Mass spectrometry imaging (MSI) comprises several techniques that allow the analysis of hundreds of molecules simultaneously, sometimes with spatial resolutions as low as 1 µm. 1,2 Among MSI techniques, matrix-assisted laser desorption/ionization (MALDI) [3][4][5][6] and desorption electrospray ionization (DESI) 7,8 have been extensively used for the spatially resolved analysis of metabolites, 9,10 lipids, 11,12 peptides, 13,14 proteins, 15,16 and exogenous analytes, like drugs. [17][18][19] In comparison, laser ablation inductively-coupled plasma MS (LA-ICP-MS) imaging can provide quantitative distributions of a wide range of elements within a sample. [20][21][22] Despite the near universal detection capabilities of MS, it is difficult to detect all compounds of interest in a given MSI experiment. 23,24 Furthermore, the right combination of MSI modalities can provide complementary information, allowing optimal information to be obtained from a given sample. 25- 27 Properly combining the data from different imaging modalities can allow the strengths of each modality to be leveraged so that more in-depth information about a sample can be obtained. [28][29][30][31] For example, MALDI-MSI provides valuable biomolecule information but can suffer significant pixel-to-pixel signal variability. 32 On the other hand, LA-ICP-MS imaging is less subject to signal suppression and signal variability because samples are completely ablated, allowing it to provide quantitative information about metal distributions. 20 However, LA-ICP-MS imaging provides no molecular information due to its destructive nature. 21 By appropriately leveraging the datasets from MALDI-MS and LA-ICP-MS imaging of a given tissue section, more precise and informative images should be accessible. The comparison of results from LA-ICP-MS and MALDI-MS imaging is usually done by visually overlaying the images generated by the techniques. [33][34][35] However, simple image overlays hinder quantitative correlations because the images have different coordinate systems and orientations. More sophisticated multimodal registration methods allow pixel-to-pixel analysis of tissue sections by transforming the coordinates of one modality into the other modality, which allows statistical comparisons (e.g. Pearson's correlations). 24 Such approaches have been used by Caprioli and co-workers to register MALDI MSI images with microscopy images 36 and Van Malderen et al. to register LA-ICP-MS and microcomputed tomography images, 37 although these approaches have not been used for LA-ICP-MS and MALDI-MS image registration. Holzlechner et al. recently reported an approach to align LA-ICP-MS and MALDI-MS images 38 that was based upon a multisensor image integration method that uses fiducial markers to arrange images in the same coordinate system. 39 Since the approach is based on fiducial markers, the accuracy of the registration is limited and only linear transformations of the images are possible, making it unsuitable for registering images from adjacent tissue slices. 36 Moreover, this alignment approach is performed in the software package Epina Imagelab 39 , which is not open source or freely available.
Here, we describe a freely available computational workflow written in Python that allows the registration of LA-ICP-MS and MALDI-MS images in the same coordinate system, even for images from adjacent tissue sections, with accuracies within ± 50 µm. We illustrate the value of this registration approach by quantitatively correlating the distributions of nanomaterials with their biochemical effects. In addition, we demonstrate that the more precise imaging data from LA-ICP-MS imaging can be used to improve image segmentation in MALDI-MS images.
Nanoparticle (NP) synthesis was performed using the Brust-Schiffrin reaction, 40 followed by functionalization of the AuNP core with surface ligands, as described in previous work. [41][42][43][44] Similarly, NP-stabilized capsules (NPSC) were synthesized by mixing arginine-coated NPs with linoleic acid, followed by its functionalization with siRNA that causes knockdown of tumor necrosis factor α (TNFα), as described in detail in previous reports. [45][46][47] The methods used to synthesize and characterize these nanomaterials are described in the supporting information (SI), along with the results from these characterization measurements.
Balb/c mice were tail-vein injected with the NPs or NPSCs and euthanized after 48 h. Mice were sacrificed by carbon dioxide inhalation and cervical dislocation. All animal experiments were approved by the University of Massachusetts Amherst Institutional Animal Care and Use Committee (IACUC), which is guided by the U.S. Animal Welfare Act and U.S. Public Health Service Policy. Tissues were flash frozen and kept at -80 °C until sectioned for imaging. Frozen tissues were sectioned using a LEICA CMM1850 cryostat. Adjacent tissue sections of 12 µm thickness were thaw-mounted on indium tin oxide (ITO)-coated glass for MALDI-MSI and glass slides for LA-ICP-MS imaging experiments.
MALDI-MSI experiments were performed using 2,5-dihydroxybenzoic acid (2,5-DHB) as a matrix. Two different methods for matrix deposition were used: spraying and sublimation.
Spraying was performed using a Bruker ImagePrep device to spray a 25 mg/mL matrix solution in 1:1 methanol:water on the sectioned tissue. Sublimation was performed on a home-built sublimation apparatus similar to the setup described by Chaurand and co-workers. 48 For liver tissue sections, 200 mg of matrix were deposited at 140 °C at 7 mTorr for 9 minutes. For spleen tissue sections, 170 mg of matrix were deposited at 140 °C at 7 mTorr for 8 minutes. Data acquisition was performed on a Bruker UltrafleXetreme MALDI TOF/TOF at 50 μm resolution over a m/z range of 200 to 2000. MS/MS experiments to confirm analyte identities were performed in a collision-induced dissociation (CID) LIFT cell.
LA-ICP-MS images of 197 Au, 57 Fe, and 66 Zn were acquired on a CETAC LSX-213 G2 laser ablation system coupled with a Perkin Elmer NexION 300x ICP-MS instrument. The following laser parameters were used: 50 μm spot size, 20 μm/s scan rate, 3.65 J laser energy, 10 Hz laser frequency, and a 10 s shutter delay. The helium carrier gas from the laser ablation system was set to 0.6 L/min. Other ICP-MS experimental parameters were similar to those used in previous reports on the LA-ICP-MS analysis of nanomaterials in tissue sections. [49][50][51] Image preprocessing MALDI-MS images were normalized and exported as imzML files using FlexImaging (Bruker, Daltonics). The imzML files were imported to Python using the pyimzML parser, developed previously. 52 Peak picking was performed using SCiLS Lab 2015b, and the list of selected ions and mass tolerances were imported to Python as a text file. Images of the selected ions were rendered with the pyimzML parser. LA-ICP-MS images were reconstructed, analyzed, and segmented using a custom Python script RecSegImage-LA, which was described recently and is freely available at GitHub (https://github.com/Vachet-Lab/RecSegImage-LA). 53 Hotspot removal was performed on MALDI-MS and LA-ICP-MS images by selecting the intensities in the >0.99 quantile and replacing them with the 0.99 quantile value. 54 In some cases, t-stochastic nonlinear embedding (t-SNE) dimensionality reduction module from the scikit-learn Python library 55 was applied to selected ion images in MALDI-MSI data to obtain a single image representation of the dataset.
Image registration was performed using the SimpleElastix 56 Python extension of the Elastix C++ library. 57 The MALDI-MSI image (t-SNE or heme channel) was set as the fixed image, while the LA-ICP-MS image (Fe channel) was set as the moving image. Registration was performed using the default affine parameter map followed by the default non-linear parameter map in SimpleElastix with certain modifications as follows: 4,000 iterations for the affine parameter map, 8,000 iterations for the non-linear parameter map and 50 final grid spacing in physical units. Validation of the registration was performed using Dice similarity coefficient calculations (DSC), as described by Klein and co-workers (equation 1), where X and Y represent the binary label images. Selected regions, such as veins in the liver and white pulp spleen regions in the spleen, were manually selected in Fiji, 58 and imported to Python to calculate the DSC value.
Landmark distance analysis after registration was calculated by selecting corresponding points in the two images, followed by image overlay and calculation of their distances in Fiji.
Correlation coefficients between ionic signals in MALDI-MSI and LA-ICP-MS imaging data were calculated using the implementation of Pearson correlations in the Scipy library 59 on the vectorized, background subtracted MALDI-MS and LA-ICP-MS images.
Image registration involves transforming two or more images containing different data features into the same coordinate system. Once the images are registered, the combined information from the different imaging modalities allows deeper statistical and quantitative analyses of the images. In the process of image registration, one of the images is set as the fixed image, and the other one is the moving image (e.g., Figure S1). The moving image is transformed to maximize its similarity to the fixed image, resulting in an image that has the same coordinates and pixel number as the fixed image. In this work, LA-ICP-MS and MALDI-MS images were registered using SimpleElastix registration algorithms 56 in a custom Python workflow. Access to the scripts, examples and documentation can be found at GitHub: https://github.com/Vachet-Lab/MS-Registration.
Our approach to registration of LA-ICP-MS and MALDI-MS images relies on the use of internal signal features to drive the optimization process, as it seeks to maximize the mutual information present in both images. Ideally, these signal features should reflect the morphologic structure of the image (e.g., distinct sub-organ regions in a tissue) to ensure the best registration possible. Proper choice of the signal channels enables successful registration of LA-ICP-MS and MALDI-MS images when different MALDI matrix deposition approaches are used or even when adjacent tissue sections are imaged. For LA-ICP-MS, we find that the Fe signal channel (i.e., 57 Fe) is an effective feature to use as it indicates blood-rich regions that often define different regions in a tissue. For MALDI-MS images, we initially used the heme signal (m/z 616), as analogous indicator of blood flow. Figures 1a andb illustrate the LA-ICP-MS (red) and MALDI-MS (blue) images of liver and spleen tissue sections before and after registration. The MALDI-MS image was used as the fixed image, and the LA-ICP-MS image was used as the moving image. Visual inspection of these images shows that the registration process successfully aligns the tissue boundaries and other internal structure features. For example, in Figure 1a, a large piece of connective tissue (CT) that is devoid of heme signal in the MALDI image aligns well with the same low Fe signal in the LA-ICP-MS image. Similarly, the red pulp (RP) and white pulp (WP) regions of the spleen are very well aligned after registration (Figure 1b).
Registration of LA-ICP-MS and MALDI-MS images can also be accomplished when the matrix in MALDI-MS is deposited via sublimation. After matrix sublimation, we find that the heme signal is less reliably abundant, so we used a dimensionality reduction strategy, which leverages the many more signal channels that are present in MALDI-MSI datasets, to generate the signal features for the registration process. Dimensionality reduction was accomplished using a tdistributed stochastic neighbor embedding (t-SNE) approach that has been used successfully on MALDI-MSI data. [60][61][62] Using liver tissue sections as an example, t-SNE can be used effectively to combine nearly 40 different ions measured in MALDI-MS images (Figure S2). When the t-SNE generated features from MALDI-MS images are used together with the Fe signal from LA-ICP-MS, registration of the two images can be achieved. Our approach successfully registers the two images (Figure 1c), as indicated by the excellent overlap of the tissue boundaries and veins (V) in the images. It should be noted that the registered images shown in Figure 1 are from adjacent tissue sections. Using adjacent tissue slices allows MALDI-MS and LA-ICP-MS imaging conditions to be separately optimized. Registering adjacent tissue sections is only possible using non-linear registration approaches (Figure S3) to correct for local deformations in the tissues that can arise from placement of adjacent tissue slices.
The effectiveness of our registration approach was evaluated by two methods: Dice similarity coefficient (DSC) calculations and landmark validation. DSC values were calculated using the approach described by Rohlfin 63 (equation 1). For calculating the DSC values, regions of interest (ROI) were first chosen in both the LA-ICP-MS and MALDI-MS images (Figure 2). The chosen ROIs depended on the tissue type. For the liver, we used blood vessels, and for the spleen, we used the white pulp. were registered. 64 To further test the ability of our registration methods, we also tested tissues sections that were not immediately adjacent but were two sections apart. In one example, the DSC value increased from 0.34 to 0.69 after registration (Figure S4), indicating there is reasonable similarity between non-adjacent tissue sections.
Landmark validation 36 was also used to assess registration effectiveness. In the landmark approach, several morphologically distinct points are chosen in both LA-ICP-MS and MALDI-MS images, and the distance between these points is calculated and averaged to provide an effective registration accuracy (Figure S5). For the images shown in Figure 2, average registration accuracies of 40 ± 30 µm and 70 ± 20 µm are obtained for the liver and spleen, respectively.
Since the images were acquired at 50 µm resolution, the landmark distances show that most of the pixels are either perfectly correlated or are one pixel off. Given that the diameters of veins in the liver vary between 300 and 600 µm, and the diameters of white pulp areas are typically between 300 and 900 µm, these registration accuracies allow us to make confident conclusions about the veins and white pulp sub-organ regions. DSC values before and after registration. Green = LA-ICP-MS only pixels, Magenta = MALDI-MS only pixels, White = overlaid pixels. Segmentation of the veins and white pulp was performed manually using the Fe image in LA-ICP-MS and the t-SNE image in MALDI-MS to generate computational masks for each of the two images.
Once the images are registered, we can then compare how the signals in one image modality correlate with the signals in the other modality, which allows us to understand better the underlying biochemistry of the tissues. Using Pearson's correlations, which are one of the more accurate methods for quantifying the degree of co-localization in two images, 54 we can compare the extent to which metal distributions that are detected in LA-ICP-MS images correlate to specific biomolecule distributions that are detected in MALDI-MS images (Figure S6). As examples, we correlate Fe signals in LA-ICP-MS images of the liver and spleen with a range of lipids that are observed in MALDI-MS images of these same organs (Figure 3a and3c). For the spleen, we find that the signal levels for two classes of lipids, including ceramides (Cer) and some phosphatidylethanolamines (PE), correlate with the Fe signals as these lipids are expected to be present in the red pulp. 65 Since high Fe signals in LA-ICP-MS indicate the location of red pulp regions in the spleen, the lipids that positively correlate with the Fe are predominantly located in the red pulp. In contrast, the lipids that anticorrelate with the Fe, including lysophosphatidylcholines (LPC), phosphatidylcholines (PC), sphingomyelins (SM), 65 and carnitines (Car), are predominantly located in the white pulp, which has low Fe levels. In effect, we have identified several lipids that can act as biomarkers of the red and white pulp regions of the spleen.
The value of these correlations is even more evident when we image tissue sections from mice injected with gold nanomaterial delivery agents such as NPSCs that deliver TNFα-specific siRNA (Figure S7). These NPSCs have shown the ability to knockdown the production of TNFα in cell culture and in animals, [45][46][47] and this knockdown causes changes in the levels of various lipids. 66 In contrast, many more lipids show the opposite trend, becoming more negatively correlated with Fe and Au. For example, CAR (16:0) and SM (d18:1/17:0) have correlation values that change from 0.07 and -0.14 to -0.17 and -0.32, respectively. These anti-correlated values suggest that the presence of the NPSCs is generating changes to the levels of these lipids in places where the Fe and Au concentrations are low. That means that these lipid changes are occurring primarily in the white pulp where Fe concentrations are low and where Au accumulation is minimal (Figure S8). TNFα knockdown therapies, like these NPSCs, typically target macrophages and lymphocytes, which are highly abundant in the white pulp of the spleen, 68 likely explaining why so many lipid changes occur in the white pulp. CAR (16:0), SM (d18:1/17:0), and several of the PC lipids are signaling lipids known to undergo changes in concentrations upon TNFα suppression, 66 and the ability to correlate MALDI-MS and LA-ICP-MS images helps identify the specific sub-organ regions in which these changes are happening.
The analysis of tissue regions upon MSI often involves the division of the image into segments so that the detected molecular features can be associated with different cell types and regions of the analyzed tissue. A commonly used methods for such image segmentation is kmeans clustering, which is a statistical method that divides the image into segments that possess similar spectral characteristics, while being agnostic to the spatial structure of the data. 69 In MALDI-MS imaging, several approaches have been used to further improve segmentation, such as the implementation of more sophisticated spatially aware methods 70,71 and spatially-shrunken centroids. 72 Although segmentation algorithms for MALDI-MS imaging analysis are well developed, they highly depend on data quality, making the segmentation process challenging for noisy datasets. 72 LA-ICP-MS imaging usually produces less noisy images than MALDI-MSI primarily because the tissue section is completely ablated during the imaging process.
Consequently, we sought to leverage this higher precision in LA-ICP-MS imaging to improve segmentation in MALDI-MS images. To do this, we first segment the LA-ICP-MS image and then apply the resulting segmentation masks to the registered MALDI-MS images to improve the segmentation of the MALDI-MS data.
As an example of LA-ICP-MS-assisted segmentation of MALDI-MS images, imaging data from mouse spleen tissues were acquired by both techniques (Figures 4a andb). First, we segmented the MALDI-MS images using two methods available in SCiLS lab: (i) k-means on the normalized dataset with a cluster number of 4 (Figure 4c) and (ii) bisecting k-means (Figure 4d).
Both segmentation approaches differentiate the red and white pulp regions of the spleen, but neither method identifies a segment associated with the marginal zone, which is the ~100 µmsized region where initial immune responses occur in this organ. 73 In contrast, segmentation of the Fe image from LA-ICP-MS imaging using RecSegImage-LA 53 does classify the marginal zone of the spleen as a separate segment in the image in addition to the red and white pulp regions (Figure 4e). The three segmented areas from the LA-ICP-MS image can then be used as computational masks to classify the lipid signals from the MALDI-MS images that are most associated with each of the three different regions of the spleen (Figure 4f, g, andh). This lipid classification is accomplished by separately averaging the lipid ion signals from the pixels associated with each of the three separate regions. Statistical t-tests were then used to identify lipids that had significant differences between regions in the spleen (Table 1). From these a t-test probabilities were calculated to determine if there are statistically significant differences in localizations of each detected ion among red pulp, white pulp, and marginal zone regions. 'Yes' in green indicates significant differences between the compared areas. 'No' in orange indicates no significant difference between the compared areas. Probabilities are significant at a 99% confidence interval.
We have developed and evaluated a new computational workflow to register LA-ICP-MS and MALDI-MS images. Our workflow is written in Python and contains functions for image pre-