Femtosecond laser desorption ionization mass spectrometry imaging and multivariate analysis of lipids in pancreatic tissue Free

The imaging of lipids within intact mammalian tissue has been repeatedly demonstrated using both secondary ion mass spectrometry (SIMS),^1–3 matrix assisted laser desorption ionization mass spectrometry (MALDI-MS),^1,4–7 and desorption electrospray ionization (DESI).^8,9 However, the cluster ion beams now popular in SIMS, the solvent requirements in DESI, and the sample preparation methods required of MALDI-MS can lead to limited spatial resolution, selective detection, and/or other experimental challenges. There is general interest in new MS imaging strategies that might mitigate these challenges while incorporating the many advantages of MALDI, DESI, and SIMS.

Hanley and coauthors^10,11 have previously demonstrated the use of <100 femtosecond (fs) near-infrared laser pulses for ablation of neutrals that were postionized with a second vacuum ultraviolet (VUV) laser. This so called femtosecond laser desorption postionization mass spectrometry (fs-LDPI-MS) imaging method demonstrated ∼2 μm lateral resolution on an organic test pattern.¹² The present work uses the single laser fs-laser desorption ionization (fs-LDI-MS) configuration—minus postionization—that can sometimes generate useful signal.¹³ 400 nm fs-LDI-MS was previously used by Winograd and coworkers¹⁴ to demonstrate that intact protein molecular ions could be desorbed from ice. Dantus and coworkers¹⁵ subsequently employed 800 nm fs-LDI-MS to image onion epidermis with ∼10 μm resolution. In this paper, we explore the use of 800 nm fs-LDI-MS to perform proof-of-concept studies of the mass spectrometric imaging of lipids in mammalian tissues, via the examination of human pancreas tissue.

All methods of MS imaging generate large datasets that defy manual analysis. Thus, SIMS and other MS imaging methods often utilize multivariate analysis (MVA) to assist in the interpretation of these large datasets.^16,17 The object of such analyses is to spatially resolve the sample into separate “components” and to determine the significant spectral features associated with each. MVA methods differ in the requirements imposed on the component spectra, the numerical methods used to find the optimum components and their distributions, and the statistical distribution of noise assumed in the analysis. Two MVA methods are used in the current study. The first is the principal component analysis (PCA), which is widely used in the analysis of MS imaging data.^18–20 PCA-derived component “spectra” (formally known as loading vectors) invariably contain negative peaks and must therefore be interpreted carefully and not confused with the mass spectra of individual species or mixtures. The second method used is a maximum a posteriori (MAP) analysis that is described further below.^21–23 This method allows for the quantitative extraction of physically meaningful component spectra which can be correlated to mass spectra and can be easily extended to include matrix effects, instrument effects, sample topography, and prior knowledge about the sample composition or other parameters.^21–23 

The imaging of lipids in human pancreatic tissue is investigated in the current work by fs-LDI-MS. Bright field optical microscopy of hematoxylin and eosin (H&E) stained adipocytes (fatty tissue) is used to guide the analysis of fs-LDI-MS images. Increased pancreatic fatty tissue is associated with deleterious effects on pancreatic islet function.²⁴ Pancreatic islets are endocrine (hormone-producing) cells which are important in glucose metabolism, and so the presence of pancreatic adipocytes may have implications in the onset of type 2 diabetes.²⁴ The multivariate analyses methods of PCA (Ref. 16) and MAP (Refs. 21–23) are used to obtain further information about the tissue sections.

The fs-LDI-MS experiment has been discussed in detail previously^12,13 and is summarized here in Fig. 1. The ∼75 fs, 800 nm pulses from a Ti:sapphire laser are incident on a sample that sits in vacuum on a high precision sample motion stage held at a potential of 3.3 kV which defines the acceleration voltage. The ions that form by fs-LDI are gated into custom ion optics during the 10–100 μs period when a 2.1 kV differential potential is removed from a split extraction cone. The custom steering and focusing ion optics transport the desorbed ions into a reflectron time-of-flight mass analyzer with ∼2000 optimum mass resolution (m/Δm), although lower mass resolution was generally utilized here. At m/z 112, the mass resolution is ∼550. Only positive ions were detected. Two to five laser shots were typically required to obtain an excellent signal-to-noise ratio, and so it took ∼2 min to collect the fs-LDI-MS images over an analysis area of (630 × 360) μm². Figure 1 also shows the VUV beam that was turned off for fs-LDI-MS, but has been used for single photon ionization in separate fs-LDPI-MS experiments not discussed here.¹² 

Anonymized pancreas tissue samples were prepared from the human tissue bank at the University of Chicago and handled in accordance with Institutional Review Board guidelines. Tissue samples were microtomed into ∼10 μm thick sections for mounting on indium tin oxide coated glass slides. Adjacent sections from the same piece of tissue were prepared for complementary multimodal analyses. One sample was H&E stained for lipid content, then imaged using bright field transmission optical microscopy. An adjacent sample was stored in a dry ice cooler and transported from the University of Chicago to the University of Illinois at Chicago for mass spectrometric imaging. All samples were stored at −80 °C when not in transport. Mounted tissue sections used for fs-LDI-MS were submerged for 30 min in a 0.15M lithium chloride/formalin (37%) solution.⁴ Except for spotting a small region of the wet tissue sections with CsI for internal mass calibration, the sections were otherwise untreated prior to the fs-LDI-MS analysis. A total of four samples from two patients were analyzed by bright field optical microscopy and fs-LDI-MS.

Two different multivariate analyses were performed on the data collected. In all cases, analyses were performed on the data in “channel” units (that is, arrival times at the detector), followed by conversion to m/z units. Channel data were binned into blocks that were 50 channels wide before analysis, such that the total mass range analyzed was discretized into approximately 1300 bins; this corresponds to analysis at varying mass spacings, ranging from 0.52 at m/z 100 to 1.16 at m/z 500 to 1.65 at m/z 1000. This is more than sufficient resolution given the relatively broad peaks observed in the sample data, and preliminary calculations at finer resolution produced results in quantitative agreement with those reported below. All calculations were performed using an in-house code developed in python and making use of numpy for linear-algebra operations.

The application of PCA has been discussed in detail elsewhere.¹⁶ The data analyzed here were first scaled according to the prescription of Keenan and Kotula,²⁵ which approximately accounts for the Poisson statistics that apply to the underlying experiment, and then mean-centered. PCA was then performed, after which the scaling was removed (again, following Keenan and Kotula²⁵) prior to the production of principal component (PC) plots for analysis.

The second analysis technique performed is MAP reconstruction using an underlying bilinear model. In MAP, one finds the maximum of the posterior probability distribution of the model parameters $θ$ given the experimental data $E$ and any prior information about sample properties is obtained via Bayes' theorem

The term $P(E|θ)$ is the distribution of experimental data corresponding to the actual sample parameters $θ$ and is called the likelihood, while prior information about sample properties is encoded in the “prior” distribution $P(θ)$⁠. A uniform prior was used in this application such that $P(θ|E)∝P(E|θ)$⁠, and MAP analysis is therefore equivalent to “maximum likelihood” analysis.

The bilinear model used for these calculations assumes that the signal at a given m/z (m) and pixel (p) is a sum of contributions from each component (i) present

where $Cip$ is the concentration of component $i$ at pixel p and $Mi$ is the component mass spectrum.

The likelihood function appropriate to a particular problem depends on the details of the experiment. In the case of normally distributed data, maximizing the likelihood is equivalent to minimizing a $χ2$ statistic, the mean squared difference between the predicted (⁠$I)$ and observed (⁠$E)$ data

Here, p is an index over pixels of the image and m is an index over the mass range covered. The denominator contains the variance, which must be estimated from the data $E$⁠. The combination of the bilinear model with the assumption of normally distributed data and the constraint that the component spectra $Mi$ be non-negative is equivalent to “multivariate curve resolution” (MCR; also seen as multiple component regression), which is most commonly solved using the “alternating least squares” method (ALS-MCR), which is specific to the $χ2$ statistic. Different iterative schemes have been developed for analysis using the $Λ$ statistic.²⁶ In the case of Poisson-distributed data, which is typical for SIMS and related experiments, the quantity to minimize is²⁷ 

Note that for Poisson-distributed data, no estimate of the variance is required.

For a MAP analysis using N components, our code generates an initial guess by performing a PCA (non-mean-centered) and selecting the first N components, then subjecting them to a Promax rotation^28,29 with a power of 3 and ensuring they are non-negative by taking their absolute value. The Poisson-type $Λ$ statistic [Eq. (4)] is then minimized using Lee–Seung iteration until the $Mi$ and $Ci$ individually converge to a tolerance of $10−4$⁠, defined as the relative root-mean-squared deviation of each matrix from one iteration to the next. Note that in reanalyses with different N, this procedure is repeated from the beginning, rather than adding one or more “new” components to the previous analysis and restarting the iteration. These calculations are not expensive at the mass and spatial resolutions used, with MAP analysis at N up to four requiring only a minute or two using a desktop computer equipped with an Intel i7-7820 processor and 64 GB RAM and running the Linux operating system. However, we note that these are low-memory calculations which used less than 2GB RAM.

Figure 2 displays an example of the total ion and optical images of adjacent pancreas tissue sections. H&E staining causes lipids to appear white in the optical images while nuclear and fibrous proteins are dark. The highest ion intensity area of Fig. 2(b) is correlated with the lipid component observed in the optical image of Fig. 2(a). The lateral resolution of the fs-LDI-MS image is estimated to be ∼10 μm, based upon the 63 × 36 pixel size of the image and the 10 μm movement of the ablation laser spot between adjacent pixels. We also note that there are subtle differences in the shapes of the lipid regions in the total ion and optical images which can be attributed to the analysis of adjacent tissue slices from a given pancreas and/or mechanical distortion of the tissue slices before fs-LDI-MS analysis.

To obtain further information about the composition of biological samples, PCA is often performed on mass spectrometric data.^16,17 The results of PCA of the fs-LDI-MS data shown in Fig. 2 are displayed in Figs. 3 and 4. The scree plot¹⁶ (Fig. 3) indicates that the data can be described by 12 or 13 principal components which account for 85% of the variance. Figure 4 displays the scores of the first twelve principal components (PCs). It can be clearly seen that PC1 separates the non-lipid and lipid area of the sample. Higher principal components, in particular PCs 2 to 6, suggest that there are differences in the non-lipid areas which can be attributed to changes in the chemistry and/or topography of the sample. However, plots of the loading versus mass-to-charge ratios (m/z) do not provide clear information about the origin of these differences (see supplementary material, Figs. S1 and S2).³¹ 

In contrast, MAP analysis can distinguish between the effects of different chemical components present and/or sample topography. Using the bright field optical image as a guide [Fig. 2(a)], we first assumed that the pancreas tissue section contained only two chemically distinct areas. When two components are assumed, the data are separated into regions which are correlated with the lipid and non-lipid areas of the sample [Figs. 5 and 6(a)]. In the lipid component (Fig. 6), we observe ions which can be assigned to fragments of lithiated phosphatidylcholine [e.g., m/z 239 and 504 (Ref. 4)] and lithiated fatty acids (e.g., m/z 287). However, no ions were observed from the common headgroup fragments of phosphatidylcholine (e.g., m/z 184), or its lithium, sodium, or potassium adducts. Further, fs-LDI-MS detected no peaks at m/z 750–900 that correspond to the intact parent ions of phosphatidylcholine, sphingomyelin, and/or other lipids previously reported by SIMS and/or MALDI-MS imaging of mammalian pancreas,⁵ brain,^1,2,4 and liver tissue.⁴ Taken together with the relatively low mass resolution, more definitive peak assignments could not be made.

If a larger number of components is employed in the MAP analysis, the non-lipid component is further differentiated (Fig. 5). If three components are used, the non-lipid data splits into a component (Fig. 5: “non-lipid 2”) containing ions such as m/z 7 (Li⁺), 133 (Cs⁺) and non-lipid peaks that are slightly shifted from the “non-lipid 1” component in Fig. 5 (see supplementary material, Figs. S3 and S4). These ions can be attributed to the sample preparation, as well as the topography of the sample leading to slightly different ion flight times (and therefore different m/z).³⁰ This is because there are no other unique peaks in the non-lipid 2 spectrum; any ion that cannot be attributed to sample preparation is observed in the non-lipid 1 component with a small mass difference. For four components, this third component is further separated into a component which is attributed to the topography of the sample (Fig. 5: n = 4, “non-lipid 3”) and one a component associated with the preparation of the sample (Fig. 5: n = 4, non-lipid 2 and supplementary material, Figs. S5 and S6). We note that the topography of the sample may arise due to distortion and compound segregation during freezing, thawing, and dehydration in vacuum. The fourth component (Fig. 5: n = 4, non-lipid 3) contains ions only associated with the sample chemical treatment, e.g., Li⁺ and Cs⁺, and shows that these species are inhomogeneously distributed on the tissue sample. Indeed, a fs-LDI-MS image centered at m/z 7 matches the distribution of this fourth component in the MAP analysis (Fig. 7). The MAP analyses of the other tissue sections also show that the samples can be split into a lipid and non-lipid component (see Fig. S4). Further, the non-lipid component can be split into multiple components; one of the components can clearly be attributed to the sample chemical treatment while the other components are due to the topography of the sample (data not shown).

Femtosecond laser desorption ionization mass spectrometry (fs-LDI-MS) can be employed to obtain mass spectrometric images of tissue samples which can be separated into lipid and non-lipid areas using multivariate analysis. Poisson-weighted PCA resolved the lipid areas from non-lipid areas, but loadings and higher PCs were difficult to interpret. MAP analyses more convincingly suggest that the mass spectra of the lipid areas are composed of fragment ions of phosphatidylcholines and fatty acids. However, definitive peak assignments could not be made because none of the intact parent ions of phosphatidylcholine, sphingomyelin, and/or other lipids were observed. The non-lipid areas can be separated into components that are due to the sample chemical treatment and the topography of the specimen.

A.V.W. and L.D.G. gratefully acknowledge the support from the National Science Foundation (Nos. PHY1027781 and CHE 1709667). L.H. was supported by the National Institute of Biomedical Imaging and Bioengineering under Grant No. 1U01EB019416. G.I.B. and M.H. were supported by the National Institutes of Health Grant No. DK-020595 to the University of Chicago Diabetes Research and Training Center. This material is based in part upon work supported while L.H. was serving at the National Science Foundation.