Discovering New Lipidomic Features Using Cell Type Specific Fluorophore Expression to Provide Spatial and Biological Specificity in a Multimodal Workflow with MALDI Imaging Mass Spectrometry

Abstract

Identifying the spatial distributions of biomolecules in tissue is crucial for understanding integrated function. Imaging mass spectrometry (IMS) allows simultaneous mapping of thousands of biosynthetic products such as lipids but has needed a means of identifying specific cell-types or functional states to correlate with molecular localization. We report, here, advances starting from identity marking with a genetically encoded fluorophore. The fluorescence emission data were integrated with IMS data through multimodal image processing with advanced registration techniques and data-driven image fusion. In an unbiased analysis of spleens, this integrated technology enabled identification of ether lipid species preferentially enriched in germinal centers. We propose that this use of genetic marking for microanatomical regions of interest can be paired with molecular information from IMS for any tissue, cell-type, or activity state for which fluorescence is driven by a gene-tracking allele and ultimately with outputs of other means of spatial mapping.

Graphical Abstract

Simultaneously mapping the spatial localizations of biomolecules enables the formulation of new hypotheses and can test models related to physiology, disease pathogenesis, and clinical applications. Although a variety of technologies exist for spatial localization of metabolites, these technologies face barriers in providing full biological context to findings because biosynthesis and steady-state levels of molecular determinants of cell metabolism and function may be regulated post-translationally. Thus, complementary imaging modalities are required for correlation of molecular images with biologically relevant substructures. Matrix-assisted laser desorption/ionization (MALDI) imaging mass spectrometry (IMS) enables the mapping of thousands of unlabeled molecules, including lipids and other metabolic products, directly from tissue sections at high spatial resolution.¹ The challenge of correlating ion localization to unambiguous identification of microanatomical regions of interest (ROIs) is both computational and experimental.

Microscopy images collected from stained tissue (e.g., staining of tissues by hematoxylin and eosin (H&E) or immunofluorescence (IF)^2–4 are generally used to provide biological context to IMS data. It has been reported that with low laser fluence, immunofluorescence can be performed on the same tissue section.⁵ However, this does not enable the detection of low abundance lipids. In addition, imaging of lipid and protein distributions on the same tissue section has also been reported,⁶ but the identification of low abundance of cell type-specific proteins usually requires a technique more sensitive than MALDI IMS. The use of serial sections, the standard method of providing this bio logical context, limits the discriminant power of scoring cell identity or functional status (e.g., activity of a particular gene) for small ROIs. Moreover, differences in spatial resolution can make correlating IMS and microscopy images challenging. Routine spatial resolution of most IMS experiments is 10–30 μm but can attain 5 μm resolution using specialized instruments which is less than the thickness of tissue sections used.^7,8 These considerations highlight the need for a multimodal workflow in which biological features can be identified at a microanatomic scale in IMS analyses.⁹ The spatial colocalization of a transgenic fluorophore with IMS data provides enhanced biological specificity and advanced data-mining strategies to uncover molecular correlations with ROIs.

Every multimodal analysis has three central processes: registration (alignment of images in 2-D space¹⁰), data mining (parsing through data for relevant m/z values¹¹), and molecular identification (elucidation through MS/MS⁸). Traditionally, multimodal imaging has relied on manual interpretation of coregistered ion images,¹² which is prone to human bias. Other supervised and unsupervised approaches have been used to improve data analysis.^13–19 Each of these approaches still requires an independent benchmark to define cells or structures. Herein, we provide evidence of a new approach that enabled the identification of ROIs on the same tissue section using a cell-type specific transgenic fluorophore to provide biological specificity and the basis for fluorophore-directed data mining.

To develop this technology, we analyzed the spleens of unimmunized and immunized mice using a well-characterized tracking allele that encodes green fluorescent protein (GFP) to enable high accuracy image registration and provide biological context.²⁰ Data mining strategies, such as manual interpretation,^11,12 standard segmentation,²¹ and data-driven image fusion,²² were subsequently applied to determine whether lipids could be mapped to a feature of normal microanatomy in immune responses. The analyses show that data-driven image fusion allowed for the most robust mining of multimodal data by leveraging the correlation of fluorescence emission and IMS to identify previously unknown spatial molecular relationships.

MATERIALS AND METHODS

Materials.

MALDI matrix 1,5-diaminonapthalene (DAN) was purchased from Sigma-Aldrich Chemical Co. (St. Louis, MO, USA). Sheep red blood cells (SRBC), ammonium formate, carboxymethyl cellulose sodium salt, isopropyl alcohol, mass spectrometry grade water, chloroform, and acetonitrile were purchased from Fisher Scientific (Pittsburgh, PA, USA); streptavidin-Alexa647 antibody (Ab) and chemically conjugated monoclonal Ab (GL7-FITC, αIgD-PE and αCD35-biotin) were purchased from BD Biosciences (San Jose, CA). C57BL/6-J mice and breeding stock transgenic for a bacterial artificial chromosome that integrates a translational fusion of GFP with AID into the Aicda locus (AID-GFP mice; stock no. 018421) were obtained from Jackson Laboratory and bred with C57BL/6-J. All mice were housed in ventilated microisolators under Specified-Pathogen-Free conditions in a Vanderbilt mouse facility and used in accordance with protocols approved by the Institutional Animal Care & Use Committee.

Tissue Preparation.

AID-GFP (n = 3) and C57BL/6-J (n = 3) mice age 6–7 weeks were immunized with sheep red blood cells to compare with nonimmunized controls (C57BL6-J, n = 3) and euthanized 8 days postimmunization. Spleens were sectioned at 12 μm, and three serial sections were used for H&E, IF, and IMS with fluorescence emission including autofluorescence on all sections prior to a secondary modality (Figure 1).

Figure 1.

Workflow for multimodal analysis and data extraction. Shown are a schematic (a) and representative data (b–f) to illustrate the initial IMS analyses. (a) Mice of the indicated genotypes (bearing or lacking an Aicda BAC transgene engineered to express AID-GFP translational fusion protein) and immunization status were used starting at 6–7 weeks of age. (b–d) Spleens harvested 8 d postimmunization were used to generate triads of serial tissue sections (12 μm thickness) (b), followed by fluorescence emission (F_em) and other imaging modalities (c). After processing, immunofluorescence (IF), IMS, and hematoxylin and eosin staining (H&E) were each performed with one of the three sections (d). (e) F_em data from sections 1–3, as indicated, are shown adjacent to the IF, a single m/z from negative ion mode IMS and H&E images from the same section as the F_em. Intra- and intersection registrations were then performed using a published method in which IMS data are aligned with the post IMS laser ablation marks, and all other modalities were aligned to IMS data through fluorescence emission on each section. (f) Manual interpretation, segmentation, and data-driven image fusion were performed with publicly available software to map ions of interest, as detailed in the Materials and Methods.

Mass Spectrometry Imaging.

IMS sections were washed with ammonium formate and sprayed on a TM Sprayer (HTX, Chapel Hill, NC, USA) with recrystallized 10 mg/mL 1,5 DAN in 9:1 (v/v) acetonitrile/deionized water. Negative ion mode IMS data were acquired from m/z 200–2000 with a raster step of 30 μm with a 9.4T Bruker FT-ICR Solarix mass spectrometer (Bruker Daltonics, Billerica, MA, USA). Laser power was optimized for each sample between 18% and 20% with 500 laser shots per pixel. Smart walk of 25 μm was enabled to increase sensitivity. For image fusion analysis, a higher spatial resolution image was generated using the same 9.4T FT-ICR with similar settings except that the raster step was 15 μm without smart walk enabled, and 750 laser shots per pixel were generated at a laser power of 13%. All data sets are available at https://figshare.com/projects/Discovering_new_lipidomic_features_using_cell_type_specific_fluorophore_expression_to_provide_spatial_and_biological_specificity_in_a_multimodal_workflow_with_MALDI_IMS/70058.

Image Registration.

Image registration techniques were performed according to previously published methods;²⁰ however, rather than using only autofluorescence images for registration, we used fluorescence emission images.

IMS Data Analysis.

All data was root-mean-square (RMS) normalized, and further analyses were performed. Manual interpretation analyses were performed in SCiLS, and spatially shrunken centroid segmentation analysis was performed in R with the package Cardinal. Image fusion analyses were performed according to previously published methods,²² utilizing partial least-squares regression to link IMS images’ fluorescence emission data. Localization to germinal centers was determined using QuPath software for annotation and an R script for data extraction. Weighted averages were tested for significance with ratio-paired t tests.

Identification of Lipid Species.

LC-MS/MS of total splenocytes was performed on a Q Exactive HF mass spectrometer from m/z 375–1650 in negative ion mode in parallel reaction monitoring (PRM) mode with an isolation window of 2 Da for each ion of interest using eluates from a Vanquish UHPLC (Thermo Scientific, Waltham, MA, USA). MS/MS resolving power was 15 000 at m/z 200, while full scan resolving power was 30 000 at m/z 200. Complementary analyses were performed using MS/MS based imaging experiments using a 15T Bruker FT-ICR solariX mass spectrometer (Bruker Daltonics, Billerica, MA, USA). Data were collected in negative ion mode from m/z 250–2000 with 1000 laser shots per pixel, and a raster step was set to 60–120 μm. Ions were isolated with a 2–6 Da mass window and fragmented using collision induced dissociation (CID) with a collision energy of 17–27 V.

RESULTS AND DISCUSSION

Overview.

We used multiple microscopy-based imaging modalities registered to IMS measurements to elucidate the lipidomic differences between germinal center and splenic white pulp or lymphoid follicles (Figure 1). Specifically, the AID-GFP (activation-induced deaminase-green fluorescent protein) transgene provided a cell type and region-specific fluorophore.²³ This tracking allele highlights a microanatomical feature, the germinal center, that forms within lymphoid follicles after immunization because of a large increase in Aicda gene expression in germinal center B lymphocytes, diversifying and improving the qualities of antibody responses.²³ Fluorescence emission provided a nondestructive means of identifying germinal centers via colocalization with AID-GFP, including autofluorescence from endogenous molecules that correlated with the splenic tissue structures (e.g., red pulp and white pulp surrounding germinal centers). This fluorescence emission modality also provided a single image type that could be collected from every tissue section prior to other modalities (i.e., conventionally stained microscopy, fluorescence microscopy, and IMS) (Figure 1b–d), enabling high accuracy image registration (Figure 1e). In H&E stains, the most traditional means of providing biological context to IMS data, red pulp can be differentiated from white pulp but germinal centers are less conclusively differentiated. IF microscopy allowed for the identification of germinal centers and their light and dark zones. Because AID expression is similar in light and dark zones, both conventional and fluorescence after immunostaining were performed on serial sections. As this results in plane-of-section differences from sections used for IMS, advanced registration approaches were needed. By integrating these modalities into a single multiplanar data set, we enabled a full integration of imaging modalities to provide a unique combination of molecular coverage, spatial resolution, and biological specificity.

Registration.

We first tested whether this method allowed incorporation of fluorescence emission as an additional modality within each section to enable a high degree of spatially localized biological information. Sections were analyzed for germinal centers in spleens of mice, immunized or not, and bearing or lacking an AID-GFP transgenic fluorophore. The same tissue sections were then used for IMS, while serial sections were used for IF and H&E. This method was applied to an investigation of the differences between lipids associated with germinal centers and other regions in spleens using nondestructive fluorescence emission as a monomodal registration medium²¹ (Figure 1e). Spleens of nonimmunized controls were compared to those of immunized mice bearing or lacking the AID-GFP transgene (Figure 1a). IMS was then used to identify m/z features after collection of fluorescence emission images of the AID-GFP fluorophore.

In addition to identification of germinal centers within the section destined for IMS through fluorescence emission, we investigated lipid differences in subregions of the germinal center. Accordingly, the workflow incorporated IF staining of adjacent sections with antibodies specific for markers that not only would identify germinal centers by independent criteria (IgDneg GL7+) but also would allow subdivision of the germinal centers into functionally distinct domains termed the dark (CD35−) and light (CD35+) zones. To compare the conventional use of serial sections to intrasection registration, we quantitated the error in overlap between adjacent sections. Germinal center masks annotated for all AID-GFP mouse spleen serial section pairs (n = 5) were used to calculate a Dice–Sorenson coefficient (DSC), a statistical means of determining the similarity of two samples that were registered as described by Patterson et al.²¹ The average DSC was 0.81 (±0.3) for the five pairs, indicating that serial sections as registered can be expected to have roughly 80% germinal center overlap (Figure S1 and Table S1).

Data Mining.

Overall, 1375 m/z features were detected at a S/N > 3 by IMS, including a variety of lipids with diverse patterns of localization to substructures of spleen that included red and white pulp. In addition to these constitutive features of splenic microanatomy, germinal centers form in the white pulp after lymphocyte activation by immunization generates T cell help. Mice were immunized to increase size and numbers of germinal centers as observed in all imaging modalities when comparing immunized to nonimmunized controls (Figures 2a and S2). In fluorescence emission images, a difference in germinal centers localized GFP expression is observed between samples with and without AID-GFP (Figures 2a and S2). Autofluorescence detected in the DAPI and TRITC channels enhanced the identification of germinal centers in the FITC channel by distinguishing them from other portions of the white pulp highlighted in autofluorescence. The IMS data were first analyzed using manual interpretation (Figure 1f). Two ions of interest were selected by virtue of their association with in-section AID-GFP, m/z 752.5591 and m/z 776.5596 (Figure 2a). A ratio-paired t test applied to the ion intensity was performed to determine significance of correlation and anticorrelation throughout this work. Specifically, germinal centers were compared to nongerminal center regions. Because AID-GFP does not distinguish the germinal centers’ subregions,²⁴ IF of adjacent sections was employed to identify the light and dark zones.

Figure 2.

High accuracy registration of multimodal data. (a) Representative registered images highlighting the types of detection. Rectangular areas of immunized AID-GFP transgenic (AID-GFP Imm) mouse spleen are shown with each section, from left to right: Hematoxylin and eosin (H&E); fluorescence emission/autofluorescence (Fem/AF); immunofluorescence (IF) after staining with mAb; IMS with three ions [m/z 752.5591, m/z 791.5410, and m/z 810.5269] overlaid for context of white pulp and red pulp; and a single ion image showing m/z 752.5591 (IMS752). Intensity scales from least to greatest total ion intensity and color legends are displayed below each set of images. A 1000 μm scale bar is depicted in the H&E image. Fem was taken on the same section imaged by IMS. IF and H&E were then taken from serial sections to the IMS section. IF was used to identify microanatomic portions of lymphoid follicles, and included both indirect and direct staining of GL7, IgD, and CD35. (b) Higher magnification images of a single representative GC (germinal center; designated by a white box in 1a) are shown with the same sample order and modalities. Light and dark zones (LZ, DZ) are demarcated by a yellow and blue outline, respectively. (c) The bar graph shows the ratio of ion intensities in GC to non-GC regions for the m/z features of 776.5596 and 752.5591 [identified by IMS MS/MS in Figure 4 as PE (O-18:0 22:6) and PE (O-18:0 20:4), respectively] (p = 0.0409, p = 0.0099, n = 3). (d) Geometric mean of the ratio of LZ/DZ ion intensity of two lipids is 1.6 and 1.5 for PE (O-18:0 20:4) and PE (O-18:0 22:6) (p = 0.007, n = 65, p = <0.0001, n = 65). Replicates and magnified regions for WT Imm and WT Non-Imm samples can be found in the Figure S2

Data were further analyzed for significant differences in germinal center light and dark zones²⁵ as identified in fluorescence emission and IF microscopy images. To obtain ion intensity for statistical analysis, we used QuPath and a custom R program to extract ion intensity values for all germinal center and nongerminal center regions identified through fluorescence emission. Subregions of germinal centers, light and dark zones identified through IF were annotated in QuPath²⁶ and compared. Pairs of germinal center light and dark zones were identified based on shortest Euclidian distance (Figure S3).

Ions discovered through manual interpretation, m/z 752.5591 and m/z 776.5596, were mapped to the germinal center (~8-fold and ~5-fold enrichment; Figure 2b and c), and each of these lipid species was further enriched in the light zone compared to the dark zone (~1.6-fold and ~1.5-fold enrichment within the germinal center; Figure 2b and d).

Spatially shrunken centroids segmentation circumvents the potential for human cognitive bias introduced through manual interpretation by computationally determining ROIs²¹ (Figure 1f). This approach generated a list of four ions that localize to germinal centers, m/z 752.5591, 776.5596, 883.5360, and 887.5609 (m/z 883.5360, p = 0.0025, n = 3; m/z 887.5609, p = 0.0087, n = 3)^21,27 (Table S2). Of these, the first two (m/z 752.5591 and 776.5596) matched the ions discovered by manual interpretation, and all localized to germinal centers but not all localized to light or dark zones (m/z 883.5360, p = 0.12, n = 108, and m/z 887.5609, p = 0.070, n = 106) (Figures 2 and S6q–r and Table 1).

Table 1.

Germinal Center Lipids Revealed through All Data Mining Strategies^a

m/z	lipid ID	DB matches	P value GC vs non-GC	P value LZ vs DZ	ppm errorb	man. int.	seg.	imagefusionrelative slope in model
671.4647	PA(18:1_16:1)	6	0.09	0.0007	0.070			222.1
699.4957	PA(18:1_18:1)	6	0.03	0.0002	0.36			277.2
699.4957	PA(18:0_18:2)	7	0.03	0.0002	0.36			277.2
699.4957	PA(20:2_16:0)	7	0.03	0.0002	0.36			277.2
714.5069	PE(18:2 16:0)	4	0.04	0.2	0.053			102.5
716.5224	PE(18:0_16:1)	3	0.1	0.9	0.059			243.8
725.5120	PA(20:3_18:0)	8	0.007	0.02	0.53			63.2
740.5246	PE(18:1_18:2)	4	0.01	0.01	2.9			112.8
742.5389	PE(18:0_18:2)	5	0.04	0.0006	1.0			290.5
746.5130	PE(P-16:0_22:6)	6	0.005	<0.0001	1.5			280.0
748.5273	PE(O-16:0_22:6)	6	0.007	0.2	0.37			236.6
752.5591	PE(O-18:0_20:4)	5	0.01	<0.0001	0.32	X	X	219.0
762.5088	PE(16:0_22:6)	4	0.03	0.2	2.6			167.0
772.5314	PE(P-18:1_22:6)	5	0.03	0.01	5.0			163.3
776.5596	PE(O-18:0_22:6)	5	0.05	<0.0001	0.88	X	X	244.7
786.5303	PS(18:0_18:2)	8	0.02	0.0004	2.9			279.4
812.5460	PS(18:0_20:3)	2	0.03	0.3	3.0			37.1
857.5182	PI(16:0_20:4)	16	0.009	0.002	0.82			400.6
883.5360	PI(18:1_20:4)	6	0.003	0.1	3.3		X	565.8
887.5609	PI(18:0 20:3)	14	0.0006	0.07	3.9		X	252.9

^aFrom left to right the m/z value, identification of the lipid found via the multimodal work flow validated through MS/MS imaging, matches to the LIPIDMAPS database, p value for a t-test between germinal center (GC) and non-germinal center (non-GC) regions, p value for a ratio-paired t-test between light and dark zones (LZ, DZ), ppm error in identification, manual interpretation discovery, segmentation discovery, or data-driven image fusion discovery.

^bNote that ppm error was determined from a tune mix doped IMS experiment.

Although segmentation enabled the identification of four ions of interest localizing to germinal centers (Figure 2a and Table 1), this approach is well suited only for determining ions that directly correlate to a specific tissue subregion. Data-driven image fusion connects the spatial and informational content of two imaging modalities by constructing a mathematical cross-modality model between the two, using multivariate linear regression (Figures S4 and S5).²² In previous work, data-driven image fusion and the models it builds have been used for prediction-oriented applications such as spatial sharpening, out-of-sample prediction, and image denoising²².

In this study, we introduce a new relationship discovery application of data-driven image fusion that does not pursue prediction but rather only executes the first model-building phase of the fusion framework. The second phase, prediction using the built model, is not needed for this discovery application, essentially side-stepping the usual considerations of uncertainty that come with prediction. Instead, we hypothesized that by building a model that ties IMS images to fluorescence emission, the multivariate linear models produced this way could be used directly to empirically uncover new correlative relationships between the two modalities. Building a cross-modal model, and instead of using it for prediction, opening it up to see what relationships it has learned, could potentially enable fluorophore-directed data mining. Accordingly, we tested the use of data-driven image fusion to provide a deeper understanding of all correlative relationships between IMS and fluorescence emission data in germinal centers.

Since we used multivariate linear models, the relationship between a fluorescent channel and an m/z “channel” are encoded by slope values. High relative slope values indicate a strong relationship between the two channels. Although prediction is not a part of this modeling effort, there is always discrepancy between the phenomenon modeled and the model itself. To capture this, we report only ions with acceptable reconstruction scores as described in Van de Plas et al.²² and acknowledge that, since always some uncertainty remains, the findings should be independently cross-checked by evaluating the genuine ion distributions of the fusion-proposed ion species. The fusion-driven discovery process is simply meant as a rapid and automated means of filtering through a large set of ion species (e.g., in the thousands) and reducing it down to a more digestible panel of potentially correlating ion species (e.g., in the tens) that can be further investigated and validated.

From the fusion of a high resolution (15 μm) IMS and fluorescence emission images, 16 germinal center-specific ions were revealed (Table 1), of which four were those highlighted by segmentation-based analyses (Figure 2a and c). Integration of the image fusion algorithm into the workflow allowed identification of a far greater number of candidates for germinal center-associated ions along with species that were anticorrelated (e.g., m/z 687.5447 and m/z 788.5442, Figures 3 and S5 and Table S3).

Figure 3.

Identification of anticorrelating germinal center (non-GC) ions by image fusion. (a) Shown are representative registered images highlighting the localization of anticorrelating GC ions. From left to right, the following image types are pictured: H&E with scale bar, Fem, IF, IMS showing an overlay of non-GC ion m/z 687.5447 and GC ion m/z 752.5591, and IMS showing an overlay of non-GC ion m/z 788.5442. LZ and DZ, as identified by IF, are outlined in yellow and blue, respectively. (b) From left to right, the m/z value, identification, matches to the LIPIDMAPS database in the IMS MS/MS spectrum, statistical significance, and ppm error in mass identification are listed. These two ions were identified as PE-Cer (d36:1) by accurate mass and PS(18:1 18:0) through IMS MS/MS, respectively. A complete listing of results from image fusion is in Table 1; further information is in the extended data (Figures S4 and S5).

Germinal center areas annotated in fluorescence emission images served as a means for identifying germinal center (p = 0.0099, n = 3, slope = 219.0 for green channel) and nongerminal center regions for statistical analysis (p = 0.04, n = 3, slope= −57.5 for green channel and p = 0.04, n = 3, slope = −202.6 for green channel respectively) (Figure 3, Table S3, and Figure S5). The ion m/z 752.5591 is shown for contrast with nongerminal center ions m/z 687.5447 and m/z 788.5442 (Figure 3). In contrast to manual interpretation and segmentation, ten additional ions revealed through data-driven image fusion were higher in germinal centers and exhibited a pattern of light zone > dark zone (Table 1 and Figures S4 and S5).

Molecular Identification.

Because of the large number of potential isomers at these m/z values, mass accuracy alone was not enough to identify lipid species. For example, the phosphatidylethanolamine ether species PE(O-40:6) and PE(O-38:4) are isomers of the phosphatidylethanolamine plasmalogen species PE(P-40:5) and PE(P-38:3), respectively. Liquid chromatography-tandem mass spectrometry (LC-MS/MS) determined the presence of both ether and plasmalogen species for these ions of interest in total lipid extracts from whole spleen tissue (Figure 4a). Thus, a spatial component was needed to confirm the identity of the ions that correlate to fluorescence emission signals. IMS-MS/MS was performed with sectioned spleens of immunized transgenic AID-GFP mice (S9–27). The MS/MS-based imaging experiment determined these ions to be ether lipids PE(O-18:0_20:4) (Figure 4b–d) and PE(O-18:0_22:6) and not the isomeric plasmalogens (Figure 4e and f). In addition, the colocalization of the specific fragment ions from these ether lipids with fluorescence emission signals reveals that these species were enriched in splenic germinal centers.

Figure 4.

Identification of species localizing to germinal centers as ether linked lipids. (a) LC-MS/MS fragmentation spectra of total splenocytes show common fragments for both plasmalogen and ether lipids (enlarged) from a parent mass of m/z 752.545. (b) Shown to the left is the chemical structure of the parent ether ion and to the right the corresponding ion image. (c, d) Correlating ether fragments are depicted with the chemical structure on the left and ion image on the right. (e, f) Similarly, the plasmalogen parent ion structure and fragments are shown with chemical structure on the left and corresponding ion images on the right.

In addition to plasmalogen and ether species, image fusion enabled the identification of a variety of phosphatidylethanolamine (PE), phosphatidic acid (PA), glycerophosphoserine (PS), and glycerophosphoinositol (PI) lipids that were enriched in germinal centers, with some observed at higher intensity in germinal center light zones. Fatty acid tails of 16:0 and 18:0 were most common. We observed many repeats of fatty acid tails 20:3, 20:4, and 22:6. In germinal centers, five out of eight lipids had unsaturated fatty acid tails, whereas in germinal center light zone, all eight had at least one unsaturated fatty acid tail. Two ions, m/z 687.5447, PE-Cer(d36:1) (phosphatidylethanolamine ceramide), and m/z 788.5442, PS(18:1_18:0), were identified as anticorrelating with germinal centers (Figure 3).

Implications.

Overall, the multimodal imaging process reported here combines high spatial resolution IMS with a transgenic fluorophore to identify microanatomical regions of biological interest. Our approach incorporates high accuracy registration and various data mining tools, including data-driven image fusion, to fully integrate multiple imaging modalities collected from a single tissue section and across adjacent sections. This technology revealed an enrichment of ether and plasmalogen lipid species in substructures central to antibody response and humoral immunity known as germinal centers. Unambiguous identification of germinal centers and the assessment of lipid abundances in light and dark zones was made possible by combining fluorescence emission of the transgenic tracking allele with traditional microscopy approaches (i.e., stained and IF microscopy). While data-driven image fusion has previously been used for predictive applications, such as image sharpening and out-of-sample prediction, the evidence presented here indicates that it can also be applied to mine highly dimensional data to find correlations between modalities by interpreting the linear models constructed during the fusion process. When compared to conventional approaches, the yield of structure-associated molecules was enhanced 4- to 5-fold, as 16 germinal center-associated lipid species were determined.

We identified three key processes in multimodal imaging as (1) registration, (2) data mining, and (3) molecular identification. Histological depth differences between serial sections are becoming larger challenges as the spatial resolution of IMS increases^28–30 because of the small size of single cells within a tissue. In addition to histological depth differences, accurate data alignment correlating H&E or IF to IMS becomes central as spatial resolution increases and regions of interest approach single cells. Importantly, the technologies presented here should be applicable to fusion of IMS, fluorescence, and spatial transcriptomic or protein data.^31,32

The unexpected finding that the prevalence of a series of ether lipid species is higher in germinal centers frames new hypotheses, that is, that molecular programing of germinal center lymphocytes is tied to increased ether lipid synthesis and that these species are functionally important in humoral immunity. A higher abundance of ether lipids in the spleen and white blood cells has been reported, but the exact role of these ether lipids remains uninvestigated.³³ Ether lipid synthesis begins in the peroxisome and is completed in the endoplasmic reticulum.³⁴ Disruption of this pathway in peroxisome biogenesis disorders, such as Zellweger spectrum (PBD-ZSD), or by gene-targeting generates decreased ether lipid levels.^34–36 In this light, it was striking that image analysis of IMS uncovered germinal center PE lipids with the same tail lengths as their ether and plasmalogen counterparts. Most notably, PE(16:0_22:6) localized to germinal centers as did its ether lipid counterpart PE(O-16:0_22:6), while its plasmalogen derivative, PE(P-16:0_22:6), localized not only to germinal centers but within them to their light zone (Table 1, Figure S5o–q, and Table S3). This enrichment along a pathway suggests that germinal centers have enhanced peroxisomal activity, resulting in increased abundance of PE-ether lipids.

The peroxisome also generates reactive oxygen species (ROS).³⁷ Plasmalogen ether lipids scavenge reactive oxygen species.³⁸ This capability has not been documented for nonplasmalogen ether lipids, but the structural similarity suggests a connection in synthesis pathways and roles.³⁹ Starting 3.5 d after immunization, germinal centers form in the follicles of secondary lymphoid organs and are sites of B-cell proliferation, differentiation, and selection that are central to promoting antibody affinity increases, as well as vaccine efficacy and humoral immunity.²⁴ Substantial AID-mediated mutational²³ and nutrient^24,40 stresses appear to be present in germinal centers. This microanatomic structure consists of light and dark zones in which the native oxygen levels vary, such that hypoxia is present in an light zone > dark zone pattern.²⁴ While there is strong evidence of connections between hypoxia and inflammation,^41,42 much remains unknown as to the effect of this hypoxic microenvironment on lipid synthesis within these regions.⁴³ The role of ether lipids in the adaptive immune microenvironment has not yet been explored, and thus, management of ROS and their levels are crucial for lymphocyte physiology.⁴⁴ This point, in conjunction with known metabolic stresses in germinal centers^23,24,45 and influences of hypoxia on ROS generation,²⁴ suggests that a model in which higher plasmalogen and ether lipid abundance involves a physiological role in which ether lipid production aids in maintaining optimal ROS levels.³⁷

CONCLUSION

We demonstrated a multimodal molecular imaging workflow that integrates two key methodological advances documented here: (1) the use of engineered alleles that track gene expression by linking a fluorophore to the normal gene product, and (2) application of data-driven image fusion for data mining. We used this workflow to identify 16 germinal center-specific ions that led to the formation of a new hypothesis related to ether lipids in germinal centers.

This approach should be widely applicable to a variety of experiments in a broad range of biological systems. Geneediting technologies, such as CRISPR-Cas9, will further expand an already abundant supply of transgenes that mark specific biological pathways and cell types. Moreover, this new discovery-oriented application of data-driven image fusion as a means of elucidating ions of interest colocalizing with a specific fluorophore will enable unique applications of data mining, including applications in settings where unambiguous marking of a region of interest by other modalities exists.