MeXpose Data Analysis

MeXpose deploys a dedicated high-dimensional multiplexed image analysis pipeline combining various third-party software tools with additional newly developed Python scripts into a user-friendly, flexible containerized platform. Figure ​Figure22 outlines the workflow and key features. The elemental images obtained by LA-ICP-TOFMS are exported in single- or multichannel tiff format and submitted to the technology agnostic workflow. MeXpose data evaluation offers both an interactive and a scalable semiautomated version deployed as a single docker container. The latter was developed to meet the requirements of large-scale studies. The interactive mode was designed to enable immediate control on preprocessing, segmentation, and exploratory data analysis. This approach ensures optimal cell segmentation for different tissue types, which is a key prerequisite for quantitative image analysis. Depending on the individual needs, all included applications can be executed in “GUI-mode” (Graphical User Interface) by simply launching them using “-gui” as a suffix to the desired application name. The core functionality of both workflows (interactive and scalable) is the same and utilizes identical software with the exception being that the interactive Jupyter notebooks are replaced by command line Python scripts in the automated workflow.

Open in a separate window

Figure 2

MeXpose—a pipeline for image analysis at the single-cell level. User-defined elemental channels are stacked, preprocessed, and used to run a pretrained network to generate cell masks. Consequently, a feature matrix is extracted from the raw data using the segmentation mask. The extracted data can be used for image analysis at the single-cell level by downstream statistical analysis. The quantitative module integrates an external calibration function, which can be applied on the data to obtain metal quantities per cell.

Both MeXpose versions—i.e., the interactive and scalable workflow—offer essential functionalities to obtain tabular cellular data characterized by object-based integrated channel/isotope intensities and morphology for downstream quantitative analysis. For data preprocessing, cell segmentation, exploratory data analysis, and visualization, state-of-the-art software packages dedicated to each individual step were combined in a modular end-to-end pipeline. MeXpose deploys Fiji (v1.54f) image processing software for preprocessing and multichannel image visualization, Cellpose 2.0 (v2.2.3) for deep learning-assisted cell segmentation, Cellprofiler (v4.2.1) for feature extraction (intensities and morphology) of cellular objects as well as preprocessing, and newly developed in-house Python scripts (v3.8.8) and JupyterLab notebooks (JupyterLab v3.5.3; Jupyter Notebook v6.5.6) for data analysis and visualization.⁶⁹⁻⁷³ Each module can be replaced with other software at will, which allows for seamless integration into existing image analysis workflows. The obtained tabular single-cell data are exported for further quantitative image analysis. Scripts for absolute metal quantification and statistical analysis are integrated. The latter task involves established strategies of image analysis, such as unsupervised clustering and/or dimensionality reduction. Phenograph (v1.5.7), an unsupervised graph clustering algorithm based on the Leiden algorithm, is used to categorize segmented cells into clusters depending on their feature intensities and inter- versus intracluster connections.⁷⁴ Simplified, when represented as a graph, cells that form a subpopulation with a high number of internal connections and a lower number of external connections to other subpopulations are designated as a cluster. MeXpose allows us to visualize these high-dimensional clusters by projecting them as a two-dimensional embedding using UMAP (v0.5.4).⁷⁵ This enables visual inspection of cluster quality. By creating a cluster heat map, MeXpose permits in-depth analysis of the channel expression profile and relative intensities within each cluster, enabling the characterization of the various cellular phenotypes within tissue. Data interpretation is facilitated by the MeXpose option of projecting the corresponding segmented objects onto elemental images as a channel heatmap for visual inspection of spatial arrangements.

MeXpose uniquely allows us to apply a calibration function to the obtained data (including pixel-based and cell-based data), resulting in object-based quantitative-integrated elemental amounts. Following standardization, the color intensity code in the images can be converted to a color concentration code. Upon visualizing segmented cells on the images with a color concentration code, quantitative bioaccumulation resulting from metal exposure can be observed at the single-cell level. The images showing quantitative single-cell metal data can be generated for a user-defined cell population. Finally, the single-cell metal quantities can be plotted as histograms either for the complete imaged cell populations or again for certain subpopulations, e.g., cell phenotypes as revealed by statistical analysis. Producing histograms and plotting quantified cellular objects for distinct cellular subpopulations greatly aid in the discovery of so far missing links between cellular metal enrichment and cellular phenotypes and/or tissue architecture.

Revealing the Structural Landscape of Human Skin through Multiplexed Imaging

To make the case for single-cell metallomics, CoCl₂ permeation in human skin was studied for the first time at single-cell resolution. Skin represents one of the major metal exposure routes in occupational settings. A cutting-edge ex vivo human skin model, NativeSkin, was exposed to CoCl₂. The model closely resembles in vivo skin permeation upon exposure as it preserves tissue architecture in controlled environments.⁷⁶Figure ​Figure33 shows LA-ICP-TOFMS images of the ex vivo skin model obtained by overlay visualizations of selected image channel intensities. The overlay images unveiling the skin tissue structural architecture relied on five markers along with one marker for cell nuclei (6 channels). The selected antibody panel (Table S1) characterized the different layers of the epidermis and the dermis together with essential structural features such as blood vessels, sweat glands, and muscle tissue.

Open in a separate window

Figure 3

Skin tissue architecture. Following FFPE treatment and sectioning, the skin thin sections were labeled with metal-conjugated antibodies and analyzed by LA-ICP-TOFMS (1 μm pixel size, 200 Hz pixel acquisition rate). The applied histochemistry and the DNA intercalator allowed us to visualize the epidermal and dermal layers together with essential structural features, including blood vessels, sweat glands, and muscle tissue. The marker panel included α-SMA (specifically labels smooth muscle cells), vimentin (mesenchymal cells, encompassing fibroblasts of the papillary dermis, as well as the endothelium of blood vessels), pan-keratin (cornified layer and keratinocytes in the epidermis), CD44 (cell surface receptor for hyaluronic acid, membrane marker for both epidermal and dermal cells), collagen (extracellular matrix of the dermis), and DNA intercalator (cell nuclei). (A) shows the signal intensity maps of the individual markers, (B) visualizes an overlay of the 6 channels, and (C) shows structural features of regions of interest of the skin tissue together with H&E staining of consecutive FFPE sections for comparison.

Single-Cell Analysis in Human Skin Tissue

Following the interactive branch of the MeXpose image analysis pipeline, single-cell data were extracted for human skin tissue sections. Stacking required the selection of two channels, (1) a selective membrane and (2) a nuclei marker. In human skin, CD44 proved to be a valuable membrane marker, while the established Ir-based DNA intercalator was selected as a nuclei marker (Figure S1). Before stacking, the two channel images were preprocessed using a combination of steps, including contrast enhancement and intensity thresholding, outlier filtering, and median/Gaussian filtering. MeXpose deploys Cellpose 2.0 for cellular segmentation, leveraging its integrated pretrained “model zoo” as well as its interactive GUI (interactive workflow only). It also inherently includes segmentation quality control through visualization and the ability to interactively correct segmentation errors and add custom annotations. Upon completion, Cellpose generates PNG cell masks which are successively used to extract multiparametric tabular single-cell data (characterized by morphology and integrated intensities) from raw images using Cellprofiler.

MeXpose image analysis of human skin areas in the mm² range revealed cell numbers in the 10³ orders of magnitude with an average cell diameter of 10 μm. Figure S1 shows the obtained segmented cells in the human skin model, projected on top of the LA-ICP-TOFMS image. High cell numbers were observed in the epidermis, while the dermal layers showed fewer cells embedded in the collagenous matrix.

The applied downstream statistical analysis resorted to a marker panel of 9 metal-conjugated antibodies (Table S1) and the endogenous metal iron (Fe). Phenograph clustering revealed seven clusters, which could be assigned to cellular phenotypes, typical for the different epidermal and dermal layers. Keratinocytes (basal, spinous, and granular), macrophages, fibroblasts, and smooth muscle cells could be assigned (Figure ​Figure44). MeXpose uniquely allows the consideration of endogenous elements upon cellular phenotyping. As has already been shown, intrinsic elements such as Fe cannot be accurately quantified due to the applied staining protocols.^68,77 However, they can be used as complementary markers if their distribution is maintained. In the case of human skin, adding cellular Fe was beneficial for characterizing epidermal keratinocytes. Basal cells, which have the highest blood flow due to their proximity to capillaries, could be differentiated from other keratinocytes using Fe (Figure S2).^78,79 Larger blood vessels such as arteries and veins, sweat glands, and hair follicles also showed characteristic Fe distributions that are useful for cell pheno-/subtyping (Figure S3).

Open in a separate window

Figure 4

Downstream statistical analysis. (A) Heatmap and (B) UMAP are shown for the 1782 segmented cellular objects. The marker panel of collagen type I, pan keratin, E-cadherin, CD44, vimentin, α-SMA, CD68, KI-67, pHistone H3, and endogenous Fe revealed 7 distinct cell populations (0–6). Cluster 1 and Cluster 2 relating to the epidermis were well separated from the dermal clusters (0, 3–6). (C) Visualization of the phenograph clusters in the segmentation mask, revealing the location of the cells within the skin tissue.

Quantitative Analysis Calls for Validation

Whether the applied histochemistry allows for simultaneous phenotyping and cellular metal enrichment needs to be evaluated for each specific exposure study. As in any quantitative approach, additional experiments are required proving the method fit for purpose. The effect of sample preparation has been investigated using pixel data, as staining is required for the analysis of single cells in tissue. A dedicated script for pixel-based image analysis is therefore provided with MeXpose. As shown in Figure ​Figure55, two consecutive skin sections exposed to Co were measured, one following the immunostaining procedures and the other omitting the sample preparation steps. K-means clustering demonstrated the validity of the quantitative approach, as both samples showed excellent agreement regarding Co quantity and location.

Open in a separate window

Figure 5

Imaging of Co in consecutive sections of human skin, (A) unlabeled and (B) labeled, showing the calculated amounts of Co obtained by k-means clustering (n = 3).

Quantifying Metal Accumulation at the Single-Cell Level

The human skin model exposed to CoCl₂ showed pronounced metal permeation into the epidermis. High bioaccumulation was found down to a depth of approximately 100 μm; however, the highest Co amount was observed within the cornified layer at a permeation depth of 20 μm (Figure S4). Only small amounts of Co were found to reach the dermal layer of the skin. By applying the external calibration function (obtained from the measurement of gelatin-based microdroplets) (Figure ​Figure66A), both pixel-based and cell-based intensity data on Co were converted into absolute amounts (Figure ​Figure66B,C). Within the MeXpose pipeline, the obtained single-cell quantities can be plotted as histograms showing the distribution of cellular metal content for subpopulations or the entire cell population. For visualization, projections of segmented cells on top of the image are associated with a color-coded concentration scale. The entire imaged cell population of 1782 cells revealed a mean Co content of 2.4 fg per cell (Figure ​Figure66D). A procedural limit of detection of 0.38 fg Co per cell was estimated by studying control skin tissue samples (not exposed to CoCl₂) that were stained with the metal-conjugated antibody panel (Figure S5). The cellular Co intensities constituting the blank values were converted into quantities and treated as Poisson distributed to infer the limit of detection.⁸⁰

Open in a separate window

Figure 6

Quantitative Co imaging in human skin at the single-cell level (LA-ICP-TOFMS analysis, 1 μm pixel size, 200 Hz pixel acquisition rate). Ex vivo skin models were incubated with 100 μg cm^–2 CoCl₂ solution (dose chosen based on previous results from patch tests in Co allergic subjects) for 48 h. (A) Quantification of Co was performed using gelatin-based calibration standards which were measured along with the skin thin sections. (B) Co signal intensity map of the tissue was converted to absolute Co amounts using the calibration factors. Co was enriched in the cornified layer but penetrated the epidermis, whereas the dermis showed significantly lower Co quantities. (C) Pixel data were transformed into cellular data to better visualize the Co accumulation within cells. The histogram in (D) shows the Co distribution in all segmented cells. The average Co amount per cell is 2.4 ± 1.8 fg. (E) Histograms for each individual cell cluster revealed by unsupervised clustering (Figure ​Figure44). Cluster 2 (basal cells) in particular showed significantly higher Co content compared to the other cell types (Figure S6).

Exposure of Ex Vivo Skin Model to CoCl₂

Upon arrival of the skin model, the media (supplied by Genoskin) was added under sterile conditions and incubated for 1 h at 37 °C, 5% CO₂, 95% relative humidity (according to the user manual). A 100 mg mL^–1 stock solution of CoCl₂·6H₂O was prepared by dissolving 1 g in 10 mL of ultrapure water, followed by sterile filtration. For the target dose of 2.5 μg μL^–1 (based on prior results from patch testing of Co allergic individuals), 25 μL of the stock solution was mixed with 975 μL of sterile ultrapure water and vortexed prior to exposure.^63,81 The CoCl₂ solution was added to the model by pipetting 20 μL. The solution was homogeneously distributed on the apical side of the model. The plate was then placed in an incubator (37 °C, 5% CO₂, 95% relative humidity). The next day the media was changed. Samples were collected after 48 h, later embedded in FFPE, and sectioned at 5 μm.

Immunostaining of Skin Tissue Sections

Skin tissue sections were deparaffinized by treatment with fresh xylene for 20 min. Descending grades of ethanol (100–70%) were used for rehydration. After rinsing with ultrapure water, heat-mediated antigen retrieval was performed at 96 °C for 30 min using a Tris-EDTA buffer at pH 9. Slides were then cooled and washed with ultrapure water and TBS/0.05% Tween. To prevent nonspecific binding, SuperBlock buffer was applied to the skin tissue for 30 min at RT, followed by CD16/32 treatment for 10 min. The sections were then exposed to a cocktail of metal-conjugated antibodies overnight in a hydration chamber at 4 °C. Details of the metal-conjugated antibodies used are given in Table S1. The antibodies (1:100 dilution) were prepared in a mixture of 0.5% BSA, 1:100 CD16/32 in TBS/0.05% Tween. To prevent aggregation, the antibodies were centrifuged at 13,000 g for 2 min prior to use. Tissue sections were then stained with the Ir-based DNA-intercalator (125 μM) using a 1:100 dilution in TBS/0.05% Tween. After incubation for 5 min at RT in a hydration chamber, slides were thoroughly washed with ultrapure water and air-dried. Microscopic images were taken prior to LA-ICP-TOFMS analysis to provide an overview of the tissue sections.

Calibration Standards for LA-ICP-TOFMS Analysis

Quantitative analysis was carried out by LA-ICP-TOFMS with the use of gelatin microdroplets, as described previously.^49,82 For this purpose, multielement standard solutions were prepared gravimetrically using commercial standard stock solutions and mixed with a gelatin solution. The resulting solutions were then transferred to the wells of a 384-well plate. The plate served as the sample source for a microspotter system.

The CellenONE X1 microspotter (Cellenion, Lyon, France) was used for generating arrays of gelatin microdroplet standards on glass slides. These droplets were approximately 200 μm in diameter and had a volume of 400 ± 10 pL. The software of the instrument evaluated the size of the droplets, which was then used for normalization to determine the absolute amounts of elements in the droplets. The entire microdroplets were subjected to quantitative and selective ablation, followed by multielement analysis using LA-ICP-TOFMS.

LA-ICP-TOFMS Analysis

An Iridia 193 nm LA system from Teledyne Photon Machines (Bozeman, MT, USA) was coupled to an icpTOF 2R ICP-TOFMS instrument from TOFWERK AG (Thun, Switzerland). The LA system consisted of an ultrafast, low-dispersion cell in a cobalt ablation chamber.^65,66 The cell was connected to the ICP-TOFMS using an aerosol rapid introduction system (ARIS). An argon makeup gas flow (~0.90 L min^–1) was introduced through the low-dispersion mixing bulb of the ARIS into the He carrier gas flow (0.60 L min^–1) before entering the plasma. The NIST SRM612 glass-certified reference material (National Institute for Standards and Technology, Gaithersburg, MD, USA) was used for daily tuning. Tuning was aimed at high intensities of certain ions (⁵⁹Co⁺, ¹¹⁵In⁺, and ²³⁸U⁺), minimal oxide formation (²³⁸U¹⁶O⁺/²³⁸U⁺ ratio <2%), and low elemental fractionation (²³⁸U⁺/²³²Th⁺ ratio ~1). In addition, aerosol dispersion was minimized by optimization of the pulse response time for ²³⁸U+ with the FW0.01 M criterion. LA sampling was performed in fixed dosage mode 2 with a 200 Hz repetition rate. A circular spot size of 2 μm was used with a line spacing of 1 μm, resulting in a pixel size of 1 μm × 1 μm. By using energy densities above the ablation threshold of the samples but below that of glass, selective ablation of the samples was achieved.⁸³ Gelatin microdroplets and skin tissue were quantitatively removed using a fluence of 1.0 J cm^–2. The icpTOF 2R ICP-TOFMS instrument offers a specified mass resolution of 6000 (R = m/Δm) and detects ions in the m/z range of 14 to 256. Its integration and readout rates were in line with the LA repetition rate. The instrument was equipped with a torch injector of 2.5 mm inner diameter and nickel sample and skimmer cones (with a 2.8 mm skimmer cone insert). It was operated with a radiofrequency power of 1440 W, an auxiliary Ar gas flow rate of ~0.80 L min^–1, and a plasma Ar gas flow rate of 14 L min^–1. All measurements were performed in collision cell technology (CCT) mode, with the collision cell pressurized with a H₂/He gas mixture (93% He (v/v), 7% H₂ (v/v)) at a flow rate of 4.2 mL min^–1. Detailed instrumental parameters for LA-ICP-TOFMS measurements are given in Table S2.

Data Acquisition and Processing of LA-ICP-TOFMS data

LA-ICP-TOFMS data were acquired using TofPilot 2.10.3.0 from TOFWERK AG (Thun, Switzerland) and stored in the open-source hierarchical data format (HDF5, www.hdfgroup.org). Subsequent data processing was done using Tofware v3.2.2.1 (TOFWERK AG, Thun, Switzerland), used as an add-on to IgorPro (Wavemetric Inc., Oregon, USA). This processing included three steps: (1) correction of mass peak position drift across the spectra with time-dependent mass calibration; (2) determination of peak shapes; and (3) fitting and subtraction of the mass spectral baseline. The data were further processed using HDIP version 1.8.4.120 from Teledyne Photon Machines (Bozeman, MT, USA). A built-in script was used to automate the processing of files generated by Tofware, producing 2D elemental distribution maps. Thus, acquired LA-ICP-TOFMS data were processed in HDIP and exported as separate TIFF files for each isotope to enable single-cell-based analysis or in CSV format for pixel-based image analysis.

MeXpose Image Analysis

TIFF files or CSV files were imported for single-cell- or pixel-based image analysis, respectively.

MeXpose Implementation

MeXpose contains all the software and scripts needed to run the workflow and was designed to be used as an all-in-one Docker container. The MeXpose image can be found on the Docker Hub under the name MeXpose. It contains both the interactive and scalable versions of the workflow.

MeXpose Description

In principle, the workflow of the MeXpose data analysis consists of four steps: Data preprocessing, cell segmentation, data extraction, and downstream statistical analysis. Our pipeline provides near seamless integration into existing analysis workflows by allowing substitution and individual execution of each step. The pipeline is available in two versions: an interactive graphical user interface (GUI)-based version for initial exploratory data analysis and a scalable version using the same software but streamlined to run with minimal user input for efficient analysis of large data sets. In addition to in-depth single-cell-based analysis, basic pixel-based analysis, k-means clustering, visualization, and pixel-based quantification are provided. All analysis steps can be performed from the container’s command line.

For single-cell-based analysis, raw LA-ICP-TOFMS data is processed in HDIP. Isotope channel images are exported as individual 16-bit TIFF files for analysis with the MeXpose workflow. Single- or multichannel TIFF images can be provided from any multiplexed imaging modality. Fiji can be used to preprocess the raw images. The interactive mode allows the user to access the full range of image processing tools in Fiji and to view the effects live on their image set in real-time. The scalable mode provides user-selectable Fiji macros for outlier/hot pixel removal, Gaussian or median filtering for speckle removal, channel stacking, and image tilling. The resulting images are saved as stacks or processed copies of the raw isotope images in 16-bit TIFF format. Cell segmentation requires a 2-channel TIFF stack consisting of a nucleus and a membrane/cytoplasm channel. Cellpose 2.0 is used in both the interactive and scalable versions of MeXpose to provide cutting-edge cell segmentation performance. Cellpose provides multiple pretrained deep neural network models based on the Cellpose generalist segmentation algorithm, from which the user can freely select the best-performing model.^69,70 In addition, the interactive workflow enables the use of the Cellpose human-in-the-loop pipeline. This feature allows the user to make manual adjustments such as adding/removing object annotations and then retraining the model, resulting in a refined model with improved segmentation performance. By using this approach in conjunction with the Fiji tiling macro, it is possible to make iterative improvements to small subsets of the final data until a satisfactory set of results and a specialist model are obtained. Fine-tuning a model typically requires 3–5 iterations, as recommended in the Cellpose documentation. Successful segmentation results are saved as PNG masks and used to extract morphological features and intensity information from individual cells using Cellprofiler. Similar to preprocessing, the user has access to Cellprofiler functionality in interactive mode. A basic Cellprofiler pipeline file is provided as a template. The user is encouraged to perform the initial setup of a Cellprofiler pipeline either in interactive mode or by running Cellprofiler as stand-alone software. Marker panels and image channels are subjective to the experiment and require manual configuration. The scalable workflow uses a user-preconfigured Cellprofiler pipeline that is executed from the command line. Cellmasks are imported into Cellprofiler along with raw images of all elemental channels of interest. Within Cellprofiler, the user can assign isotopes to image channel names, filter objects that touch the image boundary, filter objects by size, measure object intensities and morphological features, and finally export the data in a tabular format.

MeXpose uses custom Python scripts to perform exploratory data analysis and quantification. These scripts can be deployed either as Jupyter notebooks for interactive analysis or as command line executable scripts for scalable analysis. There are three interactive Jupyter notebooks and three corresponding Python scripts provided, a phenotyping notebook/script containing all the necessary modules for single-cell data analysis, a quantification notebook/script, and a pixel notebook/script including modules to analyze image data at the pixel level. All scripts can be run independently. The exploratory analysis of the data includes visual inspection using heatmap overlays and histograms, as well as cell phenotyping. Phenotyping is performed by combining unsupervised phenograph clustering and heatmap-based cluster analysis on scaled and normalized single-cell data. The user can select relevant channels as needed. The heatmap visualization provides quick identification of relevant cellular phenotypes. A typical analysis workflow would start with the phenotyping notebook and later feed cell phenotype-specific data into the quantification script. In summary, the structure of the phenotyping notebook consists of the following functions:

• 1.

  Initial setting and loading of the working directory and single-cell data

• 2.

  (Optional) Inspection of raw histograms

• 3.

  Size-normalization (based on “real world” pixel size) and outlier filtering

• 4.

  Heatmap visualization of isotope intensities at the single-cell level (whole sample)

• 5.

  Data scaling for accurate clustering

• 6.

  Phenograph clustering of the data

• 7.

  Dimensionality reduction using UMAP

• 8.

  Visualization of clusters and UMAP embedding

• 9.

  Heatmap visualization of isotope intensities at the single-cell level (per cluster)

• 10.

  Data export

The quantification script can be run after clustering or as a stand-alone method without clustering, depending on the needs of the user. It uses two input files, a file containing the standard measurements and calibration function, and one or more files containing the single-cell data to be quantified. After setting the working directory and loading all necessary files, the elemental channels for quantification are selected and divided by their corresponding calibration factors. Histograms of the selected channels can then be plotted and saved, and the quantified data can be exported as CSV files.

Pixel-based analysis can be conducted using the pixel script by importing a CSV file with intensity information organized in columns, where each row represents a pixel. The user needs to specify the width and height dimensions of the image in pixels and the desired number of clusters for k-means clustering. The resulting image will show each pixel colored according to its cluster assignment. Using Spearman’s rank correlation coefficients, it is also possible to generate a correlation heatmap for user-selected channels.

This research was funded in whole or in part by the Austrian Science Fund (FWF) [10.55776/FG3]. For open access purposes, the author has applied a CC BY public copyright license to any author accepted manuscript version arising from this submission.