Single cell metabolism: current and future trends

Abstract

Background

Single cell metabolomics is an emerging and rapidly developing field that complements developments in single cell analysis by genomics and proteomics. Major goals include mapping and quantifying the metabolome in sufficient detail to provide useful information about cellular function in highly heterogeneous systems such as tissue, ultimately with spatial resolution at the individual cell level. The chemical diversity and dynamic range of metabolites poses particular challenges for detection, identification and quantification.

Aim of Review

In this review we discuss both significant technical issues of measurement and interpretation, and progress toward addressing them, with recent examples from diverse biological systems. We provide a framework for further directions aimed at improving workflow and robustness so that such analyses may become commonly applied, especially in combination with metabolic imaging and single cell transcriptomics and proteomics.

Graphical Abstract

1. INTRODUCTION

The Precision Medicine Task Group of the Metabolomics Society and MANA held a virtual Workshop on Single Cell Metabolomics on Friday February 26, 2021 (Kaddurah-Daouk, Kirwan, Larion, Lane 2021). This summary overview article builds on the presentations and discussions of that meeting, to highlight the current state of the art, and where further effort is needed to enhance metabolic coverage and spatial resolution and address issues of data quality and reliability. Numerous general and specific questions were asked; some of them are addressed directly in the following Discussion, and others remain open questions for the future. Of major interest were the technologies, software and techniques available and being developed for the field. Of special interest for the future is how reliability and quality assurance of the results can be assured, such that they are both reproducible and reflect in-vivo metabolism. In particular we address or highlight the following points that were raised.

• 1What is the impact of sorting dissociated cells on metabolism (Llufrio, Wang, Naser, Patti 2018b), and what are the technical and biological consequences? See Section 2.2
• 2
  What are the technical and biological consequences of fixed cell versus live cell metabolism? See Sections 2.2

  1. How can we expand these technologies to infectious diseases?
  2. How do we enable compatibility with biohazardous processing pipelines? This is important for clinical applications as well, to be able to handle human material under Universal Precautions. See Section 2.1
• 3How will we deal with cellular heterogeneity? Will the error be small enough to be able to determine minor metabolic chances in a single cell as it undergoes malignant transformation? See Sections 2.1, 2.3, 2.4

How will subcellular heterogeneity (cf. compartmentation) be addressed? See Sections 2.3, 2.4

• 4
  How do we ensure quality management?

  1. How to assign confidence of metabolite annotations? See Section 2.5
  2. How to include proper blanks, controls and performance measures? See Section 2.8
  3. How can we ensure throughput and robustness, to enable sufficient numbers for adequate statistical analyses?
• 5What statistical tools are available and what is needed? Is it appropriate to use the same statistics tools as single cell transcriptomics or flow cytometry for single cell metabolomics? See Section: 2.7.

1.1. What is single-cell (sc) metabolomics?

The metabolome is often (circularly) defined as the net content of metabolites in a cell, tissue or system (Beger, Dunn, Schmidt, Gross S.S., Kirwan, Cascante, Brennan et al. 2016; Oliver SG, Winson MK, Kell DB, F 1998; Tweeddale H, Notley-McRobb L, T. 1998). Metabolites may be defined as the consequences of metabolism, i.e. the transformation of one molecule into another via enzyme-catalyzed biochemical reactions. As such most molecules in a system are metabolites, either endogenous, or exogenous, and logically include the macromolecules built by anabolic processes from smaller, activated precursors, using energy derived by catabolic oxidation of a variety of substrates (which are metabolites). If metabolomics comprises the determination of the parts list and the quantitative analysis of the metabolome, then sc metabolomics are the same analyses carried out at the single cell level, which requires the ability to determine the metabolome of individual cells. Further, the cellular metabolome reflects the biochemical activity of each cell in its particular environment. In order to cover the cellular metabolome, technologies developed for analyzing small molecules should be combined with those designed for macromolecules, i.e. transcriptomics, proteomics and glycomics (Baccin, Al-Sabah, Velten, Helbling, Grunschlager, Hernandez-Malmierca, Nombela-Arrieta et al. 2019; Kearney, Vervoort, Ramsbottom, Todorovski, Lelliott, Zethoven, Pijpers et al. 2021; Lundberg, Borner 2019; Minoshima, Ozaki, Odaka, Tateno; Mistry, Greenplate, Ihrie, Irish 2019; Zilionis, Engblom, Pfirschke, Savova, Zemmour, Saatcioglu, Krishnan et al. 2019).

1.2. Why is sc metabolism/omics important?

Measuring the metabolites in a single cell is inherently more difficult than measuring the metabolites in a larger sample, and so the first and overarching question is why make such an effort to make such measurements? If a particular biological question cannot be adequately addressed by analysis of bulk samples, then additional or different approaches are necessary. Where heterogeneity of cellular interactions is fundamental to the biological problem, single-cell analyses become warranted. Tissues in particular are heterogeneous both at the level of the diversity of types of cells present (which have intrinsically different metabolic activities related to their function determined by expression profile) and positional variability of the same cell type within tissue due to cell-cell interactions, different nutrient supply within the tissue determined by proximity to capillaries and metabolic demands of neighboring cells (Cassidy, Caldas, Bruna 2015; Hensley, Faubert, Yuan, Lev-Cohain, Jin, Kim, Jiang et al. 2016; Lane, Higashi, Fan 2020) (cf Fig. 1).

Figure 1. Thin tissue slice of lung cancer- Digital Signal Processing.

Adjacent fresh slices were incubated with ²H glucose+¹³C,¹⁵N Gln and analyzed by NMR and ultra-high resolution mass spectrometry.

This slice (10 μm thickness) was stained for cancer (panCK), CD8 (T cells), CD68 (macrophages) expressing cells and DAPI (localizing nuclei). Circled regions (100 μm diameter) were further analyzed for the protein expression level of 58 different markers of cancer and immune cell functional states using oligonucleotide-barcoded antibodies using the NanoString Digital Spatial Profiling system. H&E stained sections showed the presence of other cell types including fibroblasts and endothelial cells. From Figure 4C of Fan et al. (Fan, Higashi, Song, Daneshmandi, Mahan, Purdom, Bocklage et al. 2021) under creative commons https://creativecommons.org/licenses/by/4.0/

As several exemplars, many diseases progress depending on differences between cells; some tissues such as the brain function because of differences in connections and chemistries between adjacent cells, and embryogenesis involves metabolic changes between cells. If a sufficient number of individual cells is analyzed in detail, then it is possible to generate a population profile of the metabolic properties, whereas a bulk sample analyzed as a whole can be rather difficult to decompose in the absence of spatial information or the specific origin of the averaged levels of various metabolites (Fan, Higashi, Song, Daneshmandi, Mahan, Purdom, Bocklage et al. 2021). However, regional analysis of sequential thinly sectioned tissue slices can provide valuable spatially resolved metabolic information (Dean, Gautham, Siqueira-Neto, McKerrow, Dorrestein, McCall 2021) (Fan, Higashi, Chernyavskaya, Lane 2020; Hossain, Khanam, Dean, Wu, Lostracco-Johnson, Thomas 2020) (Quinn, Melnik, Vrbanac, Fu, Patras, Christy 2020) whereas in situ sc-metabolomics can provide greater detail.

Even knowledge of the chemical variability between so-called identical cells requires SC measurements. Thus single-cell analysis provides information about the state and potentially, the activity of different cell types within a sample, and if spatial information is retained, how activities vary according to position-dependent interactions among cells and nutrient exchange.

As cells can vary greatly in their volumes, connections, and contents, the best approaches to use may be cell type specific. For example, some individual cells for much of their biological lifetime make few long term cell-cell interactions, the most obvious being circulating cells of various kinds (different leukocytes, circulating tumor cells (CTCs) for example). Indeed, CTCs are rare cells (Haber, Velculescu 2014; Keller, Pantel 2019; Zhong, Zhang, Zhu, Liang, Yuan, Li, Li et al. 2020), and they are not all necessarily functionally the same, depending also on their interactions with other tumor cells or with circulating immune cells-only sc level analysis can truly discover the variability within an individual subject. Once these cells infiltrate tissues, their state may change dramatically with a concomitant change of function and thus metabolic activity. Measuring populations of CTCs is thus less desirable than measuring them individually.

At the other extreme, a microliter (1 mm³) of a mammalian brain will contain hundreds of thousands of cells tightly entwined with each other, with the predominant cell types including astrocytes, oligodendrocytes, microglia, and a host of distinct neuronal cell types, and within a cell type, different (activity-dependent) connections underlie functions such as memory. Understanding such a process requires single cell resolution. Both types of applications are driving the development and many of the initial applications of sc metabolomics.

Current applications of sc-MS are revealing highly heterogenous drug concentrations between cells (Newman, Havelund, Passarelli, Marshall, Francis, West, Alexander et al. 2017; Pan, Rao, Kothapalli, Liu, Burgett, Yang 2014; Pan, Standke, Kothapalli, Sun, Bensen, Burgett, Yang 2019). Assessing drug uptake and drug metabolism on an individual cell level is and will be extremely valuable to generate a comprehensive understanding of mechanisms of treatment failure and selection of drug resistance.

Another field in which sc metabolomic analyses will be important is in the field of infectious diseases, either to study individual rare dormant or drug tolerant microorganisms (e.g. (Mikolajczak, Vaughan, Kangwanrangsan, Roobsoong, Fishbaugher, Yimamnuaychok, Rezakhani et al. 2015; Sanchez-Valdez, Padilla, Wang, Orr, Tarleton 2018), or to distinguish between effects of infection on pathogen-containing cells versus bystander effects on adjacent uninfected cells versus matched cells from an uninfected animal (Liu, Ulrich vonBargen, McCall 2021) (Nguyen, Lan, Kane, Haffner, Liu, McCall et al 2022).

In the context of solid tumors, there is tremendous interest in understanding the intratumoral heterogeneity, i.e. the identity of cells within the tumor microenvironment (cf Fig. 1), since heterogeneity is thought to be responsible for drug resistance and the evolution of tumors in time towards novel clones via mutational acquisition (Greaves, Maley 2012; Temko, Tomlinson, Severini, Schuster-Böckler, Graham 2018). In addition, the ability to identify single cancer cells within the normal tissue would benefit surgeons tremendously in determining tumor margins more precisely.

As with all single-cell ‘omics methods, the goals of sc metabolomics are first to identify and quantify the metabolic content of individual cells, in each cell type while maintaining spatial information. This then defines the metabolic phenotype, or functional status of the cells which may be impacted by various factors such as age, genetics, and environmental perturbations such as pathologies, or drug treatment. Such information may also provide the basis for more detailed mechanistic analyses, especially when combined with single-cell transcriptomics and proteomics.

Where metabolism differs from transcriptomics and proteomics is in the tracing of pathways that occur in different compartments, such that the same metabolite can occur at different concentrations within different compartments of the cell, which with some exceptions (Chen, Freinkman, Wang, Birsoy, D.M. 2016) is generally not captured by extraction techniques.

2. DISCUSSION

2.1. Current state of the art- what is possible with existing technology.

Single-cell metabolomics is a rapidly evolving field with a number of recent examples (Guo, Zhang, Le 2021; Liu, Yang 2021; Seydel 2021a; Steinbeck, Fuchs, Negroni, Elsasser, Lichtenauer, Stockdreher, Feitosa-Araujo et al.; Taylor, Lukowski, Anderton 2021). However, because of the scale, there are differences in what information is obtained at the individual cell level. For example, rather than the thousands of metabolites quantified in bulk samples, tens to hundreds of the most abundant metabolites are typically identified in sc experiments. Thus the advantages of single-cell resolution need to outweigh the experimental complexity and issues related to limited depth of metabolomic coverage. Perhaps not surprisingly, there have been several divergent approaches for obtaining single cell measurements. Several of the major approaches covered here include downscaling separations via capillary electrophoresis mass spectrometry, direct single cell mass spectrometry and mass spectrometry imaging at the single cell level. Current work on single-cell metabolism has focused mainly on endogenous cellular heterogeneity and cancer (e.g. (Bensen, Standke, Colby, Kothapalli, Le-McClain, Patten, Tripathi et al. 2021; Lombard-Banek, Li, Portero, Onjiko, Singer, Plotnick, Al Shabeeb et al. 2021; Rappez, Stadler, Triana, Gathungu, Ovchinnikova, Phapale, Heikenwalder et al. 2021a; Tian, Sparvero, Anthonymuthu, Sun, Amoscato, He, Bayır et al. 2021)). In all cases, the metabolites measured reflect cellular content and are distinct from the metabolome determined from biofluids for example. In contrast, there has been limited work in the context of heterogeneity in response to infection. Specifically, most pathogens will only colonize a subset of available cells. Bulk analysis will conflate the metabolome of infected cells with adjacent bystander (uninfected) cells. In contrast, single-cell analyses will enable differentiation of pathogen-containing versus uninfected cells. Applying existing technology to infection models either requires BSL-1 level infectious agents that can safely be handled in the analytical laboratory (Samarah, Khattar, Tran, Stopka, Brantner, Parlanti, Veličković et al. 2020; Samarah, Tran, Stacey, Vertes 2020; Stopka, Wood, Khattar, Agtuca, Abdelmoula, Agar, Stacey et al. 2021), or a method to inactivate pathogens prior to analytical analysis, some of which are already compatible with existing data acquisition methods (e.g. heat fixation (Cazares, Van Tongeren, Costantino, Kenny, Garza, Donnelly, Lane et al. 2015; Wang, Sarathy, Zimmerman, Kaya, Wang, Dartois, Carter 2021), glutaraldehyde fixation (Li, Cheng, Lin, Cao, Yang, Zhang, Shen et al. 2021). Recent work has demonstrated that single-cell metabolism in the context of infection is feasible and provides valuable insight into infection processes, while also enabling direct validation that observed heterogeneity is biologically mediated rather than an analysis artefact (Nguyen, Lan, Kane, Haffner, Liu, McCall et al 2022).

2.2. Experimental design and sample preparation

MS-based metabolomics measurements can be divided into the procedures and protocols related to sampling and sample preparation, the spectrometry/spectroscopy-based measurement and finally, data analyses. The sample preparation for single-cell metabolomics is distinct from larger scale metabolomics measurements. As illustrated in Figure 2, there are in principle several ways in which single-cell metabolism can sample cells including:

1. Capturing and analyzing individual free cells, such as circulating cells in the plasma, including individual immune cells from peripheral blood or lymph or rare circulating tumor cells (Abouleila, Onidani, Ali, Shoji, Kawai, Lim, Kumar et al. 2019; Haber, Velculescu 2014; Keller, Pantel 2019; Zhong, Zhang, Zhu, Liang, Yuan, Li, Li et al. 2020).

2. Selecting specific cells based on markers via manual isolation or automated cell sorting:

3. Assaying populations of individual cells after dissociating tissue into single cells and sorting;

4. Measuring metabolites in tissue slices at the single cell spatial resolution level.

Figure 2: Schematic outline of the approaches used for isolating and preparing cells for single cell measurements.

The workflow includes tissue sectioning and mounting for imaging native distributions of analytes, isolating specific cells, and cell dissociation. Adapted from the American Chemical Society (J. Am. Chem. Soc. 2017, 139, 11, 3920–3929).(Comi, Do, Rubakhin, Sweedler 2017) https://pubs.acs.org/doi/full/10.1021/jacs.6b12822 further permissions related to the material excerpted should be directed to the ACS

Many of these approaches have been used for scSeq genomic/transcriptomic methods (Al-Sabah, Baccin, Haas 2020; Asp, Giacomello, Larsson, Wu, Furth, Qian, Wardell et al. 2019; Baccin, Al-Sabah, Velten, Helbling, Grunschlager, Hernandez-Malmierca, Nombela-Arrieta et al. 2019; Close, Stead, Nsengimana, Reilly, Droop, Wurdak, Mathew et al. 2019; Kalb, Adikari, Hong-Geller, Werner 2019; Kyrochristos, Ziogas, Goussia, Glantzounis, Roukos 2019; Yang, George, Wang 2019; Zilionis, Engblom, Pfirschke, Savova, Zemmour, Saatcioglu, Krishnan et al. 2019) (Bartoschek, Oskolkov, Bocci, Lovrot, Larsson, Sommarin, Madsen et al. 2018; Cohen, Giladi, Gorki, Solodkin, Zada, Hladik, Miklosi et al. 2018; Luecken, Theis 2019; Neal, Li, Zhu, Giangarra, Grzeskowiak, Ju, Liu et al. 2018; Nguyen, Pervolarakis, Blake, Ma, Davis, James, Phung et al. 2018), as well as proteomics (Liu, Chen, Zhang, Liu 2019; Lundberg, Borner 2019; Mistry, Greenplate, Ihrie, Irish 2019). Generally, metabolomics lags these DNA and RNA based technologies (Ali, Abouleila, Shimizu, Hiyama, Emara, Mashaghi, Hankemeier 2019; Gilmore, Heiles, Pieterse 2019; Liu, Chen, Zhang, Liu 2019; Misra 2020; Neumann, Ellis, Triplett, Rubakhin, Sweedler 2019; Seth Nanda, Venkateswaran, Patani, Yuneva 2019; Thiele, Wunderling, Leyendecker 2019), though considerable progress is being made. (Seydel 2021b), (Duncan, Fyrestam, Lanekoff 2019b)

For example, at one extreme, one can select specific cells in a living tissue such as a brain slice, and characterize their electrical activity (and cell type) via the patch clamp method; briefly, this involves patching onto a cell with a pulled pipette so that the electrical activity can be measured; this pipette can also be used to withdraw some of the cytoplasm from the cell soma and introduce it into a characterization method such as CE-MS; in this way, electrical activity and metabolomics can be performed using the same living cell and the sampling does not even kill the cell (Aerts, Louis, Crandall, Govindaiah, Cox, Sweedler 2014) This is a serial sampling approach (one cell at a time) but is information rich.

At the other extreme, the Sweedler group has published a series of articles where they take a tissue and disperse the cells across a slide, label select cells with a probe and determine the cell positions on the slide (Comi, Neumann, Do, Sweedler 2017). They use the positions to determine the spatial coordinates to acquire data using a range of MALDI or SIMS MS-based approaches for direct MS acquisitions. As it only takes a second or so to probe each cell, they can acquire spectra on tens of thousands of individual cells. This has been used for rodent brain cells, islets cells and other animal models (Do, Ellis, Neumann, Comi, Tillmaand, Lenhart, Rubakhin et al. 2018; Jansson, Comi, Rubakhin, Sweedler 2016; Neumann, Ellis, Triplett, Rubakhin, Sweedler 2019). While the approach is high throughput, the number of metabolites and lipids measured is reduced but still provides enough data to differentiate similar cell types; as material is left behind after the MALDI or SIMS measurement, they are able to assay the same spot using other approaches including other types of MS(Comi, Makurath, Philip, Rubakhin, Sweedler 2017; Comi, Neumann, Do, Sweedler 2017; Neumann, Comi, Rubakhin, Sweedler 2019).

The metabolome of a cell can change rapidly to mitigate stress and perform necessary tasks based on changes in their chemical microenvironment (Llufrio, Wang, Naser, Patti 2018a). To measure the resting metabolome of an individual cell, the sample preparation method needs to conserve it without introducing additional stress. There are in principle two strategies: either keep the cell alive in an optimal environment or snap freeze it. For analysis with MALDI or SIMS cells are preferably snap frozen while the currently most popular strategy is to keep cells alive for ambient sampling and ionization (Lanekoff, Sharma, Marques 2022). Ambient sampling followed by electrospray ionization mass spectrometry does not require any particular sample preparation (Figure 3). Instead, metabolites are extracted direct from the cell using a solvent. An organic solvent will directly lyse the cell and enable metabolomics from intact individual cells using a plethora of creative strategies for cells in suspension or cells adhered to a substrate (Figure 4). For subcellular sampling, a dual channel pipette with a continuously flowing solvent can be inserted into the cell’s interior to extract the material for direct transfer to the mass spectrometer inlet (Pan, Rao, Kothapalli, Liu, Burgett, Yang 2014). Alternatively, a micropipette can be inserted into the cell and withdraw material from the cytosol and subcellular compartments for subsequent transfer for mass spectrometric analysis without liquid extraction (Mizuno, Tsuyama, Harada, Masujima 2008; Pan, Rao, Yang 2020). The choice of sampling approach will be highly dependent on the experimental design and research question.

Figure 3:

Sampling a living cell with patch clamp pipette, enabling physiology, morphology and capillary electrophoresis mass spectrometry-based metabolomics on the same cell. Adapted from the American Chemical Society (Anal. Chem. 2014, 86, 6, 3203–3208). https://pubs.acs.org/doi/full/10.1021/ac500168d. Further permissions related to this material should be directed to the ACS

Figure 4. Schematics of some strategies for ambient ionization techniques for single-cell metabolomics.

DESI- Desorption Electrospray Ionization; LDIDD- Laser Desorption/Ionization Droplet Delivery; ESI-electrospray ionization. Reproduced from Duncan et al. (Duncan, Fyrestam, Lanekoff 2019a) with permission from the Royal Society of Chemistry.

2.3. Metabolic Imaging at single cell spatial resolution

2.3.1. Mass spectrometry-based metabolic imaging

Imaging mass spectrometry (IMS)

Imaging mass spectrometry (IMS) enables cellular metabolite analysis while preserving the spatial architecture of the specimen of interest and has significant potential to be impactful in measuring metabolic processes at single- and subcellular resolutions. Since the inception of laser desorption ionization (LDI) by Koichi Tanaka (Tanaka, Waki, Ido, Akita, Yoshida, Yoshida, Matsuo 1988), through the advances by Michael Karas and Franz Hillenkamp (Hillenkamp, Karas, Beavis, Chait 1991; Karas, Bachmann, Hillenkamp 1985), who demonstrated that small molecules could be used as a matrix to assist laser desorption ionization (MALDI), the analytical focus of these laser-based ionization methods has focused on molecules with higher mass-to-charge (m/z) ratio (e.g., peptides, polymers, and proteins). Although the amino acids tryptophan and alanine, along with another small-molecule (Vitamin B12, cobalamin), were among the first analytes ionized and analyzed by Karas and Hillenkamp using MALDI, the sensitivity for detection of molecules with low m/z (70–500 Da) (Hillenkamp, Karas, Beavis, Chait 1991; Karas, Bachmann, Hillenkamp 1985), such as metabolites involved in central metabolism has been limited. In large part this is attributable to spectral interference of the matrix compounds required for ionization or other ion-suppressive compounds (such as lipids) in the low m/z range. Recent advances in analytical chemistry and instrumentation have facilitated the development of MALDI-IMS-based methods on cryosectioned specimens while achieving spatially resolved metabolite measurements at 5–100 μm lateral resolution with significantly higher metabolic coverage.

High spatial resolution MS imaging is beginning to show some promise beyond protein analysis, including with stable isotope tracers (Wang L, Xing X, Zeng X, Jackson SR, TeSlaa T, Yang L, McReynolds M et al. 2022). MS imaging using MALDI -FT MS is especially well suited analyzing for surface protein glycosylation, with spatial resolution determined by the laser spot size which is typically in the range 10–50 μm, close to individual cellular resolution (Hawkinson, Sun 2022; Powers, Neely, Shao, Tang, Troyer, Mehta, Haab et al. 2014) Using longer wavelength lasers (in the near IR), more labile sugars such as sialic acid were preserved in paraffin embedded blocks (Pace, Angel, Drake, Muddiman 2022). As indicated in Figure 4, there are several techniques for ionization under ambient conditions in addition to MALDI-based methods. In addition to MALDI, direct Laser Ablation Electrospray Ionization (LAESI) using near IR lasers has been used for metabolite imaging (Nemes, Barton, Li, Vertes 2008; Shrestha, Patt, Vertes 2011) (Roman, Walsh, Oh, Dana, Hong, Jo, Alleyne et al. 2018) which typically use a large spot size (≈200 μm) though more recent work using fiber optics greatly increased resolution to the single cell level (Kiss, Hopfgartner 2016).

2.3.2. Iso-imaging.

Whereas most metabolomics analyses focus on measuring absolute or quasi steady-state abundances of metabolites, additional information about pathway utilization and flux analysis can be obtained by stable isotope tracing either in vitro or in vivo (DeBerardinis, Mancuso, Daikhin, Nissim, Yudkoff, Wehrli, Thompson 2007; Le, Lane, Hamaker, Bose, Gouw, Barbi, Tsukamoto et al. 2012; Sellers, Fox, Bousamra, Slone, Higashi, Miller, Wang et al. 2015) (Niedenfuhr, Wiechert, Noh 2015). To date, this approach has been limited to bulk samples, but spatial information would greatly enhance the information retrieval and modeling.

Infusions of stable-isotope-labeled nutrients have been coupled to MALDI-IMS (“Iso-imaging”) to determine spatial metabolic activity (Wang L, Xing X, Zeng X, Jackson SR, TeSlaa T, Yang L, McReynolds M et al. 2022). Coupling stable-isotope tracing to MALDI-IMS avoids several problems commonly associated with making metabolic measurements using MALDI-IMS. First, spatial differences in tissue composition result in differential ion suppression across the same tissue, a situation analogous to the ‘[tissue-] matrix effect’ in conventional liquid-chromatography coupled to mass spectrometry. Therefore, without isotope labeling, it is difficult to know whether observed metabolite changes across a tissue are biologically meaningful or the result of regional ion suppression. Second, stable-isotope labeling patterns can be compared to those obtained by LCMS and offer advantages for compound validation if isotopologue distributions match between methods.

Isotopically enriched metabolites can also be isolated and fragmented by MS2, and fragmentation patterns can be compared across regions or tissues or with other mass-spectrometry modalities. Third, stable-isotope tracing enables the potential of performing additional calculations to either determine relative nutrient contribution or perform a more rigorous flux analysis (Niedenfuhr, Wiechert, Noh 2015) (Bartman, Shen, Lee, TeSlaa, Jankowski, Wang, Yang et al. 2021; Selivanov, Marin, Tarragó-Celada, Lane, Higashi, Fan, de Ataur et al. 2020). Although stable-isotope labeling approaches can add additional information and analytical rigor to MALDI-IMS experiments, unlabeled experiments that assess regional metabolite or lipid levels have also demonstrated utility in various settings. Combining MALDI-IMS data obtained in isotope-labeled or -unlabeled studies with other existing and emerging ‘spatial-omics’ methods promises to address critical biological questions surrounding regional nutrient use.

2.3.3. The Future of MALDI-IMS.

Important advances in MALDI-IMS technology and accompanying analytical chemistry-based strategies will facilitate the successful implementation of MALDI for broad coverage of the metabolome at the single-cell level. Atmospheric pressure (AP) MALDI has recently exhibited 1.4 μm lateral spatial resolution (Kompauer, Heiles, Spengler 2017). The application of this technology for small molecules is still hindered by the general sensitivity for molecules in the low m/z range. Improving metabolite sensitivity to sufficient levels will likely result from strategies that will enhance ionization or reduce ion suppression from interfering compounds (Yang, Ji, Guan, Li, Zhang, Zhao, Mao 2018), along with hardware advances that better select small molecules in the gas phase. Metabolome coverage using any single matrix is relatively incomplete compared to conventional chromatography-based methods. The further development of different matrices (Qiao, Lissel 2021), including reactive matrices (matrices that act as derivatizing agents and absorb the UV laser pulse essential for MALDI) (Shariatgorji, Nilsson, Fridjonsdottir, Vallianatou, Kallback, Katan, Savmarker et al. 2019) and those matrices that may be compatible with mixed-mode (positive-negative switching) may improve the metabolome coverage and throughput. The relative speed at which data are acquired will be crucial as lateral resolution decreases. Fourier-transform Ion Cyclotron Resonance (FT-ICR)-based mass spectrometers provide higher mass accuracy, which is essential for distinguishing isotope labeling patterns but is relatively slow compared to Time-of-flight (TOF) based mass-spectrometry. Without chromatographic separation, molecule identification is challenging, and isobaric (molecules with the same mass) are indecipherable. Ion mobility, the process of separating ions in the gas phase based on their collisional cross-section (CCS) with a carrier gas, would theoretically enhance sensitivity and aid in identifying molecules based on an added time dimension (similar in principle to chromatography). While ion-mobility has been particularly useful for peptides and lipids, it is unclear whether the utility will be extended to enhance small-molecule metabolites. For certain application, targeted imaging by MALDI or DESI by tandem quadrupole MS can identify predefined metabolites of interest, albeit not at single cell resolution (Lamont, Eijkel, Jones, Flinders, Ellis, Siegel, Heeren et al. 2018; Lamont, Hadavi, Viehmann, Flinders, Heeren, Vreeken, Siegel 2021). Breakthroughs in any of these areas could lead to significant improvements in making spatially-resolved metabolite measurements.

2.3.4. Optical Imaging

Optical imaging techniques have been explored to quantify tumor cell metabolism non-destructively at cellular level resolution by using endogenous contrast or appropriate probes (Madonna, Fox, Crouch, Lee, Zhu, Martinez, Alvarez et al. 2019b; Walsh, Cook, Sanders, Aurisicchio, Ciliberto, Arteaga, Skala 2014b; Zhu, Li, Vincent, Martin, Crouch, Martinez, Madonna et al. 2019; Zhu, Martin, Crouch, Martinez, Li, Palmer, Dewhirst et al. 2018). For example, two endogenous fluorophores, reduced nicotinamide adenine dinucleotide (NADH) and flavin adenine dinucleotide (FAD) have been explored (Zhu, Chen, Chui, Tan, Liu 2014) to report the reduction-oxidation (redox) state in the electron transport chain of cancer cells (Hou, Wright, Chan, Tran, Razorenova, Potma, Tromberg 2016). Multi-photon microscopes have been explored for cellular level optical redox ratio imaging of in vitro cancer cells (Hou, Wright, Chan, Tran, Razorenova, Potma, Tromberg 2016) and tumor organoids (Walsh, Cook, Sanders, Aurisicchio, Ciliberto, Arteaga, Skala 2014a) for cancer research. These cutting-edge microscopes quantify the auto-fluorescence of both NADH and FAD, and then provide an indirect measurement of the balance between glycolysis and OXPHOS by looking at the ratio of the two. Although it is an indirect measure of redox balance, it has been firmly validated with Seahorse assays in cancer research studies (Hou, Wright, Chan, Tran, Razorenova, Potma, Tromberg 2016). Being able to simultaneously quantify both glycolysis and mitochondrial metabolism directly and explicitly would further facilitate cancer research and translational applications. To achieve this capability, several metabolic probes have been explored for intracellular metabolic measurements. The 2-[N-(7-nitrobenz-2-oxa-1, 3-diazol-4-yl) amino]-2-deoxy-D-glucose (2-NBDG) has been used in cancer cells to report glucose uptake, similar to the clinically available FDG-PET (Yamada, Saito, Matsuoka, Inagaki 2007; Zhu, Li, Vincent, Martin, Crouch, Martinez, Madonna et al. 2018). TMRE has been utilized to quantify mitochondrial membrane potential (MMP) to study OXPHOS (Perry, Norman, Barbieri, Brown, Gelbard 2011; Zhu, Martinez, Martin, Li, Crouch, Carlson, Haystead et al. 2017). BODIPY has been explored to report fatty acid uptake in cancer cells (Megan C Madonna 2021). All of these three key metabolic parameters are associated with human cancer therapy resistance development (Tang, Wei, Wu, He, Shi, Xiong, Gong et al. 2018). When these metabolic probes were used, a labeling procedure was required but it was non-destructive and lasted less than 1 hour (Madonna, Fox, Crouch, Lee, Zhu, Martinez, Alvarez et al. 2019a). Optical metabolic imaging enables repeatable non-destructive measurement on various cancer models and it provides spatial information, while these techniques have poor coverage which is limited by usable metabolic probes. However, an increasing number of fluorescence labeled metabolites are becoming available, substantially enhancing the metabolic coverage of optical microscopy (Benson, Fernandez, Barth, de Moliner, Horrocks, Herrington, Abad et al. 2019; Hong, Pawel, Pei 2021).

Redox biochemistry, specifically the determination of NAD(P)⁺, NAD(P)H and their ratios, in different compartments within live cells is an important area of research that has seen the development of numerous fluorescent probes that bind the free nucleotides and can be imaged by confocal microscopy. These studies enable the redox responses to experimental conditions in different cell types to be determined as a function of time (Bilan, Belousov 2017; Cambronne, Stewart, Kim, Jones-Brunette, Morgan, Farrens, Cohen et al. 2016; Hu, Wu, Walker, Wang, Tian, Wang 1800; Sallin, Reymond, Gondrand, Raith, Koch, Johnsson 2018; Steinbeck, Fuchs, Negroni, Elsasser, Lichtenauer, Stockdreher, Feitosa-Araujo et al. 2020; Zhao, Yang 2016). For example, although the total dinucleotide concentration is of the order 0.5 mM in mammalian cells, the concentrations and redox ratios are significantly different in the cytoplasm versus the mitochondria, and indeed, the free concentrations are much lower than the total amounts, owing to binding to the large number of dehydrogenases present. This also leads to NAD⁺/NADH and NADPH/NADP⁺ ratios quite different from the ratios of the totals. Typically, the free NAD⁺/NADH ratio in the cytoplasm is 500–1500 (Christensen, Karlsson, Winther, Jensen, Lerche 2014; Zhao, Yang 2016) compared with <10 in mitochondria (Williamson, Lund, Krebs 1967). In contrast, the free NADPH/NADP⁺ ratios are in the range 20–80 (cytoplasm) and ca. 2-fold higher in mitochondria, whereas the free NAD⁺ concentration is lower in the cytoplasm (ca. 100 μM) than in the mitochondria (Sallin, Reymond, Gondrand, Raith, Koch, Johnsson 2018).

2.3.5. Raman Imaging Spectroscopy

Owing to recent progress in the diode lasers and notch filters, the study of biological systems became possible in the form of Raman Imaging Microscopy. Focusing the laser through the microscope allows the collection of Raman spectra for each pixel and subsequent movement of the stage permits the acquisition of spectra in a new position (Shen, Hu, Min 2019). This approach provides a chemical and biochemical map of the entire cell without the need to remove it from the cellular milieu. Therefore, the technique is non-destructive, non-invasive, and can accept a wide range of samples from bulk to microscopic, from solids to liquids or gasses. The differences in the intensity and positioning of peaks in the spectra are attributed to the compositional changes between the different cells or organelles. The variation in the intensity of these peaks is due to the variation in the amount of lipids, proteins and DNA contents present in that location of the cell. In the past, the lack of methodology has hampered the use of Raman in biological samples. Recently, the development of a single organelle approach for the quantification of DNA, proteins, RNA, and lipids has opened a new way to study spatial metabolism (Lita, Kuzmin, Pliss, Baev, Rzhevskii, Gilbert, Larion et al. 2019; Lita, Pliss, Kuzmin, Yamasaki, Zhang, Dowdy, Burks et al. 2021; Pliss, Kuzmin, Lita, Kumar, Celiku, Atilla-Gokcumen, Gokcumen et al. 2021). With the introduction of a deconvolution method and the use of external markers, it is now possible to accurately quantify biomolecular components such as total DNA, RNA, lipids, and proteins in organelles of live cells. Lipids could be characterized using the lipid unsaturation parameter (LSU) which describes the ratio of double bonds to single bonds; the TCP parameter which describes the ratio of trans/cis in the lipid isomer; phosphatidylcholine sphingomyelin (PC); cholesterol and cholesteryl ester (CL); and triglycerides (TG). In addition, the spatial resolution of 500 nm positions this technique at the frontiers of metabolic analysis.

Since the Raman microscope is not able to visualize organelles and distinguish one organelle from another, as some other methods can, Figures 5 and 6, show how Lita et al. (Lita, Pliss, Kuzmin, Yamasaki, Zhang, Dowdy, Burks et al. 2021) circumvented this. They labeled one organelle with a fluorescent dye, and with the help of a fluorescence filter they visualize the desired organelle is in the field of view. Once they have the organelle identified, they can start recording the metabolic information that is intrinsically hidden in the Raman spectra. Lita et al., (Lita, Pliss, Kuzmin, Yamasaki, Zhang, Dowdy, Burks et al. 2021) applied this approach to gliomas cell that contains either IDH1^wt or IDH1^mut and identified a unique lipid profile associated with IDH1^mut, which could be exploited for therapeutic purposes.

Figure 5. Overview of organelle level metabolic characterization of cells.

Using the Raman Microscope, a single cell can be selected from the tumor microenvironment (A) and a more detailed image of organelle inside this cell can be obtained (B). Using our methodology, different classes of lipids in one organelle in different cell types can be quantified. Each organelle measurement is displayed with a different color (lysosomes in blue, mitochondria in green, Golgi apparatus in red and Endoplasmic reticulum in black. R132H is an active site mutation in isocitrate dehydrogenase, very prevalent in lower grade gliomas. From (Lita, Pliss, Kuzmin, Yamasaki, Zhang, Dowdy, Burks et al. 2021) with permission.

Figure 6. Protocol for Raman-based quantification of lipids at the organelle level.

In Step 1, a green, fluorescent label is applied to the cells. In Step 2, organelles are localized and focused under microscope using this tag. In Step 3 and 4, Raman spectra are collected and BCAbox algorithm is applied to extract lipid profiles. In the last step, the unsaturation of lipid (LSU) is computed based upon the area ratio of 1655 cm⁻¹ to 1443 cm⁻¹. Via the same procedure, TCP parameter, which characterize trans/cis C=C bonds ratio in lipid species can be obtained from the intensities at 1666 cm⁻¹ and 1655 cm⁻¹. AG, Golgi apparatus, LD, lipid droplets. With permission from (Lita, Kuzmin, Pliss, Baev, Rzhevskii, Gilbert, Larion et al. 2019)

Another way to image metabolites is to use a special type of Raman spectroscopy named Stimulated Raman Spectroscopy (SRS) as done by Zhang et. al. (Zhang, Shi, Shen, Miao, Wei, Qian, Liu et al. 2019). Contrary to the spontaneous Raman, this method uses two synchronized pulsed lasers, and the difference frequency between the lasers, accelerates the transitions of molecular vibrations that match that difference. By this method the signal can be enhanced up to 10⁸ fold, therefore allowing detection of low intensity signals.

This approach was used in conjunction with deuterated glucose to image the incorporation of glucose into proteins, lipids, DNA, and glycogen in tissues of live animals and cells. The use of deuterium allows for specific pathways to be visualized as more enriched in glucose-derived than others, while the spatial resolution of the technique allows for tracing of this isotope in different types of cells or organelles.

While these efforts are pushing the limit of current detection with Raman spectroscopy both in terms of metabolite identification as well as signal sensitivity, more developments are needed for this method to increase the number of metabolites that it can detect. In the future, the use of Raman spectroscopy in combination with mass spectrometry such as Nano ESI could increase the potential of subcellular metabolomics (Ali, Abouleila, Shimizu, Hiyama, Watanabe, Yanagida, Germond 2019). Combining with sc proteomics and transcriptomics and the limited Stable Isotope Resolved Metabolomics (SIRM) information obtainable (see above)(Lima, Muhamadali, Goodacre 2022) and improved computational modeling will make the in situ metabolic imaging more powerful

2.4. Technical bottlenecks and new approaches

There are several technical bottlenecks for deep metabolic analysis at single cell resolution (see for example (Shrestha 2020)), that are being addressed with new approaches as discussed below.

2.4.1. Spatial dispersion and sensitivity

Nuclear magnetic resonance (NMR) and high-performance liquid chromatography mass spectrometry (HPLC-MS) are the “gold standards” for metabolite identification. However, the low sensitivity of NMR means that comparatively large amounts of compound are needed, such that Magnetic Resonance Imaging (MRI) cannot achieve single cell or near single cell resolution. Sensitivity enhancement via dynamic nuclear polarization (DNP) for example allows for localized kinetic analysis of a few reactions, but is still limited to mm³ sized volumes (Ehrhardt, Gallagher, McLean, Schönlieb 2022). The great advantage of MRI is that the imaging is in vivo and can be carried out at greater depth than optical imaging (Ahn, Hwang, Nam, Jin, Woo, Park 2019; Dang, Bardhan, Qi, Gu, Eze, Lin, Kataria et al. 2019). Mass spectrometry is a far more sensitive technique but since it consumes material the sensitivity is inversely proportional to the square of the probed area, and being destructive, does not readily lend itself to kinetic studies at the same sampling point. From an analytical perspective, this creates what can be termed the “Metabolite Uncertainty Principle” where the more certain we are about a metabolite’s identity, the less certain we are about its localization. Many of the pioneering efforts over the last decades including sample preparation for spatial dispersion or enhancements in sensitivity for imaging have focused on escaping this frustrating confine. The evolution of these techniques is illustrated in Fig. 7 with the width of each evolutionary line indicating the approximate number of instruments in the field and the color of each label categorizing the certainty of metabolite identity from gold (fully identified, level 1) through silver (putative annotation, levels 2,3) to bronze (class of metabolite, level 4) according to the classifications in to bronze (class of metabolite, level 4) according to the classifications in (Blaženović, Kind, Ji, Fiehn 2018; Sumner, Amberg, Barrett, Beger, Beale, Daykin, Fan et al. 2007).

Figure 7. Analytical techniques for metabolite identification and their evolution towards single cell metabolomics using either spatial dispersion methodologies or by imaging.

The width of each evolutionary line indicates the approximate number of instruments in the field (see key) and the color categorizes the certainty of metabolite identify from gold (full identification), silver (annotated) to bronze (class of metabolite, e.g. not specifically identified). LDI Laser Desorption Ionization; LAESI ablation electrospray ionization; SIMS secondary ion mass spectrometry; TERS tip-enhanced Raman spectroscopy.

Innovation has been driven by the needs to increase sensitivity and simultaneously reduce the probe size. In electrospray MS this has been achieved extracting material with a fine capillary as a probe guided by a stereo microscope followed by nanospray ionization (Masujima 2009; Mizuno, Tsuyama, Masujima 2014). Whilst not an imaging method per se, it is possible to precisely target locations at the organelle scale and from live cells (Fujii, Matsuda, Tejedor, Esaki, Sakane, Mizuno, Tsuyama et al. 2015; Mizuno, Tsuyama, Date, Takanori Harada, Masujima 2008). Laser capture micro-dissection (LCMD) may also be combined with electrospray and is increasingly used to give multi-omic information with spatial localization. Desorption electrospray ionization (DESI) (Takats, Wiseman, Gologan, Cooks 2004) uses an electrospray as the probe with MS analysis of analytes captured in secondary droplets. This has the advantage of no sample preparation requirement and gentle electrospray ionization enabling improved confidence in metabolite identification but with a spatial resolution of above 50 μm (Gilmore, Heiles, Pieterse 2019).

In the desorption mass spectrometries, the probability of a neutral molecule being ionized is low (typically < 1 × 10⁻³) so that much effort has gone into increasing this probability through combining with a laser to post-ionize neutral compounds. The conversion of a neutral molecule in the gas phase to an ion requires an energy of around 6 – 8 eV (579–769 kJ mol⁻¹) so photons in the UV and vacuum UV (120–210 nm) are needed. Niehaus et al. developed a MALDI-2 instrument where a second 266 nm laser was used to post-ionize neutral molecules enabling a resolution of 600 nm to be achieved (Niehaus, Soltwisch, Belov, Dreisewerd 2019) and Walker et al used VUV post-ionization femtosecond laser desorption/ionization mass spectrometry (LDI-MS) to image the distribution of lipids in pancreas tissue (Walker, Gelb, Barry, Subanajouy, Poudel, Hara, Veryovkin et al. 2018).

Reducing the uncertainty of metabolite identify requires high-performance mass spectrometers which includes high m/z resolving power to separate confounding peaks, high mass accuracy for searching databases of calculated or known mass peaks and MS/MS capabilities for structural identification. These attributes, generally, require Fourier Transform (FT)-MS using magnetic trapping (e.g. FT-ICR) or electrostatic trapping (e.g. Orbitrap). Sub-cellular imaging with both MALDI and SIMS in combination with Orbitrap Mass Spectrometry detection has been achieved in SMALDI (Kompauer, Heiles, Spengler 2017), OrbiSIMS (Passarelli, Pirkl, Moellers, Grinfeld, Kollmer, Havelund, Newman et al. 2017) and MALDI-2 (Niehaus, Soltwisch, Belov, Dreisewerd 2019). Analysis of samples in their native state is another important technological driver. Cryogenic sample preparation is also an important direction, for example Cryo-OrbiSIMS where the sophisticated sample preparation methods developed for cryo-SEM are utilized (Zhang, Brown, Scurr, Bullen, MacLellan-Gibson, Williams, Alexander et al. 2020).

Raman and IR optical methods have the advantage of operating directly in ambient conditions. Similarly to MS, there have been tremendous advances over the last decades to boost signal intensities using non-linear methods. For example, Stimulated Raman Spectroscopy (SRS) is able to achieve high-speed imaging (usually at a selected frequency rather than broadband) with a spatial resolution of 400 nm and an axial resolution of < 1 μm. The spatial resolution of Raman and IR spectroscopies is substantially increased when combined with electric field enhancement caused by the sharp tip of an Atomic Force Microscopy (AFM probe) in Tip Enhanced Raman Spectroscopy (TERS) and IR-AFM, respectively.

2.4.2. Sensitivity

Sensitivity and dynamic range are a significant issue for metabolomics. Metabolite concentrations in cells (averaged over the entire cell) vary widely from abundant molecules such as ATP in the mM range, to amino acids and dinucleotides in the sub millimolar range, to low μM for many intermediates, and in the nM range for some signaling molecules, a range of more than 6 orders of magnitude. For a cell of total volume 0.5–5 pL, this range is equivalent to zmol to fmol amounts. However, modern mass spectrometers can easily measure sub fmol to amol quantities, and the more abundant metabolites (100 μM to mM) are accessible by confocal light microscopy on live cells (see Section 2.3.4). Many biological questions can be addressed even with moderate metabolic coverage, especially when multiple modalities are engaged on the system as a whole. The fast progress in the field attests that attomole to femtomole levels provides sufficient coverage to address may biochemical questions, and there have even been examples of individual organelle MS (Rubakhin, Garden, Fuller, Sweedler 2000) (and see above).

2.4.3. Microfluidics

The small number of single cells presents another issue for metabolomics. As mentioned above, there is a broad range of metabolite concentrations in a given cell; however, the typical mammalian cell volume is in the picolitre scale. Therefore, controlling the sampled volume throughout the analytical workflow and preventing sample loss is essential to maintain the ability to detect the less abundant, but highly relevant molecules such as signaling molecules. As it stands, traditional sample preparation and separation techniques are ill equipped to handle such small volumes. However, techniques that inherently constraint the sample volume in a microfluidics format by means of slugs or droplets are much more suited to the task. Furthermore, separation techniques that can be incorporated into microfluidic chips such as capillary electrophoresis, or that occur after the sample introduction to the MS (ion mobility) can improve confidence in metabolite identification, in contrast to methods that rely on MS measurements alone. The work done by Li et al. (Li, Zhao, Hu, Yi-Ming Liu 2016) highlights the potential of microfluidics in single cell analysis. Their work showcased an automated microfluidic platform that lyses cells using nanoelectrodes followed by a separation step using electrophoresis, and subsequent introduction into the MS instrument via nanospray.

Controlling perturbations of the cells’ microenvironment and the metabolic response to them is an important variable of any metabolomic experiment. This is especially so when targeting relatively rare cells such as circulating tumor cells, where the generated data are highly sensitive to fluctuations and outliers due to the low number of cells measured, which makes it difficult to discern the source of variation (e.g. heterogeneity, or perturbations of microenvironment). Therefore, minimizing the perturbations to the microenvironment, as well as rapid quenching of metabolic reactions ensure that the variations seen in the metabolome representative to the true cellular state, and not in response to experimental conditions. Controlling perturbations is especially important in methods where the cell is sampled while it is still alive as is the case in capillary microsampling techniques. Considering that metabolic reactions occur in seconds (Tung, Blischak, Hsiao, Knowles, Burnett, Pritchard, Gilad 2017b), quenching the sampled cells by a cold mixture of organic solvent and buffer (Reichard, Asosingh 2019) is required prior to further cell manipulation and sample preparation to ensure that the measured metabolites represent the normal cellular state as much as possible (Hoa Thi Nhu, Ang, Chevrier, Zhang, Lee, Goh, Chen 2020; Tung, Blischak, Hsiao, Knowles, Burnett, Pritchard, Gilad 2017a).

2.5. Metabolite identification.

As already described, NMR is insufficiently sensitive to detect and analyze even high abundance metabolites at the single cell level, though is important for determining metabolites in bulk samples, which may support conclusions based on high resolution experiments. The main source of information is mass spectrometry which with sufficient resolution can return accurate masses from which molecular formulae can be determined (Higashi 2011; Lane, Fan, Xie, Moseley, Higashi 2009). This does not discriminate among structural isomers, for which tandem MS at ultrahigh resolution (preferably with LC retention time) is the gold standard (Sun, Fan, Lane, Higashi 2021), but may not always be technically possible. In many cases, analysis of an additional bulk sample is still employed to assist peak assignment; the greater spectral intensity enables better fragmentation data leading to improved analyte identification. Increasing mass spectrometer sensitivity can be expected to help with tandem MS of lower abundance an analytes. Other (orthogonal) biochemical information can help in deciding plausible identification.

Stereochemical information will be lacking with most LC-MS/MS-based annotation practices, which may matter for chiral drug metabolites, although there are recent examples of probing the stereochemistry of compounds in individual cells using CE-MS coupled to ion mobility MS (Mast, Liao, Romanova, Sweedler 2021) and this is expected to become more common with the growth of ion mobility enhanced spectrometers. Even with MS/MS, annotation may not always be possible, either due to lack of diverse enough reference libraries combined with persistent (though improving) limitations of computational structure prediction approaches, or insufficiently informative MS/MS spectra (for example for glycerophosphocholines in positive ion mode, where often only fragments from the head group are observed and there is a dearth of informative fragments from the fatty acid tails).

Optical/microscopy methods provide high resolution metabolic imaging on cell models (Madonna, Fox, Crouch, Lee, Zhu, Martinez, Alvarez et al. 2019c; Shah, Beckler, Walsh, Jones, Pohlmann, Skala 2014; Shah, Heaster, Skala 2017), organoid model (Gil, Deming, Skala 2021; Walsh, Cook, Sanders, Aurisicchio, Ciliberto, Arteaga, Skala 2014b) or even tumor slices without extraction. However, it is challenging to use the technique for absolute amounts quantification as most of the existing optical techniques are semi-quantitative or ratio-metric (Gil, Deming, Skala 2021) based. It is possible to quantify the molecular amounts using optical techniques, while rigorous calibration procedures are typically needed.

2.6. Metabolite quantification

Quantification of metabolites in single cells with mass spectrometry is analytically challenging but desirable from several perspectives. In particular, quantification will enable the direct comparison of the abundance of individual metabolites in the cell despite their different ionization efficiency and detectability. Furthermore, quantification will allow for calculation of significance and fold changes between cells and groups of cells that will otherwise be highly dependent on the sensitivity of individual metabolites. For liquid extraction approaches, internal standards are included in the extraction and ionization solvent to measure the detected concentration of individual metabolites with one point calibration (Bergman, Lanekoff 2017). The internal standards are typically isotopically labeled to minimize overlap with endogenous signals. Another approach is to perform relative quantification using the ratio of two similar metabolites and assuming that they ionize similarly(Bergman, Lanekoff 2017). Although this does not provide a quantitative measure, it allows for direct comparison between the metabolome of individual cells. Absolute quantification remains a significant challenge in metabolomics.

2.7. Statistical tools-what is available and what is needed?

While the statistical analysis of single-cell metabolomics remains in its infancy, single cell transcriptomics have matured to the point where sophisticated statistical tools for high-dimensional data can be routinely applied as part of standard data processing workflows (Luecken, Theis 2019). This is largely facilitated by the widespread adoption among the scRNA community of mature open-source data analysis packages written in modern scripting languages, such as ScanPy (Wolf, Angerer, Theis 2018) for Python, its extension SquidPy for spatial omics (Palla, Spitzer, Klein, Fischer, Schaar, Kuemmerle, Rybakov et al. 2022), and Seurat (Hao, Hao, Andersen-Nissen, Mauck, Zheng, Butler, Lee et al. 2021) for R. Such tools can also be applied to MS-based single-cell metabolomics data once mass spectra are converted to intensities of annotated metabolites. Rappez et al. (Rappez, Stadler, Triana, Gathungu, Ovchinnikova, Phapale, Heikenwalder et al. 2021b) used ScanPy to perform statistical analyses on a dataset of 740 metabolites from 29,738 hepatocytes, including cell-to-cell normalization, nonlinear batch correction, UMAP dimensionality reduction (Becht, McInnes, Healy, Dutertre, Kwok, Ng, Ginhoux et al. 2019), Leiden clustering, and pseudotime trajectory analysis (Trapnell, Cacchiarelli, Grimsby, Pokharel, Li, Morse, Lennon et al. 2014). This revealed three clusters of cells corresponding to homeostasis, steatosis, and an intermediate metabolic state, for which marker metabolites were identified via hypothesis tests of differential expression. The ScanPy package was also used (Castro, Xie, Rubakhin, Romanova, Sweedler 2021b) to carry out clustering and differential expression testing for mass spectra of single organelles, and by Yuan et al. (Yuan, Zhou, Cai, Pan, Sun, Qumu, Yu et al. 2021) for differential expression analysis, dimensionality reduction, and clustering of mass spectra of single hepatocyte nuclei in-situ.

However, single-cell metabolomics data differs from other modalities in ways that may present obstacles to further cross-application of existing statistical methods. First, MS, optical, and vibrational metabolomics data are natively continuous-valued signals, whereas sequencing data is count-valued and often sparse. The meanings of zeroes also differ between transcriptomics and MS-based metabolomics: while the technical versus biological origins of sparsity in scRNA have been the subject of considerable debate (Jiang, Sun, Song, Li 2022; Svensson 2020), metabolite peaks can fail to be detected in mass spectra for computational reasons: e.g. if they insufficiently exceed background spectral noise during peak detection (Alexandrov 2012), or are mis-annotated due to mass drift (La Rocca, Kune, Tiquet, Stuart, Eppe, Alexandrov, De Pauw et al. 2021). While such events may be digitally recorded as zeroes, they are better understood and modelled as a form of random missingness (Do, Wahl, Raffler, Molnos, Laimighofer, Adamski, Suhre et al. 2018).

Methods based upon MS imaging of single cells also risk co-sampling of neighboring cells: although Rappez et al. (Rappez, Stadler, Triana, Gathungu, Ovchinnikova, Phapale, Heikenwalder et al. 2021b) indicate this minimally affected their analysis of cultured hepatocytes, this may bias estimates of spatial autocorrelation of metabolites in tissues where cells are tightly packed (and irregularly shaped) relative to the MS spatial resolution. Finally, a significant promise of single-cell omics is its potential (through sufficiently large sample sizes) to infer regulatory relationships de-novo between biomolecules. Many algorithms toward this end have been developed for single-cell RNA data (Aibar, Gonzalez-Blas, Moerman, Van, Imrichova, Hulselmans, Rambow et al. 2017; Chan, Stumpf, Babtie 2017; Pratapa, Jalihal, Law, Bharadwaj, Murali 2020), as well as for mass cytometry (Sachs, Perez, Pe’er, Lauffenburger, Nolan 2005; Wang, Lin, Sontag, Sorger 2019). While in principle such methods could also be applied toward learning metabolic networks from data, we caution these may not be robust to sources of (co-) variation specific to MS-based single-cell metabolomics, such as in-source fragmentation (Dolatmoradi, Samarah, Vertes 2021), region-dependent ion suppression (Taylor, Dexter, Bunch 2018), or global metabolic shifts arising from cell manipulation (Llufrio, Wang, Naser, Patti 2018c): these may induce correlations between metabolites that are primarily technical rather than biological in origin.

A goal at times is to determine cell heterogeneity within a population of cells and to understand what biochemical species drive this heterogeneity. While sc MS can determine mass markers that drive cell differences, the gold standard for determining cell type remains immunohistochemistry with established markers; Neumann et al. (Neumann, Comi, Rubakhin, Sweedler 2019) performed single cell MALDI MS and then followed up on the same cells using immunohistochemical staining via GFAP and neurofilament antibodies and thus were able to correlate both datasets and determine the differences in lipids between astrocytes and neurons (Neumann, Comi, Rubakhin, Sweedler 2019) The lipid markers discovered later have been shown to be consistent with data obtained on a distinct platform (Neumann, Ellis, Triplett, Rubakhin, Sweedler 2019). Using these large datasets of thousands of individual rodent brain single cell MALDI MS spectra, Xie et al created a new statistical pipeline based on interpretable machine learning. Using these same datasets, they were able to classify cells based on their original location (e.g., hippocampus and cerebellum) and found they could classify cells 96% of the time (Xie, Castro, Bell, Rubakhin, Sweedler 2020). This does not necessarily mean that the remaining 4% of the cells are wrongly classified, but perhaps up to 4% of the cells do not have features that allow such classification. Additional studies use larger datasets such as 150,000 human brain cells (Bhaduri, Neumann, Kriegstein, Sweedler 2021).These approaches have recently been adapted to work with single organelle MS measurements (Castro, Xie, Rubakhin, Romanova, Sweedler 2021a)

2.8. QA/QC- how to include proper blanks, controls and performance measures?

The goals of QA/QC

A scientific measurement must be reliable enough that it enables robust conclusions to be drawn from the experiment. Quality management (QM) is the process of ensuring such reliability. It is divided into quality assurance and quality control. Quality assurance describes the procedures that are undertaken in the laboratory to have a reasonable confidence that quality standards will be met. Quality control describes the points in the process where specific measurements are recorded to demonstrate that the quality level has been achieved (Barwick 2020). QM gives the researcher confidence in their biological results and enables experimental results to be reproducible in other laboratories. As such, it is an essential part of good scientific practice. However, the exact requirements for QM will be method specific. Given the novelty of much of the latest single cell technologies, QM is still very much an active area of development in this field. Ultimately, being able to reproduce the results in a different independent biological cohort, preferably in a different laboratory is the ultimate quality assurance test, and is improved still further where a second alternative method can be applied e.g. optical imaging (see above) with Seahorse assays (Little, Kovalenko, Goo, Hong, Kerk, Yates, Purohit et al. 2020; Wu, Neilson, Swift, Moran, Tamagnine, Parslow, Armistead et al. 2007) although this is not available to most studies.

QM has some particular challenges in single cell metabolomics, not only because of new technology, but also because, by its definition, it is difficult to conduct proper tests of reproducibility on single cell data. Traditional methods in metabolomics for measuring repeatability or reproducibility across an experiment by identifying a suitable quality control sample for regular measurement is more challenging in SC metabolomics, especially given that small sample volumes often preclude replicate measurements on the same samples. However, several of the other QA practices can be taken direct from more traditional metabolomics. Readers are directed to some seminal and some more recent reviews on this e.g. (Beger, Dunn, Bandukwala, Bethan, Broadhurst, Clish, Dasari et al. 2019; Broadhurst, Goodacre, Reinke, Kuligowski, Wilson, Lewis, Dunn 2018; Sumner, Amberg, Barrett, Beger, Beale, Daykin, Fan et al. 2007).

All experiments should have a clear, easy to follow experimental protocol. This should highlight not only the important steps in the procedure, but also which steps have been previously shown to be most critical for quality and reproducibility. This reduces the risk that new technologies are only successful in the hands of a particular scientist. While written protocols are still most useful in the lab, supplementary video protocols are often a clearer way of sharing new technologies. Standard experimental practice such as randomizing the order in which samples are analyzed (and the order of analysis of rastering patterns in the case of metabolic imaging), are also to be strongly recommended.

As for any metabolomics experiments, high quality reagents, especially solvents are required. Before starting an analysis, an instrument should be calibrated, and demonstrated to be properly working with an appropriate system suitability test (SST). While manufacturers often provide SST reagents to use for such a purpose, some laboratories prefer to use their own, more complex mixtures. This enables a much better overview of the complete scan range of interest. Both calibration and the SST results can be used as points of quality control. As such, their results should be recorded (Broadhurst, Goodacre, Reinke, Kuligowski, Wilson, Lewis, Dunn 2018).

Appropriate blank samples should also be used, both to demonstrate that the instrument is free from contamination before starting, and from carryover midway through experimentation, and the software must be set up to identify impurities from the sample preparation and instrumental artefacts as extensive manual curation is impractical.

In some technologies such as optical microscopy, a background fluorescence needs to be measured as a starting point (Sun, Zhu 2021). For mass spectrometry and similar methods, process blanks (i.e. solvent mixtures that have followed the exact processing steps of the samples) enable identification and removal or concentration adjustment of peaks that appear to come from the processing steps themselves.

Since single cell analyses are likely to be subject to the same intra and inter-batch effects seen in more traditional metabolomics analysis, this should be monitored where possible. This has typically been achieved with some form of representative quality control (QC) sample which can be repeatably analyzed across and between batches. There are already proven batch effects in most currently used methods for single cell analysis (Balluff, Hopf, Siegel, Grabsch, Heeren 2021; Rappez, Stadler, Triana, Gathungu, Ovchinnikova, Phapale, Heikenwalder et al. 2021b). The ideal QC sample is a pooled sample consisting of aliquots of all, or a representative cohort, of the samples to be analyzed. This mimics better the matrix (and therefore the matrix effects) of the samples themselves and should also enable detection of a very similar range of metabolites. For obvious reasons, such a QC sample is difficult to achieve in most single cell analyses with the exception of some techniques which extract cell contents before analysis. Alternatives to a traditional QC have included the analysis of technical replicates (Choi, Munoz-Llancao, Manzini, Nemes 2021; Nemes, Knolhoff, Rubakhin, Sweedler 2012; Onjiko, Morris, Moody, Nemes 2016), as well as several of the approaches already described above including the use of larger cellular populations to obtain robust results, the measurement of the cells using more than one platform and monitoring reproducibility and precision of internal standards.

Some methods have additional requirements for QM, such as addressing the effects of variation in slice or matrix thickness on MALDI results for example. For many technologies, such considerations still need to be defined and appropriate solutions found.

So far, the discussion has concentrated on technical validation; that an instrument is carrying out reliable and reproducible measurements. However, biological degradation as a result of the process itself needs also to be considered. As with other metabolomics measurements, single cell metabolomics typically requires multiple steps to process samples before the point of an analysis. Some methods also take many hours (e.g. MALDI imaging). Both may lead to post-harvesting changes being measured as part of the metabolome. Currently, little research has been published on how these processes may themselves be affecting the measurements. Some methods lend themselves better to testing this hypothesis than others e.g. analyzing monocultures of cells before and after FACs sorting may enable an estimate of how much the FACS process changes the metabolome. Degradation, which occurs over the time course of an analysis, is easier to measure, either in technical replicates (Nemes, Knolhoff, Rubakhin, Sweedler 2012) of samples, or by analyzing the trend of analyses over a single slide in the case of MALDI imaging. Some recent developments such as fluidic force microscopy (Guillaume-Gentil, Rey, Kiefer, Ibanez, Steinhoff, Bronnimann, Dorwling-Carter et al. 2017) claim to circumvent this problem by extracting intracellular fluid directly from living cells in a two-step process. Obviously, short lived metabolites (e.g., nitric oxide) would be difficult to measure with all of these approaches.

Many single cell metabolomics methods rely on relative differences between groups of cells. However, as the technologies move toward true quantification, quantitative methods need themselves to be validated. Existing guidelines from the EMA and FDA will probably require some adaptation to allow for the unique challenges of sc metabolomics.

3. Future Directions: prospects for integration with large scale ‘omics

Major advances have been made in metabolic analyses of individual cells that provide biological mechanisms and complement the information available from metabolomics studies of biofluids. Nevertheless, there remain several technological issues. In addition to numerous technical developments to increase sensitivity and reliability of identification and quantification discussed above, it is desirable to integrate metabolic information with protein and gene expression data. As changes in metabolite abundances can have more than one origin (altered synthesis routes and/or altered utilization routes), for mechanistic purposes (and overall confidence), it may be important to determine the levels of the relevant enzyme isoforms (enzyme catalyzed rates are usually proportional to enzyme concentration), and any post translational modifications that may impact the intrinsic enzymic activity independent of the expression level. If enzyme abundances are important determinants of the fluxes and metabolite levels, then transcriptomics can determine whether this is transcriptionally controlled versus post translationally (protein degradation). The metabolite levels corresponding to enzyme activities (or groups of enzymes) may also shed light on the possible importance of non-covalent enzyme modifications of intrinsic activity (cf. allosteric interactions). There have been attempts to link, for example, metabolomics information with transcriptomics (Al-Sabah, Baccin, Haas 2020; Baccin, Al-Sabah, Velten, Helbling, Grunschlager, Hernandez-Malmierca, Nombela-Arrieta et al. 2019; Fan 2012; Zhang, He, Tan, Budhu, Gaedcke, Ghadimi, Ried et al. 2013), such as using the confirmation from isotope tracer methods to interrogate RNAseq data sets with metabolic modeling (Fan, Bruntz, Yang, Song, Chernyavskaya, Deng, Zhang et al. 2019). As it is now possible to apply metabolomics methods at the single cell level, at least for relatively abundant metabolites, as well as proteomics and near single cell resolution spatially resolved transcriptomics (e.g. the Visium system from 10X Genomics and single cell resolution via CosMx from NanoString) major issues are corresponding coverage of the different omics levels, and to what extent these ‘omics can be acquired on the same cells, or at least similar cells from (for example) adjacent regions of tissue. Different workflows might be envisaged, such that tissues are first interrogated by non-destructive imaging modalities prior to fixing, staining and subsequent proteomic/transcriptomic processing. Incorporating metabolic tracer information even for a few metabolites greatly increases the information content in the context of interacting metabolic pathways.

To do justice to the very large amounts of quantitative data that such multiomics approaches imply, a major growth area will be in the construction of detailed integrative models of cells in complex (tissue) environments, incorporating large data bases (Lorenzi, Chiu, Cobos, Gross, Volders, Cannoodt, Nuytens et al. 2021; Muraro, Dharmadhikari, Grun, Groen, Dielen, Jansen, van Gurp et al. 2016; Robinson, Kocabas, Wang, Cholley, Cook, Nilsson, Anton et al. 2020; Thul, Lindskog 2018; Wang, Robinson, Kocabas, Gustafsson, Anton, Cholley, Huang et al. 2021; Wishart, Guo, Oler, Wang, Anjum, Peters, Dizon et al. 2022) and informatics approaches to enhancing overall throughput.