FC 017DEEP-LEARNING ENABLED QUANTIFICATION OF SINGLE-CELL SINGLE-MRNA TRANSCRIPTS AND CORRELATIVE SUPER-RESOLVED PODOCYTE FOOT PROCESS MORPHOMETRY IN ROUTINE KIDNEY BIOPSY SPECIMEN

Abstract Background and Aims Although high-throughput single-cell transcriptomic analysis, super-resolution light microscopy and deep-learning methods are broadly used, the gold-standard to evaluate kidney biopsies is still the histologic assessment of formalin-fixed and paraffin embedded (FFPE) samples with parallel ultrastructural evaluation. Recently, we and others have shown that super-resolution fluorescence microscopy can be used to study glomerular ultrastructure in human biopsy samples. Additionally, in the last years mRNA in situ hybridization techniques have been improved to increase specificity and sensitivity to enable transcriptomic analysis with single-mRNA resolution (smFISH). Method For smFISH, we used the fluorescent multiplex RNAscope kit with probes targeting ACE2, WT1, PPIB, UBC and POLR2A. To find an on-slide reference gene, the normfinder algorithm was used. The smFISH protocol was combined with a single-step anti-podocin immunofluorescence enabled by VHH nanobodies. Podocytes were labeled by tyramide-signal amplified immunofluorescence using recombinant anti-WT1 antibodies. Slides were imaged using confocal laser scanning, as well as 3D structured illumination microscopy. Deep-learning networks to segment glomeruli and cell nuclei (UNet and StarDist) were trained using the ZeroCostDL4Mic approach. Scripts to automate analysis were developed in the ImageJ1 macro language. Results First, we show robust functionality of threeplex smFISH in archived routine FFPE kidney biopsy samples with single-mRNA resolution. As variations in sample preparation can negatively influence mRNA-abundance, we established PPIB as an ideal on-slide reference gene to account for different RNA-integrities present in biopsy samples. PPIB was chosen for its most stable expression in microarray dataset of various glomerular diseases determined by the Normfinder algorithm as well as its smFISH performance. To segment glomeruli and to label glomerular and tubulointerstitial cell subsets, we established a combination of smFISH and immunofluorescence. As smFISH requires intense tissue digestion to liberate cross-linked RNAs, immunofluorescence protocols had to be adapted: For podocin, a small-sized single-step label approach enabled by small nanobodies and for WT1, tyramide signal amplification was used. For enhanced segmentation performance, we used deep learning: First, a network was customized to recognize DAPI+ cell nuclei and WT1/DAPI+ podocyte nuclei. Second, a UNet was trained to segment glomeruli in podocin-stained tissue sections. Using these segmentation masks, we could annotate PPIB-normalized single mRNA transcripts to individual cells. We established an ImageJ script to automatize transcript quantification. As a proof-of-principle, we demonstrate inverse expression of WT1 and ACE2 in glomerular vs. tubulointerstitial single cells. Furthermore, in the podocyte subset, WT1 highly clustered whereas no significant ACE2 expression was found under baseline conditions. Additionally, when imaged with super-resolution microscopy, podocyte filtration slit morphology could be visualized The optical resolution was around 125 nm and therefore small enough to resolve individual foot processes. The filtration slit density as a podocyte-integrity marker did not differ significantly from undigested tissue sections proving the suitability for correlative podocyte foot process morphometry with single-podocyte transcript analysis. Conclusion Here we present a modular toolbox which combines algorithms for multiplexed, normalized single-cell gene expression with single mRNA resolution in cellular subsets (glomerular, tubulointerstitial and podocytes). Additionally, this approach enables correlation with podocyte filtration slit ultrastructure and gross glomerular morphometry.