This section presents a comprehensive review of the applications of MVA and ML to the ToF-SIMS images of organic, bioorganic, and biological systems. We opted to categorize the literature broadly according to their application areas, rather than grouping them based on the applied MVA or ML algorithm(s). This was done to provide a focus for readers in any of the covered fields, by presenting a comprehensive summary of the types of analyses that have been performed in their field of research.

A. Organic films, residues, and coatings

Organic (including lipid, polymeric and protein) films, residues and coatings have been studied extensively over the past decades using ToF-SIMS imaging.61,76,77–8261. W. Gardner, S. M. Cutts, B. W. Muir, R. T. Jones, and P. J. Pigram, Anal. Chem. 91, 13855 (2019). https://doi.org/10.1021/acs.analchem.9b0332276. M. C. Biesinger, P. Y. Paepegaey, N. S. McIntyre, R. R. Harbottle, and N. O. Petersent, Anal. Chem. 74, 5711 (2002). https://doi.org/10.1021/ac020311n77. C. R. Anderton, B. Vaezian, K. Lou, J. F. Frisz, and M. L. Kraft, Surf. Interface Anal. 44, 322 (2011). https://doi.org/10.1002/sia.380678. M. C. Biesinger, D. J. Miller, R. R. Harbottle, F. Possmayer, N. S. McIntyre, and N. O. Petersen, Appl. Surf. Sci. 252, 6957 (2006). https://doi.org/10.1016/j.apsusc.2006.02.12179. C. M. McQuaw, A. G. Sostarecz, L. Zheng, A. G. Ewing, and N. Winograd, Appl. Surf. Sci. 252, 6716 (2006). https://doi.org/10.1016/j.apsusc.2006.02.21080. Y. Yokoyama, S. Aoyagi, T. Shimanouchi, M. Iwamura, and H. Iwai, Biointerphases 11, 02A314 (2016). https://doi.org/10.1116/1.494070681. L. Zheng, C. M. McQuaw, M. J. Baker, N. P. Lockyer, J. C. Vickerman, A. G. Ewing, and N. Winograd, Appl. Surf. Sci. 255, 1190 (2008). https://doi.org/10.1016/j.apsusc.2008.05.25582. M. K. Passarelli and N. Winograd, Biochim. Biophys. Acta, Mol. Cell Biol. Lipids 1811, 976 (2011). https://doi.org/10.1016/j.bbalip.2011.05.007 ToF-SIMS is particularly useful for imaging lipids, owing to its high sensitivity for lipid structures.8282. M. K. Passarelli and N. Winograd, Biochim. Biophys. Acta, Mol. Cell Biol. Lipids 1811, 976 (2011). https://doi.org/10.1016/j.bbalip.2011.05.007 More generally, the high spatial resolution of ToF-SIMS can offer unique insights into the spatial arrangements of various types of organic films. Several MVA and ML techniques have been used for these applications over the past two decades.ToF-SIMS imaging and MVA were used by Biesinger et al.7676. M. C. Biesinger, P. Y. Paepegaey, N. S. McIntyre, R. R. Harbottle, and N. O. Petersent, Anal. Chem. 74, 5711 (2002). https://doi.org/10.1021/ac020311n to explore different lipid films in one of the earliest studies using PCA for relatively large (at the time) ToF-SIMS images (256 × 256 pixels). This study elucidated the type of chemical information obtained from the spatial distributions of principal component scores (and their corresponding loadings) across the analysis area (Fig. 2). The authors investigated Langmuir–Blodgett films composed of dipalmitoylphosphatidylcholine (DPPC) or a mixture of DPPC and palmitoyloleoylphosphatidylglycerol (POPG). PCA distinguished the lipid phases in the film and highlighted the topographical effects. Later, Biesinger et al.7878. M. C. Biesinger, D. J. Miller, R. R. Harbottle, F. Possmayer, N. S. McIntyre, and N. O. Petersen, Appl. Surf. Sci. 252, 6957 (2006). https://doi.org/10.1016/j.apsusc.2006.02.121 also combined selective deuteration, ToF-SIMS imaging, and PCA to discriminate between the lipid components of artificial lipid membranes. This isotopic labeling improved the MVA workflow: The deuteration of DPPC and POPG allowed successful discrimination of these lipids from nondeuterated DPPC and the phases in the film using PCA. Similar results were reported using deuterated cholesterol. Note that discrimination was also relatively straightforward using individually selected ion peaks due to the distinctive peaks associated with deuteration. Nevertheless, the PCA results demonstrated higher contrast images with less noise than the corresponding single ion images.In a later work, Baker et al.8383. M. J. Baker, L. Zheng, N. Winograd, N. P. Lockyer, and J. C. Vickerman, Langmuir 24, 11803 (2008). https://doi.org/10.1021/la802582f investigated the distribution of glycophorin A in Langmuir–Blodgett films comprised of cholesterol and either DPPC or DPPE. These films were designed to replicate the compositions of the outer and inner leaflets of a cell membrane, respectively. ToF-SIMS images showed homogeneity of the lipids and protein in the DPPC film and the heterogeneity of distributions in DPPE films. Again, the PCA produced higher contrast images of these heterogeneities, which were consistent with the individual ion images. Similarly, Anderton et al.8484. B. Vaezian, C. R. Anderton, and M. L. Kraft, Anal. Chem. 82, 10006 (2010). https://doi.org/10.1021/ac101640c applied PCA to visualize phase separation in DSPC/DLPC and DPPC/DLPC membranes. The same group later showed that PCA provided more interpretable results if a refined peak list that more accurately captures the chemical differences in the surface is used.7777. C. R. Anderton, B. Vaezian, K. Lou, J. F. Frisz, and M. L. Kraft, Surf. Interface Anal. 44, 322 (2011). https://doi.org/10.1002/sia.3806In polymer research, Lau et al.8585. Y. T. R. Lau, L. T. Weng, K. M. Ng, and C. M. Chan, Surf. Interface Anal. 43, 340 (2010). https://doi.org/10.1002/sia.3404 used ToF-SIMS imaging and PCA to investigate poly(bisphenol-A-etheralkane) polymer films. Comparing ToF-SIMS images with AFM height images, the authors showed that PCA (specifically component 1) discriminated between the amorphous and flat-on lamellar areas of the semicrystalline polymer. It should be noted that the total ion image was also able to discriminate between these regions, although normalization to total ion count was performed, so the discrimination by PCA was not influenced by this. As the authors note, other components did not capture any further meaningful chemical differences in the images, such as local variations within each region.The combination of PCA and ToF-SIMS imaging has also found promising applications in studying polymeric materials for medical care. Aoyagi et al.8686. S. Aoyagi, K. Abe, T. Yamagishi, H. Iwai, S. Yamaguchi, and T. Sunohara, Anal. Bioanal. Chem. 409, 6387 (2017). https://doi.org/10.1007/s00216-017-0578-1 used PCA with ToF-SIMS and near-field infrared microscopy (NFIR) to analyze blood adsorption on dialysis membranes composed of polysulfone or asymmetric cellulose triacetate. Combined univariate and multivariate analyses showed that blood adsorption was low in both cases but was significantly higher for the polysulfone membranes.Utilizing more than one MF technique has proven useful in studying multicomponent polymeric systems. For example, Ravati et al.8787. S. Ravati, S. Poulin, K. Piyakis, and B. D. Favis, Polymer 55, 6110 (2014). https://doi.org/10.1016/j.polymer.2014.09.013 used a combination of PCA and MCR to study ternary polymer blends by ToF-SIMS imaging. PCA identified polymer blends with high spectral similarity for further investigation and characteristic ions associated with each polymer. They selected a blend of polybutylene succinate (PBS), polylactic acid (PLA), and polycaprolactone (PCL) for further analysis as these had high chemical similarities according to the PCA. ToF-SIMS imaging was then performed at high (∼200 nm) and low (∼2 μm) spatial resolutions, and MCR was used to process the images of the ternary PBS/PLA/PCL blend. In both cases, MCR could successfully segment the images into distinct phases, yielding higher contrast images compared to univariate analyses (Fig. 3). MCR also revealed information about the interfaces between the phases. The combination of PCA and MCR also proved useful for Yokoyama et al.,8888. Y. Yokoyama, T. Kawashima, M. Ohkawa, H. Iwai, and S. Aoyagi, Surf. Interface Anal. 47, 439 (2015). https://doi.org/10.1002/sia.5731 who used ToF-SIMS images to discriminate between three polymers, polyethylene terephthalate (PET), polystyrene (PS), polycarbonate (PC). As has been reported more generally, components of the MCR factorization correlated better with the known spatial distributions of the different polymers than those of PCA. However, it should be noted that Poisson scaling was used prior to MCR, but only mean-centering was used prior to PCA. Both MCR and PCA also identified Na+ and K+ contamination across the polymers. Similarly, Stroka and Goacher8989. J. R. Stroka and R. E. Goacher, J. Vac. Sci. Technol. B 34, 03H101 (2016). https://doi.org/10.1116/1.4939465 studied wood-polymer composites (WPCs) using ToF-SIMS imaging, PCA, and MCR analysis, in conjunction with other surface science techniques. This work was presented as an initial investigation into the merits of using TOF-SIMS and MVA in the field of WPC research, outlining the advantages of this approach with respect to other more commonly applied techniques, such as scanning electron microscopy (SEM).8989. J. R. Stroka and R. E. Goacher, J. Vac. Sci. Technol. B 34, 03H101 (2016). https://doi.org/10.1116/1.4939465In addition to PCA and MCR, Trindade et al.37,38,4037. G. F. Trindade, M. L. Abel, C. Lowe, R. Tshulu, and J. F. Watts, Anal. Chem. 90, 3936 (2018). https://doi.org/10.1021/acs.analchem.7b0487738. G. F. Trindade, M. L. Abel, and J. F. Watts, Chemom. Intell. Lab. Syst. 163, 76 (2017). https://doi.org/10.1016/j.chemolab.2017.02.01240. G. F. Trindade, D. F. Williams, M. L. Abel, and J. F. Watts, Surf. Interface Anal. 50, 1180 (2018). https://doi.org/10.1002/sia.6378 reported the application of NMF to ToF-SIMS imaging in several publications, also disclosing the software simsMVA for performing PCA, NMF, and k-means clustering of the ToF-SIMS data in general.3939. G. F. Trindade, M. L. Abel, and J. F. Watts, Chemom. Intell. Lab. Syst. 182, 180 (2018). https://doi.org/10.1016/j.chemolab.2018.10.001 The group described how to apply NMF to large ToF-SIMS datasets where a training set was selected via a subsampling procedure. This approach was applied to the ToF-SIMS image of a chemically contaminated fingerprint,3838. G. F. Trindade, M. L. Abel, and J. F. Watts, Chemom. Intell. Lab. Syst. 163, 76 (2017). https://doi.org/10.1016/j.chemolab.2017.02.012 successfully discriminating the fingerprint residue from the substrate. An interesting result from this work was that the NMF algorithm separated pixels based on the misalignment of the spectra from different regions in the image. While the cause of this misalignment was uncertain, the authors showed that the alignment could be corrected using results from the NMF analysis. This significantly reduced computation time for the correction compared to aligning individual spectra on a per-pixel basis. Trindade et al.3737. G. F. Trindade, M. L. Abel, C. Lowe, R. Tshulu, and J. F. Watts, Anal. Chem. 90, 3936 (2018). https://doi.org/10.1021/acs.analchem.7b04877 also applied a related method to cross sections of layered, highly similar resin materials. In this work, rather than analyzing a single ToF-SIMS image independently, several images of the layered resins were stitched together and analyzed simultaneously using NMF. The ToF-SIMS fingerprint images of resin standards were also included in the analysis, providing a reference or key to aid an interpretation of the component images. A total of six components were selected for the NMF algorithm (based on an optimization procedure described in the original article). To visualize the distribution of these components in a single image, a unique color was applied to each; then a single image was produced, in which, the intensities of the components in each pixel determined the opacity of the respective coloring. This approach yielded good separation of the resins in the cured and uncured states.Tyler et al.9090. B. J. Tyler, C. Bruening, S. Rangaranjan, and H. F. Arlinghaus, Biointerphases 6, 135 (2011). https://doi.org/10.1116/1.3622347 demonstrated the capabilities of MAF in discriminating similar and dissimilar proteins in ToF-SIMS images. The authors investigated spheres coated with one of the three proteins: human serum albumin (HSA); bovine serum albumin (BSA); or hemoglobin. HSA and BSA are structurally similar, whereas hemoglobin is very different. In each case, MAF produced clear contrast between all three proteins (confirmed by fluorescence studies), regardless of whether the full spectrum or only specific amino acid peaks was used. This suggested that the ToF-SIMS data were sufficient for discriminating highly similar proteins using single-pixel data when appropriate MVA methods are used.Outside of MF, multicomponent polymeric and protein samples have spurred the development of new ML techniques, such as was reported by Tuccitto et al.91,9291. N. Tuccitto, G. Capizzi, A. Torrisi, and A. Licciardello, Anal. Chem. 90, 2860 (2018). https://doi.org/10.1021/acs.analchem.7b0500392. N. Tuccitto, A. Bombace, A. Torrisi, A. Licciardello, G. Lo Sciuto, G. Capizzi, and M. Woźniak, Chemom. Intell. Lab. Syst. 191, 138 (2019). https://doi.org/10.1016/j.chemolab.2019.07.002 They developed an approach for discriminating single-pixel ToF-SIMS spectra using statistical pattern recognition. The method could analyze an entire ToF-SIMS dataset without the need for peak-selection or mass binning. Their approach used a Fourier transformation of the raw data to reveal fragmentation patterns by the skewness and kurtosis of the distribution in the Fourier transform domain. The authors successfully used this technique to discriminate between several two-component systems of proteins and polymer films.9191. N. Tuccitto, G. Capizzi, A. Torrisi, and A. Licciardello, Anal. Chem. 90, 2860 (2018). https://doi.org/10.1021/acs.analchem.7b05003 Recent work extended this approach to multicomponent systems. A supervised probabilistic neural network was used to classify pixels from each polymer component according to their skewness and kurtosis values.9292. N. Tuccitto, A. Bombace, A. Torrisi, A. Licciardello, G. Lo Sciuto, G. Capizzi, and M. Woźniak, Chemom. Intell. Lab. Syst. 191, 138 (2019). https://doi.org/10.1016/j.chemolab.2019.07.002 While this method demonstrated the separation of four unique polymers, it is not clear how robust this approach is for more complex datasets, for example, ToF-SIMS images of polymer microarray systems comprising tens or hundreds of unique surface chemistries. Nevertheless, the results provided a tentative proof-of-principle for this approach.We have used ToF-SIMS and machine learning to study drug-loaded liposomes, as well as lipid films composed of the same lipid components.6161. W. Gardner, S. M. Cutts, B. W. Muir, R. T. Jones, and P. J. Pigram, Anal. Chem. 91, 13855 (2019). https://doi.org/10.1021/acs.analchem.9b03322 This was part of a research study investigating the incorporation of an antibiotic prodrug, cefditoren pivoxil (CP), into liposomes to improve its bioavailability and to target drug delivery. It was based on research on the use of prodrugs for enhancing chemotherapy treatments for breast cancer. We used SOMs to visualize the molecular similarities of each pixel—via the construction of a color similarity map—based on their mass spectra. This method definitively identified the prodrug-loaded and empty liposomes and showed which fragment ions were specific to each of the liposome populations.

B. Biological systems

Studying biological systems directly using ToF-SIMS imaging can be challenging due to the complexity and size of the acquired datasets. Although ToF-SIMS has valuable spatial resolution, it generates high levels of molecular fragmentation compared to other techniques such as MALDI and DESI. While using cluster ion sources can mitigate this to some extent,93–9593. C. Bich, D. Touboul, and A. Brunelle, Mass Spectrom. Rev. 33, 442 (2014). https://doi.org/10.1002/mas.2139994. S. Rabbani, A. M. Barber, J. S. Fletcher, N. P. Lockyer, and J. C. Vickerman, Anal. Chem. 83, 3793 (2011). https://doi.org/10.1021/ac200288v95. J. Xu, S. Ostrowski, C. Szakal, A. G. Ewing, and N. Winograd, Appl. Surf. Sci. 231-232, 159 (2004). https://doi.org/10.1016/j.apsusc.2004.03.104 the issue still remains to varying degrees. Often, different molecular structures can produce similar and overlapping mass spectral fingerprints. This makes ToF-SIMS data extremely difficult to interpret, especially using the univariate analysis. Hence, ToF-SIMS community is increasingly using MVA and ML to interpret and analyze complex images of the biological samples.In 2004, Nygren and Malmberg9696. H. Nygren and P. Malmberg, J. Microsc. (Oxford) 215, 156 (2004). https://doi.org/10.1111/j.0022-2720.2004.01374.x reported an early study of ToF-SIMS and PCA being used to image cells. They studied freeze-dried leucocytes sputter-coated with silver, and some typical score images from this work are shown in Fig. 4. The PCA revealed membrane components and increased levels of silver-labeled cholesterol in the attachment areas of the cells, with lower levels in the cell body. Barnes et al.9797. C. A. Barnes, J. Brison, M. Robinson, D. J. Graham, D. G. Castner, and B. D. Ratner, Anal. Chem. 84, 893 (2012). https://doi.org/10.1021/ac201179t also used MVA to interpret the ToF-SIMS images of cells. However, rather than using the score images directly, they used a combination of PCA and partial least-squares discriminant analysis (PLS-DA) to identify peaks associated with two different cell types. Prominent peaks from each class of cells were then summed to produce single images of each cell type (Fig. 5). Later, Tyler et al.9898. B. J. Tyler, M. M. Takeno, and K. D. Hauch, Surf. Interface Anal. 43, 336 (2010). https://doi.org/10.1002/sia.3679 showed how isotopic labeling can improve the MVA workflow, which they demonstrated using MAF. The group analyzed ToF-SIMS images of labeled and nonlabeled 3T3 fibroblasts and showed that 15N labeling, along with automated peak selection and MAF, resulted in high-contrast images of the cells.Outside of single cell imaging, Kotze et al.9999. H. L. Kotze, E. G. Armitage, J. S. Fletcher, A. Henderson, K. J. Williams, N. P. Lockyer, and J. C. Vickerman, Surf. Interface Anal. 45, 277 (2013). https://doi.org/10.1002/sia.5055 used PCA to analyze ToF-SIMS images of spheroids, either untreated or treated with the chemotherapeutic agent doxorubicin. PCA was applied in the investigation of hypoxia and associated chemoresistance. Tyler et al.100100. B. J. Tyler, S. Rangaranjan, J. Moller, A. Beumer, and H. E. Arlinghaus, Appl. Surf. Sci. 252, 6712 (2006). https://doi.org/10.1016/j.apsusc.2006.02.278 investigated the diffusion of the antimicrobial agent chlorhexidine digluconate (CHG) in C. albicans biofilms. The MAF analysis of ToF-SIMS images of the biofilm cross sections successfully discriminated the biofilm from CHG that may be transported through the biofilm in microchannels.In tissue imaging, Börner et al.101101. K. Borner, H. Nygren, P. Malmberg, E. Tallarek, and B. Hagenhoff, Appl. Surf. Sci. 252, 6777 (2006). https://doi.org/10.1016/j.apsusc.2006.02.200 applied PCA to discriminate regions of a rat cerebellum using the distributions of Na+ and K+. While the individual ion images also readily distinguished between these regions, as one of the earlier demonstrations of PCA applied to the ToF-SIMS images of tissue sections, this work provided a valuable proof-of-principle of the unsupervised PCA methodology. A similar exemplification was provided by Wu et al.,102102. L. G. Wu, X. H. Lu, K. S. Kulp, M. G. Knize, E. S. F. Berman, E. J. Nelson, J. S. Felton, and K. J. J. Wu, Int. J. Mass Spectrom. 260, 137 (2007). https://doi.org/10.1016/j.ijms.2006.09.029 who used ToF-SIMS to study mouse embryo tissue sections. Here, PCA was used to differentiate regions of the embryo using summed spectra, and to image the skull section by analyzing single-pixel spectra. The PCA results from this work, shown as an RGB overlay of the first three components and their respective loadings, are presented in Fig. 6. As with Börner et al.,101101. K. Borner, H. Nygren, P. Malmberg, E. Tallarek, and B. Hagenhoff, Appl. Surf. Sci. 252, 6777 (2006). https://doi.org/10.1016/j.apsusc.2006.02.200 in this work, the authors presented similar images using univariate analysis prior to PCA. Nevertheless, these results again demonstrated a workflow for imaging without a priori knowledge of the sample, and for obtaining valuable single-pixel molecular information through the loadings. Park et al.103103. J. W. Park, H. K. Shon, B. C. Yoo, I. H. Kim, D. W. Moon, and T. G. Lee, Appl. Surf. Sci. 255, 1119 (2008). https://doi.org/10.1016/j.apsusc.2008.05.102 used PCA and ToF-SIMS to differentiate between healthy and cancerous human colon tissue. PCA score images were produced for each type of tissue and compared by analyzing the respective loadings of each principal component. However, it should be noted that the way in which the images were presented could be misleading (Fig. 7): PC 1 and 2 for each type of tissue are presented adjacent; however, the components were not related since PCA was performed on each image separately. This was addressed to a degree by considering the respective loadings; however, visualizing images in this way should be done with caution.An excellent application-focused (as opposed to algorithm-focused) demonstration of the use of PCA in ToF-SIMS imaging of biological tissues was reported by Brulet et al.5656. M. Brulet, A. Seyer, A. Edelman, A. Brunelle, J. Fritsch, M. Ollero, and O. Laprevote, J. Lipid Res. 51, 3034 (2010). https://doi.org/10.1194/jlr.M008870 As part of a larger study of a mouse model of cystic fibrosis (CF), the group used ToF-SIMS imaging to explore a CF colon section. The authors suggested that the first four principal components discriminated between several known histological features in the tissue, as shown in Fig. 8. This was validated by the loadings associated with each region. While this work clearly demonstrated the advantages of PCA, and more generally MF techniques, over univariate analysis of ToF-SIMS images, it should be noted that the authors preselected 15 mass fragments known to capture differences between these regions. This method may introduce experimenter bias and also highlights how the results of PCA and many ML algorithms can depend on the selection of mass peaks, an issue that has been discussed recently by Teresa et al.104104. T. Murta et al., Anal. Chem. 93, 2309 (2021). https://doi.org/10.1021/acs.analchem.0c04179 Note that, in the same study, the authors also used the k-means clustering with varying numbers of clusters to produce similar results to PCA (Fig. 9), with regard to identifying and distinguishing different spatial features in the image. In addition to the selection of mass peaks influencing PCA results, tissue pretreatment can also have an effect. Angerer et al.105105. T. B. Angerer, M. D. Pour, P. Malmberg, and J. S. Fletcher, Anal. Chem. 87, 4305 (2015). https://doi.org/10.1021/ac504774y reported that the treatment of mouse brain tissue sections with trifluoracetic acid improved the lipid imaging capabilities of ToF-SIMS by improving the contrast between brain regions from PCA scores.While PCA score images and loadings can be analyzed directly, as has been discussed, the algorithm can also be used to identify spatial regions of interest (ROIs) for further exploration. This was ably demonstrated by Bluestein et al.,106106. B. M. Bluestein, F. Morrish, D. J. Graham, J. Guenthoer, D. Hockenbery, P. L. Porter, and L. J. Gamble, Analyst 141, 1947 (2016). https://doi.org/10.1039/C5AN02406D who described a workflow in which PCA was applied to ToF-SIMS images of human breast tumor tissue to generate spatial masks for segmenting the tissue sections into either cellular or stromal regions. These masks were then used to select ROIs for subsequent spectral PCA analysis of tissue sections before and after chemotherapy. The results indicated that, compared to no segmentation or segmentation using H&E staining and optical imaging, the PCA-based segmentation resulted in different characteristic ion peaks being associated with chemotherapy. More recent work by the same group107107. B. M. Bluestein, F. Morrish, D. J. Graham, L. Huang, D. Hockenbery, and L. J. Gamble, Biointerphases 13, 06D402 (2018). https://doi.org/10.1116/1.5038574 used PCA and ToF-SIMS imaging to study the metabolic microenvironment of pancreatic and cell tumors. They investigated metabolic changes associated with the overexpression of the oncogene Myc. PCA revealed differences between tumor and control islets, as well as between tumors and surrounding tissue. Among other differences, the authors noted an abundance of heme fragments spatially associated with the tumor tissue. Further PCA analysis of isolated pixels from the tumor tissue revealed heterogeneity within the tumor itself, highlighting “blood lakes” containing high levels of the heme fragment ions, thought to arise due to defective blood vessel linings. This work provided a valuable example of a comprehensive analysis of biological tissue microenvironments using ToF-SIMS and PCA.In addition to PCA, other MVA and ML techniques have shown promise in ToF-SIMS imaging of the tissue. Aoyagi et al.108108. S. Aoyagi, T. Matsuzaki, M. Takahashi, Y. Sakurai, and M. Kudo, Surf. Interface Anal. 44, 772 (2012). https://doi.org/10.1002/sia.4822 used MCR to analyze mouse skin sections treated with various reagents, including hair restoring agents. MCR helped to identify the distribution of active components in the skin. Hanreider et al. presented a series of studies in which MAF and ToF-SIMS imaging were used to discriminate anatomical features of biological tissues, including human spinal cord tissue and rat brain cryosections.50,51,54,10950. J. Hanrieder and A. G. Ewing, Sci. Rep. 4, 5266 (2014). https://doi.org/10.1038/srep0526651. J. Hanrieder, L. Gerber, A. P. Sandelius, E. B. Brittebo, A. G. Ewing, and O. Karlsson, ACS Chem. Neurosci. 5, 568 (2014). https://doi.org/10.1021/cn500039b54. J. Hanrieder, P. Malmberg, O. R. Lindberg, J. S. Fletcher, and A. G. Ewing, Anal. Chem. 85, 8741 (2013). https://doi.org/10.1021/ac401830m109. J. Hanrieder, K. Oskar, E. B. Brittebo, P. Malmberg, and A. G. Ewing, Surf. Interface Anal. 46, 375 (2014). https://doi.org/10.1002/sia.5418 Using this information, additional MVA was applied using spectra extracted from each region, similar to the method of Bluestein et al.106106. B. M. Bluestein, F. Morrish, D. J. Graham, J. Guenthoer, D. Hockenbery, P. L. Porter, and L. J. Gamble, Analyst 141, 1947 (2016). https://doi.org/10.1039/C5AN02406D For example, the group5151. J. Hanrieder, L. Gerber, A. P. Sandelius, E. B. Brittebo, A. G. Ewing, and O. Karlsson, ACS Chem. Neurosci. 5, 568 (2014). https://doi.org/10.1021/cn500039b successfully used MAF to discriminate anatomical regions in rat brains to study neonatal exposure to the neurotoxin β-N-methylamino-l-alanine (BMAA). After segmenting the imaging data using MAF, spectra from specific anatomical regions (cornus ammonis 1 and dentate gyrus) were extracted and further analyzed using an orthogonal projection to latent structures by partial least squares−discriminant analysis (OPLS-DA). This revealed changes to specific cellular components that correlated with BMMA exposure. MAF was also used by Nilsson et al.110110. K. Dimovska Nilsson, N. Neittaanmäki, O. Zaar, T. B. Angerer, J. Paoli, and J. S. Fletcher, Biointerphases 15, 041012 (2020). https://doi.org/10.1116/6.0000340 to study tumor heterogeneity in basal cell carcinoma (BCC) sections. MAF better discriminated tumor from healthy tissue, compared to H&E-stained images (Fig. 10). Fragment ions were associated with both highly aggressive and less aggressive tumor islands, suggesting that chemical signatures (mostly from lipids) might be used to grade tumor aggressiveness.We have shown that the NLDR techniques offer unique analytical capabilities when studying tissue with ToF-SIMS, compared with traditional MF techniques. For example, we used ToF-SIMS images of a mouse tumor tissue section (as part of broader work investigating chemotherapy treatments using a mouse model of breast cancer) to exemplify the novel SOM-RPM technique, described earlier.6767. W. Gardner, R. Maliki, S. M. Cutts, B. W. Muir, D. Ballabio, D. A. Winkler, and P. J. Pigram, Anal. Chem. 92, 10450 (2020). https://doi.org/10.1021/acs.analchem.0c00986 This work was primarily algorithm-focused and demonstrated a comparison between similarity maps produced by SOM-RPM, t-SNE, and UMAP [Fig. 11(a)], using both Euclidean and Cosine distance metrics. Each technique produced visually similar results, as shown in Fig. 11(a); however, a more careful evaluation of the distance preservation [Fig. 11(b)] showed differences between the techniques. Most notably, SOM-RPM and UMAP more faithfully represented global distances than t-SNE, and SOM-RPM better preserved local and intermediate distances when Cosine similarity was used. In the Supporting Information associated with this work, we also presented the same analyses of a spheroid section, which showed that all three techniques performed similarly well in terms of the modeling distance information.ToF-SIMS imaging, combined with MVA and ML, has also had limited yet successful application in plant science. Kulkarni et al.111111. P. Kulkarni, M. Dost, O. D. Bulut, A. Welle, S. Bocker, W. Boland, and A. Svatos, Plant J. 93, 193 (2018). https://doi.org/10.1111/tpj.13763 studied Populus trichocarpa leaf surfaces using a combination of ToF-SIMS and PCA, MCR, hierarchical clustering (HC), and spatially aware k-means clustering. They identified regions on the leaf surface with different chemical compositions, revealing segregation patterns for different hydrocarbons, alcohols, and wax esters. Tyler et al.112112. B. J. Tyler, R. E. Peterson, T. G. Lee, F. Draude, A. Pelster, and H. F. Arlinghaus, Biointerphases 11, 02A327 (2016). https://doi.org/10.1116/1.4945326 identified localization of capsaicinoids in Scotch Bonnet peppers using a combination of ToF-SIMS and MCR. They compared their results to previous MALDI-MSI and fluorescence microscopy studies, identifying the high spatial resolution and label-free capabilities of ToF-SIMS as key advantages.

C. Spatially structured organic, bioorganic, and biological systems

Spatially structured systems are defined as those with a well-known spatial distribution that introduces ground truth information to the analysis. This is an important characteristic—so we have provided an extended discussion of these systems. Various categories of spatially structured systems exist that allow the controlled study of materials’ structure-property relationships (SPRs). For example, microarrays of different materials provide a rapid way of generating SPRs for surface interactions with biology. These data are well suited to train ML models that capture SPRs and allow prediction of the properties of new materials (supervised learning); however, unsupervised exploratory analyses are also well documented. These fields have been reviewed recently.113,114113. S. Baudis and M. Behl, Macromol. Rapid Commun. 43, 2100400 (2021). https://doi.org/10.1002/marc.202100400114. J. Kerner, A. Dogan, and H. von Recum, Acta Biomater. 130, 54 (2021). https://doi.org/10.1016/j.actbio.2021.05.053Recently, topography (micropatterning) as well as surface chemistry have been used to instruct biology using surfaces. Topographical biomaterials provide a new avenue for the control of cell fate.115,116115. H. M. Rostam et al., Matter 2, 1564 (2020). https://doi.org/10.1016/j.matt.2020.03.018116. M. J. Vassey et al., Adv. Sci. 7, 1903392 (2020). https://doi.org/10.1002/advs.201903392ToF-SIMS is one of several techniques employed in this field, and MF algorithms have commonly been used to analyze the ToF-SIMS images of these systems. Lee et al.117117. C. Y. Lee, G. M. Harbers, D. W. Grainger, L. J. Gamble, and D. G. Castner, J. Am. Chem. Soc. 129, 9429 (2007). https://doi.org/10.1021/ja071879m used ToF-SIMS and PCA to investigate the heterogeneity and chemical composition of printed spots in a DNA microarray. This approach detected damage to the underlying polymer coating caused by the DNA printing process. Here, unlike many studies in this review, PCA provided information not readily obtainable from single ion imaging. Wickes et al.118118. B. T. Wickes, Y. Kim, and D. G. Castner, Surf. Interface Anal. 35, 640 (2003). https://doi.org/10.1002/sia.1580 combined PCA with ToF-SIMS imaging to investigate polymers and proteins printed onto a Si-based substrate. The primary purpose of this work was to show that denoising algorithms (binning, wavelet filtering, and boxcar filtering) improved PCA performance. No discrimination between different proteins or polymer materials was achieved; however, PCA was able to clearly show the localization of the printed regions on the substrate. Subsequently, Lee et al.119119. T. G. Lee, H. K. Shon, K. B. Lee, J. Kim, I. S. Choi, and D. W. Moon, J. Vac. Sci. Technol. A 24, 1203 (2006). https://doi.org/10.1116/1.2206191 investigated arrays of micropatterned streptavidin and biotinylated cells using ToF-SIMS. PCA was used prior to the incorporation of biotinylated cells to visualize the distribution of streptavidin on the surface. While PCA revealed certain heterogeneities in the streptavidin distribution with higher contrast than the single C9H8N+ ion image, this work was more of an exemplification of PCA, rather than direct application. On the contrary, Shon et al.120120. H. K. Shon, J. G. Son, K. B. Lee, J. Kim, M. S. Kim, I. S. Choi, and T. G. Lee, Bull. Korean Chem. Soc. 34, 815 (2013). https://doi.org/10.5012/bkcs.2013.34.3.815 used PCA to directly compare Ga+, Au1+, and Au3+ primary ions for imaging micropatterns of streptavidin and human epidermal carcinoma A431 cells. Score images from PCA showed that the Au3+ cluster ions produced more accurate representations of the micropatterns, where fluorescence images were used as ground truth references.Aside from PCA, MCR has proven particularly valuable in these fields, given the interpretability of the results and the ability of the algorithm to identify real latent surface chemistry from ToF-SIMS imaging data.32,5332. M. S. Wagner, D. J. Graharn, and D. G. Castner, Appl. Surf. Sci. 252, 6575 (2006). https://doi.org/10.1016/j.apsusc.2006.02.07353. B. J. Tyler, Appl. Surf. Sci. 252, 6875 (2006). https://doi.org/10.1016/j.apsusc.2006.02.160 For example, Scoutaris et al.121121. N. Scoutaris, A. L. Hook, P. R. Gellert, C. J. Roberts, M. R. Alexander, and D. J. Scurr, J. Mater. Sci. Mater.Med. 23, 385 (2012). https://doi.org/10.1007/s10856-011-4474-5 demonstrated how ToF-SIMS imaging and MCR can aid the development of polymeric drug delivery systems. The authors investigated polymer/drug combinations printed onto a glass substrate. ToF-SIMS and MCR successfully discriminated the drug from one of the polymers, revealing a lack of homogeneous mixing of the two components. MCR was also used by Scurr et al.122122. D. J. Scurr, T. Horlacher, M. A. Oberli, D. B. Werz, L. Kroeck, S. Bufali, P. H. Seeberger, A. G. Shard, and M. R. Alexander, Langmuir 26, 17143 (2010). https://doi.org/10.1021/la1029933 to explore the ToF-SIMS images of carbohydrate microarrays. In this case, MCR was first used to discriminate the entire set of pixels, then applied to each single spot in the array. The results provided insights into the heterogeneity of the spots and enabled the formulation of a model describing the process of spot formation. Later, the same group4545. D. J. Scurr, A. L. Hook, J. Burley, P. M. Williams, D. G. Anderson, R. Langer, M. C. Davies, and M. R. Alexander, Surf. Interface Anal. 45, 466 (2013). https://doi.org/10.1002/sia.5040 demonstrated how individual ToF-SIMS images can be stitched together and the combination subjected to MCR analysis. A series of 500 × 500-μm images of eight unique polymer spots from a 572 spot microarray were analyzed using a high-performance computer (HPC) facility. MCR compared each image directly and simultaneously. Component 1 identified the substrate across all eight images, whereas other components identified commonalities in the surface chemistry of specific polymers. Hook et al.4747. A. L. Hook, P. M. Williams, M. R. Alexander, and D. J. Scurr, Biointerphases 10, 019005 (2015). https://doi.org/10.1116/1.4906484 also used MCR to process the ToF-SIMS images of an acrylate polymer microarray consisting of 70 unique printed polymer spots. In order to reduce the computation time and memory requirements of the algorithm, the authors first reduced the 920 × 920-pixel data (containing 706 spectral dimensions) by a factor of 64, then implemented an iteratively expanding algorithm to build the MCR model of the low spatial resolution data, with each iteration introducing an additional factor. The results were used as a starting point for the analysis of the original, high-spatial resolution data, thus reducing the overall computation time associated with random initialization of the scores and loadings matrices. Examples of MCR results from this work are shown in Fig. 12. Iteratively expanding approaches to MCR were reported prior to Hook et al. by Gallagher et al.4444. N. B. Gallagher, J. M. Shaver, E. B. Martin, J. Morris, B. M. Wise, and W. Windig, Chemom. Intell. Lab. Syst. 73, 105 (2004). https://doi.org/10.1016/j.chemolab.2004.04.003 for ToF-SIMS and Raman data.We used a SOM-based workflow to explore ToF-SIMS images of the same acrylate polymer microarray analyzed by Hook et al.47,6247. A. L. Hook, P. M. Williams, M. R. Alexander, and D. J. Scurr, Biointerphases 10, 019005 (2015). https://doi.org/10.1116/1.490648462. W. Gardner, A. L. Hook, M. R. Alexander, D. Ballabio, S. M. Cutts, B. W. Muir, and P. J. Pigram, Anal. Chem. 92, 6587 (2020). https://doi.org/10.1021/acs.analchem.0c00349 We demonstrated almost complete, single-pixel molecular discrimination of the 70 polymers in the array by the SOM using the same automatically generated peak list as in the earlier MCR work. We also provided a comprehensive analysis of the heterogeneities in surface chemistry within a single polymer spot (Fig. 13). By combining the output of the SOM with protein fluorescence data acquired by Hook et al.,121121. N. Scoutaris, A. L. Hook, P. R. Gellert, C. J. Roberts, M. R. Alexander, and D. J. Scurr, J. Mater. Sci. Mater.Med. 23, 385 (2012). https://doi.org/10.1007/s10856-011-4474-5 protein adsorption for each polymer spot was also visualized for this microarray. This approach provided a new method for identifying trends in polymer performance in the context of data topology.6262. W. Gardner, A. L. Hook, M. R. Alexander, D. Ballabio, S. M. Cutts, B. W. Muir, and P. J. Pigram, Anal. Chem. 92, 6587 (2020). https://doi.org/10.1021/acs.analchem.0c00349