Matrix-Assisted Laser Desorption/Ionization Imaging Mass Spectrometry

Abstract

Matrix-assisted laser desorption/ionization imaging mass spectrometry (MALDI IMS) is a powerful molecular imaging technology that is capable of imaging a wide range of molecular classes at cellular resolution. This Primer includes an overview of basic imaging mass spectrometry concepts, instrumentation, processing approaches, and advanced applications. We also address key challenges and considerations with an eye on how to optimize instrumentation and methods to overcome issues with spatial resolution, sensitivity, and specificity.

INTRODUCTION

Matrix-assisted laser desorption/ionization imaging mass spectrometry (MALDI IMS) is a cutting-edge imaging technology that allows for molecular mapping and spatially resolved analysis of biological tissues and other complex samples.¹ This unique molecular imaging technology is untargeted, not requiring any antibody markers, probes, or stains. It is highly multiplexed, allowing for the detection and mapping of hundreds to thousands of molecules in a single experiment. Since MALDI imaging uses a mass spectrometer to generate imaging data, it enables highly specific structural characterization and spatial differentiation of molecular species. Imaging mass spectrometry has been used to analyze a wide range of biomolecules, including proteins,^2,3 glycans,^4,5 lipids⁶, small metabolites,⁷ and pharmaceuticals^8,9. The MALDI imaging experiment (Fig. 1) is performed by first mounting thin tissue sections onto a substrate such as a glass slide or metal sample target. A chemical matrix [G] is applied to assist with desorption and ionization during laser irradiation. A measurement region is defined, and mass spectrometric data is collected in a pixel-wise manner. Typically, a single mass spectrum is collected at every pixel location, and the intensity of any ion detected can be represented as a spatial heat map over the sample area to create an ion image.

Fig. 1 |. MALDI Imaging Mass Spectrometry workflow.

(1) Tissues are sectioned and mounted onto a substrate, often an ITO-coated glass slide. (2) A chemical matrix is applied in a way to minimize the delocalization of endogenous molecules. (3) The surface is irradiated with a laser, desorbing and ionizing endogenous molecules from each pixel location, which are then analyzed by mass spectrometry. (4) The resulting data includes mass spectrometric data for each pixel where the intensity of any peak can be represented as a heat map (5) covering the sampled area, producing an ion image. Annotation of detected peaks is often performed by matching ions to theoretical databases or (6) orthogonal LC-MS/MS data based on mass accuracy.

Recent years have seen significant advancements in both instrumentation and computational methods that are driving modern applications of MALDI IMS. State-of-the-art instrumentation is providing higher spatial resolution MALDI sources, improved mass analyzers, and high throughput data acquisition systems. This enables finer spatial resolution and increased sensitivity, allowing for more detailed molecular characterization of tissue. Computational methods have also played a crucial role in improving the capabilities of MALDI IMS. Data pre-processing and analysis algorithms have been developed to enhance data quality, reduce noise, and extract meaningful information from complex datasets that are often distinctly large and high-dimensional. Machine learning techniques are increasingly being employed for dimensionality reduction, clustering, image segmentation, sample classification, and biomarker discovery. Together, these advancements are leading to enhanced reproducibility and performance, allowing MALDI IMS to play a prominent role as a tool for spatial biology and biomedical research.

In this Primer, we explain the overall workflow of the MALDI IMS experiment, common instrument platforms, including specific developments for high spatial resolution imaging, sample preparation, and data collection methods for tissue analysis, as well as typical pre-processing and analysis pipelines applied to these complex data, followed by highlighted advanced applications of the technology. Importantly, for new users of MALDI IMS, potential limitations such as spatial resolution, sensitivity, quantitation, and specificity are also presented, along with approaches and techniques for addressing these challenges. We end by looking to the future, including emerging approaches that build on previous integration efforts to now bring MALDI IMS into alignment with other advanced molecular imaging technologies, such as multiplexed immunofluorescence microscopy and spatial transcriptomics to provide a systems view of tissue microenvironments. With emerging instrumental and computational technologies, MALDI IMS is poised to become a leading bioanalytical tool in both the basic sciences, providing capabilities that are approaching in situ single-cell imaging, and clinical arenas, supporting precision medicine and molecular pathology. For additional details, we refer our readers to the excellent method and standardization-focused reviews cited throughout this Primer.^10–15

EXPERIMENTATION

Instrumentation for MALDI Imaging

MALDI IMS platforms, like all mass spectrometric instrumentation, incorporate (i) an ion source [G] (including the laser and transfer ion optics in some cases), and (ii) the mass analyzer [G] and detector [G] (Fig. 2). While the precise nature of each component varies between different instrument geometries and vendors, the optimization of each is essential to enable direct tissue sampling and maximize spatial resolution, sensitivity, and throughput for imaging applications.

Fig. 2 |. MALDI imaging mass spectrometry instrumentation.

(A) MALDI imaging platforms most commonly consist of three primary components: (i) the ion source consisting of the MALDI source and transfer ion optics, (ii) the mass analyzer and detector, and (iii) the digitizer and computer. (B) MALDI ion sources are found in three geometries. Most imaging platforms use a traditional ‘front-side’ geometry (i). Alternative MALDI sources have been developed utilizing both transmission (ii) and co-axial (iii) geometries. (C) Although any mass analyzer can be used for MALDI imaging, the most common analyzers used on modern, commercial platforms include reflectron and linear time-of-flight analyzers (i), quadrupole (i.e., orthogonal) TOF instruments (ii), and FTMS platforms including both orbitrap (iii) and Fourier transform ion cyclotron resonance (iv) mass analyzers.

MALDI Ion Source

Key components of the ion source include the laser, visualization camera, and sample stage. The vast majority of MALDI IMS sources, including Q-TOF, axial-TOF FTICR, and some Orbitrap instruments, operate under vacuum, typically 10⁰-10⁻⁷ mbar, depending on the precise instrument design. Automated sample docking mechanisms are often used to transfer the sample – typically a tissue section on a microscope slide – into the ion source without venting the mass analyzer. Atmospheric pressure (AP) ion sources¹⁶ are also available and can sustain a more straightforward sample loading mechanism. For many IMS experiments, the laser beam position is held constant whilst the sample stage is moved to acquire mass spectra at each location. Diode-pumped solid-state lasers based on frequency-tripled Nd:YAG lasers emitting at 355 nm with nanosecond pulse durations are widely used for MALDI IMS due to their stability, high beam quality, high repetition rates (nowadays up to 10 kHz for MALDI), and long lifetimes. Other UV laser systems suitable for MALDI include nitrogen lasers (337 nm) and frequency-tripled Nd:YLF lasers (349 nm). Infrared lasers have also been used for MALDI IMS, but are not widely applied or commercially available.^17–19 Laser optics cannot be placed directly above the sample due to interference with the ion path. Instead, a focusing lens (the final optical element before the sample) is positioned some distance away from the sample (typically 10–20 cm), and the laser is introduced at an angle, producing slightly elliptical spot sizes of ~5–50 μm in diameter. Other designs allow the use of shorter focal length lenses (higher numerical aperture) closer to the sample to focus the laser to spot sizes of 1 μm in diameter, enabling higher resolution imaging.²⁰ Such designs include the transmission mode [G], whereby the laser is introduced from the back of the sample^16,21, and co-axial geometries [G] for AP MALDI²².

Although the precise details of the MALDI process are still under debate, several general steps are involved. Irradiation of a matrix-coated sample leads to rapid heating of the matrix and embedded analytes and photo-excitation of the matrix. This, in turn, promotes the desorption and ionization of analytes via a series of charge transfer processes catalyzed by photo-excited matrix molecules/clusters. Alternatively, some ions may be pre-formed in the matrix crystals prior to desorption.²⁶ Careful optimization of the laser pulse energy is needed to give the highest ion yield while minimizing fragmentation and clustering effects.

Following MALDI, ions are extracted and transferred through a series of DC and/or RF voltage ion optics towards the mass analyzer, sometimes involving ion trapping and accumulation stages. Proper optimization of ion optical design and voltages is necessary, as they may influence the observable mass range and m/z-dependent sensitivity. Furthermore, as ion optics can control the kinetic energy of ions entering the mass analyzer, they also play a key role in acquiring spectra with the highest possible mass resolving power [G] and accuracy [G].

Mass Analyzer

The final stage of a MALDI IMS instrument is the mass analysis of the generated ions. The most common mass analyzers used for MALDI IMS include orthogonal and axial time-of-flight, Fourier transform-based orbitrap, and Fourier transform ion cyclotron resonance (FTICR). Each has its strengths and limitations in terms of speed, mass resolving power, mass accuracy, and sensitivity. Mass resolving power is defined as $\frac{m}{\Delta m}$ where m is the m/z of a given peak and Δm is the peak width.²⁷ This metric also relates to the ability to discern between two peaks with very similar m/z values, a property often referred to as mass resolution, and it is an important metric in IMS experiments. Resolving peaks with similar m/z values is essential for ensuring that generated images are not reporting a mixture of unresolved ion species signals, al be there an upper bound to the resolving power and mass resolution that can be achieved in a practical setup. Both orthogonal and axial TOF analyzers have a mass resolving power largely independent of m/z value. Also, because of the high speed of TOF scans (microseconds), the resolving power of these analyzers also has marginal effects on the time spent per pixel. In contrast, the mass resolving power of Orbitrap and FTICR analyzers is proportional to the analyzer scan time and decreases with increasing m/z due to the inverse relationship between m/z and frequency. Typical scan times for Orbitrap and FTICR analyzers for MALDI imaging experiments lie between 0.5–4 scans per second, which is also typically the rate limiting step for the time spent per pixel. This also depends on the required mass resolving power with typical values of 50,000–400,000 (at m/z 400). As such, IMS experiments may take many hours, and a compromise may be required to balance mass resolving power, spatial resolution, and acquisition time for FT-based systems. TOF systems can acquire spectra faster with acquisition rates of up to 50 pixels/second on axial TOF systems^23,28,29 and 20 pixels/second on orthogonal TOF systems³⁰ whilst maintaining resolving powers of ~20,000–50,000 (when using ion reflectrons).

Sample Preparation

MALDI IMS is most often performed on fresh frozen tissues, as they have simplified preparation requirements while preserving a wide range of molecular classes. Tissues removed from animals or patients are flash-frozen in a cryogen, for example, by placing the tissues on a metal weigh boat on cryogenic liquid such as liquid nitrogen or a slurry of isopentane and dry ice.³¹ It is critical to process tissues as quickly as possible to minimize warm ischemia times, reduce the degradation of endogenous molecules, and to ensure that reproducible data can be obtained.^32,33 This is especially important for small metabolites such as adenosine triphosphate that degrade in very short time scales.^34,35 Various preservation methods are available to help retain the original molecular content of certain frozen tissues, including focused microwave irradiation³⁶, funnel freezing³⁷, and heat stabilization^38,39. Although typically used for paraffin-embedded tissues, formalin fixation can also be applied to fresh frozen tissues (see below). It may lead to loss of signal for amine-containing molecules such as various phospholipids.^40,41 For more consistent tissue sectioning, mass spectrometry-compatible embedding media such as carboxymethyl cellulose⁴² or gelatin⁴³ can help preserve tissue structure during sectioning. Optimal Cutting Temperature (OCT) compound⁴⁴ can also be used, but runs the risk of contaminating the tissue with polymer from the OCT.

Thin sections of frozen tissue are prepared using a cryomicrotome with the precise temperature and thickness depending on the tissue; typically between −10°C to −25°C and a thickness of 5–20 μm. For OCT-embedded tissues, it is advisable to cut away as much of the OCT as possible prior to sectioning and to clean the blade between sections to minimize the risk of smearing polymer on the sample. Tissue sections are then thaw-mounted onto glass slides, often coated with indium tin oxide (ITO). ITO-coated glass slides, or other conductive substrates, are required for high voltage ion sources (e.g., axial TOF platforms) and preferred for most MALDI systems to maximize sensitivity.

In some cases, the tissues undergo a washing step prior to matrix application. Washing is most commonly used for imaging of intact proteins or before on-tissue enzymatic digestion to remove interfering salts and small molecules (e.g., lipids) from the tissue.⁴⁵ The exact washing solvent(s) is/are optimized for the tissue and target analytes, but typically involve organic solvent-containing solutions.⁴⁵ In some cases tissue washing using aqueous solutions, such as ammonium acetate, ammonium formate, and alkali salts, can be used to either remove endogenous salts, leading to the preferential formation of protonated ions⁴⁶ or for forming favourable alkali adducts for lipids such as triacylglycerols with a high affinity for alkali cations, but poor proton affinity.

Formalin-fixed paraffin-embedded (FFPE) tissues are increasingly being used for MALDI IMS due to the vast array of biobanked tissue samples stored this way.^47–50 FFPE is the gold standard for storing clinical tissues, as it allows long-term storage at room temperature. To prepare for MALDI IMS measurement, the tissue must first be deparaffinized (wax removal) and undergo an antigen retrieval step where it is heated in a suitable buffer solution such as citric acid, Tris or Tris-EDTA that renders molecules such as proteins more accessible for subsequent enzymatic digestion. Due to both the fixation and paraffin embedding, many small molecules (e.g., metabolites and lipids) are lost from the tissue during the fixation and washing steps. MALDI IMS analysis of FFPE tissue thus often focuses on either tryptic peptides or N-glycans released from proteins following enzymatic digestion using trypsin or PNGase F, respectively, although other enzymes can also be used depending on the application and the target analyte(s).^51–53 The on-tissue digestion process involves spraying an enzyme over the tissue and incubating it, typically for at least several hours, to release the desired analytes.

In some cases, the digestion step can be skipped, and analytes such as small metabolites⁵⁴ or endogenous neuropeptides⁵⁵ can be detected directly from FFPE tissue. Due to the additional sample preparation steps, greater variability between experiments can sometimes be observed when analyzing FFPE tissues. The use of quality controls to monitor key parameters such as digestion efficiency is an important consideration for experiment design.⁴⁰

Matrix Application

The next step in the sample preparation workflow is coating the sample with a suitable MALDI matrix. While numerous MALDI matrices have been tried and tested, there are common traits that make a good matrix, including (i) strong absorption at the wavelength of the MALDI laser; (ii) an ability to ionize the target analytes following laser irradiation of analyte-matrix co-crystals; (iii) high solubility in the solvent system used to deposit the matrix; (iv) sufficiently low vapor pressure to minimize sublimation in vacuum ion sources; and (v) ideally forming small homogenous crystals. 2,5-dihydroxybenzoic acid (DHB) and α-cyano-4-hydroxycinnamic acid (CHCA) are the two most common matrices, as they are suitable for many analyte classes. It is important to note that different matrices may favor the ionization of different compounds or work better in either positive ion mode [G] or negative ion mode [G]. For this reason, matrix optimization is a key step when using MALDI IMS for a new analyte. Figure 3 summarises common matrices and the analyte classes they are well suited for.

Fig. 3 |. Common matrices used for MALDI IMS.

The compatibility of each matrix with various molecular classes is indicated, and the color of the check mark denotes the relative sensitivity for each.

Matrix application must be performed in a manner that minimizes analyte delocalization whilst extracting molecules from the tissue to form analyte-matrix co-crystals. The most common and versatile approach for matrix deposition is a robotic sprayer that deposits an aerosol of matrix solution over the tissue.⁵⁶ Contemporary sprayers offer control of various parameters such as spray height, flow rate, gas pressure, and even temperature to allow the user to finely optimize the deposition (the same sprayers are also widely used for enzyme deposition for on-tissue digestion). The matrix is usually dissolved in a mixture of organic solvent (e.g., MeOH, EtOH or ACN) and water, sometimes with 0.1–1% of acid to provide an additional source of protons and aid in the solubilization of some compounds. It is common to optimize parameters by evaluating different solvents, matrices, and deposition conditions for a new sample and/or target analyte. If the spray is too wet, then analyte delocalization can occur, while if it is too dry, some molecules may not be extracted efficiently from the tissue. Another critical parameter is the size of the crystals - usually investigated using scanning electron microscopy. A general rule of thumb is that the finest pixel size achievable can be no smaller than the size of the largest crystals. Crystal size depends heavily on the type of matrix, the substrate, and the deposition conditions. Nowadays, spray-based matrix depositions are suitable for pixel sizes of 20 μm, and in some cases lower, is possible.^16,21,57

Sublimation is also commonly used for matrix deposition and is best suited when high spatial resolution is required.⁵⁸ Sublimation is a solvent-free method in which matrix powder is heated inside a vacuum chamber (typical temperatures are 120–200°C depending on the matrix). As the matrix is heated, it sublimates and condenses onto the cooled sample positioned above the matrix. Sublimation produces crystal sizes in the low μm to nm range, and tends to result in lower analyte delocalization than spraying methods. It can be used to apply a variety of matrices, including DHB, DAN, 2,5-DHA and norharmane, and works particularly well for lipids. However, sublimation alone is less effective for other molecular classes, such as proteins, tryptic peptides, N-glycans, and many small metabolites, presumably due to the lack of solvent-mediated extraction of analytes from the tissue into the matrix layer. Recrystallization steps (incubation of matrix-coated samples with solvent vapor) can help increase the analyte classes detected from matrix-coated samples, but may introduce delocalization.^59,60 Combining sublimation and recrystallization can also improve the sensitivity of lipid detection compared to sublimation alone while preserving analyte distribution on at least the low micrometer scale.

RESULTS

Data Structure and Visualization

MALDI IMS measures mass spectra at discrete coordinates (i.e., pixels) across the sampled area, one mass spectrum for each pixel.^61,62 The peaks in a pixel’s mass spectrum report the ions detected at that location, whereby a peak’s position along the independent (horizontal) axis relates the mass-to-charge ratio (m/z) of the detected ion species, and the height or integrated intensity of the peak along the dependent (vertical) axis corresponds to the abundance of the detected ion species. A typical MALDI IMS experiment contains thousands to millions of pixels (i.e., mass spectra), with each spectrum associated to a specific spatial location within the sample from which the spectrum is acquired. A MALDI IMS dataset essentially consists of millions of ion intensity values, each value tied to a particular m/z-coordinate and specific spatial coordinates (e.g., (x,y) in 2-D IMS experiments). The measurements can be represented as a three-dimensional data array, also referred to as a data cube or tensor, where the first two dimensions or modes of the tensor correspond to the spatial x and y coordinates, and the third mode corresponds to the m/z values, and each element reports the intensity recorded for a unique (x,y,m/z)-combination. Different types of visualizations can be gleaned from this data tensor by ‘slicing’ it along different axes. For example, extracting the intensity values at a particular m/z-index regardless of spatial coordinates yields an array that reports the spatial distribution of a particular ion species, (i.e., ion image) where false color scales tend to be used to report ion intensities as different color hues.^10,63,64 On the other hand, retrieving intensity values that match a specific (x,y)-index results in a vector that reports the mass spectrum at that location in the sample. The data can also be represented as a two-dimensional data array, i.e., a matrix or table, where rows correspond to spatial positions (x, y) and columns correspond to m/z values, with each element reporting the intensity value measured for a particular m/z-value or ion species at a specific spatial location. This representation lends itself for straightforward data analysis as many signal processing and machine learning approaches assume such an observations-by-features matrix as input.

Pre-processing and Removal of Technical Variation

Modern MALDI IMS platforms can generate massive amounts of data. Images usually have thousands to millions of pixels (i.e., raw spectra), and each spectrum has tens-of-thousands to millions of ion intensity data points (i.e., an intensity value per m/z-bin) reporting hundreds to thousands of peaks, depending on the sample and mass analyzer. In-depth analysis of such large quantities of complex mass spectrometric measurements, sometimes even across sample cohorts, requires rigorous pre-processing to minimize measurement variability, to attenuate potential noise sources such as electronic noise, chemical noise, instrument drift, and inconsistencies in sample preparation, and to ensure comparability of ion intensity values between pixels and samples. By removing these undesired (non-biological) variations from the measurements prior to analysis, data pre-processing maximizes image quality, increases the signal-to-noise-ratio of the measurements and the detectability of low-abundant species, and better captures genuine biological changes in molecular abundance between pixels and samples. Since each pre-processing method tends to focus on a particular type of undesirable variation or noise, the pre-processing phase of mass spectrometry signals can entail multiple methods and can be optimized for a particular experimental design, sample, or instrument. While additional steps can be taken, key pre-processing steps for MALDI IMS tend to include m/z peak alignment and calibration, intensity normalization, and peak picking (Fig. 4).

Fig. 4 |. Common strategies for pre-processing and analyzing MALDI IMS data.

(A) Data pre-processing aims to remove non-biological variation and improve the comparability of the data. Most workflows involve peak alignment (i) and calibration (ii), intensity normalization (iii), and peak detection and picking (iv). Highly complex IMS data are often analyzed using machine learning (ML). (B) Classification is a supervised ML approach that is routinely utilized on MALDI IMS data to perform a specific recognition task on the basis of the chemical content reported by the mass spectra. It entails providing labeled pixels (i.e., mass spectra) that serve as examples of tissue features or classes one wants to recognize (e.g., regions of normal and diseased tissue) in order to allow a computational model to be trained (i). Once a sufficiently performant model is trained, it can be used to predict labels for new pixels or locations and thus (predictively) annotate images automatically (ii). (C) Unsupervised machine learning algorithms are also commonly applied to MALDI IMS data, often in more exploratory scenarios such as aiding in human interpretation of very large complex datasets, to denoise, to reduce the dimensionality of the dataset prior to subsequent supervised ML analysis, or to reduce the data size footprint of MALDI IMS data. Common classes of unsupervised approaches applied to MALDI IMS include factorization (i), clustering (ii), and manifold learning (iii) methods.

m/z Alignment and Calibration

While both methods seek to project a set of mass spectra to a common m/z axis, m/z peak alignment and calibration are distinct pre-processing steps. If several pixels (i.e., mass spectra) are measured, instrumental drift and sample variability can introduce slight pixel and sample-specific shifts along the m/z axis. To avoid that such discrepancies propagate into subsequent analyses (e.g., influencing ion image generation or peak-picking), re-aligning empirically measured spectra onto a common and comparable m/z axis is often needed and improves the consistency of peak positions and accuracy over the course of an imaging experiment. Alignment is aimed at casting different spectra to a consensus m/z axis, increasing the comparability and internal consistency of a MALDI IMS dataset. This consensus m/z axis is not necessarily equal to the theoretical reference m/z axis, but this means that alignment can be applied even without external references or information on theoretical masses. Alignment tends to correct m/z shifts from one pixel to another and between tissue samples by either projecting spectra to a consensus spectrum or aligning them using a recurring set of m/z features (i.e., landmarks). Various algorithms that have been applied to IMS data, for example, iterative peak alignment, correlation-based methods, wavelet-based alignment, among others. Calibration, on the other hand, involves relating experimentally determined m/z values detected by the mass spectrometer to known reference (theoretical) m/z values. If calibration is applied in combination with alignment, the resulting consensus m/z axis is effectively made to also line up with the theoretical m/z axis. Calibration can be performed externally, by analyzing known standards separate from the imaging experiment to calculate a calibration curve relating detected m/z values to theoretical m/z values, or internally, by performing the same process using reference ions present in the imaged sample itself. The link with theoretical masses offered by calibration makes it a critical step for accurately annotating peaks with molecular identifications based on molecular databases or orthogonal fragmentation (i.e., MS/MS) experiments (See Limitations and Optimizations for more on this topic).

Intensity Normalization

Normalization methods are used to project a set of mass spectra to a common ion intensity axis, adjusting the intensity values of individual mass spectra while preserving the relative intensities of detected molecular ions. In most cases, normalization improves pixel-to-pixel comparability of intensity values substantially, both within and between images, by mitigating intensity variations introduced by e.g. instrumental variation, electronic noise, sample heterogeneity, and chemical matrix effects (e.g., “matrix hot spots”). The choice of normalization method depends on the specific characteristics of the IMS data and the research objectives. The most common methods for MALDI IMS include (i) Total Ion Count (or Total Ion Current; TIC) normalization; (ii) Root Mean Square (RMS) normalization; and (iii) Internal Standard normalization.

Peak Picking

Automated peak picking involves the recognition and extraction of peak-shaped signals along the measured mass spectral profile, with those peaks presumably corresponding to molecular ion species. This pre-processing step allows molecular signals to be separated from spectral background and noise variation, typically retaining only their peak m/z centroid value and intensity, and enables the data footprint to be reduced substantially, minimizing data storage requirements and computational resources needed for analysis. Commonly used peak detection algorithms for MALDI IMS data analysis include thresholding, wavelet transform, and Gaussian fitting-based approaches.

MALDI IMS Analysis

The highly multiplexed nature of MALDI IMS data has made computational analysis an integral part of most studies. While one could approach IMS data as a collection of individual ion images and apply a traditional univariate analysis to each ion species separately, this tends to ignore the (often biologically relevant) interconnections between molecular species and leaves the multivariate potential of IMS severely underutilized. Also, in a high-dimensional setting such as IMS data, a large number of univariate analyses could invite in statistical difficulties such as spurious correlations and the multiple testing problem.⁶⁵ Therefore, most IMS studies utilize multivariate statistical and computational approaches to explore and analyze the biological relationships encoded in IMS measurements. While multivariate methods bring their own challenges (e.g., the “curse of dimensionality”⁶⁶), recent years have shown extensive growth in signal processing and machine learning-based methods to extract meaningful insights from complex IMS datasets.

The specifics of a MALDI IMS analysis depend strongly on the particular question one wants to address or the practical goal one has in mind. While an exhaustive treatment of experimental designs and corresponding computational methods would exceed the scope of this paper, we will discuss two sub-areas, unsupervised and supervised machine learning, that are commonly applied to MALDI IMS data.

Unsupervised Machine Learning

After the pre-processing phase, unsupervised analysis methods are often applied to IMS spectra with an eye towards exploration, with the goal of revealing latent structures, patterns, or relationships that might be underlying the spectral measurements. Such structures can report biological as well as sample preparation or instrument-induced variations in the data and they can include dataset-wide trends as well as more localized correlations or associations, both along the spatial and spectral domains. These methods are particularly useful when the researcher has no particular spatial areas or ion species of interest in mind. The latent signatures that are found can be used to aid in human interpretation and discovery, summarizing, e.g., which ion species correlate in abundance and where in the tissue this happens. However, they can also be used as a means of reducing the dimensionality of an IMS dataset without losing much information or for denoising. Dimensionality reduction has both practical applications (e.g., reducing resource usage and memory requirements in subsequent computational analyses) as well as analytical worth (e.g., reducing collinearity prior to building a classification model, reducing susceptibility to “curses of dimensionality”).¹⁰

Clustering approaches group together pixels with similar mass spectral profiles, effectively defining clusters of sample locations that exhibit similar chemical signatures.^67,68 In IMS use cases, a cluster is usually assigned a distinct color and pixels are colored according to their cluster membership, providing a single false color image that segments the sampled area into subareas of homogeneous chemical content and providing a low-dimensional view into the high-dimensional molecular content of an IMS dataset. Clustering results depend on the pre-determined definition of what “similar” means. This is usually accomplished by a user-chosen distance or (dis)similarity metric that rates the similarity between measurements as a number.⁶⁹ Since each metric has its own sensitivities, this choice will often need to be made in a manner congruent with the application at hand. Furthermore, while most IMS studies tend to cluster spectral signatures, yielding a spatial grouping or segmentation, there are also IMS examples where ion images with similar spatial expressions are clustered and thus a grouping of ion species or spectral segmentation is the output.⁷⁰

Factorization methods seek to decompose an IMS dataset into a limited number of trends such that each recorded spectrum can be represented and approximated as a linear combination of those trends. Each trend or component consists of a spectral signature, reporting which ion species correlate or associate with each other, and a spatial signature, reporting where in the tissue sample that spectral association takes place. Unlike clustering methods where groupings are decided by taking a full mass spectrum into account, factorization-based components of IMS data have the ability to capture certain spectral correlations in one tissue area and describe another ion species correlation in another area, often providing a more nuanced view into the IMS dataset. Since the goal is to represent the dataset using a number of components that is smaller than the original number of measurements, the lower-dimensional representation offered by these methods tends to reveal spatial and molecular signals that tend to colocalize and correlate under the assumption of linear mixing. Principal component analysis (PCA)^71,72 and nonnegative matrix factorization (NMF)⁷³ are the most prominent examples of these methods for IMS analysis.

Manifold learning techniques are specifically focused on casting the high-dimensional spectra/pixels of an IMS dataset into a lower-dimensional representation such that chemical content similarity is encoded as distance in this latent space. These low-dimensional representations tend to aid in discerning which pixels are chemically similar, and are often used as preludes to more classical clustering approaches. While factorization methods also supply a low-dimensional representation, they usually assume linear mixing of latent components and are thus not necessarily well-suited to capture nonlinear relationships in IMS data. Manifold learning techniques, however, facilitate the learning of nonlinear latent representations of IMS data, effectively enabling stronger reduction of dimensionality, but this usually comes at the cost of diminished interpretability of the found latent structures. Manifold learning also requires decisions regarding the employed distance measure and other parameters, leading to application-specific considerations as in clustering. Manifold learning approaches used on IMS data include t-Distributed Stochastic Neighborhood Embedding (t-SNE)^74,75, autoencoders⁷⁶, and uniform manifold approximation and projection (UMAP)^77,78.

Supervised Machine Learning

Besides assumptions on the nature of their latent structure, unsupervised analyses tend to be largely untargeted and will try to capture the majority of variation present in an IMS dataset. However, IMS experiments are often part of studies where the relevant ion species are not necessarily known, but where the biological or medical phenomenon of interest is clearly and narrowly defined. In these cases, the goal is to have the IMS measurements inform on a specific recognition task (e.g., predicting tumor tissue areas, discovering biomarker candidates for functional tissue units) or on a non-IMS variable (e.g., predicting drug abundance, predicting another modality’s observation). Building a computational model that ties IMS observations to the non-IMS-observation of interest, while learning to ignore IMS variation that seems unrelated, is common practice in these more targeted scenarios and supervised machine learning approaches play a central role in establishing such models. Two major supervised datamining approaches that are regularly applied to MALDI IMS data are discussed below.

Classification methods seek to build computational models that relate IMS observations to a categorical variable or label, such as “healthy”, “disease”, “cortex”, “medulla”, “beta cells”, “alpha cells”, etc. These supervised MALDI IMS analyses often entail a two-step process. First, IMS-measured spectra and accompanying labels for those spectra are employed as training (and test) data to train a model such that it can predict the label variable when presented with an IMS measurement as input. Then, if the model has demonstrated sufficient performance, it can be applied to unlabeled or new IMS measurements to predict their corresponding labels. Classification approaches are commonly applied to MALDI IMS in a digital pathology context to predict tissue classes and disease labels after having been shown representative example measurements annotated by a pathologist. Recently, developments in interpretable machine learning have also enabled a different use case for classification models in IMS data analysis, namely the automated discovery of biomarker candidates.⁶⁴ There, the goal is not to use the trained model to predict new measurements but rather to use a model interpretability method such as Shapley additive explanations⁷⁹ to analyze the decision process of the model. If one can establish which ion species are most relevant for recognizing a particular tissue structure and differentiating it from other tissue areas, these estimates can be used to re-order the list of often hundreds to thousands of IMS-reported ion species in order of relevancy to a labeled area of interest. This effectively yields, at the top of that list, a panel of ion species with increased biomarker potential for the tissue area of interest. Both use cases have been demonstrated in unimodal as well as multimodal studies, where in the latter case, e.g., a model is built between IMS-supplied intensity values and microscopy-supplied functional tissue unit masks.⁸⁰

Regression methods also aim to build a computational model that relates IMS observations to an external variable, albeit a numerical rather than a categorical variable. Like classification models, regression models are also trained and subsequently used for prediction in a two-step scenario. In MALDI IMS studies focused on absolute quantitation (e.g., organ-specific drug quantitation estimates), regression models are commonly trained and used to tie molecular abundance estimates to IMS observations (here, IMS is the independent variable). More advanced multimodal use cases of regression can also be found in data-driven image fusion⁶³ and automated anatomical interpretation of ion distributions (where IMS is the dependent variable)⁸¹. For example, in data-driven image fusion, partial least squares regression is used to learn a cross-modality model that ties microscopy observations to IMS observations. The fusion model is subsequently utilized in sharpening scenarios, using microscopy measurements to predict ion images to a finer pixel size than they were originally measured at, and in out-of-sample prediction scenarios, predicting ion distributions in tissue areas where only microscopy was collected and no IMS measurements were acquired.

APPLICATIONS

MALDI IMS has found widespread application due to its capability to perform highly multiplexed measurements and generate spatial distribution maps of a diverse set of molecular classes in an anatomical context without the need for labeling or a priori knowledge of the tissue sample. Here, we present several major molecular classes that MALDI imaging has been used to probe and map within various biological tissues. The examples range from exogenous compounds, such as pharmaceutical drugs and medications, and small molecules, such as metabolites and lipids, to larger biopolymers such as proteolytic and endogenous peptides, proteins, and released glycans. This discussion is not meant to be an exhaustive review of the field, which has been detailed previously by others in the mass spectrometry community^14,15,82; rather it is meant to survey the breadth of experiments that can be undertaken using MALDI imaging mass spectrometry. A few specific case studies are highlighted to illustrate the scope of applications explored by the field.

Exogenous Compounds

The detection of exogenous compounds by MALDI IMS represents a diverse set of applications covering a swath of scientific sectors including forensics, the cosmetic industry, agriculture and plant biology, biochemistry and cellular biology, and the pharmaceutical industry. Equally diverse are the types of compounds analyzed, such as illicit drugs, ingredients from personal care products, pesticides, and pharmaceutical drugs, and medications. MALDI imaging has seen increased use in the forensic evaluation of latent fingerprints due to its ability to provide investigators with information beyond simple views of the ridge details within the print.^83–85 The molecular dimension afforded by mass spectrometry enables ridge detail images to be produced using chemical signals from exogenous compounds in the fingerprint in the same analysis. These compounds can reveal information about the personal lifestyle of an individual, providing additional evidentiary value in forensic investigations.⁸⁶ Examples of chemical compounds detected during fingerprint analyses include the active ingredients in bug sprays and sunscreens,⁸⁶ petroleum-derived polymers in cosmetics and creams,^86,87 triacylglycerols from cooking oils,⁸⁶ organic acids from alcoholic beverages,⁸⁶ sugars in citrus fruits,⁸⁶ nicotine from cigarettes,⁸⁸ pharmaceutical compounds,⁸⁹ peroxides from explosives,^89,90 and illicit drugs such as cocaine, heroin, and amphetamines^91–93.

Recent developments have enabled the spatial analysis of agricultural chemicals for various toxicological applications.⁹⁴ MALDI analysis has been performed on a variety of sample types, including plant material such as leaves and roots,^94–100 plant seeds,⁹⁴ insects,¹⁰¹ fish,⁹⁴ and paper imprints of fruits and vegetables.¹⁰² Monitoring the uptake, translocation, and degradation of pesticides, herbicides, and insecticides in these ecological systems allows for valuable insight into agricultural engineering and evaluating environmental pollutants.

A major area of development over the past two decades involves analyzing the distributions of pharmaceutical drugs in biological tissues. MALDI imaging mass spectrometry has provided valuable insight in establishing pharmacokinetic-pharmacodynamic relationships in the early phases of drug discovery.^103,104 These analyses have traditionally been performed on rodent animal models but have also been employed on organoids and three-dimensional cell cultures (Fig. 5) and can include small molecules as well as large therapeutic antibody drugs.^105–110 Biodistribution analysis can also involve quantitative measurements, which requires the generation of tissue-based calibration curves and the careful application of appropriate internal standards to ensure quantification accuracy, precision, and stability.^111–115 In addition to small molecules, MALDI IMS can be used to study novel therapeutics such as nanoparticle- and polymer-based drug delivery systems.¹⁰³ Understanding the biodistribution of drug candidates and their metabolites plays an important role in drug discovery and development.^{109,116–118} Increasingly, many pharmaceutical companies have devoted resources to developing in-house imaging mass spectrometry groups staffed with experts in MALDI and mass spectrometry technologies.

Fig. 5 |. Pharmaceuticals imaging by MALDI IMS.

(a) MALDI IMS ion-intensity maps, (b) summed mass spectra, and (c) intensity box plots of HT-29 spheroids treated with 1 mg/mL cetuximab for 24 or 72 h. (d,e) Immunofluorescence (IF) study of cetuximab localization. Data were normalized to DAPI intensities. Statistical significance was tested using Student’s t-test (n = 6, *p < 0.05). (f) IF analysis of EGFR expression in HT-29 control spheroids. Reproduced from Liu, et al.⁹⁸

Attractive features of MALDI IMS in the detection of exogenous compounds include the label-free nature of the detection and the facile discrimination of metabolized products from the parent compound.^{119–121 106,110} By comparison, QWBA would likely not have been able to differentiate the parent drug from the product metabolites.¹²² The main challenges associated with MALDI IMS analyses surround the limits of detection. Issues of low abundance and poor ionization efficiency can compromise the ability to detect targeted xenobiotics in complex chemical environments in situ. Sample preparation methods, such as alternative matrices and chemical derivatization (described above), have been developed to help combat these issues.^123–126

Metabolites

The metabolome is increasingly recognized as a more dynamic measure of molecular phenotype when compared to genomics and proteomics strategies. MALDI imaging mass spectrometry of metabolites has been applied in pharmaceutical, clinical, and biochemical settings.^127–129 In the pharmaceutical space, metabolites of both exogenous (vide supra) and endogenous compounds are of interest. In addition to drug metabolites, endogenous metabolites can be used to assess the pharmacological effect of drugs, allowing for insight into the mechanisms of action and characterization of toxicological side effects.

The analysis of tissue biopsies, cores, microarrays, and sections in clinical settings is often used to aid in the diagnosis of pathological conditions. Metabolite analyses can be used to decipher molecular mechanisms of disease pathophysiology and serve as biomarkers in disease classification studies.^130–133 MALDI IMS in the clinical arena is most commonly used in cancer studies, such as in tumor classification and tumor margin analyses.^134,135 A wide variety of cancers have been studied, including breast, prostate, skin, brain, and gastrointestinal.¹³⁶ Similarly, a wide variety of analytes have been studied, including metabolites involved in carnitine and fatty acid metabolism, beta-oxidation, the Krebs cycle, and zinc-related metabolism.^137–140 Segmentation and clustering approaches can be used to build classification models using supervised and unsupervised analysis to detect, identify, and stage cancer progression.^{130,131,141–143} Accurate identification of the metabolite features used to build these classification models can allow for pathway analyses and biochemical insight into disease mechanisms.¹⁴⁴ For example, MALDI IMS of cancer progression often involves assessing classical metabolic signatures of the Warburg effect, energy metabolism, and cell motility.¹³⁸ Cao and coworkers have used stable isotope labeling of amino acids phenylalanine and tyrosine to monitor metabolic substrate usage to study metabolic reprogramming in tumorigenesis and tumor progression (Fig. 6).¹⁴⁵ By monitoring the incorporation of the stable isotopes into energy-related metabolic pathways, metabolic heterogeneity, proliferation, and therapeutic response and resistance can be tracked in tissue.¹⁴⁶

Fig. 6 |. MALDI metabolite imaging.

Distributions of ¹³C₆-Phe (a) and ¹³C₆-Tyr (c) in mouse-xenograft lung tumor tissues (n_total = 11) at 10, 30, and 60 min after tracer injection and at 10 min of control group (n = 3). Tracer-to-tracee (TTR) images for phenylalanine (a) and tyrosine (c) were calculated by normalizing the labeled amino acid signals to the intensities of their respective unlabeled versions. TTR values of Phe (b) and Tyr (d) were then differentiated annotated viable tumor and non-viable tumor regions for every time point. Reproduced from Cao, et al.¹³³

MALDI IMS has also been used to perform in situ metabolic analyses of non-mammalian samples, including plants and bacterial colonies.^94,147–151 Plant metabolic workflows have evolved to analyze many different plant tissues, including seed capsules,¹⁵² skin,¹⁵³ vegetable pods,¹⁵⁴ leaves,¹⁵⁵ bulbs,¹⁵⁶ berries,¹⁵⁷ embryos,¹⁵⁸ stems,¹⁵⁹ and roots.^160,161 A wide variety of plant species have been studied, including those from the genera Arabidopsis,¹⁶² Solanum,¹⁶⁰ Oryza,¹⁶³ Musa,^164,165 Citrus,¹⁶⁵ Hordeum,¹⁶⁶ Zea,¹⁵⁵ and Eucalyptus.¹⁵⁹ Metabolites such as amino acids,¹⁶⁷ organic acids,¹⁶⁸ glycosides,¹⁶⁹ alkaloids,¹⁷⁰ flavonoids,¹⁷¹ lignin, and anthocyanins¹⁷² have been analyzed to monitor a range of properties, including plant defense against microbial and fungal infections as well as pollinator attraction. A detailed understanding of plant metabolism is useful in the pursuit of metabolic engineering strategies to generate sustainable and safe agricultural products. MALDI IMS is also well suited to analyze microbial systems, which typically form surface-attached communities, such as in biofilms and motile colonies.¹⁵⁰ Bacteria such as S. aureus,^173,174 P. aeruginosa,¹⁷⁵ B. subtilis,¹⁷⁶ E. faecalis, and C. difficile¹⁷⁷ have been widely studied in biofilms, in tissue, and in colonies grown on agar.¹⁷⁸ Bacteria have a profound impact on nearly every aspect of life on the planet, influencing growth, vitality, and disease across biological kingdoms. Quinolones, bile acids, amino acids, and surfactins have been mapped in bacterial samples to examine pathogenicity, growth, metabolism, antibiotic inhibition and response, as well as microbial cooperation and competition.^174,179,180 Major areas of microbiology research are currently focused on MALDI imaging mass spectrometry to accurately identify bacterial strains,^181–183 elucidate the mechanisms of antibiotic resistance, develop new therapeutic strategies to combat infection, understand microbe-microbe interactions, and decipher host immune responses to infection.

As with all metabolomics experiments, arresting metabolism in the tissue in order to enable the accurate measurement of metabolite identities and abundances can be difficult. Proper tissue handling, stabilization, storage, and fixation techniques have been developed to minimize analyte degradation in MALDI imaging experiments.^184–195 As discussed above, many tissue repositories contain FFPE tissue samples, posing a challenge for metabolite (and lipid) MALDI IMS analysis.^{151,193–199} Even with sample preparation that ensures spatial fidelity, metabolite detection efficiency can be challenging for a number of other reasons, including low ionization efficiency, poor ion transmission in the low mass range, and spectral overlap with ion signal from the MALDI matrix.^{184,185,190,200} On-tissue derivatization strategies have been developed to improve ionization and transmission efficiency.^{123,127,201,202} Appropriate selection of the MALDI IMS instrument platform can also alleviate some of these detection challenges.^203–205 Ion mobility-mass spectrometry (IM-MS) and tandem mass spectrometry approaches can enable the separation and identification of isobaric and isomeric ions.

Lipids

Lipids represent a structurally diverse class of biomolecules that play essential roles in cell signaling, energy homeostasis, and cell membrane structures.²⁰⁶ Several classes of lipids, such as glycerophospholipids (GP) and fatty acids (FA), are widely studied via MALDI IMS due to their high abundance in biological tissues and high ionization efficiency.²⁰⁷ Lipids that contain fixed charge quaternary ammonium functionalities, such as phosphatidylcholines (PC) and sphingomyelins (SM), are abundant in many tissues and are readily ionized in positive ion mode. Lipids with highly acidic, polar headgroups, such as phosphatidylethanolamines (PE), phosphatidylserines (PS), and phosphatidylinositols (PI), are readily ionized in negative ion mode. Cardiolipins (CL), ceramides (Cer), and sulfatides are also readily observed as anions in MALDI imaging mass spectrometry experiments. In general, the presence of either phosphate anions or nitrogen-centered cations in the lipid structure results in high ionization efficiency.²⁰⁸ Other classes of lipids may be more difficult to detect. For example, neutral lipids such as triacylglycerols (TAG) and sterols have lower ionization efficiencies, requiring alternative sample preparation strategies such as cationization, tissue washing, and alternative MALDI matrices to overcome ionization suppression.^209–212 Lipids involved in cell signaling pathways, such as polyphosphoinositols, tend to be present in lower abundance and are thus more difficult to detect. Overall, the molecular weights of many ionized lipids fall between m/z 200 and 2,000, which represents a mass range that is typically transmitted with high efficiency on most mass spectrometers and that can be readily detected on most mass analyzers.

MALDI imaging mass spectrometry lipid analyses have been applied in environmental, pharmaceutical, clinical, and biochemical research settings.²¹³ Similar to metabolites, a variety of plant species, components, and analytes have been studied by environmental chemists and biologists. Lipid analyses have been used to study root growth, water transport, and microbial respiration across the rhizosphere to demonstrate how these compounds influence plant growth, development, reproduction, pathology, and responses to biotic and abiotic stresses.²¹⁴ Significant research in lipid MALDI IMS has also been performed in areas of human health. For example, global changes in phospholipids, lysophospholipids, sphingolipids, and sulfatides have been studied to determine their role in the architecture of amyloid β (Aβ) plaques in Alzheimer’s disease.^215–219 Retinal lipids have been studied in models of age-related macular degeneration and other degenerative retinal diseases.^220–222 Glycerophospholipids and ceramides have been mapped to monitor areas of damage in rodent models of traumatic brain injury.^58,223 Lipids are also commonly used in tumor classification and other cancer studies.^13,224,225 Some examples include the analysis of PIs in breast cancer tumors,²²⁶ glycerophospholipids in A549 lung cancer cell spheroids,²²⁷ glycerophospholipids and sphingomyelins in colon cancer, ceramides and fatty acids in lung cancer, phosphatidic acids and diacylglycerols in brain cancer, glycerophospholipids in pancreatic cancer,²²⁸ phospholipids in prostate cancer,²²⁹ cardiolipins in lymphatic cancer,²³⁰ phosphatidic acids and phosphatidylglycerols in breast cancer,²³¹ and ceramides and sphingomyelins in liver cancer.²³² MALDI IMS has also been used to study lipids in tissues affected by diabetes and diabetes-related diseases.^233–236

In contrast to lipidomics analyses performed using LC-MS/MS, the spatial context afforded by MALDI IMS provides a new dimension to the pathophysiological investigation of biological tissues. However, the immense structural diversity of lipids can make unambiguous structural identification challenging. When accounting for the multitude of head groups, fatty acyl groups, number of double bonds, position and stereochemistry of double bonds, alkyl branching, oxidation, and other variations, it is estimated that there are well over 100,000 individual isoforms in the cellular lipidome.²³⁷ IM-MS can help deconvolute this complex mixture by separating isobaric (i.e., same nominal mass) and isomeric (i.e., same exact mass) lipids.^24,238,239 Tentative lipid identifications are made in many instances using HRAM measurements, which afford better than parts per million (ppm) mass accuracies and enable the assignment of lipid elemental composition.^{240 241,242} However, a sum composition lipid assignment can represent dozens of individual isomeric lipid compounds and thus requires additional experiments to complete the identification.²⁴³ For example, acyl chain length and double bond position have both been shown to play an important role in biomarker differentiation between normal and cancerous tissues.^244,245 A variety of dissociation methods have been used in tandem mass spectrometry experiments to elucidate lipid structures in imaging experiments, including high and low-energy collision-induced dissociation, infrared multiphoton dissociation (IRMPD), ultraviolet photodissociation (UVPD), and electron-induced dissociation (EID).^246–249 A complete discussion of the chemistry and physics underlying these approaches and their application to lipid structural analysis is beyond the scope of the current discussion; the reader is directed to several recent reports on this subject.^250–252 New condensed-phase and gas-phase derivatization methods have been developed to improve lipid structural identification and are also being applied to IMS workflows.^{245,253–261} Overall, accurate structural identification is critical to understanding the biochemical implications of lipids observed by MALDI IMS.

Endogenous Peptides and Neuropeptides

Neuropeptides represent an important class of signaling molecules that are chemically diverse, subjected to rapid degradation, and often present in low abundance with highly heterogeneous distributions in the nervous system.^12,262,263 There are several technical challenges associated with MALDI MS imaging of neuropeptides; therefore, a variety of sample preparation strategies have been explored and developed. Because neuropeptides can undergo rapid degradation in vivo by catabolic enzymes, extra caution must be exercised to either immediately deactivate degradation enzymes by proper tissue preservation or by performing rapid analysis by MALDI IMS. To this end, a snap heat stabilization device was developed and shown to preserve tissue morphology of mouse brain tissue section for neuropeptide IMS analysis, whereas non-heat treated tissue samples exhibited degradation products.^264,265 Furthermore, due to the smaller size and general hydrophilic properties of neuropeptides, tissue washing protocols need to be carefully optimized to prevent delocalization and the loss/reduction of neuropeptide signals.^266–268

MALDI-FTMS has been used to study and characterize a variety of model organisms for neuropeptide analysis. Most notably, the crustacean neuropeptidome has been profiled with MALDI-FTMS in various studies.^275–284 To further increase chemical information from these in situ imaging mass spectrometry experiments, a multiplex imaging mass spectrometry method combining data-dependent acquisition (DDA) tandem MS analysis with gas-phase fractionation enabled more comprehensive peptidome coverage with sequence validation and novel neuropeptide discovery via de novo sequencing.²⁸⁴ In recent years MALDI imaging mass spectrometry has also shown significant promise in clinical analysis of neuropeptides and will facilitate the development of improved disease therapeutics.^285,286 Figure 7 shows representative images of neuropeptides in a primate brain section. Using a unilateral 6-hydroxydopamine rat model of Parkinson’s disease, spatial and intensity changes of more than 20 neuropeptides, including several opioid peptides associated with L-DOPA treatment, were evaluated. These studies highlight the importance of studying the presence of specific peptides, their locations, and dynamic changes to discover potential biomarkers and better understand disease progression and therapeutic response.

Fig. 7 |. High specificity neuropeptide mapping with MALDI IMS.

Neuropeptide distributions in the primate brain determined by MALDI IMS. Images of neuropeptides in a single hemisphere of the coronal primate brain at level − 2.5 from the anterior commissure. Through MALDI-TOF IMS analysis, 11 enkephalins (a), eight dynorphins (b), neurokinin A (NK A), two substance P neuropeptides (c), and four other neuropeptides (d) were detected. Annotations: Amg, amygdala; Put, putamen; GPe, external segment of globus pallidus; GPi, internal segment of globus pallidus; HCd, head of caudate nucleus; LH, lateral hypothalamus; StT, nucleus of stria terminalis; and TCd, tail of the caudate nucleus. Scale bar = 10 mm. Colour scale bars are shown as percentage of maximum intensity. Reprinted from Hulme, H. et al.²⁷⁴

Proteins and MALDI-Enabled Spatial Proteomics

Proteins are major molecular players in many essential biological processes and signaling pathways. Therefore, their localization and identification in a spatially resolved manner is one of the first applications of MALDI IMS^{1,28,287–292}, and have since been a major focus of IMS method developments and applications. Compared to the traditional immunohistochemical approaches for tissue-based studies, MALDI IMS offers simultaneous measurement of hundreds of protein species with isoform or modification information that is inherently available with the chemical specificity and selectivity offered by MS approaches.

Extensive research efforts have been devoted to in situ characterization of proteins from either fresh frozen or FFPE tissues.^{144,290,292,293} This can be done employing strategies to produce multiply charged ions^294–296 and by combining intact protein mass measurement in MALDI with LC-ESI-MS/MS workflows. The latter can be achieved by either using adjacent tissue sections or by performing MALDI-IMS-guided LC-MS/MS on intact masses.²⁹⁷ Alternatively, numerous bottom-up approaches have been developed to perform on-tissue enzymatic digestion to enhance protein identifications.^{290,292,298 287,289,299–303}

Because of the heterogeneity and complex structure of brain tissue and central nervous system, MALDI IMS has found extensive applications in neuroscience and the study of neurodegenerative diseases, where abnormal aggregation and deposition of proteins are often hallmarks of disease onset and progression. Proteins have also been studied for their roles in neurobiology, which include facilitating neuronal communication and regulating transport through the blood-brain barrier. For example, MALDI IMS has been employed to image intact protein species in postmortem spinal cord sections from amyotrophic lateral sclerosis (ALS) patients.^{285,304–306} However, it is not always possible to utilize human tissues for such MALDI imaging mass spectrometry-enabled biomarker discovery research. Although no animal model can completely recapitulate the complexity of human diseases, animal models provide a valuable proxy to enable linking candidate biomarkers to the molecular mechanisms underlying neurodegeneration and provide possible therapeutic strategies.

Another major application of MALDI IMS in cancer research is the clinical and pathological classification of tumor tissues based on characteristic protein profiles or expression pattern changes.^307–317 The ability to combine histological staining, MALDI IMS, and machine learning enables the classification of cancer subtypes.³¹⁸ Furthermore, spatial mapping of protein species in tissue samples coupled with biomarker identification and post-translational modification elucidation provides characterization of tumor tissue on a molecular level. This knowledge can lead to better diagnosis and individualized treatment strategy and possibly evaluation of therapy response.^319–324 For example, a recent study enabled the identification of protein biomarkers in head and neck cancer using an integrated approach combining MALDI IMS, LC-MS/MS, and immunohistochemistry.³²⁵ MALDI IMS has played an increasingly important role as a discovering tool for proteomic signatures and protein-based markers using archived material in pathology such as FFPE tissues or tissue microarrays (TMA), which enables high-throughput screening of pathological tissues to elucidate molecular patterns. This strategy has been implemented to survey TMAs of pancreatic cancer, prostate cancer, lung cancer, colon cancer, and renal cancer, among others.^326–334

Glycans and MALDI-Enabled Spatial Glycomics

Protein glycosylation is one of the most common protein posttranslational modifications (PTMs) and can change the structures and functions of the modified proteins. Modification of proteins by glycans can play diverse functional roles in protein folding and stability, cellular adhesion and interactions, signal transduction, enzymatic activity, and cancer metastasis.^335–337 Thus, the aberrant glycosylation patterns have been implicated in a variety of diseases, including many types of cancer, cardiovascular diseases, and neurodegenerative diseases.^{332,338–342} The high degree of chemical heterogeneity of glycans presents unique challenges in their identification and mapping of their spatial distribution in an anatomical context.

The combination of MALDI IMS and exoglycosidase cleavage offers a viable approach to profile and image glycans in tissue samples. Typically, N-linked glycans can be released from glycoproteins using enzyme Peptide-N-Glycosidase F (PNGase F). Drake and coworkers developed effective protocols that combined antigen retrieval and PNGase F on-tissue digestion with MALDI IMS and demonstrated in situ spatial mapping of the N-glycome in FFPE clinical specimens.^{337,343–345} Glycan imaging via MALDI IMS has since been demonstrated using a variety of FFPE tissues ranging from mouse kidney, human pancreatic, and prostate cancers to human hepatocellular carcinoma tissue microarray.¹⁹⁷ Using this strategy, MALDI IMS of tumor tissue has revealed molecular markers of tumor progressions and morphology; a recent study examined tumors from a cohort of pancreatic ductal adenocarcinoma (PDAC) patients.³⁴⁶ Released N-glycans were mapped across pancreatic and pancreatic tumor tissues using MALDI qTOF and MALDI FTICR that offer complementary features of tandem MS and high mass accuracy measurements, respectively. Figure 8 shows N-glycans from representative stage 3 PDAC tumor tissue sections visualized using MALDI IMS. Guided with histology staining, MALDI IMS-revealed glycan distributions can mirror cancer and necrosis distributions in pancreatic tumors, which may complement pathological annotations. The results of these imaging experiments were combined with immunohistochemistry to improve PDAC identification, highlighting the improved predictive power offered by orthogonal multimodal approaches. Despite being a powerful tool for mapping spatial distribution of N-glycans, MALDI IMS identifies glycans primarily based on MS1 accurate mass measurement, which suffers from relatively low confidence of assignment due to lack of tandem MS capability, necessitating alternative multi-omics approaches. 336336 It is worth noting that the MALDI ionization process often leads to the loss of labile sialylation of glycan species, necessitating approaches to maintain these moieties.^{337–339,346,351}

Fig. 8 |. Mapping specific N-glycans in pancreatic tumor tissue.

H&E stain and MALDI-FT-ICR MS images of N-glycans released from stage 3 pancreatic tumor tissue. Reproduced from McDowell et al. (2020).³⁴⁶

REPRODUCIBILITY AND DATA DEPOSITION

Reproducibility

Reproducibility in MALDI imaging mass spectrometry studies is achievable but requires thoughtful experimental design. Therefore, before any measurements begin, it is critical to consider aspects such as experimental design, calibration curves, matrix selection and application, and data pre-processing.³⁵⁴

One of the first considerations for the experimental design is to determine how many replicate samples will be used. The inclusion of biological replicates to increase the robustness of the findings is desirable, but not always achievable, depending on sample availability. Besides biological replicates, technical replicates (multiple sections from a single sample) can also be acquired, though the latter will not guard against sample-specific biological variability.³⁵⁵ In MALDI IMS studies, samples are usually prepared by sectioning with a cryostat into 5–20 mm sections, and these sections can be used in different ways. While every section could be used for MALDI imaging if needed, a common approach is to use alternating sequential sections to collect complementary data such as immunohistochemical or hematoxylin and eosin (H&E) stained views of the sample. These images can be registered and overlaid with the MALDI ion images to correlate molecular observations with substructures within the sample delineated by microscopy. It is important to section a sample into consistent widths, as differences in section thickness affect signal intensities.³⁵⁶ While replicates are valuable to guard against non-biological variation, batch effects can still predominate if care is not taken. Batch effects are systematic, non-biological factors that can alter the data and affect reproducibility. In IMS, batch effects can appear at multiple levels,³⁵⁷ including at the section, slide, and time (of sample acquisition, sectioning, or data acquisition) levels. The cumulative non-biological variation can mask true biological variation and confound conclusions. Additional challenges to reproducibility include ion suppression, homogeneity of matrix and trypsin application, analyte delocalization affecting spatial resolution, and lack of standardized protocols.³⁵⁷ In particular, the matrix selection and application can have a large impact on the eventual data.²⁹³ The “wetness” of the matrix solution following application influences analyte extraction, which is needed for successful ionization. However, if the sample is too “wet” during matrix application, it can result in analyte diffusion and loss of spatial resolution. A more in-depth discussion of sample preparation steps affecting resulting data is provided by Buchberger 2018²⁹³.

Sample storage conditions can also have a substantial impact on the quality of the data collected. In a recent study, sections of human kidney tissues were stored under five different conditions: −80 °C (open stored or under a nitrogen atmosphere, vacuumed sealed, with matrix pre-applied before storage) or at room temperature in a desiccator, and the tissues were analyzed by MALDI IMS for lipids distributions.³⁵⁸ While all samples displayed some deterioration in lipid composition, the samples stored under a nitrogen atmosphere at −80 °C sustained the least alteration, while those stored at room temperature were the most degraded.

As the spatial distribution of molecules in the sample is critical, maintaining the orientation of the sample from the outset, while preserving the integrity of the molecules, is furthermore key. A common trick when imaging smaller samples like murine organs is to affix the sample to the bottom of a 6-well dish as soon as it is dissected to preserve the orientation. When working with very limited samples, a single section can be probed first by MALDI IMS, followed by histological imaging.³⁵⁹ However, with more abundant samples, a more common approach is to generate replicate (serial) sections by cryosectioning. These sections can all be used for IMS or alternating sections can be utilized for a complementary imaging approach, as described above.

Just as the same section can be used for both MALDI and immunohistochemical staining, a single section can also be used to image multiple classes of molecules by MALDI. Recent work by Kaya and colleagues showed that a single tissue section could be used to generate dual MALDI analysis of lipids and proteins³⁶⁰. The investigators first analyzed lipid distributions with dual polarity lipid imaging at 35 μm pixel size, using 1,5-DAN as a matrix. Following lipid imaging, the matrix was rinsed away and another MALDI matrix (2,5-DHA) was applied to the same section for protein imaging. In this way, the same section could be analyzed for multiple molecular classes. With this experimental design, care must be taken to minimize ablation of the sample during the initial lipid imaging and to thoroughly remove the lipid matrix before application of the second matrix for protein analysis.

Data Deposition

IMS datasets can be very large and high-dimensional, and several repositories have been developed to make datasets accessible. Such repositories need to ensure that IMS datasets are appropriately archived and searchable. METASPACE is an example of a community-populated knowledge base for imaging datasets (https://metaspace2020.eu)³⁶¹, serving both as a repository for high-resolution imaging datasets and an engine for metabolite annotation. Datasets can be either public or accessible only to the investigator and/or reviewers of a manuscript. After publication, the dataset can be released to the public and made available for searching by other users as well.

Other data repositories are also gaining traction in the field. The Human BioMolecular Atlas Program (HuBMAP) is run by the National Institutes of Health (https://hubmapconsortium.org) and is tasked with examining the 37 trillion cells in the human body, mapping their spatial location as well as their function and = relationship to each other. Along those lines, the Kidney Precision Medicine Project (https://www.kpmp.org) is building a kidney tissue atlas to generate maps of different renal cell types. The project goal is to define disease subgroups and help identify cells that can be targeted by therapies.⁷¹ MALDI-IMS is a critical part of the Tissue Interrogation portion of KPMP.

MetaboLights is a repository for metabolomics datasets (https://www.ebi.ac.uk/metabolights)³⁶², including IMS data. It serves as resource for cross-species, cross-technique data and is the required repository for several journals. The biological roles, locations and concentrations of the various metabolites are detailed in its experimental data.

LIMITATIONS AND OPTIMIZATIONS

Like all analytical technologies, MALDI IMS has challenges and limitations that can be overcome through careful method and instrument optimization. The spatial dimension is the defining characteristic of IMS relative to other applications of mass spectrometry, and pixel size is one of the primary limitations of the technology. In a MALDI IMS experiment, its spatial resolution is, e.g., related to its pixel size. In microprobe instruments (i.e., those that collect data from discrete pixels), pixel size is defined by the diameter of the focused irradiating laser at the sample surface and the distance between adjacent pixels, often referred to as the pitch. It is possible to decouple laser spot size from pixel size using oversampling.³⁶³ In this mode, sufficient laser shots are fired at a sampling position, ensuring all matrix has been depleted so it cannot produce more ion signals. The stage is then moved by a distance smaller than the laser spot diameter such that only the fraction of the beam irradiating fresh sample can generate new signal. The pixel size is then solely determined by the stage step size. However, in some cases oversampling can provide lower sensitivities, likely, in part due to reduced laser fluence arising from using only the flank of the laser beam.³⁶⁴ Due to the quadratic relationship between pixel size and experiment-wide acquisition time (assuming square pixels in a 2-D experiment, total acquisition time ∝ (pixel size)²), a compromise between spatial resolution and acquisition times must often be made. Other systems use a continuous raster sampling approach where the laser is fired continuously as the sample stage is moved across the area of interest. Here, pixels are not necessarily square and the spatial resolution in the direction of the stage movement (e.g., horizontal) is defined by the laser repetition rate, the sample stage velocity, and the number of laser shots that are averaged to generate a single spectrum (i.e., pixel). Vertically, the pixel size is defined by the pitch between laser raster rows. This is approaches operating continuously in oversampling mode, leading to challenges in tuning the source conditions (e.g., laser energy, laser repetition rate, and stage velocity) properly to maintain sufficient and reproducible sensitivity for IMS experiments.

As highlighted above (see Experimental), MALDI IMS has the challenge of requiring the integration of both laser optics and molecular ion optics (e.g., electrodes) into the same source region of the instrument. This makes focusing MALDI systems more challenging than, for example, microscopes because the laser-focusing lenses must be positioned at greater distances from the sample surface. Although other geometries are used in special cases (see Fig. 2), these introduce additional complexity to the experiment and have yet to demonstrate the robustness and reproducibility that traditional ‘front-side’ systems have achieved. As such, significant development efforts have focused on optimizing pixel size for front-side MALDI IMS platforms.

Minimizing the laser spot size at the sample surface requires proper laser alignment to ensure that the beam is aligned to the center of the optical axis of the focusing lens and maximizing the collimated diameter of the laser. An unfocused laser will require higher pulse energy to obtain comparable data and generate ions from a larger sample area, potentially compromising spatial resolution. Optimizing laser focus becomes particularly important for small spot sizes as the depth of focus also reduces, meaning changes in the z-position of the sample (e.g., due to stage tilt) may lead to defocusing of the laser and, thus, changes in laser spot size. To circumvent this, some modern IMS systems incorporate auto-focusing methods to compensate for such shifts.³⁶⁵ One means to optimize laser focus for MALDI is to iteratively adjust the focus position and laser energy to acquire the highest signal intensity with the lowest possible pulse energy. Modern instruments also include methods for adjusting for stage tilt, where an initial scan of the surface height is collected across the sample area, and the position of the stage is adjusted from pixel-to-pixel to account for any detected high differences.

Laser energy – more specifically, laser fluence (energy per unit area) – also strongly influences the quality of MALDI data and its spatial resolution. If the laser pulse energy is too low, sensitivity can be poor. If it is too high, sensitivity can again be compromised as the in-source fragmentation increases, and matrix-related signals (especially matrix clusters) can dominate the spectrum. The optimal range of laser energy is relatively narrow and depends on the matrix and the target analytes; thus, it is a key parameter to optimize. Laser energy is typically optimized by setting the energy slightly higher than the ionization threshold for the analytes of interest. This will minimize the size of the laser spot at the sample surface and maximize ion intensities across the sample surface.

MALDI imaging mass spectrometry, like all mass spectrometric technologies, relies on the ability to convert neutral analyte molecules to ions for mass analysis, making it sensitive to varied ionization efficiencies of different molecular classes and ion suppression effects. On-tissue derivatization strategies can be employed to detect analytes that are not efficiently ionized using MALDI. This involves the application of a reagent – often containing a fixed charge – to the tissue section, where it reacts with target functional groups and covalently attaches a fixed charge (e.g., quaternary nitrogen) or another readily ionized functional group. Application of the derivatization reagent is often performed using the pneumatic sprayers that are typically used for matrix deposition and, like the matrix, must be done in a manner that minimizes analyte delocalization. Some derivatization strategies also require incubation in a suitable solvent/vapor to maximize the yield of the derivatized analyte. Depending on the nature of the reagent, it may be necessary to coat the sample in a suitable matrix following derivatization. However, in some cases, the derivatization reagent can also act as the MALDI matrix (so-called reactive matrices).³⁶⁶ Further details of specific applications of on-tissue derivatization can be found in the following references.^11,367

Accounting for ionization efficiency and ion suppression is also important for quantitation. Other than in extreme cases where the overall molecular landscape is completely different between regions of tissue, MALDI IMS provides relative quantitative data without any special methodologies. Most notably, significant changes in salt concentrations, mostly Na+ and K+, across the tissue can significantly affect ion intensities in positive ion mode. To overcome this, many MALDI IMS protocols include a tissue washing step to remove salt, maximizing sensitivity by driving ionization exclusively towards protonated species. It is also possible to achieve absolute quantitation. Like any analytical technique, internal standards can be introduced at varying concentrations to account for the complex chemical nature of the tissue (often referred to as chemical matrix effects). Internal standards at known concentrations are used to determine the relationship between per-pixel ion intensity and the actual concentration in the tissue microenvironment. Typically, standards are mixed with a slurry of tissue to create a series of mimetic tissues, each with different concentrations of the internal standards, that can be sectioned and analyzed to generate a calibration curve that can be applied to data from the target sample. Relative quantitation is simpler and is useful for exploratory studies or for comparing the spatial distributions of molecular species. Absolute quantitation is necessary for determining molecular concentrations; however, creating mimetic tissues for intensity calibration adds complexity to the MALDI IMS workflow. The choice between the two methods ultimately depends on the specific scientific research question and the goals of the study.

OUTLOOK

The molecular revolution in basic science and medical research has brought a new understanding of the elements of life, illuminating health and disease. Overall, imaging technologies have allowed us to peer deeper into molecular events both in space and time within tissues and cells. One of the imaging technologies that has brought in a new wave of information and understanding is imaging mass spectrometry. The molecular fidelity of this technology is truly outstanding in identifying the spatial aspects of proteins, peptides, lipids, and metabolites. When taken together, they reveal the complex dynamic state of cells and the tissues they lie within.

A transformative capability of IMS is its ability for integration and fusion with data from other imaging technologies. Precise alignment of image pixels between images obtained from different imaging modalities is critical when considering multi-modal imaging experiments. This approach brings exciting possibilities of taking the best attributes of any imaging modality and combining them with the molecular specificity of IMS to produce insights and fused images that have more data content than any technology individually could provide. We can anticipate algorithms dedicated for data processing and registration of images, especially for visualization the tremendous data content of the many thousands of molecular images that can be produced by this approach. Fast and effective computational tools will be critical as IMS improves in terms of speed and molecular depth obtained. This will provide a well-validated and integrated multimodal approach for the registration of multiple imaging modalities, for cross-modal data mining, and for the interactive visualization of the combined datasets.

IMS technology is also becoming a transformative platform in molecular pathology whereby molecular biomarkers can be spatially identified in tissue biopsies. IMS is multimolecular in nature, establishing combinations of protein, lipid, and metabolomic signatures as robust aids in disease diagnosis, prediction of a patient’s response to therapy and for their overall prognosis. Moreover, IMS in conjunction with other image technologies can clearly bring the art and practice of molecular pathology to a new and more effective level. For example, a signature of multiple protein and lipid tumor biomarkers can be used to identify and validate a molecular tumor margin that goes well beyond current microscopic protocols. A major advantage is its independence from target-specific reagents such as antibodies. Further improvements in the technology will bring an increased ability to measure biomarkers at low expression levels in tissue samples.

Looking forward, we can anticipate greatly improved technological advances in image resolution and sensitivity, creating smaller laser spot sizes on target. This will be effectively targeted to the molecular spatial analysis of molecules within individual cells. IMS is now a mature technology extensively used to study disease. MALDI mass spectrometry is currently used to classify and identify bacterial strains, a technology that was approved by the FDA many years ago. It is also employed for other clinical assays for the identification of biomarkers in patient samples. With respect to basic biological research, new approaches in machine learning have provided robust discriminatory analysis for large datasets that would be challenging if not impossible using manual methods. Combining MALDI IMS’ multiplexed data acquisition with powerful machine learning-driven data analysis and visualization will provide non-experts the ability to adopt this exciting technology for routine use in basic science/biology labs and in the clinical space.