key: cord-0320594-lwd09euj
authors: Siegerist, Florian; Hay, Eleonora; Dikou, Juan Saydou; Büscher, Anja; Oh, Jun; Ribback, Silvia; Zimmermann, Uwe; Bräsen, Jan Hinrich; Lenoir, Olivia; Endlich, Karlhans; Tharaux, Pierre-Louis; Endlich, Nicole
title: scoMorphoFISH: A Deep-Learning enabled toolbox for single-cell single-mRNA quantification and correlative (ultra-)morphometry
date: 2021-09-27
journal: bioRxiv
DOI: 10.1101/2021.09.27.461916
sha: aebf4420d1d704e452d9c810c648f23769a83e20
doc_id: 320594
cord_uid: lwd09euj

Increasing the information depth of single kidney biopsies can improve diagnostic precision, personalized medicine and accelerate basic kidney research. Until now, information on mRNA abundance and morphologic analysis has been obtained from different samples, missing out on the spatial context and single-cell correlation of findings. Herein, we present scoMorphoFISH, a modular toolbox to get spatial single-cell single-mRNA expression data optimized for routinely generated kidney biopsies. Deep-Learning was used to virtually dissect tissue sections in tissue compartments and cell types to which single-cell expression data was assigned. Furthermore, we show correlative and spatial single-cell expression quantification with super-resolved podocyte foot process morphometry on the same histological section. In contrast to bulk analysis methods, this approach will help to identify local transcription changes even in less frequent kidney cell types on a spatial single-cell level with single-mRNA resolution. As this method performs well with standard formalin-fixed paraffin-embedded samples and we provide pretrained DL-networks embedded in a comprehensive image analysis workflow, this method can be applied immediately in a variety of settings.

High-precision analysis of kidney biopsies is key to providing diagnosis and targeted therapies for patients. Recently, several methods have been established to improve the analysis depth of formalin-fixed paraffin-embedded (FFPE) kidney biopsies: The filtration barrier can be morphometrically analyzed by 3D structured illumination microscopy (3D-SIM) [1] [2] [3] . The determination of the filtration slit density by PEMP (podocyte exact morphology measurement procedure) emerged as a tool that can be combined with co-staining of multiple proteins 4 . However, antibody-based quantification of local protein abundance has limitations as it depends on antibody availability and performance. Additionally, locally secreted factors are typically not captured by immunofluorescence techniques. The use of bulk proteomics and transcriptomics is limited since the expression of frequent cells like proximal tubule cells can mask transcriptional changes in less frequent cell populations. To circumvent this, tissue is manually dissected (e.g. glomeruli from tubulointerstitium) 5 , cell types are enriched by flow cytometry 6 or single-cell-RNA sequencing is performed 7 .

Unfortunately, the contextual and/or morphological information is lost in all approaches due to mechanical dissociation. As biopsy material is typically limited and interpretation in a spatial context required, correlation of multiple techniques on single sections could increase the depth of information.

An antibody-independent way to investigate spatial RNA abundance is in situ hybridization (ISH) which has been substantially improved in terms of sensitivity and multiplexing. Recently, diverse methods for single-mRNA visualization and quantification (smFISH) are available [9] [10] [11] [12] . A problem for inter-sample comparability is that smFISH highly depends on preparation-dependent RNA integrity. To rule out this problem, a stable on-slide in-cell reference gene would be required to normalize expression data for different parts of the same biopsy or even over different samples.

To assign transcripts to tissue compartments and individual cell types, reliable identification and segmentation of cellular regions of interest (ROIs) is required.

Unfortunately, correlative antibody-based cell classification is challenging as smFISH requires tissue digestion to liberate fixed mRNAs. Additionally, segmentation tasks are typical bottlenecks in image analysis workflows. To overcome this, Deep-Learning (DL) has been used for segmentation and morphometry of kidney biopsies 13 .

Herein, we present scoMorphoFISH (single-cell correlative Morphometric single-mRNA FISH), a Deep-Learning-accelerated approach for imaging-based and digital single-cell single-mRNA quantification. For the first time, we combined spatial singlecell expression data with antibody-based super-resolved podocyte foot process morphometry. We integrate spatial single-cell transcriptomic, (ultra-)morphometric, and classic histology over scales as large as whole FFPE sections down to individual foot processes.

To account for RNAintegrity differences, we wanted to identify an onslide reference gene.

Such genes should be constantly expressed and not regulated themselves. 

To assign transcripts to tissue compartments or cell types, we To establish virtual DL-segmentation-based tissue-microdissection, we custom-trained the two DL networks UNet 14, 15 and StarDist with datasets of 200 manually segmented glomeruli and 1033 cell nuclei, respectively (Suppl. Figure 1 ). As shown in Figure 3 , outlines of glomeruli and cell nuclei of raw images were predicted with high reliability and accuracy by the trained networks. 

SARS-CoV2 has been found in both tubular cells and podocytes 17 . This is surprising as podocytes express massively lower levels of the respective entry receptor ACE2 (Supplemental Figure 3 ).

Since it is not clear whether ACE2 can be locally upregulated in podocytes, we investigated its expression in different glomerulopathies. As shown in Figure 6 , we Interestingly, besides tubular ACE2 expression (arrows in Figure 7b ), glomerular ACE2 expression in this glomerulus was present but restricted to podocytes in an area with cuboidal parietal epithelial cells (arrowheads in Figure 7c and e). sclerotic glomeruli showed podocyte foot process effacement but preserved VEGFA expression (Supplemental Figure 6 ).

Supplemental Figure 4 Classic PAS histology after tissue digestion showed tissue integrity sufficient for pathohistological assessment (a). Podocyte filtration slits were labelled with a primary-conjugated podocin (NPHS2) antibody (b). Shown in b is the correlative imaging of local podocyte ultrastructure and associated smFISH transcripts As shown in b, filtration slit density was quantified as the total length of the filtration slit per glomerular capillary area. Optical resolution as determined as the full width at half maximum of the sub-diffraction filtration slit was 125 nm±12 nm, sufficient to resolve individual foot processes (d). Digestion required for smFISH had no influence of filtration slit density (e). Although intense tissue digestion was required for smFISH, super-resolved podocyte filtration slit morphology was largely unaffected and filtration slit density was in line with previously published values 2 . Even though we developed scoMorphoFISH focusing on glomerular diseases, it can be instantly applied to other tissue compartments and even other organs. Ideally, an antibody can be used to segment respective tissue compartments, which chances we significantly improved when using tyramide signal amplification. In cases in wherein such antibody is available, antibody-independent smFISH expression levels can be used to identify respective cell types.

Typically, tissue segmentation tasks are demanding in heterogeneous sample sets (like kidney biopsies) and therefore traditional bottlenecks in image analysis workflows.

DL has great potential to accelerate segmentation tasks and has already been applied to classify glomerulosclerosis 19 

Human kidney biopsies of the Departments of Pathology of Paris and Hannover and of the Department of Pediatric Nephrology Essen were used. Additionally, we used excess healthy kidney tissue of tumor nephrectomies from the Department of Urology of the University Medicine Greifswald. After immersion fixation in 3% PFA overnight at room temperature, kidneys were embedded in paraffin using standard protocols. Care was taken that the temperature did not exceed 60°C. 5 µm FFPE tissue sections were mounted on superfrost slides and air-dried at room temperature. were stored at 4°C in the dark.

After being whole-slide imaged by widefield and 3D-SIM on an N-SIM-E 3D-SIM setup (Nikon), slides were immersed in 37°C 1xPBS for 1h. Coverslips were gently removed and Mowiol mounting medium washed out in 3 changes of 1xPBS. After that, routine PAS staining was performed as described before 2 . Sections were mounted in Eukitt (Carl Roth).

To obtain confocal-laser scanning micrographs, a Leica TCS-SP5 system was used.

Micrographs were acquired using a 40x 1.2 NA oil immersion objective with a voxel size of 189x189x500 nm (xyz). Stacks over 4 µm were acquired. For super-resolving 3D-structured illumination microscopy, a Zeiss Elyra PS.1 system (Carl Zeiss Microsystems) or a Nikon N-SIM-E was used as described before 2 . Whole slide images of PAS-stained sections were acquired on a Leica SCN400 slidescanner. SCN files were imported and processed in QuPath (v0.3.0).

Using the Google Colab-based ZeroCostDL4Mic notebooks, we trained a U-Net, a 

The script can perform different tasks. If glomerular outlines are stained by immunofluorescence:

1. Glomerular vs. tubulointerstitial transcript counter; For two-channel smFISH + NPHS2 IF + DAPI. The script asks first for the source folder of the multichannel tiff stacks The script uses the trained UNet to predict the glomerulus segmentation mask from the glomerular staining, takes the outlines of the mask as an ROI, calculates the intra-and extraglomerular area. It then differentially segments intra-and extraglomerular cells and counts transcript in the single-cell ROIs. Output is two-channel intra-and extraglomerular single-cell transcripts, total transcripts intra-and extracellular, intra-and extraglomerular area.

If no glomerular staining is present. The script will ask if glomerular outlines should be segmented manually. If nuclear IF is present, immunofluorescence positive cells are segmented using the trained StarDist network, and expression is differentially measured IF positive and negative cells. This part is established for podocytes but works with every strong nuclear marker. Output data is exported in one accumulated .xls file per folder analyzed. The script works in batch processing mode and processes all tiff files in a directory.

For the creation of 2D-density plots, the R spatstat package was used: After thresholding and binarization of the smFISH channels, xy-positions of the transcript spots were detected using RS-FISH and exported as csv files from ImageJ. XY positions were loaded to R-studio (version a1.3.1093) and plotted using the density plot function of the SpatStat library.

All source codes of scripts used in this manuscript as well as example data is available online: http://www.github.com/Siegerist

Nanoscale imaging of clinical specimens using pathology-optimized expansion microscopy

Structured illumination microscopy and automatized image processing as a rapid diagnostic tool for podocyte effacement

Visualization of podocyte substructure with structured illumination microscopy (SIM): a new approach to nephrotic disease

Super-resolved local recruitment of CLDN5 to filtration slits implicates a direct relationship with podocyte foot process effacement

Quantitative gene expression analysis in renal biopsies: A novel protocol for a high-throughput multicenter application

Tripartite Separation of Glomerular Cell-Types and Proteomes From Reporter-Free Mice

Single-cell transcriptomics of the mouse kidney reveals potential cellular targets of kidney disease HHS Public Access. Science (80-)

Formation and detection of RNA-DNA hybrid molecules in cytological preparations

Autonomous combinatorial color barcoding for multiplexing single molecule RNA visualization. bioRxiv

Third-generation in situ hybridization chain reaction: Multiplexed, quantitative, sensitive, versatile, robust. Dev

A novel in situ RNA analysis

U-net: Convolutional networks for biomedical image segmentation

U-Net: deep learning for cell counting, detection, and morphometry

RS-FISH : Precise , interactive and scalable smFISH spot detection using Radial Symmetry. bioRxiv Prepr

Visualization and analysis of gene expression in tissue sections by spatial transcriptomics. Science (80-)

Therapeutic effects of lisinopril and empagliflozin in a mouse model of hypertension-accelerated diabetic kidney disease

ZeroCostDL4Mic: an open platform to use Deep-Learning in Microscopy

An open-source platform for