Abstract

   High-resolution spatial imaging is transforming our understanding of
   foundational biology. Spatial metabolomics is an emerging field that
   enables the dissection of the complex metabolic landscape and
   heterogeneity from a thin tissue section. Currently, spatial metabolism
   highlights the remarkable complexity in two-dimensional (2D) space and
   is poised to be extended into the three-dimensional (3D) world of
   biology. Here we introduce MetaVision3D, a pipeline driven by computer
   vision, a branch of artificial intelligence focusing on image workflow,
   for the transformation of serial 2D MALDI mass spectrometry imaging
   sections into a high-resolution 3D spatial metabolome. Our framework
   uses advanced algorithms for image registration, normalization and
   interpolation to enable the integration of serial 2D tissue sections,
   thereby generating a comprehensive 3D model of unique diverse
   metabolites across host tissues at submesoscale. As a proof of
   principle, MetaVision3D was utilized to generate the mouse brain 3D
   metabolome atlas of normal and diseased animals (available at
   [57]https://metavision3d.rc.ufl.edu) as an interactive online database
   and web server to further advance brain metabolism and related
   research.
     __________________________________________________________________

   Spatial biology has emerged as a pivotal discipline for decoding the
   spatial organization and interactions of biomolecules within cells and
   tissues. Building upon the foundational pillars of spatial
   transcriptomics^[58]1,[59]2, spatial proteomics^[60]3,[61]4 and spatial
   metabolomics^[62]5,[63]6, the field is substantially advancing our
   understanding of biological systems within multicellular organisms.
   Specifically, spatial metabolomics, performed through matrix-assisted
   laser desorption/ionization (MALDI) mass spectrometry imaging (MSI),
   enables the high-resolution mapping of metabolites in tissues at the
   mesoscale^[64]6. This methodology allows the identification,
   quantification and spatial distribution of multiple classes of
   metabolites such as small molecules^[65]7,[66]8, lipids^[67]9,[68]10,
   glycogen and glycan-related complex carbohydrates^[69]11,[70]12,
   thereby providing critical insights into cellular metabolism and
   disease mechanisms. Computational workflows are being developed to
   align and co-analyse high-dimensional datasets produced from spatial
   metabolomics datasets^[71]5,[72]13,[73]14, bringing pathway enrichment
   and network analyses into spatial metabolism and enabling future
   advancements that promise a more comprehensive and nuanced
   understanding of cellular and tissue architecture. The field is thus
   poised for important contributions to elucidating disease mechanisms
   and potential therapeutic interventions. For example, spatial
   metabolomics has provided insights into new skeletal myofiber
   subtypes^[74]15, dopaminergic lipid homeostasis during Parkinson’s
   disease^[75]16, and metabolic dependencies of glycogen in both Ewing
   sarcoma^[76]11 and pulmonary fibrosis^[77]17.

   Biological organisms and tissues operate in three-dimensional (3D)
   space. While a number of new technologies are presented for spatial
   transcriptomics^[78]18–[79]20, current methodologies and data generated
   for spatial metabolomics are predominantly confined to two-dimensional
   (2D) analyses. The usefulness of 3D metabolomics has been demonstrated
   through manual co-registration and alignment of brightfield
   images^[80]21–[81]23, but substantial manual input remains a major
   barrier to widespread adaptation and application of 3D metabolism data
   into an integrated workflow^[82]24. This limitation stands as a barrier
   to understanding the biological function of metabolic networks in 3D.
   One existing approach to 3D metabolic imaging is through magnetic
   resonance spectroscopy; however, this technique is constrained by a
   limited set of detectable metabolites and offers insufficient spatial
   resolution^[83]25. This is especially problematic in heterogeneous
   structures such as the rodent brain in which the cellular composition
   of distinct brain regions changes at a spatial gradient that is
   substantially finer than the resolution offered by magnetic resonance
   spectroscopy. Advancing to a 3D metabolome at the mesoscale represents
   the next major milestone, serving to bridge the gap between 2D and 3D.
   Achieving 3D metabolomics at this scale would enable the mapping of
   metabolic interactions and networks, reflecting the spatial
   complexities inherent in biological systems and bridging the gap in
   understanding from cellular to organismal levels. Such a shift from 2D
   to 3D would provide unprecedented insights into the spatial
   heterogeneity of metabolic activities within cells and tissues,
   facilitating a deeper understanding of the connection between molecules
   and physiology.

   Here, we introduce MetaVision3D, a computational framework for the
   automated generation of 3D metabolome models from serial tissue
   sections. MetaVision3D represents a major advancement in spatial
   metabolomics, offering algorithms designed to reconstruct the 3D
   metabolome of complex diseases or tissues ([84]Fig. 1). MetaVision3D
   uses an automated alignment framework through computer vision
   techniques (MetaAlign3D) and normalization steps to account for
   intraslice signal variabilities (MetaNorm3D) and then performs
   imputation (MetaImpute3D) and interpolation algorithms (MetaInterp3D)
   to improve overall 3D rendering visualization ([85]Fig. 1). The
   innovation lies both in the precision and resolution of the spatial
   data as well as in the workflow’s ability to compensate for
   experimental and technical variabilities inherent to MALDI imaging.
   MetaVision3D serves as a critical bridge between existing imaging
   modalities, enabling mesoscale spatial resolution and a comprehensive
   characterization of metabolic distributions within the brain.
   Leveraging MetaVision3D, we created the 3D mouse brain metabolome atlas
   in an interactive online database and web server
   ([86]https://metavision3d.rc.ufl.edu).

Fig. 1 |. MetaVision3D, a computer vision pipeline for the generation of
spatial metabolism in 3D.

   Fig. 1 |
   [87]Open in a new tab

   Computational framework for MetaVision3D for the generation of spatial
   metabolome of mouse hemibrain in 3D. MetaVision3D uses an automated
   alignment framework through computer vision and normalization steps to
   account for intraslice signal variabilities and then performs
   imputation and interpolation algorithms to improve overall 3D rendering
   visualization. Created with [88]BioRender.com.

Results

AI-driven alignment of 2D MALDI imaging datasets

   The first step towards building a 3D spatial metabolomics atlas of the
   brain is the ability to seamlessly align serial tissue sections
   subjected to MALDI imaging in an automated fashion. Here, we developed
   a powerful MALDI imaging alignment module (MetaAlign3D) in MetaVision3D
   based on the principle of enhanced correlation coefficient (ECC)
   maximization^[89]26, which has been adopted for image alignment in the
   field of computer vision, a type of artificial intelligence (AI)
   focusing on image registration tasks. Using this algorithm, MALDI
   images are systematically aligned through an iterative process that
   maximizes the correlation coefficient between adjacent sections
   directly from data array of MALDI images. Utilizing the first section
   as the starting point, each subsequent image was subjected to
   parametric transformations—comprising rotations, translations and, in
   some instances, scaling and skewing—to identify the optimal overlay
   ([90]Fig. 2a). The ECC maximization principle functioned as the guiding
   metric in this process, providing a quantitative measure of similarity
   that was optimized until the alignment of molecular patterns across
   serial sections reached the peak of the correlation function
   ([91]Methods). To test MetaAlign3D, we applied it to align in five
   sequential sagittal-cut MALDI image sections from the medial plan of a
   mouse hemibrain ([92]Fig. 2b and [93]Extended Data Fig. 1a). Compared
   with manual fit ([94]Methods), MetaAlign3D achieved superior alignment
   ([95]Fig. 2b and [96]Extended Data Fig. 1a). This was indicated by the
   enhanced overlap accuracy of the anatomical landmarks highlighted by
   the metabolite distribution across the brain section ([97]Fig. 2b and
   [98]Supplementary Video 1).

Fig. 2 |. MetaAlign3D for the automated alignment of serial MALDI imaging
tissue sections.

   Fig. 2 |
   [99]Open in a new tab

   a, Schematics of the computation workflow for the automated alignment
   of sequential MALDI imaging tissue powered by MetaAlign3D. Created with
   [100]BioRender.com. b, MetaAlign3D versus manual fit for five serial
   sagittal sections cut from the medial side of a mouse hemibrain.
   Created with [101]BioRender.com. c, MetaAlign3D versus manual fit for
   six sagittal sections cut from the medial side to lateral side of a
   mouse hemibrain with major shift in tissue size. Created with
   [102]BioRender.com. d, Statistical measures of alignment and fit
   quality between manual fit and MetaAlign3D, ECC, SSIM and MSE. The
   box-and-whisker plot shows the median (line), interquartile range (box)
   and variability (whiskers extending to the maximum and minimum data
   points) for 11 different lipid classes across n = 79 brain sections. e,
   Schematics of the experimental workflow MALDI imaging of mouse
   hemibrain at 50 μm resolution. A total of 79 sections were acquired.
   Created with [103]BioRender.com. f, MetaAlign3D versus manual fit for
   the 79 serial sagittal sections cut from a mouse hemibrain for the
   creation of spatial metabolome atlas in 3D. Schematics created with
   [104]BioRender.com.

   To test the ability of MetaAlign3D to fit serial tissue sections of
   different sizes, we performed MALDI imaging on tissue sections derived
   from the medial and lateral side of the mouse sagittal hemisected brain
   ([105]Fig. 2c and [106]Extended Data Fig. 1b). MetaAlign3D corrected
   for distortion due to size differences between lateral and medial brain
   sections and maintained anatomical alignment highlighted by metabolic
   features ([107]Fig. 2c). This improvement is particularly evident when
   juxtaposed against manual alignment efforts, where notable
   imperfections were observed ([108]Fig. 2c). Statistical measures of
   alignment and fit quality included ECC for geometric
   transformation^[109]26, structural similarity index measure (SSIM) to
   compare image similarities^[110]27 and mean squared error (MSE) to
   quantify differences between serial images^[111]28. MetaAlign3D of the
   entire volume demonstrates substantial improvement over the manual
   fitting technique for all parameters assessed ([112]Fig. 2d).

Construction of the 3D spatial metabolome of the brain

   We further demonstrate the scope and utility of MetaAlign3D in an
   ambitious task of constructing the first mesoscale mouse brain 3D
   metabolomics atlas. To achieve this, we performed cryosectioning of
   10-μm-thick brain sections from a mouse hemibrain. Importantly,
   sections are 50 μm apart (z axis), corresponding to the MALDI imaging
   spatial resolution of 50 μm (x and y axes). This strategy allows the
   creation of a 3D brain atlas at 50 μm resolution ([113]Fig. 2e). A
   total of 79 serial brain sections were collected from cryosectioning
   and subjected to MALDI imaging. We used MetaAlign3D and manual fit to
   connect all 79 brain sections and created a 3D rendering of the
   metabolic features ([114]Methods). We rendered
   lysophosphatidylethanolamine (LPE) 16:1 [M – H]^− in 3D view using
   ImageJ ([115]Fig. 2f and [116]Methods) to allow visualization of
   detailed brain anatomical structures ([117]Fig. 2f). For example, the
   structure of the corpus collosum is much more pronounced after
   MetaAlign3D compared with manual fit ([118]Fig. 2f and
   [119]Supplementary Video 1).

   Upon completion of the alignment, a series of challenges became
   apparent. First, due to the inherent characteristics of the MALDI
   workflow, discernable variabilities were observed in certain metabolic
   features both within and between slides even after normalization to
   total ion current ([120]Extended Data Fig. 2). Such variations pose
   challenges for ensuring consistent representations of the metabolic
   landscapes across sections. In addition, the imperfections associated
   with handling and sectioning frozen brain tissue at 10 μm became
   evident, as several sections exhibited missing regions—probably
   attributed to the fragile nature of the frozen brain ([121]Extended
   Data Fig. 4a). These findings underscore the complexities involved in
   the experimental aspects of such atlas construction and point towards
   areas requiring further refinement and methodological adaptation.

   Although MALDI imaging analyses were normalized to total ion current,
   interslide normalizations are not commonly performed. Interslide
   variability exists, leading to inconsistencies in the representation of
   metabolites across serial sections ([122]Extended Data Fig. 2a). Such
   disparities could potentially skew the interpretative outcomes of the
   atlas. To account for these issues, we instituted a normalization
   strategy, referred to as MetaNorm3D, for each metabolite, using the
   median value of the section as the reference point ([123]Methods). By
   normalization to the tissue section median, we markedly improved the
   intersection consistency ([124]Extended Data Fig. 2b). This ensured a
   smoother transition of metabolite intensities between serial sections.
   As a result, the normalized dataset provided a more contiguous and
   coherent display of the metabolic landscape, substantially reducing the
   technical variabilities associated with the MALDI process.
   Phosphatidylinositol phosphate (PIP) 38:4 [M – H]^− and
   phosphatidylinositol (PI) 36:4 [M – H]^− are shown as examples
   ([125]Extended Data Fig. 3).

   To improve visualization of the 3D spatial distribution, including
   challenges stemming from the physical handling of brain sections, which
   occasionally led to missing regions or gaps in the tissue, we developed
   an advanced AI-driven imputation module named MetaImpute3D in
   MetaVision3D to fill in these gaps ([126]Fig. 3a). For these processes,
   we utilized adjacent sections—two sections from either side of the
   compromised section—as references. Drawing data from these intact