An isotopically-labelled temporal mass spectrometry imaging data analysis workflow to reveal glucose spatial metabolism patterns in bovine lens tissue

Abstract

Application of stable isotopically labelled (SIL) molecules in Matrix-Assisted Laser Desorption/Ionization Mass Spectrometry Imaging (MALDI-MSI) over a series of time points allows the temporal and spatial dynamics of biochemical reactions to be tracked in a biological system. However, these large kinetic MSI datasets and the inherent variability of biological replicates presents significant challenges to the rapid analysis of the data. In addition, manual annotation of downstream SIL metabolites involves human input to carefully analyse the data based on prior knowledge and personal expertise. To overcome these challenges to the analysis of spatiotemporal MALDI-MSI data and improve the efficiency of SIL metabolite identification, a bioinformatics pipeline has been developed and demonstrated by analysing normal bovine lens glucose metabolism as a model system. The pipeline consists of spatial alignment to mitigate the impact of sample variability and ensure spatial comparability of the temporal data, dimensionality reduction to rapidly map regional metabolic distinctions within the tissue, and metabolite annotation coupled with pathway enrichment modules to summarise and display the metabolic pathways induced by the treatment. This pipeline will be valuable for the spatial metabolomics community to analyse kinetic MALDI-MSI datasets, enabling rapid characterisation of spatio-temporal metabolic patterns from tissues of interest.

Introduction

Matrix-assisted laser desorption/ionisation mass spectrometry imaging (MALDI-MSI) surveys the spatial complexity of biomolecules to reveal insights into normal tissue physiological function and disease mechanisms¹. Hundreds of signals that represent biomolecules are detected simultaneously, with spatial information at a resolution approaching single-cell level^2,3. Diseases such as cancer⁴, neurological disorders⁵, metabolic disorders⁶, infectious diseases⁷ and those that affect special senses such as vision⁸ have been studied. In ocular lens research, MALDI-MSI has been used to map the spatial distribution of lipids⁹, proteins¹⁰ and metabolites¹¹, and has previously revealed distinct metabolic activity regions. A major limitation is that MALDI-MSI often only captures a single point in time. Since the metabolome is highly dynamic, this limits the information that can be gathered from a single experiment. Furthermore, detection of only endogenous metabolites provides limited information on the underlying physiology of either normal tissue or that with aberrant metabolism and tissue pathology. To address this, an exogenously administered stable isotopically-labelled (SIL) molecule can be introduced, and its corresponding downstream metabolites detected since the masses of the metabolites are directly related to the incorporated labelled compound^6,12,13.

While this ‘iso-imaging’ approach is often used to assess the metabolism of a SIL molecule at a single timepoint, ‘kinetic’ MALDI-MSI data can utilise multiple timepoints to generate time-series data. Similar to a standard MALDI-MSI experiment, either ‘kinetic’ or ‘iso-imaging’ approaches produce a large amount of data. The efficient distinction of labelled compounds from other signals in the data can be particularly challenging. There are commercially available software packages to discriminate mass spectral signals between treatment groups, and automated open-source solutions to annotate endogenous metabolites detected in MALDI-MSI data at the tissue level^14,15 and at the single cell level¹⁶. More recently, an open-source approach has been used to quantify SIL tracer incorporation and metabolic pathway activity¹⁷, and spatially-resolve metabolic flux in tumours¹³. However, annotation of labelled compounds and the reconstruction of metabolic pathways from either iso-imaging or kinetic MALDI-MSI data has remained a manual process⁶. A potential limitation of this approach is that not all SIL compounds are detected from the high dimensional data. In addition, despite the availability of tools such as Metaspace for annotating metabolites based on MALDI-MSI data, these tools do not integrate SIL and/or unlabelled metabolites into metabolic pathways¹⁴. Conversely, existing empirical knowledge-based annotation tools such as the mummichog algorithm are able to annotate endogenous metabolites and assign them to activated pathways, but they cannot annotate downstream SIL metabolites. To address this growing need to identify enriched SIL metabolites within biological pathways in kinetic MSI data, a novel bioinformatics pipeline is required.

Recently, a kinetic MSI approach was developed¹⁸ and used to investigate uptake, transport and metabolism of glucose in the bovine ocular lens¹⁹. Glucose is the primary nutrient in the lens from which cellular energy is derived and is utilised to actively maintain lens tissue transparency over many decades of life. The unique structure of the lens (Fig. 1a)²⁰ contains lens cells of different ages that have different phospholipid components^21,22, crystallin proteins²³ and types of metabolism. This structure leads to inherently different, yet relatively spatially predictable, regions of metabolic activity, where a metabolically active outer cortex that is capable of aerobic metabolism due to the presence of mitochondria surrounds a less metabolically active central region. Moreover, the anterior lens epithelium features some different enzyme activities and metabolic pathways to the differentiating fibre cells. These spatial metabolic differences were visualised recently by exposing ex vivo bovine lenses to (U¹³–C₆) glucose via incubation in artificial aqueous humour. This set of experiments showed that (U¹³–C₆) glucose was taken up initially in the equatorial region of the lens, and that over time it distributed first to the lens cortex, and then to the core (Fig. 1b)¹⁹. In addition, these data showed the metabolic profiles of the lens anterior and posterior surfaces are not symmetric, where due to the presence of the anterior epithelium, the anterior region sees more rapid consumption of glucose than the posterior region.

Figure 1.

(a) Schematic diagram of the bovine lens. The anterior surface of the lens is covered by a single layer of cuboidal epithelial cells that divide at the lens equator and elongate to form secondary lens fibre cells. These fibre cells populate the bulk of the lens and through a cell differentiation process degrade their nuclei and organelles such as mitochondria in order to eliminate these potential light-scattering particles from the path of light transmission^24,25. Since the cell differentiation process continues throughout life, fibre cells are laid down in a series of concentric growth layers around an internalized core to produce an inherent cell age gradient across the lens. The figure is made by Biorender. (b) A sequence of ion images to visualise biological dynamics of SIL glucose uptake in incubated bovine lenses¹⁹. All the lenses in the figure are orientated with the anterior surface towards right. A sequence of snapshots was captured from axial sections of bovine lenses that were incubated in AAH with 5 mM SIL glucose for a duration ranging from 5 min to 20 h. The images reveal distinct regional variations in the behaviour of SIL glucose (m/z = 221.0528).

Since known spatial metabolic differences could be visualised with this kinetic imaging approach, the lens is an ideal model system to develop an automated annotation pipeline for kinetic MSI studies. Such a pipeline would ideally process large volumes of data in a fraction of the time it would take for manual interpretation and automatically identify labelled compounds and metabolic pathways detected in the tissue. In the study presented here, the development of a data analysis pipeline for kinetic MSI to facilitate the efficient construction of metabolic dynamic models is described. This pipeline integrates MALDI-MSI data pre-processing, tissue spatial overlay, machine learning algorithms to find potential spatial clusters and the mummichog algorithm to achieve identification of labelled and unlabelled metabolites with metabolic pathway enrichment analysis.

Results

The data analysis pipeline consists of three main modules: spatial overlay, dimensionality reduction, and a combined metabolite annotation/pathway mapping step (Fig. 2). The resulting output consists of annotated m/z features and ion images that represent both unlabelled (endogenous) and SIL (exogenous) metabolites grouped by metabolic pathway. Each module utilises existing R scripts to curate the data for optimised and accurate annotation and is presented in turn.

Figure 2.

Schematic diagram of the computational pipeline. (i) Lens coordination position alignment and rescaling; (ii) Lens-to-lens hexagonal binning; (iii) Data table reshaping; (iv) t-SNE and UMAP analysis; (v) Principle component analysis and top loading feature selection; (vi) K means segmentation; (vii) SIL metabolite database construction; (viii) 1st round metabolomic annotation and pathway enrichment; (ix) Feature scoring and FDR controlled filtering; (x) 2nd round metabolomic annotation and pathway enrichment within filtered features; (xi) Feature visualization (ion images and kinetic images).

Module 1: Spatial alignment and data reshaping

The time-series iso-imaging experiment comprised three biological replicates for each of the ten incubation time points, totalling thirty sampled lens tissue sections. Despite using a consistent, stable supply of tissue, there was variability in the size and shape of both the incubated lenses and the analysed tissue sections. This necessitated a spatial alignment process to allow the subtly different shapes of lens sections to be overlaid for metabolism to then be monitored in spatially consistent tissue regions. Raw pixel data and coordinates could not be used for lens-to-lens alignment since the different sizes of lens sections gave a variable number of pixels in each sample section.

To address this, all tissue sections were first standardised to fit a common x and y coordinate frame. The length of the equatorial poles on the lens sections were normalised to a reference lens section shape while preserving the original aspect ratios. The lens tissue section data from each time point were then loaded into a shared coordinate frame. Then, a lens-to-lens hexagonal binning (hexbin) approach was employed. Hexbins were chosen based on their ability to tessellate and form a grid across any polygon²⁶. In this approach, hexbins were uniformly applied to the coordinate frame for spatial alignment with a median pixel count of 9 in each hexagon. These hexbins followed a consistent indexing approach across all samples. This systematic procedure guaranteed that hexbins occupying the same raster position always share the same index and closely matched the same anatomical position throughout the entire time series. Additionally, the indexes probed spatial coordinates, indicating the location on the lens sections. Once aligned, an average spectrum was computed for each hexbin, resulting in a dataset combining the spatial indexes of the hexbin with average mass spectra representing biological replicates and incubation timepoints. By averaging raw mass spectra into each hexbin spectrum, more stable spectral signals were produced for each hexbin, which can mitigate the effects of noise or variability in the MALDI data. Each hexbin was also assigned to an anatomical lens region (anterior, posterior, equator, inner cortex and core) based on conventions in the lens field (see Supplementary Fig. 1). These manually assigned anatomical labels have previously been used to quantify SIL metabolite signal levels over time¹⁹, and allowed temporal signal intensity patterns to be plotted for each region (see Supplementary Fig. 2). In addition, this approach allowed preliminary assessment of the performance of our automated annotation pipeline (see Module 2, Fig. 4a,b).

Figure 4.

Visualisation of the spatial variation of the lens MALDI data at reduced data space. (a) Schematic diagram indicating the lens anatomical labels to enhance the interpretability of the dimensionality reduced data. The lens data is classified mainly into 5 groups, including anterior (red), posterior (green), equator (purple), core (blue). Regions coloured with grey are not defined. (b) Dimensionality reduction for visualising the kinetic lens MALDI data using t-SNE, UMAP and PCA with predefined lens anatomical labels. (c) Unsupervised machine learning algorithm K means to automatically segment the dimensionality reduced lens data based on specification of k = 6 (colours show individual segments). (d) Mapping K-means clusters onto the standardized lens to visualize the spatial clustering pattern.

At this point in the workflow, output data was structured such that each row represents a detected m/z feature, and each column represents a spatial position (i.e. hexbin index) from a single MS scan. However, the inclusion of spatial and temporal data in the columns of a general dataset poses a challenge in distinguishing whether changes in signals arise from spatial positioning or evolve over time during the incubation process. Consequently, to identify the metabolites responsible for the variations in spatial metabolic patterns, the dataset underwent a transformation into a reshaped data table where each column contains the full time profile data for a single hexbin.

Module 2: Dimensionality reduction

While the reshaped data table accommodated changes to m/z features due to incubation time and spatial position, it led to a substantial number of m/z features within each hexbin index, because the m/z features in this dataset are associated with time points. This implies a tenfold increase in the number of features (due to the 10 incubation timepoints), resulting in a high-dimensional dataset that necessitates dimensionality reduction for improved data interpretability.

t-SNE (t-Distributed Stochastic Neighbour Embedding) and UMAP (Uniform Manifold Approximation and Projection) have been used in metabolomics analysis to map the high-dimensional data onto a 2D space while preserving the local structure and relationships between the data points^27,28. In our study, the time profile data underwent separate executions of these two algorithms. The outputs were then projected to a standardised lens shape to visualise the spatial variation profiles within the reduced data space (Fig. 3a,b). A symmetric pattern was observed in t-SNE_1 and UMAP_1, while spatial differences between anterior and posterior are evident in the other dimensions (t-SNE_2, UMAP_2). PCA (Principal Component Analysis) was then applied to the reshaped data table and the first 5 principal components (PCs) were used to project to the standardised lens shape to visualise the spatial variation profile (Fig. 3c). The lens projections indicate that PC1 (11.51%), PC2 (3.57%), and PC3 (1.81%) exhibit a closely symmetric profile, whereas PC4 (1.41%) and PC5 (1.33%) clearly display an asymmetric pattern between the anterior and posterior lens.

Figure 3.

Single dimension projection of t-SNE (a), UMAP (b) and PCA (c) to bovine lens based on the transformed coordinates of each hexbin in the lower-dimensional spaces. The PCA results are PC1 to PC5 from left to right. All lens diagrams are orientated with anterior surface to the right. The Blue-White-Red continuous color scale was employed to show the hexbin loading image from the lowest (blue) to the highest (red) in the dimension.

To further investigate whether the reduced data space could profile the difference between the metabolic patterns of the lens anterior and posterior regions, biplots were used to show the combined effects of the dimensionality (Fig. 4, Supplementary Fig. 5). Since these data reduction methods (t-SNE, UMAP and PCA) are all unsupervised methods, the temporal hexbin data could undergo dimensionality reduction regardless of the lens anatomical labels that were previously assigned. To aid visualisation of the lens anatomical regions that were grouped by the algorithm in biplots, outer cortical lens regions were shrunk to 30% of their original sizes (Fig. 4a). These manual segmentation labels were added to the data to assist in interpretation of the clustering of the data. From the t-SNE and UMAP in Fig. 4b, the core (blue) and equator (purple) clusters can be well separated, while the anterior (red) and posterior (green) clusters do not have a very clear boundary but could be resolved in one dimension (t-SNE_2, UMAP_2). From the PCA biplot, the subplot of PC1 versus PC2 explained most of the variation within all the components, where the equatorial region (purple colour) and core region (blue colour) were well classified. While hexbins labelled ‘anterior’ (red colour) and ‘posterior’ (green colour) were highly overlapped in PC1 versus PC2, they were more tightly clustered in PC4 versus PC5 (Fig. 4). Together these results suggested that data reduction methods could be used to automatically highlight m/z signals that changed over time and were potentially derived from (U–¹³C₆) glucose introduced during lens incubations.

Previously, manual labelling was used to group spectra into anatomical regions (see Supplementary Fig. 1b) for m/z feature time profile generation and interpretation. However this approach can introduce biases due to the arbitrary nature of the labelling. Therefore, segmentation of the dimensionality-reduced data was used to enhance the accuracy of the spatially-grouped spectra. The automatically generated spatial clusters can then be projected onto the lens to compare with the manually assigned anatomical regions (see Supplementary Fig. 1). For this procedure, the K-means algorithm was employed to spatially cluster the reduced data space (Fig. 4c). Importantly, this process utilised only the feature space and did not directly incorporate hexbin spatial information. The value of K in the algorithm was carefully considered, and K = 6 was an appropriate value that ensured that spatial variations could be discovered while minimising over-segmentation (see Supplementary Fig. 3). Projection of spatial K-means analysis onto each individual lens section (n = 30) following hexbin application showed very similar clustering (see Supplementary Fig. 4), suggesting that this approach was able to align lens anatomical regions across all timepoints.

From the lens projections, the anterior and posterior lens regions were separated in t-SNE, UMAP and PCA dimensions PC4 versus PC5, but not PC1 versus PC2 (Fig. 4d). The primary variability in the data (represented by PC1 and PC2) showed a symmetric concentric ring pattern, with the lens outer cortex, inner cortex, and core regions evident. This matches known lens anatomical structures and patterns of endogenous small molecules, such as primary metabolites and phospholipids²⁹. In contrast, PC4 versus PC5 fit the expected asymmetric pattern that was previously observed in lenses incubated in SIL glucose¹⁹. While SIL glucose signal derived from ex vivo incubation and the changes over time did not contribute a large proportion of the total variability in the dataset, the automated segmentation of the reduced data space aligns with both manual segmentation and the anticipated results derived from the anatomical structure of the lens and ex vivo incubation. Together, these results gave confidence that the spatial alignment approach was valid and that the observed asymmetry in the data-reduced space could be utlised in subsequent signal annotation.

t-SNE and UMAP do not explicitly consider information on the relative contribution of each feature to the resulting dimensionality-reduced data. Therefore, while they may not be ideal for tasks such as feature selection and pathway enrichment, these methods consider full variation in the data to effectively capture the most prominent patterns or information present in the data (e.g. symmetric pattern from lens natural development and asymmetric pattern led by ex vivo SIL glucose incubation). This results in a more complete overview of variation reduction results where PCA shows a limited proportion of variations in each PC. Compared to PCA, t-SNE and UMAP are more straightforward to demonstrate the spatial variation, which provides insights for the interpretation of PCA results. The reliability of PCA results with similar patterns is enhanced with corroboration by t-SNE and UMAP support. Apart from the ability of dimensionality reduction, PCA also highlights the feature contribution in each PC. This characteristic facilitates the identification of key m/z features that underly the spatial segmentation of the lens, which could therefore be utilised in Module 3 for annotation and metabolic pathway analysis. In the case of the lens, there is an expectation of metabolic differences, leading to spatial clustering differences between the anterior and posterior regions. Therefore, the m/z features that significantly influenced PC4 and PC5 were likely the features responsible for the observed metabolic asymmetry between the anterior and posterior lens regions. This suggests that our automated pipeline is able to identify distinct metabolic activity regions in lens tissue. We leveraged the m/z features that contribute the most to the observed variation PC4 and PC5 in the following module to automatically annotate metabolites and determine enriched metabolic pathway activity in a spatial context.

Module 3: Metabolite annotation, pathway enrichment, and visualization

In an iso-imaging experiment, the biological tissue can simultaneously consume both SIL and naturally occurring metabolites, resulting in a mixture of labelled and unlabelled metabolites whose levels change over time and can therefore be key features of PCA dimensions. Therefore, to implement the automated metabolic activity analysis pipeline, the top 30 features from every dimension underwent the mummichog algorithm using a library that included both labelled and unlabelled metabolites that we previously generated (see “Metabolite library construction” in “Methods”). The mummichog algorithm is very sensitive while performing metabolite annotation and pathway enrichment, and led to many false positives following the first round of the annotation (Table 1). To remove false annotations in the 1st round annotation list, feature scoring and FDR-controlled filtering were employed to select only high-quality m/z features and assign them to potentially enriched pathways. Under this restriction, there is more confidence in the pathways that were identified as enriched since they are based on high-quality m/z feature annotations (Table 2). A comparison of annotations and pathways resulting from both rounds of annotation are shown in Supplementary Fig. 6. The number of enriched pathways was reduced from 14 to 5, which resulted in more biologically accurate information being produced. Finally, the automated pipeline allows for generation and display of the individual annotated ion images that are determined to change over time in a metabolic pathway context (see Fig. 5, for other pathway ion images see Supplementary Figs. 8–11).

Table 1.

1st round pathway enrichment result.

Pathway	Category	Dim.1	Dim.2	Dim.3	Dim.4	Dim.5
Amino sugar and nucleotide sugar metabolism	m/z Normal	13	9	19	15	12
	m/z SIL	5	7	11	10	8
	Sig level	*		***	***	**
Ascorbate and aldarate metabolism	m/z Normal	4	8	4	4	2
	m/z SIL	0	3	2	3	1
	Sig level		*
Fructose and mannose metabolism	m/z Normal	5	8	10	7	3
	m/z SIL	0	4	3	5	2
	Sig level		**	*
Galactose metabolism	m/z Normal	10	7	13	11	8
	m/z SIL	6	6	9	8	8
	Sig level	**		***	**	*
Inositol phosphate metabolism	m/z Normal	6	5	6	7	3
	m/z SIL	0	5	5	6	3
	Sig level				**
Metabolism of xenobiotics by cytochrome P450	m/z Normal	11	14	21	19	18
	m/z SIL	10	4	12	8	10
	Sig level					**
One carbon pool by folate	m/z Normal	1	2	3	3	3
	m/z SIL	7	3	6	3	5
	Sig level	**				*
Pentose phosphate pathway	m/z Normal	5	7	8	6	3
	m/z SIL	0	4	3	3	2
	Sig level		*
Phosphatidylinositol signaling system	m/z Normal	6	4	6	7	3
	m/z SIL	0	5	5	6	3
	Sig level			*	**
Purine metabolism	m/z Normal	18	19	29	24	25
	m/z SIL	9	12	19	15	16
	Sig level	*		*		**
Pyrimidine metabolism	m/z Normal	11	12	19	20	14
	m/z SIL	3	6	9	6	7
	Sig level		*		*
Starch and sucrose metabolism	m/z Normal	7	4	8	8	5
	m/z SIL	3	6	7	8	6
	Sig level	**	**	***	***	***
Sulfur metabolism	m/z Normal	3	2	3	3	3
	m/z SIL	1	0	1	0	1
	Sig level	*
Tryptophan metabolism	m/z Normal	2	2	2	2	2
	m/z SIL	12	9	15	12	11
	Sig level	*

Significance level: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’(No symbol) 1. All the counted compounds are significant hits.

Table 2.

2nd round pathway enrichment result (All m/z features passed 10% FDR).

Pathway	Category	Dim.1	Dim.2	Dim.3	Dim.4	Dim.5
Fructose and mannose metabolism	m/z Normal	5	5	8	6	3
	m/z SIL	0	3	2	4	1
	Sig level				*
Glycolysis or Gluconeogenesis	m/z Normal	5	3	5	5	2
	m/z SIL	0	2	2	3	1
	Sig level			*	*
Phosphatidylinositol signaling system	m/z Normal	5	3	6	7	2
	m/z SIL	0	4	4	5	2
	Sig level				*
Purine metabolism	m/z Normal	14	10	15	13	14
	m/z SIL	4	5	6	6	5
	Sig level	*				***
Starch and sucrose metabolism	m/z Normal	6	3	6	7	3
	m/z SIL	2	4	5	6	4
	Sig level				**

Significance level: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’(No symbol) 1. All the counted compounds are significant hits.

Figure 5.

Time-series metabolite maps for annotated metabolites in starch and sucrose metabolism. All lenses are orientated with anterior to the right. PC = the principal component number(s) that the annotated metabolite shows significant loading in.

One prominent pathway that was identified in the dataset was “Starch and sucrose metabolism”. The m/z images that were grouped in this pathway are shown in Fig. 5 showing it is most active in the lens outer cortex. While MALDI images of the m/z features from this metabolic pathway were relatively uniform, with most abundance observed in the lens cortex relative to the core in PC 1, additional MALDI images that represented SIL metabolites were significant in PC 4. These SIL MALDI images were found to change substantially over time and had spatial distributions that in several circumstances were asymmetrical (see SIL Glucose, SIL G6P and SIL F16BP). These MALDI images have previously been manually selected and their identities validated by GC–MS¹⁹, showing that our spatial metabolomics pipeline is able to return results similar to manually curated data.

Another display available using the pipeline is via regional analysis. To achieve this, K means segmentations were manually assigned with a lens anatomical region (i.e. anterior/posterior/equator/inner cortex/core). The signal intensities of hexbins within each cluster were summarised, and intensities plotted against time (Fig. 6a). With the perfusion of SIL glucose, the anterior region exhibited a relatively quicker signal increase than the posterior region. A notable increase in SIL S7P intensity was observed at the initial stage of the incubation, with higher signal levels in the anterior compared to the posterior region. SIL glucose level in the anterior region declined from two hours to 8 h, while the posterior region showed a delayed response in comparison. SIL G6P in anterior region fluctuated after 1 h while the SIL F16BP constantly increased. In the posterior region, SIL G6P did not have a sharp drop but the rate of increase was not as quick as that observed in anterior.

Figure 6.

Plotted time profiles of target SIL molecules (a) and kinetic MALDI images (b) to show appearance of SIL metabolites over time. SIL glucose ([M + Cl]⁻ = m/z 221.0521), SIL sorbitol ([M + Cl]⁻ = m/z 223.0679), SIL G6P ([M – H]⁻ = m/z 265.0425), SIL S7P ([M – H]⁻ = m/z 295.0538, threshold = 0.12), SIL F16BP ([M – H]⁻ = m/z 345.0090), and UDP-glucose ([M – H]⁻ = m/z 571.0667). The line plots are automatically generated using the pipeline. The metabolic pathways annotation here are mapped to KEGG pathway database.

In addition to data visualisation in metabolic pathway and tissue regional formats, the dynamics of the metabolites of interest can also be depicted through kinetic images of specific labelled compounds (Fig. 6b). In this study, SIL glucose is primarily metabolised to F16BP and S7P through G6P. After a few hours, SIL sorbitol signal is also detected. These observations reflect the main metabolic pathways active in the lens³⁰, which are represented in the pathway enrichment analysis (see Table 2). Additionally, this method enables the level of background noise to be determined for each metabolite by applying a signal threshold filter (Supplementary Fig. 7). Signal thresholding allows the display of real signal only in the kinetic images. Together, this pipeline supplies a comprehensive set of tools for automatic assignment of naturally-occurring and labelled metabolites, and a variety of display tools to better understand tissue function.

Discussion

The data analysis pipeline presented here aims to enable a physiological approach to spatial metabolomics. In the bovine lens glucose example, this tool gives a realistic view of how the ex vivo lens utilises glucose over a 20 h timeframe. The results encode both sites of preferential glucose uptake and enzyme activity that drives its metabolism in the different lens regions. This therefore complements previous MALDI-MSI-based approaches that have effectively eliminated the SIL compound delivery component by either spraying substrate directly on to tissue sections to map enzyme activity²⁶, or exposed tissue for 48 h to allow for steady-state to be reached and (U¹³–C₆) glucose metabolic flux to be measured in tumours¹³.

The pipeline showed that tissue regions with differentially active metabolic pathways could be spatially segmented based on SIL metabolites. In this example the anterior and posterior lens regions, which differ in their metabolic activity primarily due to the presence of the epithelial cell monolayer on the anterior lens surface, were separated. Moreover, using this pipeline we identified metabolic pathways such as ‘Starch and sucrose metabolism’ and ‘Glycolysis or gluconeogenesis’ were active in specific lens regions. This is consistent with our previous manual interpretation of the (U¹³–C₆) glucose MALDI-MSI data¹⁹.

For this pipeline, our goal was to automatically annotate labelled metabolites and map altered metabolic pathways, which, due to the computational challenge, was optimally achieved through the described process that combined and averaged biological replicates at each timepoint. Despite effectively decreasing the n number, this approach led to robust annotation and generated a generalized data-driven spatial map representing experimentally consistent patterns of glucose uptake and utilization in the bovine lens. After the identification of spatial metabolic change, the user can choose to use the segmentation result (see Fig. 6) to statistically assess the level of compound/pathway change¹⁹.

Correct tissue orientation and accurate tissue overlay are important for the pipeline to return biologically meaningful results. In the current approach we kept the original lens tissue section shapes and rescaled them to the same 2D coordinate system, followed by application to a universal lens shape by lens hexagonal binning to allow comparison of equivalent spatial regions across biological replicates and timepoints. Also, this spatial binning method output more stable spectra compared with raw pixel level spectra. This improvement is attributed to its ability to eliminate background ions arising from tissue cracks or missing tissue regions, which enhanced the signal-to-noise ratio. Previous MALDI-MSI studies have developed advanced approaches, such as fiducial markers, auto fluorescent imaging, and pixel coordinate data from instrument metadata to co-register images. The aim of these strategies is to co-register the same tissue that has been imaged by complementary modalities, with the accuracy of co-registration becoming increasingly important as the spatial resolution of the component imaging modalities increases. Since our goal was to overlay data from the same imaging modality (i.e. MSI) but different temporal treatment and biological replicates, imaged at relatively low spatial resolution (150 µm), we avoided a more complicated image registration routine. This approach was clearly successful since it automatically identified lens tissue regions known to have different levels and types of metabolic activity. For application to other tissue types with a more heterogeneous tissue architecture than the bovine lens, an alternative image registration approach could be incorporated³¹, or utilised as a separate pre-processing step prior to data input to the presented pipeline (at module 2). In addition, endogenous signals that are temporally stable, or internal standard MALDI images to establish image transformation parameters could be used to co-register tissue section datasets from different treatment and biological replicates to a tissue-specific anatomical standard. While the hexbinning approach results in a small loss of spatial resolution, this would scale with the original MALDI sampling spatial resolution and in our opinion is necessary to ensure stable, high quality mass spectral input to subsequent data dimensionality reduction and compound annotation modules.

One intrinsic limitation of mass spectrometry is its inability to separate isobaric species based on their mass. This is particularly evident in MALDI-MSI experiments which typically operate in MS mode only. While some instrumentation now includes the ability to separate ions based on their collisional cross-section using ion mobility approaches, separation of, for example, central carbon metabolites in a MALDI-MSI experiment is not yet routine. This limitation leads to the potential for misassignment of compounds identity by m/z alone, even using isotopologue matching and scoring approaches. To address this, we used the mummichog algorithm, which uses established organisation of metabolic networks to predict functional activity and will therefore limit misassignment of features detected in the MALDI-MSI data. Continued advances in instrumentation with enhanced ability for gas-phase separation in an imaging context will no doubt further improve the performance of our spatial metabolomics pipeline. In addition, the amalgamation of multiple datasets that utilise additional MALDI matrices, such as 9-AA and DAN which are known to detect different subsets of metabolites, could lead to even better characterisation of spatial metabolic pathway activity.

The most significant pathways detected (Table 2) were in general agreement with previous analysis. For example, “starch and sugar metabolism”, “glycolysis or gluconeogenesis” and “fructose and mannose metabolism” were all found, with their corresponding regional enrichment displayed in their respective PCA dimension images (Fig. 6, Supplementary Fig. 1). These data represent the metabolism of glucose in the lens using a bathing medium containing a glucose concentration that mimics the level of glucose found in aqueous humour from normal healthy individuals. However, the activity of the polyol pathway, which is included here under the metabolic pathway label “fructose and mannose metabolism” and active at low levels under normal conditions, is known to be elevated in metabolic diseases such as diabetes³². This can lead to hyperglycemia and the subsequent formation of diabetic cataract which typically affects the cortical region of the lens. This is the same spatial region that “fructose and mannose metabolism” compounds were found in the current example. In the future, the workflow presented here could be applied to an ex vivo high glucose exposure model to correlate any observed cataract phenotype to potentially reveal additional underlying metabolic activity changes that occur during diabetic cataract formation. This approach could also be broadened by utilising alternative SIL compounds to probe the activity of other metabolic pathways in both the normal lens and other cataract models, and further to other tissues which have metabolic dysregulation as an underlying cause of tissue pathology.

Interestingly, in addition to the expected metabolic pathways identified in the lens, the analysis revealed additional SIL compounds not originally identified during manual interpretation. For example, one pathway that was not previously found was “Phosphatidylinositol signaling system” (see Supplementary Fig. 8). The role of this pathway in the lens is not clear, however, this result clearly shows the potential of an automated kinetic MSI annotation pipeline to discover new data in comparison to manual interpretation. While new annotations can lead to new research directions, validation of signal identification using a complementary GC–MS method remains an important step that should be performed¹⁸. In addition, the continuing development of MSI hardware in combination with multi-modal MSI approaches should provide additional data for incorporation into future annotation engines, further improving our ability to spatially resolve tissue metabolism.

Conclusion

The combination of spatial metabolomics and dynamic metabolic activity gained through an iso-imaging approach is revealing new insights into normal and aberrant tissue physiology. Using MALDI-MSI data and the bovine lens as a model system, we have developed an automated data analysis pipeline that through PCA analysis rapidly determines distinct metabolic regions in a tissue, and characterises the major metabolic pathways present in each region. Pathways containing labelled downstream metabolites from isotopically labelled glucose are prioritized for temporal analysis. This approach was cross-validated with higher dimension reduction techniques UMAP and t-SNE. With continued advances in tissue preparation techniques, instrument resolution, sensitivity, and gas-phase ion separation techniques, this approach will become even more powerful in understanding tissue physiology and aid in the development of novel treatments for diseases with underlying metabolic dysfunction.

Methods

Running environment

The pipeline was developed in a Windows 10 system. The minimum system requirements are 4-core CPU, 64 GB RAM, 1 TB hard disk drive. The pipeline was scripted using R (version 4.3.1) using common R functions, and specific packages where stated below. Detailed R function information is available by inspecting the source code.

Bovine lens MALDI-MSI data generation

We utilised data generated in a previous study¹⁹, collected using the following methodology. The dissected lenses were incubated in 5 mM (U¹³–C₆) glucose-containing artificial aqueous humor from 5 min to 20 h. Following freezing at the end of incubations, 20 µm-thick axial sections were collected on cooled double-sided carbon tape (ProSciTech, Kirwan, Australia) attached glass slides (PINK COLORFROST, LabServ, NZ). N-(1-naphthyl) ethylenediamine dihydrochloride (NEDC) matrix solution containing an internal standard (IS) 3-O-Methyl-d-glucose [(M + Cl)⁻ = m/z 229.0473] was applied using a TM-Sprayer (HTX Technologies, Carrboro, NC). The IS was used to control for ion suppression effects in the various anatomical regions of the bovine lens. All MALDI data acquisition was performed on a Bruker SolariX XR 7 T FT-ICR mass spectrometer using FTMS Control v2.2.0 and flexImaging v5.0 (Bruker Daltonics, Billerica, MA). The m/z values ranging from 100 to 1000 were collected in negative ion mode with a resolving power (m/Δm) of 66,000 at m/z 400. Two hundred laser shots were accumulated per position at a laser repetition rate of 1 kHz and spatial resolution of 150 μm. While laser power was optimized for each dataset, the laser beam dimensions were matched to the raster step size by setting it to ’medium’ for all images collected. Raw data files were first imported into SCILS lab (version 2020 Pro) for manual image trimming to remove the off-tissue regions and then converted into generic “.imzML” format. The R script package Cardinal³³ was utilized for the management and pre-processing of .imzML data and another R package HiTMaP³⁴ was employed to rectify mass drift discrepancies between runs and amalgamate biological replicates into a unified dataset using a consensus mass list. MALDI datasets were normalised to the IS signal prior to analysis through the reported pipeline.

Lens coordinate rescaling, alignment, binning, and data reshaping

Each bovine lens section was orientated with the anterior pole to the right using HiTMaP built-in functions³⁴. The following transformations were then applied for image alignment:

1. Rescaling of Y Coordinates: Y coordinates were rescaled to one standard unit within each run while preserving the original lens aspect ratio. This transformation ensured that all lenses shared the same coordinate frame despite variations in size.

2. Hexagonal Binning (Hexbin): After rescaling, hexagonal bins (hexbin) were applied to the coordinate frame³⁵. This was done using the R script package hexbin, ensuring a consistent indexing approach for all lens samples. The hexbin transformation resulted in a median pixel count of 9 within each hexbin.

3. Generation of Average Mass Spectra: Average mass spectra for each hexbin index within the same treatment condition were generated using standard R functions. This step aggregated the spectral data within each hexagonal bin to produce a representative mass spectrum.

4. Data Reshaping: The R script package reshape2³⁶ was used to convert the data into a tabular format. In this format, individual rows represent distinct m/z values associated with specific time points, and each column corresponds to a particular hexbin. This transformation facilitated spatiotemporal analysis and visualization of the data.

Dimensionality reduction, top loading feature selection and K-means segmentation

Three dimensionality reduction methods, PCA³⁷, UMAP³⁸, and t-SNE³⁹ were individually applied to the time-scale reshaped dataset to reduce the data dimensionality. The 1/30 of all m/z features with the highest loadings from every PCA dimension (PC 1 to 5) were selected as the important features that drive the variation. K-means segmentation was performed using the reduced data space to visualise the spatial clustering result. ggplot2 was employed to plot the visualization result⁴⁰.

Metabolite library construction

Metabolite annotation and pathway enrichment were performed through the mummichog framework⁴¹. Metaboanalyst (version 5.0)⁴² was used to perform the mummichog pathway enrichment analysis. The original mummichog framework does not include SIL compounds. Therefore, SIL information was added in the bovine KEGG pathway library available from https://www.metaboanalyst.ca/resources/libs/mummichog/bta_kegg.qs. Since (U–¹³C₆) glucose was used, we initially limited SIL metabolites in the new library to + 6 amu compounds. For NEDC matrix, [M + Cl]⁻ and [M − H]⁻ ions were used.

1st round metabolomic annotation and pathway enrichment for all detected m/z features

The top 1/30 of all m/z features selected from every PC were deemed as significantly changed features from all detected m/z features during the time course incubation. These selected features from every PC (6432 features) were separately fed into the mummichog algorithm. Mass tolerance was set to 10 ppm, and the new metabolite library containing both endogenous and SIL compounds was employed for further m/z feature annotation and pathway enrichment analysis via mummichog.

Feature scoring and FDR controlled filtering

Isotopic envelope simulation was incorporated into the HiTMaP package to generate the theoretical isotopic envelope based on element composition. The compounds mapped to enriched pathways in this first round were processed in the updated HiTMaP package to calculate precursor m/z and generate their formulae according to the isotopic labelling type and assigned adduct³⁴. Because the score distribution of endogenous and SIL formulae are different, the FDR calculation and filtering was conducted separately on unlabelled formula and SIL formula based on the comparison between theoretical and observed isotopic envelopes. For unlabelled metabolites, the simulation of natural compound isotopic envelopes utilized the carbon isotope composition found in nature, with 98.89% being the ¹²C isotope and 1.11% being the ¹³C isotope. In contrast, the simulation of isotopic envelopes for SIL compounds used a carbon isotope composition where 99.5% consisted of the ¹³C isotope and 0.05% of the ¹²C isotope. Each of six molecular species carrying a single ¹³C atom account for 1.1%⁴³. Therefore, the naturally occurring glucose (U¹³–C₆) isotopologue is expected to be present at a level of ~ 10⁻¹⁰ mol %. The exogenously administered (U¹³–C₆) glucose (5 mM) is in much higher abundance than natural (U¹³–C₆) glucose level, and is therefore ignored during the simulation.

2nd round metabolomic annotation and pathway enrichment within FDR filtered features

Features that passed through the 10% FDR threshold were extracted from all detected m/z features for a 2nd round of metabolomic annotation and pathway enrichment. Top features that failed to pass the FDR cutoff were removed from the top feature list. A 2nd round of metabolomic annotation and pathway enrichment was performed using the FDR filtered features, with the same parameters as the 1st round.

Feature visualization

Annotated features can be visualized in two ways. The first is a pathway-based ion image group visualization, in which annotated m/z features are visualised as ion images presented in chronological order. The second is via kinetic images, which encode the time of detection to a colour. For kinetic images, one biological replicate for each time point is selected and the mean spectrum for every hexbin is applied followed by reshaping the data to a time-merged format. The molecules of interest from all time points are extracted. A signal threshold is applied to reduce background noise. Distinct colors are assigned to hexbins from each time point. Those hexbins are projected onto the standardised lens geometry to create signal distributional maps for each time point. Finally, the maps are overlaid in chronological order to form time-coded images using the R script package magick⁴⁴.

Author contributions

G.G. and A.G. conceptualised the work. D.S. and G.G. designed the pipeline, conducted data analysis, generated all figures and drafted the manuscript. D.S., G.G., and A.G. conducted interpretation and discussion of the data, revision and finalisation of the manuscript.

Funding

The authors thank the Health Research Council of New Zealand (Consolidator Grant #20/872, Programme Grant #20/692) for supporting this work.

Code availability

The raw data used in this study can be downloaded from https://metaspace2020.eu/project/guo-2023. An temporary link was generated for review: https://metaspace2020.eu/api_auth/review?prj=54e09ade-8cb1-11ee-adab-831686c45448&token=pwemUgFOuNiB. The analysis pipeline is freely available for download from GitHub at https://github.com/MASHUOA/Spatially_temporally_resolved_MSI. The README file contains additional information on all of the available parameters of this pipeline, including examples of use.

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-024-69507-z.