spectrum along this trajectory were plotted in a heatmap (Fig. 5b). We additionally supplied the analysis with an intensity map of the scanned area for C-H vibrations in proteins (2940 cm−1) to verify there is no general intensity shift along the large scan area. Employing spatial trajectories, we were able to spatially reconstruct metabolic changes from healthy myocardium towards the ischemic region. Three characteristic peaks for NADH34,35 and glucose34,36 were tracked along the trajectory and a specific pattern when crossing the infarct border and entering the ischemic heart region (Fig. 5c) was observed: NADH was decreased at the infarct border and elevated in the ischemic region, suggesting a mitochondrial compensatory mechanism for energy delivery in the I/R region and marked tissue damage at the infarct border. In contrast, glucose was increased in the peri-infarct area, but decreased in adjacent areas. We hypothesized that glucose is shifted to the inflammatory site of the myocardial infarction, where inflammatory cells massively infiltrate the diseased cardiac tissue and consume large amount of energy, as well as increased glycolysis of cardiomyocytes. In contrast, sham-operated mice didn’t show these metabolic alterations (Fig. 5d–f).
To further validate our findings and to supplement Raman spectroscopic data with more specific metabolic information, we combined our Raman “spectromics” approach with spatial metabolomics (Fig. 5g). Directly adjacent paraffin sections of 3 individual infarcted hearts were used to perform both Raman and MALDI imaging on the same region. Next, both datasets were integrated into the Seurat workflow and cluster analysis of the individual datasets was performed. Projecting the found clusters back to their spatial context showed for both Raman and MALDI datasets a distinct organization (Fig. 5h, left). Next, both datasets were spatially aligned (Methods), resulting in pixels containing both Raman and MALDI information. Dimension reduction and cluster analysis was then performed for both datasets individually and as multimodal analysis based on a weighted combination of the datasets. Strikingly, when translating the analyzed pixels back to their spatial context, we found a different spatial organization compared to those found by Raman or MALDI alone. The tissue organization identified by the multimodal analysis was to the best to reflect the ground truth boundaries of the infarcted region, compared to the H&E staining (Supplementary Fig. 12). Multimodal analysis on n = 3 individual infarcted hearts identified the clusters assigned to remote (healthy) and I/R (ischemia/reperfusion) border region and the spectral differences were visualized in volcano plots for both Raman (Fig. 5j) and MALDI (Fig. 5k) imaging. Both methods demonstrated an overall downregulation of metabolites in the infarct border region, as can be seen as a shift to the left in both plots. Significant and strong regulations of wavenumbers (Raman) or m/z values (MALDI) were plotted in a color-coded fashion (yellow: up, pink: down). We also plotted the average Raman spectrum from all hearts derived from the healthy remote region and the I/R border region (Fig. 5l).
Both Raman and MALDI imaging demonstrated to provide helpful information to define metabolic alterations in infarcted hearts. However, some of these information can technically or due to incomplete evidence only be identified by Raman or MALDI (Fig. 5m). As an example, Raman spectroscopy only identify a degradation of the α-helical structure of Amid I in proteins, together with an overall reduction of Amid I intensity in the ischemic region. For Amid III we found the same pattern for α-helical structures, but an intensity increase. Taken together with our previous results of changes in protein secondary structure towards ß-sheets in remodeling myocardium, this finding in the acute pathology provides further evidence on mechanisms of cardiac remodeling in the early stage of myocardial tissue damage.
Other metabolites could only be identified by MALDI analysis. Histidinyl-Serine as an example is a breakdown product of protein catabolism and cell signaling (HMDB0028894), which was increased at the infarct border. 6-Keto-Prostaglandin F1 is a marker of platelet activity (HMDB0002277) and was increased at the infarct region. The Lysophosphatidylcholines (LysoPCs) enlisted have anti-inflammatory effects (HMDB0010380, HMDB0010383) and were reduced in the infarcted area.
We also used the combination of Raman and MALDI to validate the identified metabolites by different methods. As an example, we compared 3 characteristic glucose peaks from Raman spectroscopy, as also used previously, and compared their intensity to that of Glucose-6-Phoshpate identified by MALDI imaging. We found comparable results for Glucose, and also Inositol-Phosphate compounds and DNA bases.
Thus, sophisticated usage of Raman spectromics data can be used to spatially resolve metabolic changes, multimodal analysis as well as molecular alterations along clusters identified by unsupervised algorithms of multiple datasets.
Defining the surrounding cellular (immune-) landscape in acute myocardial infarction
Following ischemia and reperfusion of the infarcted heart, blood-borne immune and inflammatory cells are a prominent feature in the diseased tissue. To explore the cellular environment of the infarcted heart, we transferred our strategy of Raman spectromics to the inflammation site of infarcted hearts. In order to define cellular subtypes by spectral data, the periinfarcted region was identified by H&E staining and consequently Raman scans from this region were overlayed by cyclic multicolor immunofluorescence staining (MACSima™ Imaging Cyclic Staining) of the identical specimen at the exact same position (Fig. 6a–d). This approach allowed direct identification of cell types by immunolabeling and allocating the underlying spectral fingerprint of the corresponding cell.
After selecting pixels with positive immunolabeling, dimension reduction of the underlying Raman spectra was performed and visualized using t-distributed stochastic neighbor embedding (tSNE) (Fig. 6f). Cardiomyocytes represented the largest cellular cohort of the analysis and presented the most heterogenous spectral composition. α-SMA+ Myofibroblasts or vascular smooth muscle cells (vSMCs) made up an own cluster, which did not clearly separate from the cardiomyocyte spectrum. In contrast, hemin+ erythrocytes and CD45+ leukocytes showed a clear clustering and separation from other cell types. CD41+ platelets ware mainly located within the immune cell cluster, however, this could also be the results of platelet-leukocyte-coaggregates. We also performed a subphenotyping of the immune cells (Fig. 6g): Ly6G+ neutrophils represented the largest and most heterogenous cluster, containing also Raman spectra from CD11b+ cells. MHC II+ professional antigen presenting cells (pAPCs) and CD68+ macrophages by contrast showed a specific clustering. We used a donut chart and venn diagram to visualize the frequency distribution of detected cells and intersection of immune cell surface markers (Fig. 6h, i). Characteristic average spectra for each identified cell type are presented in Fig. 6j, together with the spatial location of each cell type. The same approach was performed with a FFPE section, showing comparable results in spectral signatures from the identified cell types (Supplementary Fig. 14). We further assessed accuracy on how well Raman spectroscopy can delineate individual cell types found in the disease myocardium. We used the cell type-specific average spectra found in our previous analysis as “reference spectra” and calculate the similarity between the reference spectrum and each pixel spectrum of the scan. The computationally identified cell types were then compared to the corresponding immunofluorescence staining as ground truth measure. Our analyzed data suggest that Raman spectroscopy has a high specificity but low sensitivity when detecting and delineating cell types (Supplementary Fig. 16). Except for erythrocytes, which have strong characteristic spectra and low heterogeneity, sensitivity was overall around 65%. Hence, Raman spectroscopy has a good potential to exclude the existence of specific cell types in a scan, but cannot securely define if these occur in the scan.
Thus, by this proof-of-concept analysis we could demonstrate two things: 1) Raman spectromics can be easily combined with further downstream analyses of the very same histological. 2) By manual selection of cell type-specific Raman spectra, a database of characteristic spectra can be derived which can in turn be used for future identification of cell types without need for additional immunofluorescence labeling and imaging. Although Raman spectroscopy provides high sensitivity, its efficacy in accurately detecting and distinguishing different cell types within a Raman scan remains an ongoing challenge.