Perspective on Multimodal Imaging Techniques Coupling Mass Spectrometry and Vibrational Spectroscopy: Picturing the Best of Both Worlds

Studies on complex biological phenomena often combine two or more imaging techniques to collect high-quality comprehensive data directly in situ, preserving the biological context. Mass spectrometry imaging (MSI) and vibrational spectroscopy imaging (VSI) complement each other in terms of spatial resolution and molecular information. In the past decade, several combinations of such multimodal strategies arose in research fields as diverse as microbiology, cancer, and forensics, overcoming many challenges toward the unification of these techniques. Here we focus on presenting the advantages and challenges of multimodal imaging from the point of view of studying biological samples as well as giving a perspective on the upcoming trends regarding this topic. The latest efforts in the field are discussed, highlighting the purpose of the technique for clinical applications.


S-3
Despite being such powerful tools, these techniques have some limitations. Generally, MSI instruments are expensive, they require vacuum-compatible samples (except for DESI), the mass range detection is limited (depending on the type of ionization hard vs. soft ionization), and some methods are constrained to reduced lateral resolution (10 µm for MALDI and 40 µm for DESI due to heterogeneous co-crystallization, analyte delocalization, the spray tip-to-surface distance, and the spray tip and nebulizer orifice diameters). 1,[3][4][5] Lastly, the MSI data collected demands large storage capabilities and powerful computational tools for data preprocessing and analysis.
Sample preparation. Sample preparation is a critical step for any imaging technique.
Normally biological samples for mass spectrometry imaging are fresh-frozen tissues 6 although some applications have reported the use of paraffin embedded tissues. 7,8 The typical sample processing before analysis consists of: tissue sectioning, tissue section handling, choosing the right ionization agent and solvent, deposition of the ionization agent, sample transportation and storage. 1 Sometimes, extra steps are included, such as deparaffinization or other specific tissue treatments (e.g. on-tissue washes, enzymatic digestion, chemical derivatization, etc.) when necessary. 1 Data acquisition. The typical MS image measurement starts with the selection and optimization of the acquisition parameters such as mass range, mass resolution, laser power, laser spot size, number of shots per pixel, pixel size, and area of acquisition, but it also includes mass analyzer calibration. Optimizing acquisition parameters is crucial for a successful experiment, as the sensitivity of MSI depends strongly on lateral resolution and ionization efficiency while sample viability can suffer during long experiments. 2 To avoid the acquisition of big datasets in the range of tens to hundreds of gigabytes, all parameters should be optimized to preserve the image quality while also minimizing data size: fewer pixels, precise area of measurement, reduced mass range, etc. Data analysis. The raw data collected from MSI studies goes through several preprocessing algorithms to ensure high quality MS images and optimal statistical analysis. Usually, MSI data consists of a large volume of mass spectra that presents experimental variability -such as chemical S-4 noise and mass spectra shifts -due to sample preparation and small changes during image acquisition.
Preprocessing algorithms improve image reconstruction and spectral quality. Rafols et al. described in great detail each step in the general pipeline of preprocessing MSI data. 9 After preprocessing, the data is usually analyzed first by the univariate analysis -or simply spatial visualization -of one specific peak (or ion), and then by more sophisticated analysis such as supervised or unsupervised multivariate analysis. Rafols et al. outlined all aspects necessary for powerful bioinformatics tools reviewing data handling strategies with both commercial and open-source software. 9 Fortunately the mass spectrometry imaging community developed a common data format called imzML, 10 which facilitates progress in MSI processing algorithms, as well as inter-laboratory collaborations in which instruments are not from the same manufacturer. 11 Figure S1. Schematic illustration of different ionization mechanisms.

Spectroscopic Imaging Methods
Fundamentals. Vibrational Spectroscopy is based on light interacting with molecules from a sample. Specifically, the recorded spectrum represents a collection of the molecular vibrations of chemical bonds from all molecules within the illuminated area of the sample. 12 This information gives the fingerprint signature of the sample. For Raman spectroscopy, the incident light of a fixed S-5 wavelength interacts with the sample and the frequency-shifted scattered light is detected and represented in the Raman shift spectrum; 13 for IR spectroscopy the infrared incident light is absorbed by the sample and the transmitted (or reflected) light is detected, resulting in a spectrum of absorption ( Figure S2). 14 Just as for MSI, each pixel of an area of the sample is represented by a single spectrum, while the full image contains a data cube of pixel position and spectral information.
Characteristics. Raman and IR spectroscopy describe the physical and chemical properties of molecules by collecting signals that represent the stretching, bending and rotating vibrations of their chemical bonds. Vibrational spectroscopy is label-free and non-destructive, enabling imaging with lateral resolutions from several mm (for mid-IR) to below 1 µm (for Raman). Hence, Raman can be used for single-cell and intracellular imaging 15 while IR is employed for imaging larger areas on tissues 16 . The resulting spectral information gives valuable insights into tissue organization, secondary structure or molecules, lipid and protein content, cell metabolism, drug delivery and even in-vivo prediction of diseases in clinical research. This information is typically obtained from the strong vibrations of specific bonds: for example, lipid droplets and the myelin sheath of neurons are represented by CH 2 vibrations of lipids and CH 3 vibrations of proteins, respectively. 17,18 The lateral resolution in vibrational spectroscopy depends on the optical configuration of the microscope (objective numerical aperture, NA) and the incident laser wavelength (λ), following the laws of physics and optics: spatial resolution = 0.61 λ / NA. 19 Therefore, submicron lateral resolution is easily accessible for various combinations of laser wavelengths and objectives. However, VSI techniques have some limitations. Raman has low sensitivity, long acquisition time per pixel due to its lack of signal strength (ca. 1 scattered photon in 109 incident photons) and low chemical specificity. 5 SERS signal is more sensitive than Raman but it is substrate dependent with reduced spot-to-spot homogeneity, which limits reproducible imaging experiments. 20 IR techniques present strong interference from water absorption, which impedes data acquisition and analysis. 15

S-6
Sample preparation. Raman samples can be either liquids or solids: fresh frozen (e.g. tissues on calcium fluoride) or immersed in liquid (e.g. cell cultures in aqueous media). Frozen samples follow a similar procedure to MSI: tissue sectioning, tissue section handling, optional deparaffinization or other specific treatments, transportation and sample storage. The tissue sections are typically placed on calcium fluoride slides because their low refractive index and absorption mean that their natural fluorescence is also low. 21 The case of live samples such as cell cultures requires an immersion objective that permits live-cell analysis in real time. 22,23 On the other hand, IR imaging samples are usually solid (e.g. tissue sections 7,8,24,25 ), and they need to be placed on transparent slides with low absorption in the IR range, such as IR-reflecting microscope slides or calcium fluoride substrates..
Sample preparation for vibrational spectroscopy can be as simple as choosing the sample substrate which does not interfere with the sample signals. Thus, nanostructured materials such as silver 26 and gold 20,27 nanoparticles have been used as sample substrates, mostly because they are signal enhancing agents but also because of their low interference.
Data acquisition. Raman image acquisition parameters strongly depend on the optical configuration of the microscope and the lasers used with the instrument. Laser parameters (wavelength, power, spot shape, etc.), the objective's numerical aperture, the type of measurement (in air or immersed in liquid), the spectral grating (for spectral resolution), and the exposure time are the typical acquisition parameters that need to be optimized for spectroscopy measurements. Due to the weakness of Raman scattering and the intrinsic fluorescence of biological samples, the laser wavelength and the objective have to be chosen properly for each experiment so that the spot size, laser power and exposure time provide the best Raman signal without "burning" the samples. 28 Autofluorescence can be reduced by photobleaching 29 and high-resolution images can be achieved with small laser spot sizes (<1µm). Unfortunately, these conditions together with high laser power damage samples by the localized heat induced on the sample surface. SERS substrates allow using less powerful lasers 5 which ensure sample viability however, long acquisition times are still a threat.

S-7
Additionally, SERS imaging measurements struggle with reproducibility, 5 although nanostructured surfaces with high enhancement and homogeneous distribution of hotspots can overcome this limitation. 20 For IR measurements -in transmission, transflection and ATR modes -the spatial resolution depends on the configuration of the IR instrument (single aperture resolution: 2λ/3, and confocal arrangement resolution: λ/2) 30 and on measuring the right background signal. IR spectroscopy is sensitive to substrate transparency (in transmission mode), sample thickness, and water content (reflected in the spectra by the OH band).. 31 IR measurements are mostly held in atmospheric environments, so sample viability over time is also an issue, which is why both Raman and IR spectroscopy seek substrates that allow the temperature to be controlled.

Data analysis.
Data analysis for spectroscopic datasets consists of a pre-processing step that prepares the Raman and IR data for analysis 32 . Lasch described in detail the aims of signal preprocessing for both IR and Raman data: (i) robust and accurate spectra; (ii) comprehensible data for both humans and machines; (iii) outlier and trend removal and (iv) dimensionality reduction. 32 Similarly, Vidal et. al highlighted the importance of first removing background, dead pixels, spikes and outliers and then pre-processing the remaining spectral data. 33 This approach eliminates all the unwanted effects during acquisition that are both intrinsic (e.g. autofluorescence, cell media or water content, substrate, etc.) and extrinsic (e.g. detector noise, calibration errors, cosmic rays, laser power fluctuations, etc.). Gautam et al. described all the common pre-processing algorithms regarding spectral axis alignment, cosmic ray removal, background correction (or baseline removal), smoothing, normalization and outlier removal. 34 However, some data processing steps are specific for each type of data: Raman imaging spectra need to be aligned through wavelength calibration and cleaned of cosmic ray artifacts; IR spectra need to pass a quality test, undergo water vapor correction and finally go through a first or second derivative filter for interpretation. 33 Data processing usually consists of univariate analysis in which the spatial visualization of one specific band generates a heatmap, and then of supervised or unsupervised multivariate analysis which finds important structural information, S-8 image segmentation, and tissue classification. 34 Unfortunately, unlike for MSI data, there is no standard or common file format for VSI data storage and processing. Figure S2. Vibrational spectroscopy methods.
S-9  Changes in the lipid content observed by a high correlation of the Raman spectral region with MALDI mass spectra; data fusion increases reliability not only for the spectral features but also for the spatial features present in the data